diff --git a/analysis/current/analysis-report-hybrid.json b/analysis/current/analysis-report-hybrid.json
index 37f37cddfec0031fab09f27fcb802ec214e5547a..25898d28b3d77aae20d9f5db6ecb5534f659c2ea 100644
--- a/analysis/current/analysis-report-hybrid.json
+++ b/analysis/current/analysis-report-hybrid.json
@@ -1,1004 +1,11 @@
 {
   "schema_version": "1.0",
   "repo": "huggingface/transformers",
-  "snapshot_id": "20260707T180018Z",
-  "generated_at": "2026-07-07T18:14:24Z",
+  "snapshot_id": "20260708T000615Z",
+  "generated_at": "2026-07-08T00:12:27Z",
   "evidence_quality": "full",
   "llm_enrichment": true,
   "meta_bugs": [
-    {
-      "cluster_id": "cluster-43451-10",
-      "summary": "Cluster of 10 related pull requests linked by soft_similarity.",
-      "status": "open",
-      "confidence": 0.5,
-      "canonical_issue_number": null,
-      "canonical_pr_number": 43451,
-      "issue_numbers": [],
-      "pr_numbers": [
-        43451,
-        46658,
-        46912,
-        47003,
-        47034,
-        47035,
-        47092,
-        47096,
-        47137,
-        47143
-      ],
-      "evidence_types": [
-        "soft_similarity"
-      ],
-      "pr_comparisons": [
-        {
-          "left_pr_number": 43451,
-          "right_pr_number": 46658,
-          "code_similarity": 0.322,
-          "size_similarity": 0.946,
-          "file_overlap": 0.079,
-          "area_overlap": 0.185,
-          "patch_similarity": 0.299,
-          "shared_filenames": [
-            "src/transformers/conversion_mapping.py",
-            "src/transformers/models/auto/auto_mappings.py",
-            "src/transformers/models/auto/modeling_auto.py"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": "src/transformers/conversion_mapping.py",
-              "left_ranges": [
-                [
-                  737,
-                  769
-                ]
-              ],
-              "right_ranges": [
-                [
-                  739,
-                  807
-                ]
-              ]
-            },
-            {
-              "filename": "src/transformers/models/auto/auto_mappings.py",
-              "left_ranges": [
-                [
-                  391,
-                  398
-                ],
-                [
-                  824,
-                  830
-                ],
-                [
-                  945,
-                  951
-                ],
-                [
-                  1068,
-                  1074
-                ],
-                [
-                  1202,
-                  1208
-                ]
-              ],
-              "right_ranges": [
-                [
-                  362,
-                  368
-                ],
-                [
-                  591,
-                  599
-                ],
-                [
-                  798,
-                  804
-                ],
-                [
-                  884,
-                  892
-                ]
-              ]
-            },
-            {
-              "filename": "src/transformers/models/auto/modeling_auto.py",
-              "left_ranges": [
-                [
-                  331,
-                  338
-                ],
-                [
-                  1094,
-                  1100
-                ],
-                [
-                  1913,
-                  1919
-                ]
-              ],
-              "right_ranges": [
-                [
-                  1085,
-                  1091
-                ]
-              ]
-            }
-          ]
-        },
-        {
-          "left_pr_number": 43451,
-          "right_pr_number": 46912,
-          "code_similarity": 0.108,
-          "size_similarity": 0.441,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.134,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 43451,
-          "right_pr_number": 47003,
-          "code_similarity": 0.188,
-          "size_similarity": 0.805,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.18,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 43451,
-          "right_pr_number": 47034,
-          "code_similarity": 0.038,
-          "size_similarity": 0.112,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.106,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 43451,
-          "right_pr_number": 47035,
-          "code_similarity": 0.001,
-          "size_similarity": 0.002,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.001,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 43451,
-          "right_pr_number": 47092,
-          "code_similarity": 0.001,
-          "size_similarity": 0.002,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.003,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 43451,
-          "right_pr_number": 47096,
-          "code_similarity": 0.036,
-          "size_similarity": 0.099,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.108,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 43451,
-          "right_pr_number": 47137,
-          "code_similarity": 0.001,
-          "size_similarity": 0.005,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.002,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 43451,
-          "right_pr_number": 47143,
-          "code_similarity": 0.152,
-          "size_similarity": 0.392,
-          "file_overlap": 0.115,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.261,
-          "shared_filenames": [
-            "src/transformers/models/__init__.py",
-            "src/transformers/models/auto/auto_mappings.py",
-            "src/transformers/models/auto/modeling_auto.py"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": "src/transformers/models/__init__.py",
-              "left_ranges": [
-                [
-                  290,
-                  296
-                ]
-              ],
-              "right_ranges": [
-                [
-                  305,
-                  311
-                ]
-              ]
-            },
-            {
-              "filename": "src/transformers/models/auto/auto_mappings.py",
-              "left_ranges": [
-                [
-                  391,
-                  398
-                ],
-                [
-                  824,
-                  830
-                ],
-                [
-                  945,
-                  951
-                ],
-                [
-                  1068,
-                  1074
-                ],
-                [
-                  1202,
-                  1208
-                ]
-              ],
-              "right_ranges": [
-                [
-                  412,
-                  418
-                ],
-                [
-                  975,
-                  981
-                ]
-              ]
-            },
-            {
-              "filename": "src/transformers/models/auto/modeling_auto.py",
-              "left_ranges": [
-                [
-                  331,
-                  338
-                ],
-                [
-                  1094,
-                  1100
-                ],
-                [
-                  1913,
-                  1919
-                ]
-              ],
-              "right_ranges": [
-                [
-                  348,
-                  354
-                ],
-                [
-                  1960,
-                  1966
-                ]
-              ]
-            }
-          ]
-        },
-        {
-          "left_pr_number": 46658,
-          "right_pr_number": 46912,
-          "code_similarity": 0.104,
-          "size_similarity": 0.418,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.137,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46658,
-          "right_pr_number": 47003,
-          "code_similarity": 0.183,
-          "size_similarity": 0.762,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.202,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46658,
-          "right_pr_number": 47034,
-          "code_similarity": 0.036,
-          "size_similarity": 0.106,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.101,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46658,
-          "right_pr_number": 47035,
-          "code_similarity": 0.001,
-          "size_similarity": 0.002,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.001,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46658,
-          "right_pr_number": 47092,
-          "code_similarity": 0.001,
-          "size_similarity": 0.002,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.004,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46658,
-          "right_pr_number": 47096,
-          "code_similarity": 0.035,
-          "size_similarity": 0.093,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.11,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46658,
-          "right_pr_number": 47137,
-          "code_similarity": 0.001,
-          "size_similarity": 0.005,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.002,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46658,
-          "right_pr_number": 47143,
-          "code_similarity": 0.132,
-          "size_similarity": 0.371,
-          "file_overlap": 0.067,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.254,
-          "shared_filenames": [
-            "src/transformers/models/auto/auto_mappings.py",
-            "src/transformers/models/auto/modeling_auto.py"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": "src/transformers/models/auto/auto_mappings.py",
-              "left_ranges": [
-                [
-                  362,
-                  368
-                ],
-                [
-                  591,
-                  599
-                ],
-                [
-                  798,
-                  804
-                ],
-                [
-                  884,
-                  892
-                ]
-              ],
-              "right_ranges": [
-                [
-                  412,
-                  418
-                ],
-                [
-                  975,
-                  981
-                ]
-              ]
-            },
-            {
-              "filename": "src/transformers/models/auto/modeling_auto.py",
-              "left_ranges": [
-                [
-                  1085,
-                  1091
-                ]
-              ],
-              "right_ranges": [
-                [
-                  348,
-                  354
-                ],
-                [
-                  1960,
-                  1966
-                ]
-              ]
-            }
-          ]
-        },
-        {
-          "left_pr_number": 46912,
-          "right_pr_number": 47003,
-          "code_similarity": 0.129,
-          "size_similarity": 0.548,
-          "file_overlap": 0.012,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.109,
-          "shared_filenames": [
-            "src/transformers/models/gemma3n/modeling_gemma3n.py"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": "src/transformers/models/gemma3n/modeling_gemma3n.py",
-              "left_ranges": [
-                [
-                  1537,
-                  1543
-                ],
-                [
-                  1553,
-                  1559
-                ]
-              ],
-              "right_ranges": [
-                [
-                  2173,
-                  2179
-                ]
-              ]
-            }
-          ]
-        },
-        {
-          "left_pr_number": 46912,
-          "right_pr_number": 47034,
-          "code_similarity": 0.064,
-          "size_similarity": 0.253,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.087,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46912,
-          "right_pr_number": 47035,
-          "code_similarity": 0.001,
-          "size_similarity": 0.004,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.002,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46912,
-          "right_pr_number": 47092,
-          "code_similarity": 0.001,
-          "size_similarity": 0.004,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.004,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46912,
-          "right_pr_number": 47096,
-          "code_similarity": 0.06,
-          "size_similarity": 0.224,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.1,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46912,
-          "right_pr_number": 47137,
-          "code_similarity": 0.003,
-          "size_similarity": 0.012,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.002,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46912,
-          "right_pr_number": 47143,
-          "code_similarity": 0.199,
-          "size_similarity": 0.887,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.142,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47003,
-          "right_pr_number": 47034,
-          "code_similarity": 0.04,
-          "size_similarity": 0.139,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.082,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47003,
-          "right_pr_number": 47035,
-          "code_similarity": 0.001,
-          "size_similarity": 0.002,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.001,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47003,
-          "right_pr_number": 47092,
-          "code_similarity": 0.001,
-          "size_similarity": 0.002,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.003,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47003,
-          "right_pr_number": 47096,
-          "code_similarity": 0.069,
-          "size_similarity": 0.123,
-          "file_overlap": 0.073,
-          "area_overlap": 0.019,
-          "patch_similarity": 0.103,
-          "shared_filenames": [
-            "src/transformers/exporters/exporter_dynamo.py",
-            "src/transformers/exporters/exporter_executorch.py",
-            "src/transformers/exporters/utils.py",
-            "tests/exporters/test_export.py"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": "src/transformers/exporters/exporter_dynamo.py",
-              "left_ranges": [
-                [
-                  43,
-                  49
-                ],
-                [
-                  449,
-                  470
-                ],
-                [
-                  487,
-                  507
-                ],
-                [
-                  548,
-                  559
-                ],
-                [
-                  594,
-                  606
-                ]
-              ],
-              "right_ranges": [
-                [
-                  350,
-                  357
-                ],
-                [
-                  381,
-                  388
-                ]
-              ]
-            },
-            {
-              "filename": "src/transformers/exporters/exporter_executorch.py",
-              "left_ranges": [
-                [
-                  38,
-                  44
-                ],
-                [
-                  61,
-                  74
-                ],
-                [
-                  81,
-                  99
-                ],
-                [
-                  354,
-                  360
-                ],
-                [
-                  365,
-                  407
-                ],
-                [
-                  414,
-                  429
-                ],
-                [
-                  463,
-                  474
-                ],
-                [
-                  488,
-                  519
-                ],
-                [
-                  530,
-                  535
-                ],
-                [
-                  590,
-                  595
-                ],
-                [
-                  623,
-                  628
-                ],
-                [
-                  646,
-                  770
-                ],
-                [
-                  790,
-                  795
-                ],
-                [
-                  812,
-                  1154
-                ],
-                [
-                  1246,
-                  1275
-                ],
-                [
-                  1411,
-                  1463
-                ]
-              ],
-              "right_ranges": [
-                [
-                  418,
-                  442
-                ]
-              ]
-            },
-            {
-              "filename": "src/transformers/exporters/utils.py",
-              "left_ranges": [
-                [
-                  232,
-                  238
-                ],
-                [
-                  444,
-                  461
-                ],
-                [
-                  621,
-                  644
-                ],
-                [
-                  647,
-                  663
-                ],
-                [
-                  761,
-                  768
-                ]
-              ],
-              "right_ranges": [
-                [
-                  432,
-                  441
-                ],
-                [
-                  444,
-                  455
-                ],
-                [
-                  464,
-                  490
-                ],
-                [
-                  708,
-                  720
-                ]
-              ]
-            },
-            {
-              "filename": "tests/exporters/test_export.py",
-              "left_ranges": [
-                [
-                  24,
-                  30
-                ],
-                [
-                  35,
-                  41
-                ],
-                [
-                  46,
-                  114
-                ],
-                [
-                  118,
-                  150
-                ],
-                [
-                  191,
-                  266
-                ],
-                [
-                  298,
-                  314
-                ],
-                [
-                  407,
-                  414
-                ],
-                [
-                  420,
-                  452
-                ],
-                [
-                  553,
-                  560
-                ],
-                [
-                  567,
-                  599
-                ]
-              ],
-              "right_ranges": [
-                [
-                  181,
-                  188
-                ]
-              ]
-            }
-          ]
-        },
-        {
-          "left_pr_number": 47003,
-          "right_pr_number": 47137,
-          "code_similarity": 0.002,
-          "size_similarity": 0.006,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.002,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47003,
-          "right_pr_number": 47143,
-          "code_similarity": 0.124,
-          "size_similarity": 0.486,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.178,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47034,
-          "right_pr_number": 47035,
-          "code_similarity": 0.405,
-          "size_similarity": 0.017,
-          "file_overlap": 0.167,
-          "area_overlap": 1.0,
-          "patch_similarity": 0.013,
-          "shared_filenames": [
-            "src/transformers/training_args.py"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": "src/transformers/training_args.py",
-              "left_ranges": [
-                [
-                  1563,
-                  1570
-                ]
-              ],
-              "right_ranges": [
-                [
-                  1563,
-                  1570
-                ]
-              ]
-            }
-          ]
-        },
-        {
-          "left_pr_number": 47034,
-          "right_pr_number": 47092,
-          "code_similarity": 0.004,
-          "size_similarity": 0.017,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.007,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47034,
-          "right_pr_number": 47096,
-          "code_similarity": 0.19,
-          "size_similarity": 0.883,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.09,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47034,
-          "right_pr_number": 47137,
-          "code_similarity": 0.01,
-          "size_similarity": 0.047,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.005,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47034,
-          "right_pr_number": 47143,
-          "code_similarity": 0.076,
-          "size_similarity": 0.286,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.123,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47035,
-          "right_pr_number": 47092,
-          "code_similarity": 0.2,
-          "size_similarity": 1.0,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.0,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47035,
-          "right_pr_number": 47096,
-          "code_similarity": 0.004,
-          "size_similarity": 0.019,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.004,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47035,
-          "right_pr_number": 47137,
-          "code_similarity": 0.073,
-          "size_similarity": 0.364,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.0,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47035,
-          "right_pr_number": 47143,
-          "code_similarity": 0.001,
-          "size_similarity": 0.005,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.002,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47092,
-          "right_pr_number": 47096,
-          "code_similarity": 0.005,
-          "size_similarity": 0.019,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.011,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47092,
-          "right_pr_number": 47137,
-          "code_similarity": 0.073,
-          "size_similarity": 0.364,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.0,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47092,
-          "right_pr_number": 47143,
-          "code_similarity": 0.002,
-          "size_similarity": 0.005,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.005,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47096,
-          "right_pr_number": 47137,
-          "code_similarity": 0.011,
-          "size_similarity": 0.053,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.004,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47096,
-          "right_pr_number": 47143,
-          "code_similarity": 0.068,
-          "size_similarity": 0.252,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.115,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 47137,
-          "right_pr_number": 47143,
-          "code_similarity": 0.003,
-          "size_similarity": 0.013,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.003,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        }
-      ]
-    },
     {
       "cluster_id": "cluster-43979-11",
       "summary": "Cluster of 1 issues and 10 PRs centered on issue #43979.",
@@ -2043,226 +1050,264 @@
       ]
     },
     {
-      "cluster_id": "cluster-46894-6",
-      "summary": "Cluster of 2 issues and 4 PRs centered on issue #46894.",
+      "cluster_id": "cluster-41211-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #41211.",
       "status": "open",
-      "confidence": 0.8,
-      "canonical_issue_number": 46894,
-      "canonical_pr_number": 46903,
+      "confidence": 0.55,
+      "canonical_issue_number": 41211,
+      "canonical_pr_number": 44339,
       "issue_numbers": [
-        46894,
-        46897
+        41211
       ],
       "pr_numbers": [
-        46895,
-        46903,
-        46958,
-        47090
-      ],
-      "evidence_types": [
-        "closing_reference",
-        "shared_issue_target",
-        "soft_similarity"
+        41356,
+        44339
+      ],
+      "evidence_types": [
+        "closing_reference"
       ],
       "pr_comparisons": [
         {
-          "left_pr_number": 46895,
-          "right_pr_number": 46903,
-          "code_similarity": 0.533,
-          "size_similarity": 0.404,
-          "file_overlap": 0.333,
-          "area_overlap": 0.933,
-          "patch_similarity": 0.17,
+          "left_pr_number": 41356,
+          "right_pr_number": 44339,
+          "code_similarity": 0.155,
+          "size_similarity": 0.094,
+          "file_overlap": 0.273,
+          "area_overlap": 0.078,
+          "patch_similarity": 0.183,
           "shared_filenames": [
-            "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
-            "src/transformers/models/cohere_asr/modular_cohere_asr.py"
+            "docs/source/en/model_doc/deimv2.md",
+            "src/transformers/models/deimv2/__init__.py",
+            "src/transformers/models/deimv2/configuration_deimv2.py",
+            "src/transformers/models/deimv2/modeling_deimv2.py",
+            "tests/models/deimv2/__init__.py",
+            "tests/models/deimv2/test_modeling_deimv2.py"
           ],
           "shared_file_areas": [
             {
-              "filename": "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
+              "filename": "docs/source/en/model_doc/deimv2.md",
               "left_ranges": [
                 [
-                  637,
-                  650
+                  1,
+                  132
                 ]
               ],
               "right_ranges": [
                 [
-                  637,
-                  651
+                  1,
+                  65
                 ]
               ]
             },
             {
-              "filename": "src/transformers/models/cohere_asr/modular_cohere_asr.py",
-              "left_ranges": [
-                [
-                  500,
-                  513
-                ]
-              ],
-              "right_ranges": [
-                [
-                  500,
-                  514
-                ]
-              ]
-            }
-          ]
-        },
-        {
-          "left_pr_number": 46895,
-          "right_pr_number": 46958,
-          "code_similarity": 0.779,
-          "size_similarity": 0.915,
-          "file_overlap": 1.0,
-          "area_overlap": 0.5,
-          "patch_similarity": 0.81,
-          "shared_filenames": [
-            "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
-            "src/transformers/models/cohere_asr/modular_cohere_asr.py",
-            "tests/models/cohere_asr/test_modeling_cohere_asr.py"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
+              "filename": "src/transformers/models/deimv2/__init__.py",
               "left_ranges": [
                 [
-                  637,
-                  650
+                  1,
+                  15
                 ]
               ],
               "right_ranges": [
                 [
-                  22,
-                  28
-                ],
-                [
-                  638,
-                  645
+                  1,
+                  29
                 ]
               ]
             },
             {
-              "filename": "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+              "filename": "src/transformers/models/deimv2/configuration_deimv2.py",
               "left_ranges": [
                 [
-                  500,
-                  513
+                  1,
+                  74
                 ]
               ],
               "right_ranges": [
                 [
-                  16,
-                  22
-                ],
-                [
-                  501,
-                  508
+                  1,
+                  266
                 ]
               ]
             },
             {
-              "filename": "tests/models/cohere_asr/test_modeling_cohere_asr.py",
+              "filename": "tests/models/deimv2/test_modeling_deimv2.py",
               "left_ranges": [
                 [
-                  35,
-                  41
-                ],
-                [
-                  176,
-                  214
+                  1,
+                  15
                 ]
               ],
               "right_ranges": [
                 [
-                  15,
-                  21
-                ],
-                [
-                  180,
-                  221
+                  1,
+                  1734
                 ]
               ]
             }
           ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-46196-3",
+      "summary": "Cluster of 3 related pull requests linked by soft_similarity.",
+      "status": "open",
+      "confidence": 0.5,
+      "canonical_issue_number": null,
+      "canonical_pr_number": 46196,
+      "issue_numbers": [],
+      "pr_numbers": [
+        46196,
+        46836,
+        47143
+      ],
+      "evidence_types": [
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 46196,
+          "right_pr_number": 46836,
+          "code_similarity": 0.205,
+          "size_similarity": 0.936,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.12,
+          "shared_filenames": [],
+          "shared_file_areas": []
         },
         {
-          "left_pr_number": 46895,
-          "right_pr_number": 47090,
-          "code_similarity": 0.819,
-          "size_similarity": 0.932,
-          "file_overlap": 1.0,
-          "area_overlap": 0.783,
-          "patch_similarity": 0.389,
+          "left_pr_number": 46196,
+          "right_pr_number": 47143,
+          "code_similarity": 0.124,
+          "size_similarity": 0.513,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.145,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 46836,
+          "right_pr_number": 47143,
+          "code_similarity": 0.169,
+          "size_similarity": 0.48,
+          "file_overlap": 0.111,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.266,
           "shared_filenames": [
-            "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
-            "src/transformers/models/cohere_asr/modular_cohere_asr.py",
-            "tests/models/cohere_asr/test_modeling_cohere_asr.py"
+            "src/transformers/models/__init__.py",
+            "src/transformers/models/auto/auto_mappings.py",
+            "src/transformers/models/auto/modeling_auto.py"
           ],
           "shared_file_areas": [
             {
-              "filename": "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
+              "filename": "src/transformers/models/__init__.py",
               "left_ranges": [
                 [
-                  637,
-                  650
+                  452,
+                  458
                 ]
               ],
               "right_ranges": [
                 [
-                  637,
-                  649
+                  305,
+                  311
                 ]
               ]
             },
             {
-              "filename": "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+              "filename": "src/transformers/models/auto/auto_mappings.py",
               "left_ranges": [
                 [
-                  500,
-                  513
+                  628,
+                  638
+                ],
+                [
+                  905,
+                  914
+                ],
+                [
+                  1112,
+                  1118
+                ],
+                [
+                  1238,
+                  1244
                 ]
               ],
               "right_ranges": [
                 [
-                  500,
-                  512
+                  412,
+                  418
+                ],
+                [
+                  975,
+                  981
                 ]
               ]
             },
             {
-              "filename": "tests/models/cohere_asr/test_modeling_cohere_asr.py",
+              "filename": "src/transformers/models/auto/modeling_auto.py",
               "left_ranges": [
                 [
-                  35,
-                  41
+                  501,
+                  507
                 ],
                 [
-                  176,
-                  214
+                  1107,
+                  1113
                 ]
               ],
               "right_ranges": [
                 [
-                  175,
-                  222
+                  348,
+                  354
+                ],
+                [
+                  1960,
+                  1966
                 ]
               ]
             }
           ]
-        },
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-46894-5",
+      "summary": "Cluster of 2 issues and 3 PRs centered on issue #46894.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 46894,
+      "canonical_pr_number": 46895,
+      "issue_numbers": [
+        46894,
+        46897
+      ],
+      "pr_numbers": [
+        46895,
+        46958,
+        47090
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target",
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
         {
-          "left_pr_number": 46903,
+          "left_pr_number": 46895,
           "right_pr_number": 46958,
-          "code_similarity": 0.325,
-          "size_similarity": 0.369,
-          "file_overlap": 0.333,
-          "area_overlap": 0.364,
-          "patch_similarity": 0.156,
+          "code_similarity": 0.779,
+          "size_similarity": 0.915,
+          "file_overlap": 1.0,
+          "area_overlap": 0.5,
+          "patch_similarity": 0.81,
           "shared_filenames": [
             "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
-            "src/transformers/models/cohere_asr/modular_cohere_asr.py"
+            "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+            "tests/models/cohere_asr/test_modeling_cohere_asr.py"
           ],
           "shared_file_areas": [
             {
@@ -2270,7 +1315,7 @@
               "left_ranges": [
                 [
                   637,
-                  651
+                  650
                 ]
               ],
               "right_ranges": [
@@ -2289,7 +1334,7 @@
               "left_ranges": [
                 [
                   500,
-                  514
+                  513
                 ]
               ],
               "right_ranges": [
@@ -2302,20 +1347,44 @@
                   508
                 ]
               ]
+            },
+            {
+              "filename": "tests/models/cohere_asr/test_modeling_cohere_asr.py",
+              "left_ranges": [
+                [
+                  35,
+                  41
+                ],
+                [
+                  176,
+                  214
+                ]
+              ],
+              "right_ranges": [
+                [
+                  15,
+                  21
+                ],
+                [
+                  180,
+                  221
+                ]
+              ]
             }
           ]
         },
         {
-          "left_pr_number": 46903,
+          "left_pr_number": 46895,
           "right_pr_number": 47090,
-          "code_similarity": 0.519,
-          "size_similarity": 0.433,
-          "file_overlap": 0.333,
-          "area_overlap": 0.867,
-          "patch_similarity": 0.192,
+          "code_similarity": 0.819,
+          "size_similarity": 0.932,
+          "file_overlap": 1.0,
+          "area_overlap": 0.783,
+          "patch_similarity": 0.389,
           "shared_filenames": [
             "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
-            "src/transformers/models/cohere_asr/modular_cohere_asr.py"
+            "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+            "tests/models/cohere_asr/test_modeling_cohere_asr.py"
           ],
           "shared_file_areas": [
             {
@@ -2323,7 +1392,7 @@
               "left_ranges": [
                 [
                   637,
-                  651
+                  650
                 ]
               ],
               "right_ranges": [
@@ -2338,7 +1407,7 @@
               "left_ranges": [
                 [
                   500,
-                  514
+                  513
                 ]
               ],
               "right_ranges": [
@@ -2347,6 +1416,25 @@
                   512
                 ]
               ]
+            },
+            {
+              "filename": "tests/models/cohere_asr/test_modeling_cohere_asr.py",
+              "left_ranges": [
+                [
+                  35,
+                  41
+                ],
+                [
+                  176,
+                  214
+                ]
+              ],
+              "right_ranges": [
+                [
+                  175,
+                  222
+                ]
+              ]
             }
           ]
         },
@@ -2425,209 +1513,6 @@
         }
       ]
     },
-    {
-      "cluster_id": "cluster-41211-3",
-      "summary": "Cluster of 1 issues and 2 PRs centered on issue #41211.",
-      "status": "open",
-      "confidence": 0.55,
-      "canonical_issue_number": 41211,
-      "canonical_pr_number": 44339,
-      "issue_numbers": [
-        41211
-      ],
-      "pr_numbers": [
-        41356,
-        44339
-      ],
-      "evidence_types": [
-        "closing_reference"
-      ],
-      "pr_comparisons": [
-        {
-          "left_pr_number": 41356,
-          "right_pr_number": 44339,
-          "code_similarity": 0.155,
-          "size_similarity": 0.094,
-          "file_overlap": 0.273,
-          "area_overlap": 0.078,
-          "patch_similarity": 0.183,
-          "shared_filenames": [
-            "docs/source/en/model_doc/deimv2.md",
-            "src/transformers/models/deimv2/__init__.py",
-            "src/transformers/models/deimv2/configuration_deimv2.py",
-            "src/transformers/models/deimv2/modeling_deimv2.py",
-            "tests/models/deimv2/__init__.py",
-            "tests/models/deimv2/test_modeling_deimv2.py"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": "docs/source/en/model_doc/deimv2.md",
-              "left_ranges": [
-                [
-                  1,
-                  132
-                ]
-              ],
-              "right_ranges": [
-                [
-                  1,
-                  65
-                ]
-              ]
-            },
-            {
-              "filename": "src/transformers/models/deimv2/__init__.py",
-              "left_ranges": [
-                [
-                  1,
-                  15
-                ]
-              ],
-              "right_ranges": [
-                [
-                  1,
-                  29
-                ]
-              ]
-            },
-            {
-              "filename": "src/transformers/models/deimv2/configuration_deimv2.py",
-              "left_ranges": [
-                [
-                  1,
-                  74
-                ]
-              ],
-              "right_ranges": [
-                [
-                  1,
-                  266
-                ]
-              ]
-            },
-            {
-              "filename": "tests/models/deimv2/test_modeling_deimv2.py",
-              "left_ranges": [
-                [
-                  1,
-                  15
-                ]
-              ],
-              "right_ranges": [
-                [
-                  1,
-                  1734
-                ]
-              ]
-            }
-          ]
-        }
-      ]
-    },
-    {
-      "cluster_id": "cluster-46432-4",
-      "summary": "Cluster of 4 related pull requests linked by soft_similarity.",
-      "status": "open",
-      "confidence": 0.5,
-      "canonical_issue_number": null,
-      "canonical_pr_number": 46716,
-      "issue_numbers": [],
-      "pr_numbers": [
-        46432,
-        46504,
-        46563,
-        46716
-      ],
-      "evidence_types": [
-        "soft_similarity"
-      ],
-      "pr_comparisons": [
-        {
-          "left_pr_number": 46432,
-          "right_pr_number": 46504,
-          "code_similarity": 0.137,
-          "size_similarity": 0.667,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.026,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46432,
-          "right_pr_number": 46563,
-          "code_similarity": 0.147,
-          "size_similarity": 0.722,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.018,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46432,
-          "right_pr_number": 46716,
-          "code_similarity": 0.133,
-          "size_similarity": 0.667,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.0,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46504,
-          "right_pr_number": 46563,
-          "code_similarity": 0.185,
-          "size_similarity": 0.923,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.0,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46504,
-          "right_pr_number": 46716,
-          "code_similarity": 0.2,
-          "size_similarity": 1.0,
-          "file_overlap": 0.0,
-          "area_overlap": 0.0,
-          "patch_similarity": 0.0,
-          "shared_filenames": [],
-          "shared_file_areas": []
-        },
-        {
-          "left_pr_number": 46563,
-          "right_pr_number": 46716,
-          "code_similarity": 0.69,
-          "size_similarity": 0.923,
-          "file_overlap": 1.0,
-          "area_overlap": 0.571,
-          "patch_similarity": 0.033,
-          "shared_filenames": [
-            ".github/workflows/pr-ci-security-gate.yml"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": ".github/workflows/pr-ci-security-gate.yml",
-              "left_ranges": [
-                [
-                  13,
-                  19
-                ]
-              ],
-              "right_ranges": [
-                [
-                  14,
-                  17
-                ]
-              ]
-            }
-          ]
-        }
-      ]
-    },
     {
       "cluster_id": "cluster-43240-3",
       "summary": "Cluster of 1 issues and 2 PRs centered on issue #43240.",
@@ -2684,7 +1569,7 @@
       "status": "open",
       "confidence": 0.55,
       "canonical_issue_number": 46693,
-      "canonical_pr_number": 46696,
+      "canonical_pr_number": 47134,
       "issue_numbers": [
         46693
       ],
@@ -2699,11 +1584,11 @@
         {
           "left_pr_number": 46696,
           "right_pr_number": 47134,
-          "code_similarity": 0.527,
-          "size_similarity": 0.714,
+          "code_similarity": 0.464,
+          "size_similarity": 0.8,
           "file_overlap": 1.0,
-          "area_overlap": 0.186,
-          "patch_similarity": 0.123,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.029,
           "shared_filenames": [
             "src/transformers/modeling_flash_attention_utils.py"
           ],
@@ -2724,18 +1609,6 @@
                 [
                   366,
                   373
-                ],
-                [
-                  597,
-                  604
-                ],
-                [
-                  629,
-                  637
-                ],
-                [
-                  675,
-                  685
                 ]
               ]
             }
@@ -3463,6 +2336,64 @@
         }
       ]
     },
+    {
+      "cluster_id": "cluster-44001-2",
+      "summary": "Cluster of 2 related pull requests linked by soft_similarity.",
+      "status": "open",
+      "confidence": 0.5,
+      "canonical_issue_number": null,
+      "canonical_pr_number": 44004,
+      "issue_numbers": [],
+      "pr_numbers": [
+        44001,
+        44004
+      ],
+      "evidence_types": [
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 44001,
+          "right_pr_number": 44004,
+          "code_similarity": 0.087,
+          "size_similarity": 0.237,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.262,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-44002-2",
+      "summary": "Cluster of 2 related pull requests linked by soft_similarity.",
+      "status": "open",
+      "confidence": 0.5,
+      "canonical_issue_number": null,
+      "canonical_pr_number": 44002,
+      "issue_numbers": [],
+      "pr_numbers": [
+        44002,
+        44029
+      ],
+      "evidence_types": [
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 44002,
+          "right_pr_number": 44029,
+          "code_similarity": 0.135,
+          "size_similarity": 0.43,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.329,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        }
+      ]
+    },
     {
       "cluster_id": "cluster-44018-2",
       "summary": "Cluster of 2 related pull requests linked by soft_similarity.",
@@ -3741,12 +2672,12 @@
   ],
   "duplicate_issues": [
     {
-      "cluster_id": "cluster-46894-6",
+      "cluster_id": "cluster-46894-5",
       "canonical_issue_number": 46894,
       "duplicate_issue_numbers": [
         46897
       ],
-      "reason": "Issues in cluster-46894-6 are treated as duplicates because they share closing_reference, shared_issue_target, soft_similarity evidence."
+      "reason": "Issues in cluster-46894-5 are treated as duplicates because they share closing_reference, shared_issue_target, soft_similarity evidence."
     }
   ],
   "duplicate_prs": [
@@ -3768,23 +2699,6 @@
       "target_issue_number": 43240,
       "reason": "PRs in cluster-43240-3 are treated as duplicates because they converge on issue #43240 with closing_reference, shared_issue_target evidence."
     },
-    {
-      "cluster_id": "cluster-43451-10",
-      "canonical_pr_number": 43451,
-      "duplicate_pr_numbers": [
-        46658,
-        46912,
-        47003,
-        47034,
-        47035,
-        47092,
-        47096,
-        47137,
-        47143
-      ],
-      "target_issue_number": null,
-      "reason": "PRs in cluster-43451-10 are treated as duplicates because they share soft_similarity evidence."
-    },
     {
       "cluster_id": "cluster-43698-3",
       "canonical_pr_number": 43779,
@@ -3811,6 +2725,24 @@
       "target_issue_number": 43979,
       "reason": "PRs in cluster-43979-11 are treated as duplicates because they converge on issue #43979 with closing_reference, shared_issue_target, soft_similarity evidence."
     },
+    {
+      "cluster_id": "cluster-44001-2",
+      "canonical_pr_number": 44004,
+      "duplicate_pr_numbers": [
+        44001
+      ],
+      "target_issue_number": null,
+      "reason": "PRs in cluster-44001-2 are treated as duplicates because they share soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-44002-2",
+      "canonical_pr_number": 44002,
+      "duplicate_pr_numbers": [
+        44029
+      ],
+      "target_issue_number": null,
+      "reason": "PRs in cluster-44002-2 are treated as duplicates because they share soft_similarity evidence."
+    },
     {
       "cluster_id": "cluster-44018-2",
       "canonical_pr_number": 44068,
@@ -3857,21 +2789,20 @@
       "reason": "PRs in cluster-46159-3 are treated as duplicates because they converge on issue #46159 with closing_reference evidence."
     },
     {
-      "cluster_id": "cluster-46432-4",
-      "canonical_pr_number": 46716,
+      "cluster_id": "cluster-46196-3",
+      "canonical_pr_number": 46196,
       "duplicate_pr_numbers": [
-        46432,
-        46504,
-        46563
+        46836,
+        47143
       ],
       "target_issue_number": null,
-      "reason": "PRs in cluster-46432-4 are treated as duplicates because they share soft_similarity evidence."
+      "reason": "PRs in cluster-46196-3 are treated as duplicates because they share soft_similarity evidence."
     },
     {
       "cluster_id": "cluster-46693-3",
-      "canonical_pr_number": 46696,
+      "canonical_pr_number": 47134,
       "duplicate_pr_numbers": [
-        47134
+        46696
       ],
       "target_issue_number": 46693,
       "reason": "PRs in cluster-46693-3 are treated as duplicates because they converge on issue #46693 with closing_reference evidence."
@@ -3886,15 +2817,14 @@
       "reason": "PRs in cluster-46869-3 are treated as duplicates because they converge on issue #46869 with closing_reference, shared_issue_target evidence."
     },
     {
-      "cluster_id": "cluster-46894-6",
-      "canonical_pr_number": 46903,
+      "cluster_id": "cluster-46894-5",
+      "canonical_pr_number": 46895,
       "duplicate_pr_numbers": [
-        46895,
         46958,
         47090
       ],
       "target_issue_number": 46894,
-      "reason": "PRs in cluster-46894-6 are treated as duplicates because they converge on issue #46894 with closing_reference, shared_issue_target, soft_similarity evidence."
+      "reason": "PRs in cluster-46894-5 are treated as duplicates because they converge on issue #46894 with closing_reference, shared_issue_target, soft_similarity evidence."
     },
     {
       "cluster_id": "cluster-47013-3",
@@ -3907,10 +2837,10 @@
     }
   ],
   "best_issue": {
-    "cluster_id": "cluster-46894-6",
+    "cluster_id": "cluster-46894-5",
     "issue_number": 46894,
-    "reason": "Issue #46894 is the strongest global issue candidate because it is open, belongs to a cluster with 6 artifacts, and carries 1 discussion comments plus 9 inbound references.",
-    "score": 171.0
+    "reason": "Issue #46894 is the strongest global issue candidate because it is open, belongs to a cluster with 5 artifacts, and carries 1 discussion comments plus 9 inbound references.",
+    "score": 161.0
   },
   "best_pr": {
     "cluster_id": "cluster-43979-11",
diff --git a/analysis/current/analysis-report-hybrid.llm-reviews.json b/analysis/current/analysis-report-hybrid.llm-reviews.json
index b465849abe4b8576585b469bd35fcf3604135220..d2e308944cc1a88fef32f692de48f0dddece4c79 100644
--- a/analysis/current/analysis-report-hybrid.llm-reviews.json
+++ b/analysis/current/analysis-report-hybrid.llm-reviews.json
@@ -1,8 +1,8 @@
 {
   "schema_version": "1.0",
   "repo": "huggingface/transformers",
-  "snapshot_id": "20260707T180018Z",
-  "generated_at": "2026-07-07T18:14:24Z",
+  "snapshot_id": "20260708T000615Z",
+  "generated_at": "2026-07-08T00:12:27Z",
   "model": "gpt-5.4-mini?service_tier=flex",
   "reviews": [
     {
@@ -13,8 +13,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -552,8 +552,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -951,8 +951,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -1272,8 +1272,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -1593,8 +1593,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -1893,8 +1893,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -2230,8 +2230,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -2545,8 +2545,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -2874,8 +2874,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -3192,8 +3192,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -3500,8 +3500,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -3797,8 +3797,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -4104,8 +4104,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -4412,8 +4412,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -4738,8 +4738,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -5035,8 +5035,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -5326,8 +5326,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -5628,8 +5628,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -5942,8 +5942,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -6239,8 +6239,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -6533,8 +6533,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -6834,8 +6834,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -7127,8 +7127,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -7453,8 +7453,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -7751,8 +7751,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -8044,8 +8044,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -8356,8 +8356,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -8668,8 +8668,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -8966,8 +8966,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -9284,8 +9284,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -9582,8 +9582,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -9880,8 +9880,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -10166,8 +10166,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -10485,8 +10485,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -10803,8 +10803,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -11101,8 +11101,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -11384,8 +11384,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -11682,8 +11682,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -11973,8 +11973,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -12267,8 +12267,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -12572,8 +12572,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -12870,8 +12870,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -13153,8 +13153,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -13451,8 +13451,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -13741,8 +13741,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -14032,8 +14032,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -14340,8 +14340,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -14651,8 +14651,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -14980,8 +14980,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -15284,8 +15284,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -15601,8 +15601,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -15898,8 +15898,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -16210,8 +16210,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -16507,8 +16507,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -16826,8 +16826,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -17131,8 +17131,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -17450,8 +17450,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -17783,8 +17783,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -18104,8 +18104,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -18404,8 +18404,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -18772,8 +18772,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -19142,8 +19142,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -19545,8 +19545,8 @@
       "review_unit_count": 63,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-514",
       "nodes": [
@@ -19684,8 +19684,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10519-3",
       "nodes": [
@@ -19762,8 +19762,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-13244-3",
       "nodes": [
@@ -19833,8 +19833,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-26350-2",
       "nodes": [
@@ -19896,8 +19896,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-27088-9",
       "nodes": [
@@ -20162,8 +20162,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-36718-21",
       "nodes": [
@@ -20555,8 +20555,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-38740-2",
       "nodes": [
@@ -20618,8 +20618,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-41084-2",
       "nodes": [
@@ -20681,8 +20681,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-41306-2",
       "nodes": [
@@ -20744,8 +20744,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-43598-2",
       "nodes": [
@@ -20807,8 +20807,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-43684-2",
       "nodes": [
@@ -20870,8 +20870,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-43835-4",
       "nodes": [
@@ -20970,8 +20970,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-44371-6",
       "nodes": [
@@ -21114,8 +21114,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-45145-2",
       "nodes": [
@@ -21234,8 +21234,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-45762-2",
       "nodes": [
@@ -21297,8 +21297,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-45764-2",
       "nodes": [
@@ -21360,8 +21360,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-45810-2",
       "nodes": [
@@ -21423,8 +21423,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-45869-5",
       "nodes": [
@@ -21541,8 +21541,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-46439-2",
       "nodes": [
@@ -21607,8 +21607,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-46551-2",
       "nodes": [
@@ -21673,8 +21673,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-46657-2",
       "nodes": [
@@ -21736,8 +21736,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-46822-2",
       "nodes": [
@@ -21799,8 +21799,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-46906-12",
       "nodes": [
@@ -22138,8 +22138,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-46983-2",
       "nodes": [
@@ -22204,8 +22204,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-47007-2",
       "nodes": [
@@ -22267,8 +22267,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-47138-2",
       "nodes": [
@@ -22336,7 +22336,7 @@
       "source": "llm",
       "cache_hit": false,
       "model": "gpt-5.4-mini?service_tier=flex",
-      "cluster_id": "cluster-40546-42",
+      "cluster_id": "cluster-40546-44",
       "nodes": [
         "pull_request:43451",
         "pull_request:44517",
@@ -22354,7 +22354,6 @@
         "pull_request:46958",
         "pull_request:46974",
         "pull_request:46995",
-        "pull_request:47003",
         "pull_request:47025",
         "pull_request:47034",
         "pull_request:47035",
@@ -22376,7 +22375,9 @@
         "pull_request:47136",
         "pull_request:47137",
         "pull_request:47142",
-        "pull_request:47143"
+        "pull_request:47143",
+        "pull_request:47148",
+        "pull_request:47149"
       ],
       "soft_pairs": [
         "pull_request:46696|pull_request:47134",
@@ -22386,512 +22387,419 @@
         "pull_request:46958|pull_request:46974",
         "pull_request:47072|pull_request:47089",
         "pull_request:47077|pull_request:47078",
+        "pull_request:47072|pull_request:47148",
         "pull_request:46266|pull_request:46957",
+        "pull_request:47114|pull_request:47148",
         "pull_request:47072|pull_request:47134",
         "pull_request:47114|pull_request:47135",
         "pull_request:44517|pull_request:47082",
         "pull_request:47034|pull_request:47035",
         "pull_request:47069|pull_request:47089",
         "pull_request:46196|pull_request:46995",
+        "pull_request:46266|pull_request:47148",
         "pull_request:47093|pull_request:47114",
-        "pull_request:46707|pull_request:47114",
         "pull_request:46707|pull_request:47108",
-        "pull_request:46266|pull_request:46707",
+        "pull_request:46707|pull_request:47114",
+        "pull_request:47077|pull_request:47149",
+        "pull_request:46707|pull_request:47148",
         "pull_request:47089|pull_request:47113",
+        "pull_request:46266|pull_request:46707",
         "pull_request:46995|pull_request:47143",
         "pull_request:47082|pull_request:47108",
+        "pull_request:47114|pull_request:47149",
         "pull_request:47108|pull_request:47114",
         "pull_request:47077|pull_request:47114",
-        "pull_request:47093|pull_request:47143",
         "pull_request:46912|pull_request:47025",
+        "pull_request:47093|pull_request:47143",
         "pull_request:47134|pull_request:47135",
-        "pull_request:46707|pull_request:46957",
         "pull_request:46658|pull_request:47114",
+        "pull_request:46196|pull_request:47148",
         "pull_request:46707|pull_request:47082",
         "pull_request:46266|pull_request:47114",
         "pull_request:47069|pull_request:47136",
         "pull_request:46912|pull_request:47082",
+        "pull_request:47082|pull_request:47148",
         "pull_request:46912|pull_request:47078",
-        "pull_request:43451|pull_request:47082",
         "pull_request:47114|pull_request:47136",
         "pull_request:47035|pull_request:47089",
         "pull_request:46895|pull_request:47134",
-        "pull_request:43451|pull_request:46995",
         "pull_request:47100|pull_request:47134",
         "pull_request:46196|pull_request:47025",
+        "pull_request:43451|pull_request:46995",
         "pull_request:47100|pull_request:47142",
         "pull_request:46895|pull_request:47137",
         "pull_request:46836|pull_request:46995",
         "pull_request:46707|pull_request:47025",
+        "pull_request:47108|pull_request:47148",
         "pull_request:46266|pull_request:47034",
         "pull_request:47034|pull_request:47108",
-        "pull_request:46836|pull_request:47135",
         "pull_request:47035|pull_request:47069",
+        "pull_request:46836|pull_request:47135",
         "pull_request:47035|pull_request:47113",
-        "pull_request:46995|pull_request:47096",
         "pull_request:47100|pull_request:47113",
-        "pull_request:46957|pull_request:47108",
+        "pull_request:46995|pull_request:47096",
+        "pull_request:47142|pull_request:47149",
         "pull_request:47069|pull_request:47113",
         "pull_request:47025|pull_request:47114",
         "pull_request:46912|pull_request:47136",
         "pull_request:46912|pull_request:47077",
-        "pull_request:46707|pull_request:47077",
         "pull_request:47069|pull_request:47082",
-        "pull_request:43451|pull_request:46266",
-        "pull_request:47135|pull_request:47137",
         "pull_request:47089|pull_request:47134",
+        "pull_request:47135|pull_request:47137",
         "pull_request:43451|pull_request:47078",
         "pull_request:47078|pull_request:47096",
         "pull_request:43451|pull_request:47136",
         "pull_request:47096|pull_request:47136",
-        "pull_request:47003|pull_request:47078",
-        "pull_request:43451|pull_request:47135",
-        "pull_request:43451|pull_request:44517",
-        "pull_request:47100|pull_request:47143",
-        "pull_request:46912|pull_request:47100",
-        "pull_request:46912|pull_request:47113",
-        "pull_request:44517|pull_request:47143"
+        "pull_request:46707|pull_request:46957",
+        "pull_request:47100|pull_request:47149"
       ],
-      "prepared_review_unit_hash": "feed5a8ba9af9d5a289ee63759e53f204f472826203e93ba20fc4e329651045e",
+      "prepared_review_unit_hash": "447aa0f8b1af9136370bf1c82818957af07c341e6fae39b81e53a9ac9d616cc4",
       "component_budget": {
-        "node_count": 42,
-        "item_count": 42,
-        "soft_pair_count": 126,
-        "serialized_chars": 149796,
-        "estimated_input_tokens": 37449,
-        "estimated_eval_tokens": 75154
+        "node_count": 44,
+        "item_count": 44,
+        "soft_pair_count": 130,
+        "serialized_chars": 154587,
+        "estimated_input_tokens": 38647,
+        "estimated_eval_tokens": 77550
       },
       "budget": {
-        "node_count": 39,
-        "item_count": 39,
+        "node_count": 40,
+        "item_count": 40,
         "soft_pair_count": 72,
-        "serialized_chars": 113873,
-        "estimated_input_tokens": 28469,
-        "estimated_eval_tokens": 57194
+        "serialized_chars": 114038,
+        "estimated_input_tokens": 28510,
+        "estimated_eval_tokens": 57276
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "Mostly a heterogeneous mix of unrelated PRs; the only clear duplicate cluster is the CohereASR training-loss double-shift fixes. Everything else appears to target different bugs, models, or docs changes.",
-        "confidence": 0.89,
+        "summary": "This cluster is heterogeneous: model additions, documentation-only updates, and a variety of unrelated bugfixes. None of the soft edges look like true duplicates or mergeable into one PR.",
+        "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "The only real duplicate subgroup here is the CohereASR loss double-shift fixes; 47090 is the cleanest anchor because it is focused on that one bug and has an explicit issue link.",
+        "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "47090 is the strongest representative for the duplicate subgroup: it is narrowly scoped to the CohereASR loss path and presents the fix most cleanly.",
+        "best_pr_reason": null,
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46696",
             "right": "pull_request:47134",
             "accept": false,
-            "reason": "Different flash-attn issues: tracing-condition regression vs unpad max-seqlen item() performance fix."
+            "reason": "Same file, but different bugs: 46696 fixes tracing/eager int handling for FA2 decode, while 47134 fixes a separate max_seqlen .item() performance regression."
           },
           {
             "left": "pull_request:46903",
             "right": "pull_request:47090",
-            "accept": true,
-            "reason": "Same underlying CohereASR training-loss double-shift bug in the same forward/loss path; implementation differs but the fix goal is the same."
+            "accept": false,
+            "reason": "Both touch CohereASR, but 46903 is a broader multi-model loss-shift change that also affects PPFormulaNet; 47090 is a CohereASR-only fix with different implementation details."
           },
           {
             "left": "pull_request:46903",
             "right": "pull_request:46958",
-            "accept": true,
-            "reason": "Same CohereASR double-shift loss bug; both change the same loss computation to train on unshifted labels."
+            "accept": false,
+            "reason": "Same subsystem, but not the same standalone change: 46903 bundles CohereASR with PPFormulaNet, while 46958 is a separate CohereASR loss fix with a different implementation."
           },
           {
             "left": "pull_request:46954",
             "right": "pull_request:46958",
             "accept": false,
-            "reason": "Unrelated changes: GlmASR fp32-module handling vs CohereASR training-loss correction."
+            "reason": "Unrelated fixes: Pytest-v9 strict FP32-module handling in GlmASR vs CohereASR training-loss double shift."
           },
           {
             "left": "pull_request:46958",
             "right": "pull_request:46974",
             "accept": false,
-            "reason": "Different models and bugs: CohereASR loss fix vs Musicgen generate typo."
+            "reason": "Different models and code paths: CohereASR loss computation vs MusicGen generate cache-length typo."
           },
           {
             "left": "pull_request:47072",
             "right": "pull_request:47089",
             "accept": false,
-            "reason": "Unrelated: chat-template enum parsing vs AqlmConfig type-check error messages."
+            "reason": "Different issues in different utilities: chat-schema enum parsing vs AqlmConfig type-error messages."
           },
           {
             "left": "pull_request:47077",
             "right": "pull_request:47078",
             "accept": false,
-            "reason": "Both touch Japanese docs, but they are different documentation changes for different content and scopes."
+            "reason": "Both are Japanese docs edits, but they cover different documentation topics and wording changes, not the same change."
+          },
+          {
+            "left": "pull_request:47072",
+            "right": "pull_request:47148",
+            "accept": false,
+            "reason": "Different areas: schema enum extraction vs auto-mapping registration for string keys."
           },
           {
             "left": "pull_request:46266",
             "right": "pull_request:46957",
             "accept": false,
-            "reason": "Different subsystems: DINO conversion mapping cleanup vs finegrained FP8 generation changes."
+            "reason": "Different modernizations in unrelated code paths: DINOv2 conversion mapping vs generation decode switching for finegrained FP8."
+          },
+          {
+            "left": "pull_request:47114",
+            "right": "pull_request:47148",
+            "accept": false,
+            "reason": "Both relate to conversion/auto-mapping docs or mechanics, but one is documentation and the other is an auto-factory API fix; not the same bug/change."
           },
           {
             "left": "pull_request:47072",
             "right": "pull_request:47134",
             "accept": false,
-            "reason": "Unrelated bugs: schema enum parsing vs flash-attn performance regression."
+            "reason": "Unrelated: JSON schema enum parsing vs flash-attention performance fix."
           },
           {
             "left": "pull_request:47114",
             "right": "pull_request:47135",
             "accept": false,
-            "reason": "Docs-only changes in different files and topics; not the same underlying change."
+            "reason": "Different docs/utilities: custom-code conversion mapping docs vs quote rendering in the attention interface docs."
           },
           {
             "left": "pull_request:44517",
             "right": "pull_request:47082",
             "accept": false,
-            "reason": "Different model additions: Qwen3-TTS vs RoMa keypoint matcher."
+            "reason": "Different model additions in different domains: Qwen3 TTS vs RoMa keypoint matcher."
           },
           {
             "left": "pull_request:47034",
             "right": "pull_request:47035",
             "accept": false,
-            "reason": "Related to Neuron support broadly, but not the same concrete change or code-path problem."
+            "reason": "Related platform, but not the same change: a Neuron finetuning experiment vs torch.compile backend default selection."
           },
           {
             "left": "pull_request:47069",
             "right": "pull_request:47089",
             "accept": false,
-            "reason": "Unrelated: num_labels property typing vs AqlmConfig error-message wording."
+            "reason": "Completely different fixes: num_labels returning None vs AqlmConfig error-message text."
           },
           {
             "left": "pull_request:46196",
             "right": "pull_request:46995",
             "accept": false,
-            "reason": "Different attention backends and platforms: wav2vec2 SDPA dispatch vs MLU support for flash linear attention."
+            "reason": "Different attention-related capabilities: wav2vec2 SDPA dispatch vs MLU support for flash-linear-attention availability."
           },
           {
-            "left": "pull_request:47093",
-            "right": "pull_request:47114",
+            "left": "pull_request:46266",
+            "right": "pull_request:47148",
             "accept": false,
-            "reason": "Unrelated: new OpenPangu-V2 model support vs docs about custom-code conversion mapping."
+            "reason": "Different scope: DINO modernization/conversion work vs a LazyAutoMapping string-key registration fix."
           },
           {
-            "left": "pull_request:46707",
+            "left": "pull_request:47093",
             "right": "pull_request:47114",
             "accept": false,
-            "reason": "Native FSDP2 migration and doc-only conversion-mapping clarification are unrelated."
-          },
-          {
-            "left": "pull_request:47069",
-            "right": "pull_request:47136",
-            "accept": false,
-            "reason": "Different issues: num_labels None behavior vs quicktour wording around device placement."
-          },
-          {
-            "left": "pull_request:46912",
-            "right": "pull_request:47025",
-            "accept": false,
-            "reason": "Unrelated: RoPE test coverage changes vs diffusion gemma test adjustments."
-          },
-          {
-            "left": "pull_request:47134",
-            "right": "pull_request:47135",
-            "accept": false,
-            "reason": "Flash-attn performance fix vs a docs rendering fix; different code paths and intents."
-          },
-          {
-            "left": "pull_request:44517",
-            "right": "pull_request:47143",
-            "accept": false,
-            "reason": "Different new-model additions: Qwen3-TTS vs NeuCodec."
-          },
-          {
-            "left": "pull_request:47100",
-            "right": "pull_request:47134",
-            "accept": false,
-            "reason": "Unrelated: kernels API compatibility fix vs flash-attn performance regression fix."
-          },
-          {
-            "left": "pull_request:46912",
-            "right": "pull_request:47100",
-            "accept": false,
-            "reason": "RoPE test work and kernels deprecation cleanup are unrelated."
+            "reason": "Unrelated: OpenPangu-V2 model addition vs docs about custom-code conversion mapping."
           },
           {
-            "left": "pull_request:47082",
-            "right": "pull_request:47096",
-            "accept": false,
-            "reason": "Different features: RoMa keypoint matcher vs export support for Kimi."
-          },
-          {
-            "left": "pull_request:43451",
-            "right": "pull_request:47082",
-            "accept": false,
-            "reason": "Both add model support, but for different models and different functionality."
-          },
-          {
-            "left": "pull_request:47034",
+            "left": "pull_request:46707",
             "right": "pull_request:47108",
             "accept": false,
-            "reason": "Neuron training experiment vs docs-builder pin comment cleanup are unrelated."
+            "reason": "Different classes of change: native FSDP2 support vs a documentation comment tweak in doc-builder workflow pins."
           },
           {
-            "left": "pull_request:47035",
-            "right": "pull_request:47069",
-            "accept": false,
-            "reason": "Different areas: Neuron compile backend default vs num_labels property semantics."
-          },
-          {
-            "left": "pull_request:47035",
-            "right": "pull_request:47113",
+            "left": "pull_request:46707",
+            "right": "pull_request:47114",
             "accept": false,
-            "reason": "Neuron backend default is unrelated to audio feature/proc requirements cleanup."
+            "reason": "Unrelated: FSDP2 runtime migration vs conversion-mapping documentation."
           },
           {
-            "left": "pull_request:46995",
-            "right": "pull_request:47096",
+            "left": "pull_request:46707",
+            "right": "pull_request:47148",
             "accept": false,
-            "reason": "MLU flash-linear-attention availability vs exportable Kimi support are unrelated."
+            "reason": "Unrelated: native FSDP2 migration vs auto-factory string-key handling."
           },
           {
-            "left": "pull_request:47100",
-            "right": "pull_request:47142",
+            "left": "pull_request:47089",
+            "right": "pull_request:47113",
             "accept": false,
-            "reason": "Different fixes: kernels package API compatibility vs pipeline generator handling."
+            "reason": "Different issues: AqlmConfig validation messages vs audio feature/processor backend requirements."
           },
           {
-            "left": "pull_request:46895",
-            "right": "pull_request:47134",
+            "left": "pull_request:46707",
+            "right": "pull_request:46957",
             "accept": false,
-            "reason": "CohereASR loss fix vs flash-attn performance regression are unrelated."
+            "reason": "Different subsystems: distributed FSDP2 migration vs MoE decode-switching/FP8 generation changes."
           },
           {
-            "left": "pull_request:47100",
-            "right": "pull_request:47143",
+            "left": "pull_request:47096",
+            "right": "pull_request:47136",
             "accept": false,
-            "reason": "Kernels compatibility cleanup is unrelated to adding NeuCodec."
+            "reason": "Different changes: export support for Kimi vs a Quicktour inference wording clarification."
           },
           {
             "left": "pull_request:46912",
-            "right": "pull_request:47113",
+            "right": "pull_request:47025",
             "accept": false,
-            "reason": "RoPE testing changes and audio requirements cleanup do not address the same bug."
+            "reason": "Different goals: broader RoPE testing vs Diffusion Gemma test adjustments."
           },
           {
-            "left": "pull_request:44517",
+            "left": "pull_request:47093",
             "right": "pull_request:47143",
             "accept": false,
-            "reason": "Different model additions and different code paths."
+            "reason": "OpenPangu-V2 model support vs NeuCodec model support; distinct model additions."
           },
           {
-            "left": "pull_request:47082",
-            "right": "pull_request:47108",
+            "left": "pull_request:47134",
+            "right": "pull_request:47135",
             "accept": false,
-            "reason": "RoMa model support and doc-builder comment formatting are unrelated."
+            "reason": "Different fixes in different docs/utilities: flash-attention performance regression vs quote rendering in docs."
           },
           {
-            "left": "pull_request:47069",
-            "right": "pull_request:47082",
+            "left": "pull_request:43451",
+            "right": "pull_request:47078",
             "accept": false,
-            "reason": "num_labels typing fix vs new RoMa model support are unrelated."
+            "reason": "Molmo2 model addition vs Japanese documentation wording cleanup; unrelated."
           },
           {
-            "left": "pull_request:47100",
-            "right": "pull_request:47113",
+            "left": "pull_request:47078",
+            "right": "pull_request:47096",
             "accept": false,
-            "reason": "kernels API deprecation fix vs audio feature/proc requirement annotations are unrelated."
+            "reason": "Different content: Japanese docs wording vs exportability changes for Kimi."
           },
           {
-            "left": "pull_request:46957",
-            "right": "pull_request:47108",
+            "left": "pull_request:46707",
+            "right": "pull_request:47136",
             "accept": false,
-            "reason": "Finegrained FP8 generation changes vs doc-builder pin-comment cleanup are unrelated."
+            "reason": "Distributed/FSDP2 migration vs Quicktour docs wording; unrelated."
           },
           {
             "left": "pull_request:47069",
-            "right": "pull_request:47143",
+            "right": "pull_request:47113",
             "accept": false,
-            "reason": "Different topics: config property typing vs adding NeuCodec."
+            "reason": "Num_labels API tweak vs audio backend requirements; unrelated."
           },
           {
             "left": "pull_request:47025",
             "right": "pull_request:47114",
             "accept": false,
-            "reason": "Diffusion Gemma test cleanup vs docs about conversion mappings are unrelated."
+            "reason": "Diffusion Gemma test cleanup vs conversion-mapping docs; unrelated."
           },
           {
             "left": "pull_request:46912",
             "right": "pull_request:47136",
             "accept": false,
-            "reason": "RoPE testing changes and quicktour wording are unrelated."
+            "reason": "RoPE testing changes vs Quicktour documentation wording; unrelated."
           },
           {
             "left": "pull_request:46912",
             "right": "pull_request:47077",
             "accept": false,
-            "reason": "Different doc/test changes in different locales and model areas; not the same underlying change."
-          },
-          {
-            "left": "pull_request:46707",
-            "right": "pull_request:47077",
-            "accept": false,
-            "reason": "FSDP2 migration work and Japanese DistilBERT docs are unrelated."
+            "reason": "RoPE test coverage vs Japanese DistilBERT docs translation; unrelated."
           },
           {
             "left": "pull_request:47069",
             "right": "pull_request:47082",
             "accept": false,
-            "reason": "Different scope: config API typing vs RoMa model addition."
+            "reason": "Num_labels property fix vs RoMa model addition; unrelated."
           },
           {
-            "left": "pull_request:43451",
-            "right": "pull_request:46266",
+            "left": "pull_request:47089",
+            "right": "pull_request:47134",
             "accept": false,
-            "reason": "Both are model-related, but they add different models and conversion behavior."
+            "reason": "AqlmConfig error-text cleanup vs flash-attention performance fix; unrelated."
           },
           {
             "left": "pull_request:47135",
             "right": "pull_request:47137",
             "accept": false,
-            "reason": "Curly-quote rendering fix and tokenizer legacy special-token handling are unrelated."
-          },
-          {
-            "left": "pull_request:47089",
-            "right": "pull_request:47134",
-            "accept": false,
-            "reason": "AqlmConfig error-message wording vs flash-attn performance fix are unrelated."
+            "reason": "Different fixes: documentation quote rendering vs tokenizer legacy special-token handling."
           },
           {
             "left": "pull_request:43451",
             "right": "pull_request:47078",
             "accept": false,
-            "reason": "Molmo2 model addition and Japanese docs wording changes are unrelated."
+            "reason": "Molmo2 model addition vs Japanese documentation wording cleanup; unrelated."
           },
           {
             "left": "pull_request:47078",
             "right": "pull_request:47096",
             "accept": false,
-            "reason": "Japanese docs wording changes are unrelated to exportable Kimi support."
+            "reason": "Different content: Japanese docs wording vs exportability changes for Kimi."
           },
           {
-            "left": "pull_request:43451",
-            "right": "pull_request:47136",
-            "accept": false,
-            "reason": "Molmo2 model support vs quicktour inference wording are unrelated."
-          },
-          {
-            "left": "pull_request:47096",
-            "right": "pull_request:47136",
-            "accept": false,
-            "reason": "Kimi exportability and quicktour device-placement docs are unrelated."
-          },
-          {
-            "left": "pull_request:47003",
-            "right": "pull_request:47078",
-            "accept": false,
-            "reason": "OpenVINO exporter work and Japanese docs wording changes are unrelated."
-          },
-          {
-            "left": "pull_request:43451",
-            "right": "pull_request:47135",
+            "left": "pull_request:46707",
+            "right": "pull_request:46957",
             "accept": false,
-            "reason": "Molmo2 model addition and docs quote rendering fix are unrelated."
+            "reason": "Distributed/FSDP2 migration vs generation expert-kernel switching; unrelated."
           },
           {
-            "left": "pull_request:43451",
-            "right": "pull_request:44517",
+            "left": "pull_request:47142",
+            "right": "pull_request:47149",
             "accept": false,
-            "reason": "Different model additions: Molmo2 vs Qwen3-TTS."
+            "reason": "Pipelines generator handling cleanup vs MoE decode switching docs; unrelated."
           },
           {
-            "left": "pull_request:47078",
+            "left": "pull_request:47069",
             "right": "pull_request:47113",
             "accept": false,
-            "reason": "Japanese docs edits and audio requirements fixes are unrelated."
-          },
-          {
-            "left": "pull_request:47082",
-            "right": "pull_request:47096",
-            "accept": false,
-            "reason": "Different model support PRs for different architectures; not mergeable as one bug fix."
-          },
-          {
-            "left": "pull_request:47003",
-            "right": "pull_request:47142",
-            "accept": false,
-            "reason": "OpenVINO exporter support and pipeline generator cleanup are unrelated."
-          },
-          {
-            "left": "pull_request:46895",
-            "right": "pull_request:47134",
-            "accept": false,
-            "reason": "CohereASR loss fix and flash-attn performance regression are unrelated."
+            "reason": "Num_labels API tweak vs audio backend requirements; unrelated."
           },
           {
-            "left": "pull_request:47100",
-            "right": "pull_request:47143",
+            "left": "pull_request:47025",
+            "right": "pull_request:47114",
             "accept": false,
-            "reason": "Kernels compatibility fix vs new NeuCodec model support are unrelated."
+            "reason": "Diffusion Gemma test cleanup vs conversion-mapping docs; unrelated."
           },
           {
             "left": "pull_request:46912",
-            "right": "pull_request:47100",
+            "right": "pull_request:47136",
             "accept": false,
-            "reason": "RoPE tests and kernels API cleanup are unrelated."
+            "reason": "RoPE testing changes vs Quicktour documentation wording; unrelated."
           },
           {
             "left": "pull_request:46912",
-            "right": "pull_request:47113",
+            "right": "pull_request:47077",
             "accept": false,
-            "reason": "RoPE tests and audio requirement annotations are unrelated."
+            "reason": "RoPE test coverage vs Japanese DistilBERT docs translation; unrelated."
           },
           {
-            "left": "pull_request:44517",
-            "right": "pull_request:47143",
+            "left": "pull_request:47069",
+            "right": "pull_request:47082",
             "accept": false,
-            "reason": "Different new model families and different implementation scopes."
+            "reason": "Num_labels property fix vs RoMa model addition; unrelated."
           },
           {
-            "left": "pull_request:47082",
-            "right": "pull_request:47108",
+            "left": "pull_request:47089",
+            "right": "pull_request:47134",
             "accept": false,
-            "reason": "RoMa model support and doc-builder comment formatting are unrelated."
+            "reason": "AqlmConfig error-text cleanup vs flash-attention performance fix; unrelated."
           },
           {
-            "left": "pull_request:47069",
-            "right": "pull_request:47082",
+            "left": "pull_request:47135",
+            "right": "pull_request:47137",
             "accept": false,
-            "reason": "Config property typing fix vs RoMa model addition are unrelated."
+            "reason": "Different fixes: documentation quote rendering vs tokenizer legacy special-token handling."
           },
           {
-            "left": "pull_request:47100",
-            "right": "pull_request:47113",
+            "left": "pull_request:43451",
+            "right": "pull_request:47078",
             "accept": false,
-            "reason": "kernels API fix vs audio feature/proc requirements changes are unrelated."
+            "reason": "Molmo2 model addition vs Japanese documentation wording cleanup; unrelated."
           },
           {
-            "left": "pull_request:46957",
-            "right": "pull_request:47108",
+            "left": "pull_request:47078",
+            "right": "pull_request:47096",
             "accept": false,
-            "reason": "Finegrained FP8 generation changes and doc-builder pin-comment cleanup are unrelated."
+            "reason": "Different content: Japanese docs wording vs exportability changes for Kimi."
           },
           {
-            "left": "pull_request:47069",
-            "right": "pull_request:47143",
+            "left": "pull_request:46707",
+            "right": "pull_request:46957",
             "accept": false,
-            "reason": "Config typing fix and NeuCodec addition are unrelated."
+            "reason": "Distributed/FSDP2 migration vs generation expert-kernel switching; unrelated."
           },
           {
-            "left": "pull_request:47025",
-            "right": "pull_request:47114",
+            "left": "pull_request:47142",
+            "right": "pull_request:47149",
             "accept": false,
-            "reason": "Test cleanup for diffusion gemma and docs about conversion mappings are unrelated."
+            "reason": "Pipelines generator handling cleanup vs MoE decode switching docs; unrelated."
           }
         ]
       },
-      "evaluator_result": {
-        "accept": true,
-        "feedback": "Grounded and conservative overall. The CohereASR double-shift group is supported by the packet, and the non-Cohere pairs are mostly different bugs/models/docs with no clear mergeability for a single bugfix. I did not see an overstatement beyond the evidence provided."
-      },
-      "evaluator_used": true,
+      "evaluator_result": null,
+      "evaluator_used": false,
       "retried": false,
-      "accepted_nontrivial_soft_edge": true,
+      "accepted_nontrivial_soft_edge": false,
       "error_kind": null,
       "error_message": null
     },
@@ -22906,7 +22814,7 @@
       "source": "llm",
       "cache_hit": false,
       "model": "gpt-5.4-mini?service_tier=flex",
-      "cluster_id": "cluster-40546-42",
+      "cluster_id": "cluster-40546-44",
       "nodes": [
         "pull_request:40546",
         "pull_request:43451",
@@ -22925,17 +22833,28 @@
         "pull_request:47034",
         "pull_request:47035",
         "pull_request:47077",
+        "pull_request:47078",
         "pull_request:47090",
         "pull_request:47092",
         "pull_request:47096",
+        "pull_request:47100",
         "pull_request:47113",
         "pull_request:47134",
+        "pull_request:47135",
         "pull_request:47137",
         "pull_request:47142",
-        "pull_request:47143"
+        "pull_request:47143",
+        "pull_request:47148",
+        "pull_request:47149"
       ],
       "soft_pairs": [
+        "pull_request:47003|pull_request:47078",
+        "pull_request:43451|pull_request:47135",
+        "pull_request:43451|pull_request:44517",
+        "pull_request:47100|pull_request:47143",
+        "pull_request:46912|pull_request:47113",
         "pull_request:46912|pull_request:47142",
+        "pull_request:44517|pull_request:47143",
         "pull_request:47113|pull_request:47137",
         "pull_request:44517|pull_request:46836",
         "pull_request:44517|pull_request:47096",
@@ -22943,11 +22862,10 @@
         "pull_request:46957|pull_request:47096",
         "pull_request:46957|pull_request:47003",
         "pull_request:47134|pull_request:47137",
+        "pull_request:47148|pull_request:47149",
         "pull_request:46658|pull_request:46895",
         "pull_request:43451|pull_request:47142",
         "pull_request:47096|pull_request:47142",
-        "pull_request:43451|pull_request:47134",
-        "pull_request:47003|pull_request:47142",
         "pull_request:46880|pull_request:47092",
         "pull_request:46895|pull_request:47090",
         "pull_request:46958|pull_request:47090",
@@ -22962,10 +22880,9 @@
         "pull_request:46836|pull_request:47092",
         "pull_request:43451|pull_request:47077",
         "pull_request:46836|pull_request:47096",
-        "pull_request:40546|pull_request:43451",
         "pull_request:43451|pull_request:46196",
         "pull_request:46196|pull_request:47096",
-        "pull_request:46196|pull_request:47003",
+        "pull_request:40546|pull_request:43451",
         "pull_request:43451|pull_request:46658",
         "pull_request:46658|pull_request:47096",
         "pull_request:43451|pull_request:47143",
@@ -22990,154 +22907,184 @@
         "pull_request:46836|pull_request:47003",
         "pull_request:46836|pull_request:47137"
       ],
-      "prepared_review_unit_hash": "04bc256b7944e02d8de37ca423d3e361dd3131689091b1a5959e9fc588339b0b",
+      "prepared_review_unit_hash": "cdc3b4ba341ef626a5fa4f4dfcca6bbf516c8660900b12bc9605af44b385fa74",
       "component_budget": {
-        "node_count": 42,
-        "item_count": 42,
-        "soft_pair_count": 126,
-        "serialized_chars": 149796,
-        "estimated_input_tokens": 37449,
-        "estimated_eval_tokens": 75154
+        "node_count": 44,
+        "item_count": 44,
+        "soft_pair_count": 130,
+        "serialized_chars": 154587,
+        "estimated_input_tokens": 38647,
+        "estimated_eval_tokens": 77550
       },
       "budget": {
-        "node_count": 25,
-        "item_count": 25,
-        "soft_pair_count": 54,
-        "serialized_chars": 80888,
-        "estimated_input_tokens": 20222,
-        "estimated_eval_tokens": 40700
+        "node_count": 30,
+        "item_count": 30,
+        "soft_pair_count": 58,
+        "serialized_chars": 93214,
+        "estimated_input_tokens": 23304,
+        "estimated_eval_tokens": 46864
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "This cluster is mostly a false-positive mix of unrelated PRs that happen to share broad subsystems (docs/model additions, exporters, tests, training infra). The only true duplicate set is the CohereASR training-loss double-shift fixes: #46895, #46958, and #47090 all target issue 46894 and the same code path.",
-        "confidence": 0.97,
+        "summary": "The set is mostly a mix of unrelated model additions, docs changes, backend plumbing, and test tweaks. The only clear duplicate cluster is the CohereASR training-loss double-shift fixes (46895, 46958, 47090), all targeting the same issue and code path.",
+        "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #47090 is the best representative of the only real duplicate group: it directly fixes the CohereASR double-shift bug, has the explicit issue target #46894, and keeps the scope focused on the loss computation path.",
+        "canonical_pr_reason": "PR 47090 is the clearest and most complete fix for the CohereASR double-shift bug, with an explicit explanation and correct loss-call wiring; 46895 and 46958 are earlier variants of the same change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #47090 is the strongest standalone PR in this mixed cluster because it is the clearest, most explicit fix for the only confirmed duplicate bug here.",
+        "best_pr_reason": "PR 47090 is the best representative because it directly fixes the underlying CohereASR training-loss bug without the earlier ambiguity, and it cleanly subsumes the previous variants.",
         "soft_edge_verdicts": [
+          {
+            "left": "pull_request:47003",
+            "right": "pull_request:47078",
+            "accept": false,
+            "reason": "OpenVINO exporter work and Japanese DistilBERT docs are unrelated changes."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47135",
+            "accept": false,
+            "reason": "Different domains: Molmo2 model support vs a docs quote-rendering fix."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:44517",
+            "accept": false,
+            "reason": "Separate model additions for different architectures; only broad auto-mapping/docs overlap."
+          },
+          {
+            "left": "pull_request:47100",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "Kernel API cleanup is unrelated to NeuCodec model support."
+          },
+          {
+            "left": "pull_request:46912",
+            "right": "pull_request:47113",
+            "accept": false,
+            "reason": "RoPE testing updates and audio requirements fixes are unrelated."
+          },
           {
             "left": "pull_request:46912",
             "right": "pull_request:47142",
             "accept": false,
-            "reason": "Unrelated changes: RoPE-test coverage vs pipeline generator cleanup; shared infrastructure is incidental."
+            "reason": "RoPE test coverage and pipeline generator cleanup are different code paths."
+          },
+          {
+            "left": "pull_request:44517",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "Qwen3-TTS and NeuCodec are distinct model integrations."
           },
           {
             "left": "pull_request:47113",
             "right": "pull_request:47137",
             "accept": false,
-            "reason": "Different fixes in different areas; one is audio feature/processing requirements, the other is a tokenizer special-token bug."
+            "reason": "Audio requirements changes and tokenizer special-token fix are unrelated."
           },
           {
             "left": "pull_request:44517",
             "right": "pull_request:46836",
             "accept": false,
-            "reason": "Both add new model support, but for different models; similarity is only in boilerplate auto-mapping/docs."
+            "reason": "Different model additions with no same bug/change."
           },
           {
             "left": "pull_request:44517",
             "right": "pull_request:47096",
             "accept": false,
-            "reason": "Completely different changes: Qwen3-TTS model support vs making Kimi exportable."
+            "reason": "Qwen3-TTS model support and Kimi exportability are unrelated."
           },
           {
             "left": "pull_request:43451",
             "right": "pull_request:46957",
             "accept": false,
-            "reason": "Different model additions/features; shared files are generic model-registration plumbing."
+            "reason": "Molmo2 support and FP8 decode optimization are unrelated."
           },
           {
             "left": "pull_request:46957",
             "right": "pull_request:47096",
             "accept": false,
-            "reason": "Finegrained FP8/exporter work vs Kimi exportability are not the same bug or change."
+            "reason": "Different subsystems: FP8/MoE generation vs exportable Kimi."
           },
           {
             "left": "pull_request:46957",
             "right": "pull_request:47003",
             "accept": false,
-            "reason": "Different export/runtime work: FP8 kernels vs OpenVINO exporter support."
+            "reason": "FP8/MoE decode optimization is unrelated to OpenVINO exporter work."
           },
           {
             "left": "pull_request:47134",
             "right": "pull_request:47137",
             "accept": false,
-            "reason": "Flash-attention performance tweak is unrelated to tokenizer special-token handling."
+            "reason": "Flash-attention performance fix and tokenizer method regression are unrelated."
+          },
+          {
+            "left": "pull_request:47148",
+            "right": "pull_request:47149",
+            "accept": false,
+            "reason": "Auto-mapping registration typing fix is unrelated to MoE decode switching docs."
           },
           {
             "left": "pull_request:46658",
             "right": "pull_request:46895",
             "accept": false,
-            "reason": "Step-3.7 Flash implementation work is unrelated to CohereASR loss-shift fixes."
+            "reason": "Step-3.7-Flash model integration and CohereASR loss fix are unrelated."
           },
           {
             "left": "pull_request:43451",
             "right": "pull_request:47142",
             "accept": false,
-            "reason": "Molmo2 model integration and pipeline cleanup are unrelated."
+            "reason": "Model addition work and pipeline generator cleanup are unrelated."
           },
           {
             "left": "pull_request:47096",
             "right": "pull_request:47142",
             "accept": false,
-            "reason": "Exportability changes for a model are unrelated to pipeline generator refactoring."
-          },
-          {
-            "left": "pull_request:43451",
-            "right": "pull_request:47134",
-            "accept": false,
-            "reason": "Molmo2 model addition and flash-attention performance regression fix are unrelated."
-          },
-          {
-            "left": "pull_request:47003",
-            "right": "pull_request:47142",
-            "accept": false,
-            "reason": "OpenVINO exporter work is unrelated to pipeline generator cleanup."
+            "reason": "Exportability changes for Kimi and pipeline generator cleanup are unrelated."
           },
           {
             "left": "pull_request:46880",
             "right": "pull_request:47092",
             "accept": false,
-            "reason": "Both are test expectation updates, but for different models and different behavior."
+            "reason": "Both add XPU expectations, but for different models and test cases."
           },
           {
             "left": "pull_request:46895",
             "right": "pull_request:47090",
             "accept": true,
-            "reason": "Same underlying CohereASR double-shift bug; both target issue 46894 and patch the same forward/loss path."
+            "reason": "Both address the same CohereASR training-loss double-shift bug in the same files and target the same issue."
           },
           {
             "left": "pull_request:46958",
             "right": "pull_request:47090",
             "accept": true,
-            "reason": "Same CohereASR training-loss double-shift fix on the same code path, with the same explicit issue target."
+            "reason": "Same CohereASR double-shift fix, just a later refinement of the same code path."
           },
           {
             "left": "pull_request:46895",
             "right": "pull_request:46958",
             "accept": true,
-            "reason": "Same bug, same files, same issue target 46894; only the implementation style differs."
+            "reason": "Both are variants of the same CohereASR loss-shift correction."
           },
           {
             "left": "pull_request:47025",
             "right": "pull_request:47092",
             "accept": false,
-            "reason": "Different model test fixes; both touch expectations, but not the same model or code path."
+            "reason": "Diffusion Gemma test cleanup and MiniCPM3 XPU expectations are unrelated."
           },
           {
             "left": "pull_request:46707",
             "right": "pull_request:47034",
             "accept": false,
-            "reason": "Native FSDP2 migration and Neuron SFT experiment are unrelated training/runtime changes."
+            "reason": "Native FSDP2 migration and Neuron SFT experiment are different backend initiatives."
           },
           {
             "left": "pull_request:46658",
             "right": "pull_request:46836",
             "accept": false,
-            "reason": "Different model additions; shared auto-registration files are boilerplate, not evidence of the same change."
+            "reason": "Step-3.7-Flash and Unlimited OCR are different model additions."
           },
           {
             "left": "pull_request:43451",
@@ -23149,133 +23096,181 @@
             "left": "pull_request:46658",
             "right": "pull_request:47143",
             "accept": false,
-            "reason": "Step-3.7 Flash model work is unrelated to NeuCodec support."
+            "reason": "Step-3.7-Flash and NeuCodec are unrelated model support PRs."
           },
           {
             "left": "pull_request:46836",
-            "right": "pull_request:47143",
+            "right": "pull_request:47092",
+            "accept": false,
+            "reason": "Unlimited OCR model support is unrelated to MiniCPM3 test expectations."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47077",
             "accept": false,
-            "reason": "Unlimited OCR and NeuCodec are distinct model additions."
+            "reason": "Molmo2 model support is unrelated to Japanese DistilBERT docs."
+          },
+          {
+            "left": "pull_request:46836",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Unlimited OCR support and Kimi exportability are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:46196",
+            "accept": false,
+            "reason": "Molmo2 support and wav2vec2 SDPA dispatch are unrelated."
           },
           {
             "left": "pull_request:46196",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Attention SDPA dispatch and exportable Kimi are different code paths."
+          },
+          {
+            "left": "pull_request:40546",
+            "right": "pull_request:43451",
+            "accept": false,
+            "reason": "VibeVoice and Molmo2 are separate model support additions."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:46658",
+            "accept": false,
+            "reason": "Different new-model integration work, not the same change."
+          },
+          {
+            "left": "pull_request:46658",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Step-3.7-Flash and Kimi exportability are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
             "right": "pull_request:47143",
             "accept": false,
-            "reason": "Wav2Vec2 SDPA dispatch support is unrelated to adding NeuCodec."
+            "reason": "Molmo2 and NeuCodec are different model additions."
           },
           {
-            "left": "pull_request:46836",
+            "left": "pull_request:43451",
             "right": "pull_request:47092",
             "accept": false,
-            "reason": "Unlimited OCR model support and MiniCPM3 XPU test expectations are unrelated."
+            "reason": "Molmo2 model support and MiniCPM3 XPU expectations are unrelated."
           },
           {
-            "left": "pull_request:43451",
-            "right": "pull_request:47077",
+            "left": "pull_request:47092",
+            "right": "pull_request:47143",
             "accept": false,
-            "reason": "Molmo2 model support and Japanese DistilBERT docs translation are unrelated."
+            "reason": "MiniCPM3 test expectations and NeuCodec model support are unrelated."
           },
           {
-            "left": "pull_request:46836",
+            "left": "pull_request:43451",
             "right": "pull_request:47096",
             "accept": false,
-            "reason": "Different domains: Unlimited OCR model support vs Kimi exportability."
+            "reason": "Molmo2 support and Kimi exportability are unrelated."
           },
           {
-            "left": "pull_request:40546",
-            "right": "pull_request:43451",
+            "left": "pull_request:43451",
+            "right": "pull_request:47003",
             "accept": false,
-            "reason": "VibeVoice and Molmo2 are separate model additions."
+            "reason": "Molmo2 model support and OpenVINO exporter are unrelated."
           },
           {
             "left": "pull_request:43451",
-            "right": "pull_request:46196",
+            "right": "pull_request:47137",
             "accept": false,
-            "reason": "Molmo2 model support and wav2vec2 SDPA dispatch are unrelated."
+            "reason": "Model addition work and tokenizer property/test fix are unrelated."
           },
           {
-            "left": "pull_request:46196",
+            "left": "pull_request:47003",
             "right": "pull_request:47096",
             "accept": false,
-            "reason": "SDPA dispatch for wav2vec2 and Kimi exportability are unrelated changes."
+            "reason": "OpenVINO exporter and Kimi exportability touch different export stacks."
           },
           {
-            "left": "pull_request:46196",
-            "right": "pull_request:47003",
+            "left": "pull_request:47096",
+            "right": "pull_request:47137",
             "accept": false,
-            "reason": "Wav2Vec2 SDPA dispatch and OpenVINO exporter support are unrelated."
+            "reason": "Kimi exportability and tokenizer special-token fix are unrelated."
           },
           {
-            "left": "pull_request:43451",
-            "right": "pull_request:46658",
+            "left": "pull_request:47096",
+            "right": "pull_request:47143",
             "accept": false,
-            "reason": "Molmo2 and Step-3.7 Flash are different model integrations."
+            "reason": "Kimi exportability and NeuCodec support are unrelated."
           },
           {
             "left": "pull_request:43451",
-            "right": "pull_request:47143",
+            "right": "pull_request:47034",
             "accept": false,
-            "reason": "Molmo2 and NeuCodec are distinct model additions."
+            "reason": "Molmo2 support and Neuron SFT experiment are unrelated."
           },
           {
-            "left": "pull_request:47092",
-            "right": "pull_request:47143",
+            "left": "pull_request:47034",
+            "right": "pull_request:47096",
             "accept": false,
-            "reason": "Different test expectations for MiniCPM3 vs new NeuCodec support."
+            "reason": "Neuron training experiment and Kimi exportability are unrelated."
           },
           {
-            "left": "pull_request:43451",
+            "left": "pull_request:47092",
             "right": "pull_request:47096",
             "accept": false,
-            "reason": "Molmo2 model integration and Kimi exportability are unrelated."
+            "reason": "Different models and test/export concerns; not the same underlying change."
           },
           {
             "left": "pull_request:43451",
-            "right": "pull_request:47003",
+            "right": "pull_request:47035",
+            "accept": false,
+            "reason": "Molmo2 support and Neuron torch-compile default are unrelated."
+          },
+          {
+            "left": "pull_request:47035",
+            "right": "pull_request:47096",
             "accept": false,
-            "reason": "Molmo2 model support and OpenVINO exporter infrastructure are unrelated."
+            "reason": "Neuron compile default and Kimi exportability are unrelated."
           },
           {
             "left": "pull_request:43451",
-            "right": "pull_request:47137",
+            "right": "pull_request:46912",
             "accept": false,
-            "reason": "Molmo2 model addition and tokenizer special-token fix are unrelated."
+            "reason": "Molmo2 model support and RoPE test work are unrelated."
           },
           {
             "left": "pull_request:47003",
             "right": "pull_request:47137",
             "accept": false,
-            "reason": "OpenVINO exporter work and tokenizer special-token handling are unrelated."
+            "reason": "OpenVINO exporter and tokenizer property fix are unrelated."
           },
           {
             "left": "pull_request:47003",
             "right": "pull_request:47034",
             "accept": false,
-            "reason": "OpenVINO exporter infrastructure and Neuron SFT experiment are unrelated."
+            "reason": "OpenVINO exporter and Neuron SFT experiment are unrelated."
           },
           {
             "left": "pull_request:47003",
             "right": "pull_request:47143",
             "accept": false,
-            "reason": "OpenVINO exporter work is unrelated to NeuCodec model support."
+            "reason": "OpenVINO exporter and NeuCodec model support are unrelated."
           },
           {
             "left": "pull_request:47137",
             "right": "pull_request:47143",
             "accept": false,
-            "reason": "Tokenizer special-token fix and NeuCodec model support are unrelated."
+            "reason": "Tokenizer special-token fix and NeuCodec support are unrelated."
           },
           {
             "left": "pull_request:46836",
             "right": "pull_request:47003",
             "accept": false,
-            "reason": "Two different model/export feature additions; shared plumbing files are incidental."
+            "reason": "Unlimited OCR model support and OpenVINO exporter are unrelated."
           },
           {
             "left": "pull_request:46836",
             "right": "pull_request:47137",
             "accept": false,
-            "reason": "Unlimited OCR model support and tokenizer special-token fix are unrelated."
+            "reason": "Unlimited OCR support and tokenizer special-token fix are unrelated."
           }
         ]
       },
@@ -23305,7 +23300,7 @@
       "soft_pairs": [
         "pull_request:40755|pull_request:41224"
       ],
-      "prepared_review_unit_hash": "c469f4530a1e75f643eb325f300a43ef611914020815a5fbcf40d19343f74b21",
+      "prepared_review_unit_hash": "dcdf09dbc2c1b83666e5cfd824d732813dd0fa1245667936e4b27caa76c90040",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -23327,18 +23322,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "The cluster contains two unrelated pull requests: one adds TimesFM covariate forecasting support, while the other adds DINOv3 image classification support. They touch different model families, code paths, and user-facing features, so this is a false positive similarity cluster.",
+        "summary": "The two PRs are not duplicates: one adds TimesFM covariate forecasting support, while the other adds DINOv3 image classification support. They touch different models, code paths, and use cases.",
         "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": null,
+        "canonical_pr_reason": "PR 41224 is the cleaner representative of this cluster if one must be chosen: it is a self-contained feature addition for DINOv3 image classification with matching code, docs, and tests.",
         "best_issue_reason": null,
-        "best_pr_reason": null,
+        "best_pr_reason": "PR 41224 is the stronger standalone PR to surface because its change is narrowly scoped and clearly validated by the affected model and tests, whereas 40755 is a separate TimesFM feature request.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:40755",
             "right": "pull_request:41224",
             "accept": false,
-            "reason": "Different models and different changes: TimesFM forecasting with covariates vs. DINOv3 image classification support. No shared concrete bug or mergeable code-path overlap."
+            "reason": "Different subsystems and different concrete changes: TimesFM covariate forecasting vs DINOv3 image classification. They do not fix the same bug or implement the same feature."
           }
         ]
       },
@@ -23457,142 +23452,142 @@
         "pull_request:43823|pull_request:45213",
         "pull_request:43251|pull_request:43254"
       ],
-      "prepared_review_unit_hash": "7fa3587f2995b043005f952e2fa7d8e211c7411df3ed291b20bdf8d49e2f3f83",
+      "prepared_review_unit_hash": "a08b617106a1567988a5969491313c2b251a6cb3143fc75210273334fe3e1681",
       "component_budget": {
         "node_count": 12,
         "item_count": 12,
         "soft_pair_count": 17,
-        "serialized_chars": 31600,
-        "estimated_input_tokens": 7900,
-        "estimated_eval_tokens": 16056
+        "serialized_chars": 31596,
+        "estimated_input_tokens": 7899,
+        "estimated_eval_tokens": 16054
       },
       "budget": {
         "node_count": 12,
         "item_count": 12,
         "soft_pair_count": 17,
-        "serialized_chars": 31600,
-        "estimated_input_tokens": 7900,
-        "estimated_eval_tokens": 16056
+        "serialized_chars": 31596,
+        "estimated_input_tokens": 7899,
+        "estimated_eval_tokens": 16054
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "This cluster is mostly heterogeneous and does not form one duplicate group. The only strong duplicate-like pair is the cross-entropy helper fix: PRs 43251 and 43254 both patch `fixed_cross_entropy` to forward supported kwargs to `nn.functional.cross_entropy` and could plausibly be merged into one change.",
-        "confidence": 0.91,
+        "summary": "This cluster is mostly heterogeneous PRs with superficial similarity; the only clear duplicate pair is the two `fixed_cross_entropy` changes for issue 43240.",
+        "confidence": 0.89,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 43251 is the best canonical PR: it targets the explicit issue `43240`, makes the concrete loss-helper fix, and has the fuller review/discussion trail. PR 43254 is essentially the same fix in a lighter variant.",
+        "canonical_pr_reason": "PR 43251 is the clearest canonical representative of the only true duplicate pair: it targets the same issue (43240) and updates the same loss helper with the broader kwargs forwarding change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 43251 is the strongest representative of the only real duplicate pair in this cluster and is more complete than 43254, which is a near-duplicate implementation of the same bug fix.",
+        "best_pr_reason": "PR 43251 is the best overall PR candidate because it is the most explicit, issue-linked version of the one genuine duplicate change in this set; the rest are unrelated feature/fix PRs.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:41855",
             "right": "pull_request:43424",
             "accept": false,
-            "reason": "Different features: Mistral tokenizer method additions vs ExecuTorch dynamic-shape export tests."
+            "reason": "Different features: Mistral tokenizer methods vs executorch dynamic-shape export tests."
           },
           {
             "left": "pull_request:41855",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Unrelated changes: tokenizer API gaps vs adding a new MobileLLM model."
+            "reason": "Unrelated models and code paths; tokenizer utilities vs adding MobileLLM."
           },
           {
             "left": "pull_request:42668",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Processor auto-loading robustness is unrelated to MobileLLM model addition."
+            "reason": "Both are processor/model-loading related only at a very broad level, but they fix different model families and issues."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Cross-entropy kwargs fix is unrelated to MobileLLM model addition."
+            "reason": "Loss helper kwargs forwarding is unrelated to adding MobileLLM."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "Same broad repo area, but different concrete problems: loss helper kwargs vs Mistral4 test fixes."
+            "reason": "Different code paths: cross-entropy helper vs Mistral4 test/model fixes."
           },
           {
             "left": "pull_request:43424",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "ExecuTorch dynamic-shape export tests and Mistral4 fixes are different changes."
+            "reason": "Executorch export tests and Mistral4 fixes are not the same underlying change."
           },
           {
             "left": "pull_request:43424",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "ExecuTorch export testing is unrelated to MobileLLM model scaffolding."
+            "reason": "Exportability tests vs new model implementation are unrelated."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:45655",
             "accept": false,
-            "reason": "New model addition is unrelated to the config-resolution order fix."
+            "reason": "MobileLLM addition and config-resolution fix solve different problems."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:43995",
             "accept": false,
-            "reason": "MobileLLM model addition and Falcon output-interface refactor are different code changes."
+            "reason": "Different models and objectives; MobileLLM creation vs Falcon output-collection refactor."
           },
           {
             "left": "pull_request:43995",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "Falcon refactor is unrelated to the model-creation skill/documentation PR."
+            "reason": "Falcon refactor and model-creation skill are unrelated."
           },
           {
             "left": "pull_request:43995",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "Different models and different goals: Falcon interface refactor vs Mistral4 test fixes."
+            "reason": "Falcon output interface refactor is not the same as Mistral4 test fixes."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45170",
             "accept": false,
-            "reason": "Mistral4 test fixes do not match the CLIP-like `layrnorm` renaming/conversion mapping fix."
+            "reason": "Mistral4 test fixes and `layrnorm` renaming are different bug fixes."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:46048",
             "accept": false,
-            "reason": "Mistral4 test fixes and TDT loss kernel integration are unrelated."
+            "reason": "Mistral4 test fixes and TDT loss kernel addition are unrelated."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45655",
             "accept": false,
-            "reason": "Mistral4 test fixes are unrelated to class config-resolution ordering."
+            "reason": "Mistral4 test fixes and config-resolution order fix address different code paths."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "Mistral4 test fixes are unrelated to the model-creation skill PR."
+            "reason": "A model test/fix PR is not the same as a model-creation skill PR."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "Model addition for MobileLLM is unrelated to the new-model skill/docs PR."
+            "reason": "Adding MobileLLM is unrelated to the model-creation skill draft."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:43254",
             "accept": true,
-            "reason": "Both PRs fix the same concrete bug in `fixed_cross_entropy` by forwarding supported kwargs to `nn.functional.cross_entropy`; they are plausibly mergeable into one PR."
+            "reason": "Same helper function, same issue target 43240, and both add support for passing cross-entropy kwargs; they are competing implementations of the same fix."
           }
         ]
       },
       "evaluator_result": {
         "accept": true,
-        "feedback": "The summary is grounded in the packet: the cluster is mostly heterogeneous, and the only clearly duplicate-like pair is 43251/43254, which both patch `fixed_cross_entropy` to forward supported kwargs to `nn.functional.cross_entropy` and share the same issue target. The other pairwise rejections stay conservative and do not overstate similarity."
+        "feedback": "The summary is grounded in the packet and stays conservative. The only accepted pair, 43251/43254, is clearly mergeable as the same `fixed_cross_entropy` issue on the same file and issue target 43240. The rest of the false verdicts are appropriately cautious and do not overclaim duplicate status."
       },
       "evaluator_used": true,
       "retried": false,
@@ -23619,7 +23614,7 @@
       "soft_pairs": [
         "pull_request:42975|pull_request:42976"
       ],
-      "prepared_review_unit_hash": "28d5f93a3ddf87f6d4c4b9e7dd2b26580a9261c7600536965a38ffb1d83b87b1",
+      "prepared_review_unit_hash": "76c059122abc6cf823918f4037d7b61addcb7135a01d8fb7455f24526fb2ce65",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -23641,18 +23636,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Two separate maintenance PRs updating GitHub Actions dependencies: one for Node 24 compatibility across workflow actions, and one for newer Docker build/action versions. They overlap on some workflow files but target different action families and different upgrade goals.",
-        "confidence": 0.9,
+        "summary": "Two open PRs both upgrade GitHub Actions in workflow files, but they target different action families and different concrete maintenance goals: Node 24 compatibility versus docker/build-push-action bumps. They overlap at a high level, not as the same underlying change.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 42975 is the clearest canonical PR: its title and diff consistently target the Node 24 compatibility upgrade across workflow actions, with a broad but coherent dependency bump scope.",
+        "canonical_pr_reason": "PR 42975 is the better representative PR: it has the broader, more explicit goal (Node 24 compatibility) and updates several core workflow actions across many files.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 42975 is the better overall representative because it has the most specific, cohesive purpose and broader workflow coverage for the same maintenance class.",
+        "best_pr_reason": "PR 42975 is the best single PR for the cluster because it spans more of the workflow-action upgrade work and most clearly captures the main maintenance theme.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:42975",
             "right": "pull_request:42976",
             "accept": false,
-            "reason": "Both are workflow dependency updates, but they change different actions (actions/* vs docker/build-push-action) for different upgrade goals. Shared workflow filenames alone are not enough to treat them as the same underlying change or merge them into one PR."
+            "reason": "Both are GitHub Actions upgrade PRs, but they fix different concrete changes: 42975 updates actions/checkout, actions/cache, actions/upload-artifact for Node 24 compatibility, while 42976 bumps docker/build-push-action to v6 in docker build workflows. They are related maintenance, not the same underlying change."
           }
         ]
       },
@@ -23687,58 +23682,58 @@
         "pull_request:43448|pull_request:45490",
         "pull_request:45490|pull_request:46562"
       ],
-      "prepared_review_unit_hash": "a9536cec11b21387a704e49002f8a8fb71bdcb7eac6f9eaae9f79c339e2f6f03",
+      "prepared_review_unit_hash": "a230a01ccf2e14f3a5d23cdad9c42f9238c4cb04e32d4f02656bbc597714ce3d",
       "component_budget": {
         "node_count": 4,
         "item_count": 4,
         "soft_pair_count": 4,
-        "serialized_chars": 13040,
-        "estimated_input_tokens": 3260,
-        "estimated_eval_tokens": 6776
+        "serialized_chars": 13047,
+        "estimated_input_tokens": 3262,
+        "estimated_eval_tokens": 6780
       },
       "budget": {
         "node_count": 4,
         "item_count": 4,
         "soft_pair_count": 4,
-        "serialized_chars": 13040,
-        "estimated_input_tokens": 3260,
-        "estimated_eval_tokens": 6776
+        "serialized_chars": 13047,
+        "estimated_input_tokens": 3262,
+        "estimated_eval_tokens": 6780
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "All four items are separate model-addition PRs (ViT NEPA, Molmo, CTSM, Step3p7). They share boilerplate registry/docs changes, but they target different models and codepaths, so they are not duplicates.",
-        "confidence": 0.98,
+        "summary": "All four PRs are separate model-addition changes, not duplicates. They share the usual Hugging Face model-registration/docs/test pattern, but each targets a different architecture and code path.",
+        "confidence": 0.99,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No single canonical PR: each PR introduces a different model family and cannot be treated as the same change.",
+        "canonical_pr_reason": "No true canonical duplicate exists here: ViT NEPA, Molmo, CTSM, and Step3p7 are distinct model integrations rather than the same fix/change.",
         "best_issue_reason": null,
-        "best_pr_reason": "If one must be chosen as a representative, none is a true duplicate of the others; they are separate standalone additions rather than mergeable variants of the same PR.",
+        "best_pr_reason": "45490 is the strongest standalone representative because it has the highest review activity and broad model-registration changes, but it is still not a duplicate of the others.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:42978",
             "right": "pull_request:43448",
             "accept": false,
-            "reason": "Both are model-implementation PRs, but they add different models (ViT NEPA vs Molmo). Shared auto/docs files are generic registration scaffolding, not the same underlying change."
+            "reason": "Different model integrations (ViT NEPA vs Molmo); shared docs/auto-mapping edits are generic add-model boilerplate, not the same bug or change."
           },
           {
             "left": "pull_request:42978",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "These add different models (ViT NEPA vs CTSM). Overlap is limited to common model-registry plumbing, which is not enough to treat them as one PR."
+            "reason": "ViT NEPA and CTSM are unrelated architectures with different model code and conversion paths; overlap is only in registration/docs infrastructure."
           },
           {
             "left": "pull_request:43448",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "Molmo and CTSM are unrelated model additions. The shared files are broad repo-maintenance edits, not a single concrete bug or feature."
+            "reason": "Molmo and CTSM are distinct model additions touching different implementations; same subsystem is not enough to merge as duplicates."
           },
           {
             "left": "pull_request:45490",
             "right": "pull_request:46562",
             "accept": false,
-            "reason": "CTSM and Step3p7 are distinct model integrations with different modeling/processing code. Despite shared auto-registry changes, they are not the same underlying change."
+            "reason": "CTSM vs Step3p7 are different time-series/vision-model additions with separate concrete code paths, so they are not the same PR."
           }
         ]
       },
@@ -23787,130 +23782,130 @@
         "pull_request:46539|pull_request:46922",
         "pull_request:46886|pull_request:46922"
       ],
-      "prepared_review_unit_hash": "92f06f3644f366bea74eba97df55027117fc939b09d8d880c36852cac4edc996",
+      "prepared_review_unit_hash": "b6f88835bf3bac2cd8990f7fa8c687125067563baffaa6ff4098db3559087e09",
       "component_budget": {
         "node_count": 7,
         "item_count": 7,
         "soft_pair_count": 15,
-        "serialized_chars": 19604,
-        "estimated_input_tokens": 4901,
-        "estimated_eval_tokens": 10058
+        "serialized_chars": 19599,
+        "estimated_input_tokens": 4900,
+        "estimated_eval_tokens": 10056
       },
       "budget": {
         "node_count": 7,
         "item_count": 7,
         "soft_pair_count": 15,
-        "serialized_chars": 19604,
-        "estimated_input_tokens": 4901,
-        "estimated_eval_tokens": 10058
+        "serialized_chars": 19599,
+        "estimated_input_tokens": 4900,
+        "estimated_eval_tokens": 10056
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "The cluster is mostly unrelated PRs with only one true duplicate pair: the two YOLOS fixes for the same device_map='auto' mismatch. The other PRs touch different subsystems or are test/CI changes, so they should not be merged as duplicates.",
-        "confidence": 0.93,
+        "summary": "The cluster is mostly heterogeneous: several PRs are unrelated test, CI, generation, FP8, or model-specific changes. The only likely duplicate pair is the YOLOS device-mismatch fixes (46886 and 46922), which target the same issue and file; 46922 is the more complete version.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 46922 is the best canonical representative: it targets the explicit YOLOS device-mismatch issue, fixes the concrete forward-path device placement bug, and includes the model split-avoidance change as part of the same fix.",
+        "canonical_pr_reason": "46922 is the best canonical PR: it targets the concrete YOLOS device-mismatch bug, shares the explicit issue target 46869, and contains the fix plus the supporting split-module change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46922 is the strongest standalone fix: explicit issue target, concrete code-path correction in YOLOS, and more complete than the earlier 46886 patch.",
+        "best_pr_reason": "46922 is the strongest standalone PR because it directly addresses the same YOLOS runtime bug and is more complete than 46886.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43291",
             "right": "pull_request:43665",
             "accept": false,
-            "reason": "Whisper tokenizer/test fix vs unrelated CLIP/ViT test edits; no shared bug or change."
+            "reason": "Different models and different bugs; both are vague/test-focused PRs, but they do not fix the same code-path problem."
           },
           {
             "left": "pull_request:43291",
             "right": "pull_request:44815",
             "accept": false,
-            "reason": "Whisper decode/timestamp fix and FP8 dequantization fix are unrelated subsystems."
+            "reason": "Whisper tokenization/tests versus FP8 dequant/loading logic are unrelated changes."
           },
           {
             "left": "pull_request:43291",
             "right": "pull_request:46380",
             "accept": false,
-            "reason": "Different bugs: Whisper test/tokenizer behavior vs generation watermarking config validation."
+            "reason": "Whisper decode/test changes are unrelated to generation-config watermark validation."
           },
           {
             "left": "pull_request:43291",
             "right": "pull_request:46922",
             "accept": false,
-            "reason": "Whisper decoding vs YOLOS device placement; clearly different code paths and models."
+            "reason": "Whisper tokenization/test work and YOLOS device-mismatch work are different bugs in different subsystems."
           },
           {
             "left": "pull_request:43665",
-            "right": "pull_request:44815",
+            "right": "pull_request:46539",
             "accept": false,
-            "reason": "Bogus test edits are not the same as the FP8 kernel/loading fix."
+            "reason": "Vision model test edits versus Phimoe CI/memory handling are unrelated."
           },
           {
             "left": "pull_request:43665",
-            "right": "pull_request:46380",
+            "right": "pull_request:44815",
             "accept": false,
-            "reason": "Unrelated test-only changes vs generation config bug fix."
+            "reason": "CLIP/ViT test changes do not match the FP8 dequant/loading fix in core model loading."
           },
           {
             "left": "pull_request:43665",
-            "right": "pull_request:46539",
+            "right": "pull_request:46380",
             "accept": false,
-            "reason": "Unrelated test-only changes vs CI/integration-test infrastructure fix."
+            "reason": "These touch unrelated areas: vision model tests versus generation/watermark validation and Whisper code."
           },
           {
             "left": "pull_request:43665",
             "right": "pull_request:46922",
             "accept": false,
-            "reason": "Unrelated test-only changes vs YOLOS device mismatch fix."
+            "reason": "Generic vision test changes are not the same underlying bug as the YOLOS device-mismatch fix."
           },
           {
             "left": "pull_request:44815",
             "right": "pull_request:46539",
             "accept": false,
-            "reason": "FP8 dequantization/kernel fix and Phimoe CI setup are unrelated."
+            "reason": "FP8 dequant/loading changes and Phimoe CI memory tweaks are different issues and code paths."
           },
           {
             "left": "pull_request:44815",
             "right": "pull_request:46380",
             "accept": false,
-            "reason": "Different code paths: FP8 quantization/dequantization vs generation config validation."
+            "reason": "No overlap in bug or change: FP8 quantization versus generation config validation."
           },
           {
             "left": "pull_request:44815",
             "right": "pull_request:46922",
             "accept": false,
-            "reason": "Different model/fix area: FP8 loading vs YOLOS device_map mismatch."
+            "reason": "Different subsystems and different runtime failures; not the same concrete bug."
           },
           {
             "left": "pull_request:46380",
             "right": "pull_request:46539",
             "accept": false,
-            "reason": "Generation config watermarking fix vs CI runner memory workaround; not the same underlying change."
+            "reason": "Generation-config watermark validation and Phimoe CI behavior are unrelated."
           },
           {
             "left": "pull_request:46380",
             "right": "pull_request:46922",
             "accept": false,
-            "reason": "Generation configuration validation and YOLOS device placement are unrelated."
+            "reason": "Generation/watermark validation is unrelated to YOLOS device placement."
           },
           {
             "left": "pull_request:46539",
             "right": "pull_request:46922",
             "accept": false,
-            "reason": "CI infrastructure/test-environment fix vs model-specific YOLOS device mismatch."
+            "reason": "CI memory-limit patches for Phimoe are unrelated to the YOLOS runtime device-mismatch fix."
           },
           {
             "left": "pull_request:46886",
             "right": "pull_request:46922",
             "accept": true,
-            "reason": "Both address the same YOLOS device_map='auto' device mismatch in the same file and issue target; the latter is a more complete version of the same fix."
+            "reason": "Both PRs target the same YOLOS device_map='auto' mismatch in the same file and same issue target; 46922 subsumes 46886 with the additional hidden-state device placement fix."
           }
         ]
       },
       "evaluator_result": {
         "accept": true,
-        "feedback": "The cluster summary is grounded in the packet and stays conservative. The only accepted duplicate pair, pull_request:46886 and pull_request:46922, is clearly mergeable as the same YOLOS device_map='auto' bugfix: both touch the same file, share the same issue target, and 46922 simply adds an additional forward-path device placement correction on top of the split-avoidance change. The other pairwise rejections are also conservative because the PRs concern different subsystems or test/CI changes without clear evidence of a shared bug."
+        "feedback": "The summary is grounded in the packet and stays conservative about the soft edges. The only accepted pair, 46886 vs 46922, is well supported by the shared YOLOS file and explicit issue target 46869, and the two diffs are mergeable as one bugfix. The other pairwise rejections are appropriately cautious given the distinct subsystems and code paths."
       },
       "evaluator_used": true,
       "retried": false,
@@ -23940,52 +23935,52 @@
         "pull_request:43488|pull_request:43532",
         "pull_request:43532|pull_request:46747"
       ],
-      "prepared_review_unit_hash": "4d44094a26993fb90ade7dea1403b30290aa27decccf723bdc7be429e4e687fc",
+      "prepared_review_unit_hash": "4e70d9ceaae39774dc6bed193b1d3cccfb163179521399e4b5e7b5eb4bc316be",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 6348,
-        "estimated_input_tokens": 1587,
-        "estimated_eval_tokens": 3430
+        "serialized_chars": 6349,
+        "estimated_input_tokens": 1588,
+        "estimated_eval_tokens": 3432
       },
       "budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 6348,
-        "estimated_input_tokens": 1587,
-        "estimated_eval_tokens": 3430
+        "serialized_chars": 6349,
+        "estimated_input_tokens": 1588,
+        "estimated_eval_tokens": 3432
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are three unrelated '[don't merge]' PRs touching different parts of the repo (packaging/test setup, comment-driven CI, and PR security-gate workflow), so they should not be clustered as duplicates.",
-        "confidence": 0.96,
+        "summary": "These are three unrelated throwaway PRs: one tweaks packaging/versioning and a vit test, one changes several workflow/testing files to show diffs, and one alters a security-gate workflow trigger. They do not look like the same underlying bug or change.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No canonical PR: the three PRs do not describe the same underlying bug or change, so there isn\u2019t a single representative duplicate to keep.",
+        "canonical_pr_reason": "No canonical PR: the three PRs target different code paths and workflows, so there is no true duplicate representative.",
         "best_issue_reason": null,
-        "best_pr_reason": "No single PR is a good global representative for this cluster because the PRs are about different changes and would not plausibly be merged into one fix.",
+        "best_pr_reason": "43488 is the narrowest self-test/packaging PR, but it does not subsume the other two unrelated workflow changes; there is no strong cluster-level best PR.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43488",
             "right": "pull_request:46747",
             "accept": false,
-            "reason": "Different change sets: 43488 is setup/version/dependency and Vit test tweaks, while 46747 only changes the PR security-gate workflow trigger."
+            "reason": "Different files and intent: setup/version + vit test vs workflow trigger change in pr-ci-security-gate.yml."
           },
           {
             "left": "pull_request:43488",
             "right": "pull_request:43532",
             "accept": false,
-            "reason": "43488 changes packaging/versioning and a test file; 43532 changes comment CI/workflow plumbing and testing utilities. They do not fix the same code-path problem."
+            "reason": "43488 is packaging/versioning plus a test tweak; 43532 rewrites several CI/notification workflows and testing utilities. Not the same change."
           },
           {
             "left": "pull_request:43532",
             "right": "pull_request:46747",
             "accept": false,
-            "reason": "Both touch workflows, but they address different behaviors in different workflows: self-comment CI vs PR security gate. Not the same underlying bug or change."
+            "reason": "Both touch workflows, but 43532 is about comment/diff reporting plumbing while 46747 changes the security gate trigger. Different underlying code-paths."
           }
         ]
       },
@@ -24015,7 +24010,7 @@
       "soft_pairs": [
         "pull_request:43612|pull_request:43613"
       ],
-      "prepared_review_unit_hash": "974400c5bf5d445b994160b8833f17cbe95370e30389295c8719a8e3b269748d",
+      "prepared_review_unit_hash": "46f90e94225e7dd9e2983a90ca151b4d055d69fcab30b792de2503f4d54f5c79",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -24037,18 +24032,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related but distinct feature PRs: one adds a Promptable Concept Segmentation pipeline, the other adds a Promptable Visual Segmentation pipeline. They share pipeline/docs scaffolding, but the concrete models, tests, and task targets differ enough that they should not be merged as duplicates.",
-        "confidence": 0.94,
+        "summary": "Two distinct feature PRs: one adds a Promptable Concept Segmentation pipeline, the other adds a Promptable Visual Segmentation pipeline. They share documentation and pipeline scaffolding, but they target different tasks, models, and tests, so they are not duplicates.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
         "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "No single PR is a true duplicate-canonical choice here; the two PRs implement different pipelines. If forced to pick a representative, 43613 is slightly broader, but it should not subsume 43612.",
+        "best_pr_reason": "pull_request:43613 is the broader and more complete integration, touching more model backends and tracker support, so it is the better representative PR if one must be chosen.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43612",
             "right": "pull_request:43613",
             "accept": false,
-            "reason": "Different underlying changes: concept segmentation vs visual segmentation are separate pipelines with different model/test targets, so they are not the same fix/change."
+            "reason": "Different underlying features: concept segmentation vs visual segmentation. Overlap is only in shared pipeline/docs scaffolding, not the same concrete change."
           }
         ]
       },
@@ -24078,7 +24073,7 @@
       "soft_pairs": [
         "pull_request:43651|pull_request:43663"
       ],
-      "prepared_review_unit_hash": "0191764f489ed65db2f7941d607b8a7ce62eb40df680a763cdd7d8f766d4c064",
+      "prepared_review_unit_hash": "c1d436fa62a2d4dc8b92107a67e120fab9915b712c6a6e6b3f7c36067f650f9a",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -24100,18 +24095,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "The two PRs both touch `trainer.py`, but they address different trainer behaviors: one changes gradient-accumulation loss scaling, the other refactors signature-column extraction for input filtering. They are not the same underlying change.",
-        "confidence": 0.97,
+        "summary": "Two open `trainer.py` PRs, but they address different Trainer extension points: one changes gradient-accumulation loss scaling, the other refactors signature-column filtering for custom trainers. Same file, different code paths, not duplicates.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No single canonical PR: #43651 changes training loss scaling semantics in `training_step`, while #43663 extracts signature-column logic for dataset column filtering. They are separate trainer changes.",
+        "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "#43651 is the stronger standalone PR because it makes a concrete runtime behavior change in the training path. #43663 is a narrower helper extraction/refactor for column filtering.",
+        "best_pr_reason": "PR #43651 is the better representative of the cluster only in the sense that it is a concrete behavior change in training-step loss handling, but it is not a duplicate of #43663; the two PRs solve unrelated Trainer customization problems.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43651",
             "right": "pull_request:43663",
             "accept": false,
-            "reason": "Both touch `trainer.py`, but #43651 fixes gradient-accumulation loss scaling in `training_step` and #43663 adds a signature-column helper for input filtering. Different code paths and different behavior changes, so they are not duplicates."
+            "reason": "Different changes: #43651 adds `_loss_is_scaled_for_ga` and alters loss normalization during training, while #43663 adds `_get_signature_columns` to change column filtering. They touch the same file but not the same underlying bug or mergeable change set."
           }
         ]
       },
@@ -24141,7 +24136,7 @@
       "soft_pairs": [
         "pull_request:43656|pull_request:43842"
       ],
-      "prepared_review_unit_hash": "257b20b42a75416e1eb9b1eec6590deb16402fb668fd55765dd5435195b807bd",
+      "prepared_review_unit_hash": "a6a36337ad5b8dc3f5f1586b6878b623ccbe70f2d36ed9b3b3da85b48cbc6f99",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -24163,27 +24158,24 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs touch `src/transformers/cli/serve.py` and involve `TypeAdapter`-related NameErrors, but they appear to address related runtime/type-annotation problems rather than the same concrete bugfix. Their explicit issue targets also differ (`#43576` vs `#43824`), so the match should be treated cautiously.",
-        "confidence": 0.71,
+        "summary": "Both PRs touch `src/transformers/cli/serve.py` and are about `TypeAdapter`/annotation-related NameErrors, but they address different concrete failure modes with different fixes. They are related, not duplicates.",
+        "confidence": 0.84,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 43842 is the clearest representative because it directly targets the pydantic-optional `TypeAdapter` NameError with a focused fix, while 43656 looks like a broader annotation/runtime cleanup in the same file.",
+        "canonical_pr_reason": "43842 is the clearest anchor: it explicitly targets the `TypeAdapter` NameError when `pydantic` is absent and uses a narrowly scoped runtime-safety fix.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 43842 is the best overall representative for the cluster because it is narrowly scoped to the explicit `TypeAdapter` NameError path and has the more specific issue linkage.",
+        "best_pr_reason": "43842 is the best representative because its title and patch directly describe the bug and the fix is self-contained.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43656",
             "right": "pull_request:43842",
             "accept": false,
-            "reason": "They are related `serve.py` `TypeAdapter` fixes, but the packet does not firmly establish the same underlying bug or that they would be safely merged into one PR; the explicit issue targets differ and the implementation approaches are distinct."
+            "reason": "Same file and same broad `TypeAdapter` NameError theme, but different underlying problems: 43656 adds postponed annotations/type-hint cleanup, while 43842 specifically guards the `TypeAdapter` annotation for the `pydantic`-missing path. They could coexist, so this is not a duplicate."
           }
         ]
       },
-      "evaluator_result": {
-        "accept": false,
-        "feedback": "The packet supports that both PRs touch `src/transformers/cli/serve.py` and are about `TypeAdapter`-related NameErrors, but it does not firmly establish that they are the same bugfix or safely mergeable into one PR. The explicit issue targets differ (`43576` vs `43824`), so the soft-edge verdict and summary should be phrased more cautiously (e.g. \"appears related\" rather than definite)."
-      },
-      "evaluator_used": true,
-      "retried": true,
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
       "accepted_nontrivial_soft_edge": false,
       "error_kind": null,
       "error_message": null
@@ -24354,8 +24346,8 @@
         "pull_request:44010|pull_request:44073",
         "pull_request:43997|pull_request:44074",
         "pull_request:44010|pull_request:44018",
-        "pull_request:43997|pull_request:44129",
         "pull_request:44071|pull_request:44129",
+        "pull_request:43997|pull_request:44129",
         "pull_request:43996|pull_request:44161",
         "pull_request:43996|pull_request:44017",
         "pull_request:44010|pull_request:44044",
@@ -24370,403 +24362,409 @@
         "pull_request:44024|pull_request:44076",
         "pull_request:44073|pull_request:44074"
       ],
-      "prepared_review_unit_hash": "c36d3c6843cf2646e4070a79fbc2a4c2dc4a757e381bd6e424de04c88632642e",
+      "prepared_review_unit_hash": "d13eb9f09d557950649627ed98414cdd2c86b282f0414d98d95b018716e08bd7",
       "component_budget": {
         "node_count": 28,
         "item_count": 28,
         "soft_pair_count": 76,
-        "serialized_chars": 84119,
-        "estimated_input_tokens": 21030,
-        "estimated_eval_tokens": 42316
+        "serialized_chars": 84116,
+        "estimated_input_tokens": 21029,
+        "estimated_eval_tokens": 42314
       },
       "budget": {
         "node_count": 28,
         "item_count": 28,
         "soft_pair_count": 72,
-        "serialized_chars": 76941,
-        "estimated_input_tokens": 19236,
-        "estimated_eval_tokens": 38728
+        "serialized_chars": 76938,
+        "estimated_input_tokens": 19235,
+        "estimated_eval_tokens": 38726
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "This cluster is mostly model-specific output-tracing refactors tied to the standardized output-capture effort. The clearest duplicate-like subclusters are the GPT-J/CodeGen PRs and the GPT-Neo PRs; a few other pairs share the same refactor theme, but they look like related changes across different model families rather than true duplicates.",
-        "confidence": 0.84,
-        "canonical_issue_reason": "43979 is the shared umbrella tracker for the standardized output-tracing work, so it is the best issue anchor for this cluster.",
-        "canonical_pr_reason": "44722 is the strongest canonical PR because it directly implements the GPT-J/CodeGen refactor, explicitly links 43979, and best represents the clearest duplicate-like subcluster.",
-        "best_issue_reason": "43979 is the best global issue because it is the common tracking issue referenced by the same-change GPT-J/CodeGen and GPT-Neo refactors.",
-        "best_pr_reason": "44722 is the best global PR representative because it most clearly matches the shared GPT-J/CodeGen change pattern and has the cleanest explicit issue linkage.",
+        "summary": "Mostly a theme cluster around standardized output tracing, not true duplicates across models. The only strong duplicate-like overlaps are within the same model/code path: GPT-Neo (44018/44068) and GPT-J/CodeGen (44066/44085/44722).",
+        "confidence": 0.87,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 44722 is the most consolidated representative of the GPT-J/CodeGen output-tracing work and overlaps the earlier GPT-J-specific variants.",
+        "best_issue_reason": null,
+        "best_pr_reason": "44722 is the best cluster anchor: latest, broadest among the near-duplicate GPT-J-side changes, and it directly targets the shared output-tracing refactor path.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43996",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "Different model families and code paths; the shared output-tracing theme is not enough."
+            "reason": "Same umbrella output-tracing theme, but different model families; not the same concrete change."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "CVT/FNet and DeBERTa are unrelated implementations."
+            "reason": "Different models and code paths; shared issue target alone is not enough."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44085",
             "accept": true,
-            "reason": "Both are GPT-J output-tracing refactors on the same code path, so they could plausibly be merged into one PR."
+            "reason": "Both touch the GPT-J output-tracing code path in the same file and look like overlapping implementations of the same refactor."
           },
           {
             "left": "pull_request:44007",
             "right": "pull_request:44072",
             "accept": false,
-            "reason": "RegNet and EfficientNet are different vision models with different forward/output logic."
+            "reason": "RegNet/ResNet vs EfficientNet-style refactors are different model changes, not one duplicate bug."
           },
           {
             "left": "pull_request:44072",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "EfficientNet and GPT-J/CodeGen are unrelated model paths."
+            "reason": "Different model families; both are output-tracing refactors but not the same concrete change."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44072",
             "accept": false,
-            "reason": "GPT-J/CodeGen versus EfficientNet; only the broad refactor theme overlaps."
+            "reason": "Different model files and different refactors."
           },
           {
             "left": "pull_request:44013",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "MobileNetV2 and DeBERTa are different model families."
+            "reason": "MobileNetV2 and DeBERTa are unrelated implementations despite similar tracing cleanup."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-J/CodeGen and MGP-STR are different model-specific changes."
+            "reason": "Different models; same general standardized-output theme only."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-J/CodeGen and MPT are separate code paths."
+            "reason": "GPT-J/CodeGen vs MPT are different code paths."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44068",
             "accept": true,
-            "reason": "Both are GPT-Neo output-tracing refactors on the same model code path."
+            "reason": "Same GPT-Neo model file and same output-tracing refactor; these look like overlapping versions of one change."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44068",
             "accept": false,
-            "reason": "GPT-J/CodeGen and GPT-Neo are different model implementations."
+            "reason": "GPT-J/CodeGen versus GPT-Neo; same theme, different code paths."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44066",
             "accept": false,
-            "reason": "GPT-Neo and GPT-J/CodeGen are separate model code paths."
+            "reason": "Different model families; not mergeable as one concrete fix."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "ResNet output tracing is unrelated to GPT-J/CodeGen."
+            "reason": "ResNet vs GPT-J/CodeGen; too broad to be duplicates."
           },
           {
             "left": "pull_request:44086",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "MGP-STR and GPT-J/CodeGen are different models and fixes."
+            "reason": "MGP-STR and GPT-J/CodeGen are unrelated changes."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "ResNet and GPT-J are unrelated code paths."
+            "reason": "ResNet and GPT-J are different model-specific refactors."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44019",
             "accept": false,
-            "reason": "GPT-Neo and ResNet are different model families."
+            "reason": "GPT-Neo and ResNet are different code paths."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "ResNet and MPT output handling are unrelated changes."
+            "reason": "Different models; shared output-tracing theme is too broad."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44116",
             "accept": false,
-            "reason": "GPT-Neo and Flaubert are distinct models with different output logic."
+            "reason": "GPT-Neo and Flaubert are different implementations."
           },
           {
             "left": "pull_request:44068",
             "right": "pull_request:44116",
             "accept": false,
-            "reason": "GPT-Neo and Flaubert only share the broad output-tracing theme."
+            "reason": "Different models and different concrete fixes."
           },
           {
             "left": "pull_request:44068",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "GPT-Neo is not the same concrete change as GPT-J/CodeGen."
+            "reason": "GPT-Neo is not the same code path as GPT-J/CodeGen."
           },
           {
             "left": "pull_request:44071",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "MPT and GPT-J/CodeGen are separate implementations."
+            "reason": "MPT and GPT-J/CodeGen are unrelated model changes."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "GPT-Neo and GPT-J are different models despite similar refactor wording."
+            "reason": "GPT-Neo and GPT-J are distinct model-specific changes."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "Different models and files."
           },
           {
             "left": "pull_request:44068",
-            "right": "pull_request:44071",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo vs MGP-STR; not the same underlying bug."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44085",
             "accept": false,
-            "reason": "GPT-Neo and MPT are unrelated model-specific refactors."
+            "reason": "MPT and GPT-J are different code paths."
           },
           {
             "left": "pull_request:44098",
             "right": "pull_request:44101",
             "accept": false,
-            "reason": "ViLT and Flaubert/XLM are different model families."
+            "reason": "ViLT and Flaubert/XLM are separate model refactors."
           },
           {
             "left": "pull_request:44085",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-J output tracing and MGP-STR output tracing are different code paths."
+            "reason": "GPT-J and MGP-STR are unrelated implementations."
           },
           {
             "left": "pull_request:44059",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT2 and MGP-STR are separate model implementations."
+            "reason": "GPT2 and MGP-STR are different model changes."
           },
           {
             "left": "pull_request:44017",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "Segformer and LongT5 are unrelated model paths."
+            "reason": "SegFormer and LongT5 are different model-specific refactors."
           },
           {
             "left": "pull_request:44017",
             "right": "pull_request:44073",
             "accept": false,
-            "reason": "Segformer and VisualBert are different model families."
+            "reason": "SegFormer and VisualBert are unrelated model changes."
           },
           {
             "left": "pull_request:44017",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "Segformer output tracing is unrelated to GPT-J/CodeGen."
+            "reason": "Different models; only the generic tracing theme overlaps."
           },
           {
             "left": "pull_request:44068",
-            "right": "pull_request:44086",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-Neo and MPT are different code paths."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "RegNet and VisualBert are unrelated model refactors."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-Neo and MGP-STR are different changes in different models."
+            "reason": "SegFormer and MPT are different concrete changes."
           },
           {
             "left": "pull_request:44071",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "MPT and MGP-STR are unrelated code paths."
+            "reason": "Different models; not mergeable as one PR."
           },
           {
             "left": "pull_request:44073",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "VisualBert and LongT5 are different model-specific refactors."
+            "reason": "VisualBert and LongT5 are different model families."
           },
           {
             "left": "pull_request:44073",
             "right": "pull_request:44101",
             "accept": false,
-            "reason": "VisualBert and Flaubert are unrelated model implementations."
+            "reason": "VisualBert and Flaubert/XLM are separate code paths."
           },
           {
             "left": "pull_request:44059",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT2 and MPT output-tracing changes are not the same concrete fix."
+            "reason": "GPT2 and MPT are different model implementations."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44101",
             "accept": false,
-            "reason": "MPNet and Flaubert are different model code paths."
+            "reason": "MPNet and Flaubert/XLM are unrelated changes."
           },
           {
             "left": "pull_request:43997",
-            "right": "pull_request:44017",
+            "right": "pull_request:44018",
             "accept": false,
-            "reason": "RegNet and Segformer are different vision models with different forward logic."
+            "reason": "RegNet and GPT-Neo are different models."
           },
           {
             "left": "pull_request:43997",
             "right": "pull_request:44056",
             "accept": false,
-            "reason": "RegNet and MPNet are unrelated model families."
+            "reason": "RegNet and MPNet are different code paths."
           },
           {
             "left": "pull_request:44071",
             "right": "pull_request:44076",
             "accept": false,
-            "reason": "MPT and ImageGPT output refactors are separate model-specific changes."
+            "reason": "MPT and ImageGPT are different model-specific refactors."
           },
           {
             "left": "pull_request:44074",
             "right": "pull_request:44101",
             "accept": false,
-            "reason": "TextNet and Flaubert are unrelated models."
+            "reason": "TextNet and Flaubert/XLM are unrelated changes."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "MPNet and LongT5 are different code paths."
-          },
-          {
-            "left": "pull_request:43997",
-            "right": "pull_request:44018",
-            "accept": false,
-            "reason": "RegNet and GPT-Neo are unrelated implementations."
+            "reason": "MPNet and LongT5 are different model families."
           },
           {
             "left": "pull_request:43997",
-            "right": "pull_request:44161",
+            "right": "pull_request:44129",
             "accept": false,
-            "reason": "RegNet and LongT5 are different model families."
+            "reason": "RegNet and SpeechT5 are different code paths."
           },
           {
             "left": "pull_request:44074",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "TextNet and LongT5 are unrelated refactors."
+            "reason": "TextNet and LongT5 are unrelated model changes."
           },
           {
             "left": "pull_request:44010",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "SqueezeBert and MPT are different model code paths."
-          },
-          {
-            "left": "pull_request:44018",
-            "right": "pull_request:44129",
-            "accept": false,
-            "reason": "GPT-Neo and SpeechT5 are unrelated model implementations."
+            "reason": "SqueezeBert and MPT are different models."
           },
           {
             "left": "pull_request:44010",
             "right": "pull_request:44073",
             "accept": false,
-            "reason": "SqueezeBert and VisualBert are different models."
+            "reason": "SqueezeBert and VisualBert are different code paths."
           },
           {
-            "left": "pull_request:44010",
-            "right": "pull_request:44018",
+            "left": "pull_request:43996",
+            "right": "pull_request:44017",
             "accept": false,
-            "reason": "SqueezeBert and GPT-Neo are unrelated changes."
+            "reason": "FNet/CVT and SegFormer are different model-specific refactors."
           },
           {
             "left": "pull_request:44010",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "SqueezeBert and DeBERTa have different output/attention code paths."
+            "reason": "SqueezeBert and DeBERTa are unrelated changes."
           },
           {
             "left": "pull_request:44044",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "DeBERTa and TextNet are unrelated model families."
+            "reason": "DeBERTa and TextNet are different models."
           },
           {
             "left": "pull_request:44024",
             "right": "pull_request:44076",
             "accept": false,
-            "reason": "FocalNet and ImageGPT are different models; same standardized-output theme only."
+            "reason": "FocalNet and ImageGPT are different concrete implementations."
           },
           {
             "left": "pull_request:44073",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "VisualBert and TextNet are separate model-specific refactors."
-          },
-          {
-            "left": "pull_request:44066",
-            "right": "pull_request:44722",
-            "accept": true,
-            "reason": "Same GPT-J/CodeGen output-tracing refactor with highly overlapping file scope and logic."
-          },
-          {
-            "left": "pull_request:44085",
-            "right": "pull_request:44722",
-            "accept": true,
-            "reason": "Both touch the same GPT-J output-tracing change on the same file path."
+            "reason": "VisualBert and TextNet are different model-specific refactors."
           },
           {
-            "left": "pull_request:44085",
+            "left": "pull_request:44018",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "GPT-J and SpeechT5 are different model code paths."
+            "reason": "GPT-Neo and SpeechT5 are unrelated changes."
           },
           {
-            "left": "pull_request:43996",
-            "right": "pull_request:44007",
+            "left": "pull_request:44010",
+            "right": "pull_request:44018",
             "accept": false,
-            "reason": "CVT/FNet and RegNet/ResNet are different models."
+            "reason": "SqueezeBert and GPT-Neo are different model code paths."
           },
           {
-            "left": "pull_request:43996",
+            "left": "pull_request:44044",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "CVT/FNet and SpeechT5 are unrelated refactors."
+            "reason": "DeBERTa and SpeechT5 are different models."
           },
           {
-            "left": "pull_request:44044",
+            "left": "pull_request:44071",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "DeBERTa and SpeechT5 are unrelated model implementations."
+            "reason": "MPT and SpeechT5 are unrelated model refactors."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "Different models and different concrete changes."
           },
           {
             "left": "pull_request:44044",
             "right": "pull_request:44154",
             "accept": false,
-            "reason": "DeBERTa and VITS are different model code paths."
+            "reason": "DeBERTa and Vits are unrelated code paths."
           },
           {
-            "left": "pull_request:44024",
-            "right": "pull_request:44076",
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
             "accept": false,
-            "reason": "FocalNet and ImageGPT are different models; same refactor style is not enough."
+            "reason": "Different model families; shared issue target is not enough."
           },
           {
-            "left": "pull_request:44073",
-            "right": "pull_request:44074",
-            "accept": false,
-            "reason": "VisualBert and TextNet are different model-specific changes."
+            "left": "pull_request:44085",
+            "right": "pull_request:44722",
+            "accept": true,
+            "reason": "Both overlap on the GPT-J output-tracing refactor in the same file; 44722 looks like the broader consolidated version."
           }
         ]
       },
       "evaluator_result": {
-        "accept": false,
-        "feedback": "The summary is a bit too exclusive for the packet: it claims the only true duplicate-like groups are GPT-J/CodeGen and GPT-Neo, but the packet also marks other soft pairs as accepted (for example FocalNet/ImageGPT and VisualBert/TextNet). Even if you want to treat those as weaker, the word 'only' overstates what the packet supports. A more conservative summary would note the GPT-J/CodeGen and GPT-Neo clusters as the clearest duplicate-like groups without excluding the others outright."
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: it correctly distinguishes the broad cross-model output-tracing theme from the smaller same-code-path clusters. The accepted soft edges are conservative enough because they stay within the same model/file families (GPT-Neo and GPT-J/CodeGen), and the cross-model pairs are rejected appropriately. The canonical PR choice is also supported by the packet as the latest, broadest GPT-J/CodeGen variant."
       },
       "evaluator_used": true,
-      "retried": true,
+      "retried": false,
       "accepted_nontrivial_soft_edge": true,
       "error_kind": null,
       "error_message": null
@@ -24797,14 +24795,14 @@
         "pull_request:43997|pull_request:44010",
         "pull_request:44010|pull_request:44129"
       ],
-      "prepared_review_unit_hash": "326cdd8d44384e142e0d80cbfe219cdd77aeb521d31cedc86f137f537e04715c",
+      "prepared_review_unit_hash": "5025ab2700dd76a18b68676fccf735462e4a308a7df8418d770e84dcddd598a0",
       "component_budget": {
         "node_count": 28,
         "item_count": 28,
         "soft_pair_count": 76,
-        "serialized_chars": 84119,
-        "estimated_input_tokens": 21030,
-        "estimated_eval_tokens": 42316
+        "serialized_chars": 84116,
+        "estimated_input_tokens": 21029,
+        "estimated_eval_tokens": 42314
       },
       "budget": {
         "node_count": 6,
@@ -24819,36 +24817,36 @@
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "These PRs share the same general refactor pattern (migrating model output tracing to capture_outputs/can_return_tuple), but each targets a different model and code path, so they are not duplicates.",
-        "confidence": 0.92,
+        "summary": "These PRs share a broad theme\u2014migrating different model implementations to standardized output tracing/capture\u2014but they touch different architectures and code paths, so they are related maintenance changes rather than duplicates.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "44129 is the most concrete representative of the cluster: it has an explicit issue target and a clearly scoped output-tracing refactor in one model.",
+        "canonical_pr_reason": "PR 44010 is the clearest representative of the cluster: it directly frames the shared output-collection refactor and is a straightforward model-specific example of the pattern.",
         "best_issue_reason": null,
-        "best_pr_reason": "44129 is the strongest standalone PR to keep as the cluster representative because it is explicitly tied to an issue and has a well-scoped implementation.",
+        "best_pr_reason": "PR 44010 is the best overall representative because its title and diff most directly describe the common maintenance goal without extra model-specific complications or an explicit dependency on another issue.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44056",
             "right": "pull_request:44073",
             "accept": false,
-            "reason": "Both are standardized output-tracing refactors, but MPNet and VisualBert are different model implementations with distinct code paths; not the same change."
+            "reason": "Both are output-tracing refactors, but they affect different models (MPNet vs VisualBert) and different internal layers; same theme, not the same concrete change."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "Same broad migration theme, but these touch different models and include model-specific test/code adjustments; not a single mergeable duplicate PR."
+            "reason": "Different model code paths and different refactor details; they modernize output capture in separate implementations, so they are not duplicates."
           },
           {
             "left": "pull_request:43997",
             "right": "pull_request:44010",
             "accept": false,
-            "reason": "RegNet and SqueezeBert are unrelated model code paths; the similarity is only the shared output-capture refactor pattern."
+            "reason": "RegNet and SqueezeBert are unrelated model implementations; the PRs share a refactor pattern but not one underlying bug or one mergeable fix."
           },
           {
             "left": "pull_request:44010",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "Both modernize output handling, but SqueezeBert and SpeechT5 are separate model-specific changes and do not fix the same concrete bug."
+            "reason": "SqueezeBert and SpeechT5 are different subsystems with distinct output-shaping changes; they are not the same concrete fix and should not be merged as duplicates."
           }
         ]
       },
@@ -24928,268 +24926,262 @@
         "pull_request:44028|pull_request:44030",
         "pull_request:44029|pull_request:44030"
       ],
-      "prepared_review_unit_hash": "cb9766a2cea53d317f8571aabd05318052300a1a6ab3e989921903d3691435a7",
+      "prepared_review_unit_hash": "72c8633fe4ca5175fbce9537f6784749ab6fa8827b16fa4816549d4338c00c5f",
       "component_budget": {
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 40,
-        "serialized_chars": 40182,
-        "estimated_input_tokens": 10046,
-        "estimated_eval_tokens": 20348
+        "serialized_chars": 40178,
+        "estimated_input_tokens": 10045,
+        "estimated_eval_tokens": 20346
       },
       "budget": {
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 40,
-        "serialized_chars": 40182,
-        "estimated_input_tokens": 10046,
-        "estimated_eval_tokens": 20348
+        "serialized_chars": 40178,
+        "estimated_input_tokens": 10045,
+        "estimated_eval_tokens": 20346
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "This cluster is a set of independent model-specific output-tracing refactors across many architectures. They share the same refactor pattern, but not the same concrete bug or mergeable code change, so they should not be treated as duplicates.",
-        "confidence": 0.97,
+        "summary": "These are thematically similar PRs for refactoring output tracing, but they target different model implementations and code paths. They look like a family of related changes, not duplicate fixes or mergeable single changes.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 44025 is a good representative of the cluster because it shows the refactor pattern clearly across two closely related model files, but it is still just one instance of a broader series of independent changes.",
+        "canonical_pr_reason": "No true duplicate canonical PR here; 44003 is the broadest representative because it spans two related model files (mamba and falcon_mamba) and shows the core output-tracing refactor pattern.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 44025 is the clearest single representative of the repeated output-tracing refactor pattern, with a self-contained change spanning depth_anything and prompt_depth_anything.",
+        "best_pr_reason": "44003 is the best single representative of the cluster\u2019s theme, but it should not be treated as a duplicate of the others since each PR is model-specific.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43998",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different model implementations; same refactor theme, but not the same concrete change or code path."
+            "reason": "Different model code paths (timm_backbone vs upernet); same refactor theme only."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44000",
             "accept": false,
-            "reason": "Both are output-tracing refactors, but they affect different models and unrelated code paths."
+            "reason": "Different architectures and files; not the same concrete bug/change."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Separate model-specific refactors; no shared bug or mergeable fix."
+            "reason": "Distinct model implementations; only a shared output-tracing refactor pattern."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different architectures and files; this is a similar pattern, not the same underlying change."
+            "reason": "Different model families (timm_backbone vs mamba/falcon_mamba); not mergeable as one fix."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Codegen-specific forward signature change is unrelated to timm_backbone."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Same broad refactor style, but unrelated model code paths."
+            "reason": "RWKV refactor touches a different model path and cannot be treated as the same change."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Different models and implementations; not the same concrete fix."
+            "reason": "MobileNetV1 and UnivNet are unrelated code paths; same refactor idea only."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Shared output-tracing theme only; the PRs target different model code."
+            "reason": "Different model subsystems; not the same underlying bug/change."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Different model families; these are independent refactors, not duplicates."
+            "reason": "MobileNetV1 and Depth Anything are separate implementations, so this is not a duplicate."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44026",
             "accept": false,
-            "reason": "Same broad subsystem, but distinct code paths and model wrappers."
+            "reason": "Vision-encoder-decoder is a different code path from MobileNetV1."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Both adjust output tracing, but in different encoder-decoder models and code paths."
+            "reason": "Speech-encoder-decoder refactor is separate from MobileNetV1."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Unrelated model implementations; only the refactor motif is shared."
+            "reason": "Vision-text dual encoder and UnivNet are unrelated model paths."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different model files and behavior; not the same underlying bug/change."
+            "reason": "Different model families; no shared concrete fix."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "These are separate model-specific changes, not one mergeable fix."
+            "reason": "Same high-level refactor theme, but different model code paths."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "CodeGen forward/output handling is unrelated to vision-text dual encoder."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Common tracing refactor pattern, but different models and no shared code path."
+            "reason": "Different models and files; not the same change."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different model architectures; same theme does not make them duplicates."
+            "reason": "UnivNet and UperNet are unrelated implementations."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Independent model refactors; not the same code-path problem."
+            "reason": "Different model families; cannot plausibly be merged into one PR."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Different models and files; only the refactor style overlaps."
+            "reason": "UnivNet and Depth Anything are separate code paths."
           },
           {
             "left": "pull_request:44002",
-            "right": "pull_request:44029",
+            "right": "pull_request:44003",
             "accept": false,
-            "reason": "Separate model code paths with a similar output-tracing cleanup."
+            "reason": "UperNet and mamba/falcon_mamba are unrelated concrete changes."
           },
           {
             "left": "pull_request:44002",
-            "right": "pull_request:44003",
+            "right": "pull_request:44004",
             "accept": false,
-            "reason": "Different architectures and implementation details; not a duplicate fix."
+            "reason": "Different architecture and file; only same broad refactor theme."
           },
           {
             "left": "pull_request:44003",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different model families; these are unrelated refactors despite similar wording."
+            "reason": "mamba/falcon_mamba and superpoint are different model paths, not duplicates."
           },
           {
             "left": "pull_request:44003",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Distinct model implementations and diffs; not the same underlying issue."
-          },
-          {
-            "left": "pull_request:43998",
-            "right": "pull_request:44004",
-            "accept": false,
-            "reason": "Unrelated model code paths; same refactor theme only."
-          },
-          {
-            "left": "pull_request:44001",
-            "right": "pull_request:44004",
-            "accept": false,
-            "reason": "Different models and logic; not a mergeable single change."
-          },
-          {
-            "left": "pull_request:44002",
-            "right": "pull_request:44004",
-            "accept": false,
-            "reason": "Same broad cleanup, but separate concrete implementations."
+            "reason": "CodeGen output tracing is a separate change from mamba/falcon_mamba."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44026",
             "accept": false,
-            "reason": "Different model wrappers and code paths; no shared bug or exact fix."
+            "reason": "Depth Anything and vision-encoder-decoder are distinct code paths."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Both are refactors, but in different encoder-decoder stacks."
+            "reason": "Depth Anything and speech-encoder-decoder are separate implementations."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Distinct models; only the output-tracing pattern is shared."
+            "reason": "Different model families; not the same underlying fix."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different architectures and files; not the same concrete change."
+            "reason": "Depth Anything and RWKV are unrelated model code paths."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Independent refactors in unrelated model code."
+            "reason": "Depth Anything and DPR are different concrete changes."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Similar encoder-decoder refactor wording, but different modalities and code paths."
+            "reason": "Vision-encoder-decoder and speech-encoder-decoder are separate implementations despite similar wrappers."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different models; this is not one shared bug fix."
+            "reason": "Different model code paths; not a duplicate."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Separate model implementations with no concrete overlap beyond style."
+            "reason": "Vision-encoder-decoder and RWKV are unrelated concrete fixes."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Different architectures and diffs; not duplicates."
+            "reason": "Different models and files; same refactor pattern only."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different model families; same refactor motif is insufficient."
+            "reason": "Speech-encoder-decoder and superpoint do not share the same underlying bug/change."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Distinct code paths and model behavior."
+            "reason": "Different model paths; not mergeable as one PR."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "These are unrelated model-specific changes."
+            "reason": "Speech-encoder-decoder and DPR are distinct changes."
           },
           {
             "left": "pull_request:44028",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different architectures and implementations; not the same underlying issue."
+            "reason": "SuperPoint and RWKV are unrelated model implementations."
           },
           {
             "left": "pull_request:44028",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Separate model code paths with only superficial similarity."
+            "reason": "Different code paths and model types; not duplicates."
           },
           {
             "left": "pull_request:44029",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Different model implementations; not a single mergeable fix."
+            "reason": "RWKV and DPR are separate model-specific refactors."
           }
         ]
       },
@@ -25219,7 +25211,7 @@
       "soft_pairs": [
         "pull_request:44662|pull_request:45401"
       ],
-      "prepared_review_unit_hash": "539223c2fb98c173415b3aef7722b6b30cebc1cb067f642fdde7d03937c55564",
+      "prepared_review_unit_hash": "11541ef619306b165e9f825e12d8075a3ed46c188a64924f730ee720c328207e",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -25241,18 +25233,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both items are model-addition PRs, but they implement different models and different code paths: PenguinVL versus Voxtral TTS. The overlap is mostly framework boilerplate (docs and auto-registry files), not the underlying change itself.",
-        "confidence": 0.96,
+        "summary": "These are two unrelated model-addition PRs: one introduces PenguinVL, the other adds Voxtral TTS support. They share some auto/docs scaffolding, but they target different model families, code paths, and release goals, so this is not a duplicate pair.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45401 is the cleaner representative because it has an explicit issue target and is a focused, self-contained model support addition.",
+        "canonical_pr_reason": "No canonical PR: 44662 is a PenguinVL implementation, while 45401 is Voxtral TTS support with an explicit issue target; they address different underlying changes.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45401 is the strongest standalone PR candidate: it is tied to an issue, scoped to one model family, and appears more directly triageable than the broader PenguinVL implementation PR.",
+        "best_pr_reason": "No single best PR for the cluster because the two PRs are not duplicates and should not be merged into one representative change.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44662",
             "right": "pull_request:45401",
             "accept": false,
-            "reason": "They add support for different models with different architectures and implementation files. Shared auto-registration/docs are incidental and do not indicate the same underlying bug or change."
+            "reason": "Different model additions and concrete code paths (PenguinVL vs Voxtral TTS); shared auto/docs files are just framework plumbing, not the same bug or change."
           }
         ]
       },
@@ -25282,7 +25274,7 @@
       "soft_pairs": [
         "pull_request:44979|pull_request:45363"
       ],
-      "prepared_review_unit_hash": "c6f075ff1353bcad19264499ae4e1d16180bec1f946abd0f935702044f71870a",
+      "prepared_review_unit_hash": "e8522f71cc5acd9970d518dd633a561d583abb00b5a87df5869bdf114cd83875",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -25304,18 +25296,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These PRs are related to module/kernel fusion, but they target different layers of the feature: one introduces a generic Module Fusion API, while the other wires n-to-1 kernel fusion through KernelConfig and model loading. They are not duplicates and should not be merged into one PR.",
-        "confidence": 0.91,
+        "summary": "These are related fusion efforts, but they are not the same change: PR #44979 introduces a new generic Module Fusion API, while PR #45363 adds n-to-1 kernel fusion support through KernelConfig plus an example and integration changes. They overlap conceptually, but they target different layers and code paths.",
+        "confidence": 0.89,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45363 is the more concrete end-to-end feature implementation, covering KernelConfig, integration hooks, and an example; PR 44979 is a separate foundational API layer.",
+        "canonical_pr_reason": "PR #44979 is the better canonical representative because it introduces the foundational Module Fusion API that the other work can build on, whereas PR #45363 is a more specific KernelConfig-based application.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45363 is the best representative for the cluster because it is the broader user-facing fusion feature and includes the actual integration path, not just the generic API scaffold.",
+        "best_pr_reason": "PR #44979 is the strongest cluster representative: it is the core abstraction change, with tests, and is broader than the downstream kernel-fusion integration in PR #45363.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44979",
             "right": "pull_request:45363",
             "accept": false,
-            "reason": "Same broad fusion theme, but different concrete changes: 44979 adds a generic Module Fusion API, while 45363 implements kernel-config-driven n-to-1 fusion and integration points. They are related, not duplicate, and would not plausibly be merged as one PR."
+            "reason": "Reject: both are about fusion, but #44979 adds a generic module fusion API and #45363 implements a specific n-to-1 kernel-fusion path via KernelConfig; they are not the same concrete change and would not be merged as one PR."
           }
         ]
       },
@@ -25345,7 +25337,7 @@
       "soft_pairs": [
         "pull_request:45086|pull_request:45317"
       ],
-      "prepared_review_unit_hash": "440f79afb548d9464e8d797bf3387d7c32df44cacb8c9c5d080a4276c8833660",
+      "prepared_review_unit_hash": "483af6812c1a1435d48821518b83cdfad61b7cd146807c70b1c878f198f62440",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -25367,18 +25359,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both pull requests address the same AttributeError in `_patch_mistral_regex` by switching from `tokenizer.backend_tokenizer.pre_tokenizer` to `tokenizer.pre_tokenizer`; they target the same issue and the later PR adds tests, making it the stronger representative.",
+        "summary": "Both pull requests fix the same AttributeError in `_patch_mistral_regex` by switching from `tokenizer.backend_tokenizer.pre_tokenizer` to `tokenizer.pre_tokenizer`; 45317 is the more complete version because it also adds tests.",
         "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #45317 is the best canonical PR because it fixes the same concrete bug as #45086 but is more complete: it includes the same code-path correction plus test coverage in `tests/models/auto/test_tokenization_auto.py`.",
+        "canonical_pr_reason": "45317 is the better canonical PR because it addresses the same concrete bug as 45086 but includes accompanying test coverage and looks like the more complete patch.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45317 is the best overall representative for this cluster since it contains the same fix with additional validation and appears to supersede the earlier, narrower PR.",
+        "best_pr_reason": "45317 is the strongest representative for the cluster: same fix, same target issue, and it adds tests, making it the most merge-ready version.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45086",
             "right": "pull_request:45317",
             "accept": true,
-            "reason": "Same underlying bug and same code change in the same file, both explicitly target issue #45081; #45317 is effectively the same fix with added tests, so this is a true duplicate/soft match."
+            "reason": "Same underlying `_patch_mistral_regex` AttributeError fix on the same code path and shared issue target; the second PR is a superset with tests, so these are plausibly one merged change."
           }
         ]
       },
@@ -25408,7 +25400,7 @@
       "soft_pairs": [
         "pull_request:45115|pull_request:45569"
       ],
-      "prepared_review_unit_hash": "3e746dc2b838ee28a83d3356dd92f44f1862415c81a6b98aac6e01f0749a09a2",
+      "prepared_review_unit_hash": "27d236b7c7298d915e46d6a19dbb33d4c640bade82694d0faf42f6994fcc20d5",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -25430,18 +25422,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Two open PRs in the NemotronH area, but they are not duplicates: one is a structural refactor/conversion-mapping update for NemotronH, while the other splits NemotronH into dense/sparse model variants and updates docs, auto mappings, and tests.",
-        "confidence": 0.89,
+        "summary": "Both PRs touch the Nemotron H codepath, but they are not the same change: #45115 is a smaller refactor/conversion-mapping update toward GraniteMoeHybrid compatibility, while #45569 introduces the dedicated NemotronH dense/sparse model split and associated docs/tests. The overlap is mostly shared model files, not a single concrete bug fix.",
+        "confidence": 0.92,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #45569 is the more central and feature-complete NemotronH change: it adds the dense/sparse model variants, updates model registration, docs, and tests. PR #45115 is a narrower refactor focused on inheritance and checkpoint conversion mapping.",
+        "canonical_pr_reason": "No issue-linked canonical PR exists here; if one PR must represent the cluster, #45569 is the broader end-state implementation because it adds the dedicated model variants and user-facing docs/tests.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45569 is the best representative of the cluster topic because it covers the main user-facing model support work rather than just an internal refactor.",
+        "best_pr_reason": "#45569 is the best overall representative because it covers the full NemotronH family restructuring and documentation, whereas #45115 is a narrower conversion/compatibility refactor.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45115",
             "right": "pull_request:45569",
             "accept": false,
-            "reason": "Different underlying changes. #45115 is a NemotronH refactor/conversion-mapping update; #45569 introduces separate NemotronH dense/sparse implementations plus docs, auto-mapping, and tests. Shared subsystem files are not enough to treat them as the same fix."
+            "reason": "Not the same underlying change: #45115 is about conversion key remapping and inheritance/refactor details, while #45569 adds separate dense/sparse model implementations plus docs/tests. Shared files alone are not enough to treat these as duplicates."
           }
         ]
       },
@@ -25471,7 +25463,7 @@
       "soft_pairs": [
         "pull_request:45332|pull_request:45333"
       ],
-      "prepared_review_unit_hash": "70c35051c76d799a213cbf3ddd81e0dd3a69a80d5017098233cdd2c5fce21f36",
+      "prepared_review_unit_hash": "8b0722d862c2d330641bb7e183a1506f114f415aed58d03d88a7bc5597944289",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -25493,18 +25485,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related but not duplicates: PR #45333 is a narrower per-layer configuration addition, while PR #45332 is a larger heterogeneous-model feature that also includes modeling/runtime integration. They share foundation files, but they do not appear to fix the same concrete code-path problem.",
-        "confidence": 0.89,
+        "summary": "These are related PRs in the same heterogeneous-modeling effort, but they are not duplicates: #45333 appears to be the config-only slice, while #45332 includes that plus broader modeling/cache/integration changes. I would not merge them under a soft duplicate edge.",
+        "confidence": 0.91,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #45332 is the broader, more comprehensive feature PR and is the better representative of the heterogeneous-model work overall.",
+        "canonical_pr_reason": "#45332 is the broader, more complete implementation and subsumes the config work shown in #45333, so it is the better canonical representative for this cluster.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45332 is the best single representative because it covers the full end-to-end heterogeneous modeling support, not just configuration plumbing.",
+        "best_pr_reason": "#45332 is the strongest representative because it covers the larger underlying change set (per-layer config plus heterogeneous modeling/integration), whereas #45333 is a narrower subset focused on configuration only.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45332",
             "right": "pull_request:45333",
             "accept": false,
-            "reason": "Shared configuration/docs files suggest a related feature branch, but #45333 is only per-layer config support whereas #45332 adds additional modeling/cache/integration changes. They are not the same concrete change and would not naturally collapse into one PR."
+            "reason": "Related topic, but not the same concrete change: #45333 is config-only, while #45332 adds modeling/cache/integration support as well. They could coexist as separate PRs rather than being merged as duplicates."
           }
         ]
       },
@@ -25534,12 +25526,12 @@
       "soft_pairs": [
         "pull_request:45415|pull_request:45425"
       ],
-      "prepared_review_unit_hash": "1a085d8545ddcad98473fabcd49e0dff09c2747ab358a145a33396188300e4e6",
+      "prepared_review_unit_hash": "2b4daaa6f748601149058f08cbbced2fbd7c41c058b1056aa0c5b7f67d82fcb9",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5513,
+        "serialized_chars": 5514,
         "estimated_input_tokens": 1379,
         "estimated_eval_tokens": 3014
       },
@@ -25547,7 +25539,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5513,
+        "serialized_chars": 5514,
         "estimated_input_tokens": 1379,
         "estimated_eval_tokens": 3014
       },
@@ -25556,18 +25548,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both items are typing-related PRs, but they target different work: one is a broad type-checking sweep across many transformers modules, while the other adds typing protocols/helpers for specific internal interfaces. They are related in theme, not the same underlying change.",
-        "confidence": 0.92,
+        "summary": "These are related typing-focused PRs, but they address different scopes: #45415 is a broad type-checking sweep across many transformer modules, while #45425 adds specific typing protocols/aliases and updates a smaller set of typing-sensitive paths. They share a few files, but not the same concrete change or bug fix.",
+        "confidence": 0.88,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45415 is the broader umbrella change: it updates type-checking behavior across many files and the typing config, making it the most representative PR in this cluster.",
+        "canonical_pr_reason": "#45425 is the more focused typing-infrastructure PR; #45415 is a much broader repository-wide type-checking cleanup. They overlap in theme, but not in a single underlying change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45415 is the best single representative because it is the more comprehensive typing-related change; PR 45425 is narrower and focused on helper protocols rather than the same concrete fix.",
+        "best_pr_reason": "#45425 is the better representative of the cluster because it introduces concrete typing support abstractions, whereas #45415 is a sweeping type-checking pass with many unrelated file edits.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45415",
             "right": "pull_request:45425",
             "accept": false,
-            "reason": "Do not merge: both are typing-centric, but 45415 is a broad type-checking sweep and 45425 adds specific typing protocols/utilities. They do not fix the same concrete code-path problem and would not plausibly be merged as one PR."
+            "reason": "Both are typing-related and share a few touched files, but #45415 is a broad type-checking sweep while #45425 adds specific typing aliases/protocols and updates targeted code paths. They are not the same underlying fix and would not plausibly be merged as one PR."
           }
         ]
       },
@@ -25597,7 +25589,7 @@
       "soft_pairs": [
         "pull_request:45548|pull_request:45662"
       ],
-      "prepared_review_unit_hash": "528db27f97b1204b93ed9473ddc476cb7a395a6cd99325c59fb93bc78e1a794d",
+      "prepared_review_unit_hash": "0ad5f8078275230420a1efbe410f1ae63e0af1361ad1b7028acf49a9b2d3f695",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -25619,18 +25611,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are both expert-parallelism-related PRs, but they target different loading bugs and different code paths, so they should not be merged as duplicates.",
+        "summary": "The two PRs are related to expert-parallel distributed loading, but they fix different failure modes in different code paths: one targets DeepSpeed ZeRO-3/accelerate launch serialization and config handling, the other targets FSDP2/rank-0 broadcast overwriting experts during sharded parameter loading.",
         "confidence": 0.92,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #45662 is the better representative because it addresses a concrete expert-parallel loading corruption bug with a targeted fix across the relevant loading and MoE paths.",
+        "canonical_pr_reason": "PR #45662 is the better canonical representative because it addresses a more concrete, multi-file runtime bug in expert-parallel loading and includes the more direct fix path around parameter sharding/wrapping.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45662 is the stronger standalone PR for this cluster: it fixes a specific data-loss/overwrite path and has the clearer end-to-end code-path coverage.",
+        "best_pr_reason": "PR #45662 is the strongest overall artifact in this cluster since its fix is more directly tied to the observed data-loss symptom and spans the actual loading path rather than a narrower config-serialization adjustment.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45548",
             "right": "pull_request:45662",
             "accept": false,
-            "reason": "Both involve EP, but #45548 fixes serialization/loading for EP + ZeRO-3 via accelerate launch, while #45662 fixes experts being overwritten by rank-0 broadcast under FSDP2; different bugs, different mechanisms."
+            "reason": "Not the same underlying bug: #45548 focuses on DeepSpeed ZeRO-3/accelerate launch config serialization and EP detection, while #45662 fixes experts being overwritten during FSDP2/post-shard parameter assignment. Shared EP context and one common file are not enough to merge them as duplicates."
           }
         ]
       },
@@ -25660,40 +25652,40 @@
       "soft_pairs": [
         "pull_request:45639|pull_request:45645"
       ],
-      "prepared_review_unit_hash": "ac8085561ca2f7ef3a4768ad4cc20cdfe73099044aebf5e26293116fc80b035e",
+      "prepared_review_unit_hash": "446aa5b53dc5a045399cf21597918e5d3652d35ce02dab86ab04b6b6a546bb0d",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4672,
-        "estimated_input_tokens": 1168,
-        "estimated_eval_tokens": 2592
+        "serialized_chars": 4673,
+        "estimated_input_tokens": 1169,
+        "estimated_eval_tokens": 2594
       },
       "budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4672,
-        "estimated_input_tokens": 1168,
-        "estimated_eval_tokens": 2592
+        "serialized_chars": 4673,
+        "estimated_input_tokens": 1169,
+        "estimated_eval_tokens": 2594
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs address the same xdist-related captured_info collision in Transformers test debugging, with 45645 extending the core code fix from 45639 by also updating CI log collection.",
-        "confidence": 0.91,
+        "summary": "Both PRs address the same xdist-related captured_info problem around patched testing debug logs: one makes the writer use worker-specific output files, and the other adds workflow-side collection of multiple captured_info files plus the same underlying writer fix. They are close enough to be considered the same change family.",
+        "confidence": 0.92,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45645 is the broader, more complete fix: it includes the xdist-safe output file change in testing_utils and also updates the workflow to read all captured_info*.txt files, so it best represents the end-to-end solution.",
+        "canonical_pr_reason": "PR #45645 is the broader, more complete fix: it includes the xdist-safe output handling in testing_utils and also updates CI log collection so the new per-worker files are preserved and displayed.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45645 is the best representative because it fixes the same concrete artifact-collision bug while also preserving the CI debug logs that would otherwise be missed after splitting output by worker.",
+        "best_pr_reason": "PR #45645 is the best representative because it covers the full end-to-end artifact-handling change, not just the low-level write-side fix.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45639",
             "right": "pull_request:45645",
             "accept": true,
-            "reason": "These look like the same underlying xdist artifact-collision bug in the testing debug path. 45645 is an expanded version of the same fix, adding workflow handling for multiple output files, so they could plausibly be merged into one PR."
+            "reason": "Same underlying xdist collision bug for captured_info artifacts; #45645 clearly subsumes #45639 by adding the workflow-side aggregation on top of the shared testing_utils fix."
           }
         ]
       },
@@ -25789,12 +25781,12 @@
       "soft_pairs": [
         "pull_request:46126|pull_request:46269"
       ],
-      "prepared_review_unit_hash": "cb68c4023a159e559ec0e5f0712cb75c33ad4b9f6f98ddc32e65a79a6b8ec029",
+      "prepared_review_unit_hash": "deb492424e89ab074ececcd3cb6213f43b39d57547c8d05839d49071f6b0c73a",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5525,
+        "serialized_chars": 5526,
         "estimated_input_tokens": 1382,
         "estimated_eval_tokens": 3020
       },
@@ -25802,7 +25794,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5525,
+        "serialized_chars": 5526,
         "estimated_input_tokens": 1382,
         "estimated_eval_tokens": 3020
       },
@@ -25811,18 +25803,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are both distributed-system PRs and overlap on several files, but they do not look like the same concrete change. #46126 is a focused EP + Trainer integration, while #46269 is a much broader distributed branch-base PR spanning FSDP, MoE, tensor parallel, and CI infrastructure.",
-        "confidence": 0.84,
+        "summary": "These are two different distributed-feature PRs with some overlap in shared files, but they target different changes: one is a focused EP/Trainer integration, the other is a broad distributed branch base with FSDP/CI and many unrelated edits.",
+        "confidence": 0.95,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "#46269 is the broader, more central distributed branch-base change and the better anchor for this cluster if one PR must represent the shared distributed work.",
+        "canonical_pr_reason": "PR #46126 is the more specific, self-contained change with a clear feature scope (EP + Trainer integration).",
         "best_issue_reason": null,
-        "best_pr_reason": "#46269 is the best representative PR because it subsumes the shared distributed files and covers the larger underlying refactor/integration effort.",
+        "best_pr_reason": "PR #46126 is the best representative of the cluster because it has the narrowest, clearest concrete code-path change; #46269 is too broad and mixed to serve as a canonical duplicate target.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46126",
             "right": "pull_request:46269",
             "accept": false,
-            "reason": "Overlap in filenames is not enough here: #46126 is a narrow EP/Trainer integration, while #46269 is a broad distributed branch-base PR touching many unrelated areas. They do not appear to be the same underlying bug or mergeable into one change."
+            "reason": "Shared distributed filenames alone are not enough: #46126 is about EP routing/trainer integration, while #46269 is a broad distributed branch base covering FSDP, CI, and multiple subsystems. They do not look like the same underlying bug or mergeable change."
           }
         ]
       },
@@ -25890,196 +25882,232 @@
         "pull_request:46504|pull_request:46716",
         "pull_request:46432|pull_request:46716"
       ],
-      "prepared_review_unit_hash": "71453d5fbd157d417ba254d9bbee47cec82ab327951db0e91ba25c0c17df1c02",
+      "prepared_review_unit_hash": "350892a7ceab14370c304c1f5683b2170f335e83fb49e47a01b1907ee046e8e2",
       "component_budget": {
         "node_count": 10,
         "item_count": 10,
         "soft_pair_count": 31,
-        "serialized_chars": 26673,
-        "estimated_input_tokens": 6669,
-        "estimated_eval_tokens": 13594
+        "serialized_chars": 26687,
+        "estimated_input_tokens": 6672,
+        "estimated_eval_tokens": 13600
       },
       "budget": {
         "node_count": 10,
         "item_count": 10,
         "soft_pair_count": 31,
-        "serialized_chars": 26673,
-        "estimated_input_tokens": 6669,
-        "estimated_eval_tokens": 13594
+        "serialized_chars": 26687,
+        "estimated_input_tokens": 6672,
+        "estimated_eval_tokens": 13600
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "This cluster is mostly unrelated GitHub Actions workflow maintenance PRs. Several items touch the same security-gate workflow file, but they make different, non-duplicative changes, so I would not merge any of these pairs as duplicates.",
-        "confidence": 0.93,
+        "summary": "These are generic workflow-update PRs, not true duplicates. The closest cluster is the pr-ci-security-gate.yml edits, but each PR changes a different security-gate detail rather than the same concrete fix.",
+        "confidence": 0.9,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 46554 is the most substantial and central change in the cluster: it materially rewrites the PR CI security-gate workflow, while the nearby PRs are narrower follow-ups or different workflow updates.",
+        "canonical_pr_reason": "PR #46554 is the most substantive representative of the security-gate mini-cluster, since it changes the actual gate logic; the other pr-ci-security-gate PRs are separate hardening tweaks.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46554 is the strongest representative of the cluster because it contains the clearest concrete security/workflow behavior change and has the broadest overlap with the other security-gate PRs.",
+        "best_pr_reason": "PR #46554 is the best cluster representative overall because it contains the main functional workflow change and is more central than the smaller permission/comment/secrets adjustments.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46138",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflows and different fixes: stdin handling for a self-comment CI helper versus security-gate checks."
+            "reason": "Different workflows and different behavior; one updates self-comment input handling, the other changes SSH runner security setup."
           },
           {
             "left": "pull_request:46138",
             "right": "pull_request:46558",
             "accept": false,
-            "reason": "Unrelated workflow files and different behavior changes; only broad CI/security theme overlaps."
+            "reason": "Unrelated workflow updates in different files; no shared concrete bug or change."
           },
           {
             "left": "pull_request:46138",
             "right": "pull_request:46563",
             "accept": false,
-            "reason": "Self-comment CI change is unrelated to PR CI security-gate workflow changes."
+            "reason": "Different workflow files and distinct changes; not the same underlying fix."
           },
           {
             "left": "pull_request:46138",
             "right": "pull_request:46716",
             "accept": false,
-            "reason": "Different workflow and different purpose; not the same underlying bug or change."
+            "reason": "Different workflow files and separate workflow-hardening changes; not duplicates."
           },
           {
             "left": "pull_request:46138",
             "right": "pull_request:46504",
             "accept": false,
-            "reason": "Both use safer shell handling, but they affect different workflows and code paths, so they are not duplicates."
+            "reason": "Different CI workflows and different command-safety tweak; not the same change."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46503",
             "accept": false,
-            "reason": "SSH runner hardening is unrelated to benchmark workflow dispatch changes."
+            "reason": "Different workflow files and different trigger logic; no duplicate bug/change."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46501",
             "accept": false,
-            "reason": "Different workflows and different fixes; anti-slop permissions versus SSH runner hardening."
+            "reason": "Different workflow files and unrelated permission placement changes."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46502",
             "accept": false,
-            "reason": "Assign-reviewers checkout ref change is unrelated to SSH runner workflow changes."
+            "reason": "Different workflows; one changes SSH runner secrets handling, the other reviewer assignment checkout behavior."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflows and different concrete changes; no shared underlying bug."
+            "reason": "SSH-runner hardening is unrelated to PR CI security-gate logic."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46558",
             "accept": false,
-            "reason": "Runner environment cleanup is unrelated to PR CI security-gate permissions and approval flow."
+            "reason": "Different workflows and different security concerns; not the same fix."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
           },
           {
             "left": "pull_request:46501",
             "right": "pull_request:46503",
             "accept": false,
-            "reason": "Both are workflow tweaks, but they change different jobs and behavior in different files."
+            "reason": "Different workflow files; one is anti-slop permissions, the other benchmark dispatch trigger."
           },
           {
             "left": "pull_request:46501",
             "right": "pull_request:46502",
             "accept": false,
-            "reason": "Anti-slop permissions and assign-reviewers checkout behavior are different changes."
+            "reason": "Different workflow files and different behavior; not duplicates."
           },
           {
             "left": "pull_request:46501",
             "right": "pull_request:46504",
             "accept": false,
-            "reason": "Different workflows and different fixes; not the same code-path problem."
+            "reason": "Different workflow files and unrelated CI tweaks."
           },
           {
             "left": "pull_request:46501",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflow files and unrelated behavior."
+            "reason": "Different files and different workflow concerns."
           },
           {
             "left": "pull_request:46502",
             "right": "pull_request:46503",
             "accept": false,
-            "reason": "Different workflows and different logic changes."
+            "reason": "Different workflow files and separate CI updates."
           },
           {
             "left": "pull_request:46502",
             "right": "pull_request:46504",
             "accept": false,
-            "reason": "Checkout-ref fix and shell-quoting fix address different problems in different workflows."
+            "reason": "Different workflow files and unrelated changes."
           },
           {
             "left": "pull_request:46502",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "No shared underlying issue; different workflow area and change type."
+            "reason": "Different workflows and different concrete changes."
           },
           {
             "left": "pull_request:46503",
             "right": "pull_request:46504",
             "accept": false,
-            "reason": "Benchmark dispatch condition and build-image shell handling are unrelated."
+            "reason": "Different workflow files and distinct CI behavior changes."
           },
           {
             "left": "pull_request:46503",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflow and different bug/change."
+            "reason": "Different workflows and no shared concrete bug."
           },
           {
             "left": "pull_request:46504",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Same broad CI theme, but distinct workflow files and distinct fixes."
+            "reason": "Different workflow files and unrelated fixes."
           },
           {
             "left": "pull_request:46504",
             "right": "pull_request:46558",
             "accept": false,
-            "reason": "Different workflows and different concrete behavior changes."
+            "reason": "Different workflows; one is build-image tagging, the other PR CI gate permissions."
           },
           {
             "left": "pull_request:46554",
             "right": "pull_request:46558",
             "accept": false,
-            "reason": "Both touch the same workflow file, but one adds checkout/diff-based security checks while the other changes permissions and gating; these are separate changes."
+            "reason": "Same file, but different concrete changes: one adds security checks, the other adjusts permissions/approval flow."
           },
           {
             "left": "pull_request:46554",
             "right": "pull_request:46563",
             "accept": false,
-            "reason": "Same workflow file, but the changes target different concerns and are not the same underlying fix."
+            "reason": "Same workflow file, but one changes executable security-gate logic while the other only adds explanatory comments."
           },
           {
             "left": "pull_request:46554",
             "right": "pull_request:46716",
             "accept": false,
-            "reason": "Different security-hardening edits in the same workflow file; not a duplicate change."
+            "reason": "Same workflow file, but different fixes: one changes gate logic, the other removes inherited secrets."
           },
           {
             "left": "pull_request:46558",
             "right": "pull_request:46563",
             "accept": false,
-            "reason": "Different edits to the same workflow file, but not the same concrete bug or fix."
+            "reason": "Same workflow file, but different changes: permissions/guard condition versus documentation comment."
           },
           {
             "left": "pull_request:46558",
             "right": "pull_request:46716",
             "accept": false,
-            "reason": "Permission/gating changes versus secrets-handling changes are distinct."
+            "reason": "Same workflow file, but different concrete changes to the reusable workflow call."
           },
           {
             "left": "pull_request:46563",
             "right": "pull_request:46716",
             "accept": false,
-            "reason": "A review-note/comment update is not the same as changing secrets inheritance."
+            "reason": "One is a comment-only clarification; the other changes secrets inheritance. Not the same fix."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Different workflow files and different changes; not duplicates."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
           }
         ]
       },
@@ -26172,12 +26200,12 @@
       "soft_pairs": [
         "pull_request:46281|pull_request:46890"
       ],
-      "prepared_review_unit_hash": "53b1a71d6baad6b7129a71459fca1c16b0567277e649a9636bcb00dfb4091808",
+      "prepared_review_unit_hash": "0433c4cd12a70a1fe8ac748396c8f68e3f4158bab735a8726ebef6cf8caaf277",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4279,
+        "serialized_chars": 4280,
         "estimated_input_tokens": 1070,
         "estimated_eval_tokens": 2396
       },
@@ -26185,7 +26213,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4279,
+        "serialized_chars": 4280,
         "estimated_input_tokens": 1070,
         "estimated_eval_tokens": 2396
       },
@@ -26194,18 +26222,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both items are security hardening PRs against path traversal, but they target different code paths and configuration sources, so they are not duplicates.",
-        "confidence": 0.94,
+        "summary": "Two open PRs that both harden path handling against traversal, but they target different code paths and different inputs: sharded checkpoint index filenames vs the `transformers_weights` config field. They are related security fixes, not duplicates.",
+        "confidence": 0.89,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": null,
+        "canonical_pr_reason": "PR 46281 is the better cluster representative because it addresses the more specific checkpoint-shard file resolution path and directly validates index-derived filenames against directory escape.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46281 is the better representative of the cluster because it fixes a concrete sharded-checkpoint loading path traversal in `hub.py` with a clearly scoped change and test coverage.",
+        "best_pr_reason": "PR 46281 is the strongest standalone representative of the cluster theme: a concrete path-traversal fix in shard loading with a clear, narrowly-scoped code change.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46281",
             "right": "pull_request:46890",
             "accept": false,
-            "reason": "Reject: these are distinct vulnerabilities in different loaders. 46281 validates shard filenames from the checkpoint index `weight_map` in `hub.py`; 46890 validates an explicit `transformers_weights` filename in `modeling_utils.py`. Same security theme, but not the same bug or mergeable change."
+            "reason": "Both are path-traversal hardening PRs, but they fix different mechanisms in different loaders (`weight_map` shard filenames vs `transformers_weights` config). They do not look like the same underlying bug or a PR that could be merged into one change."
           }
         ]
       },
@@ -26235,7 +26263,7 @@
       "soft_pairs": [
         "pull_request:46574|pull_request:46580"
       ],
-      "prepared_review_unit_hash": "f69b8597418de81ee7360959791146afa0e11bc8c1075168b1051348db070875",
+      "prepared_review_unit_hash": "d79b388ab23b97bec0ba17354356195ce5c2b7a757f740f722db1472a25670ef",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -26257,18 +26285,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These two PRs are not duplicates: one fixes CodeLlama decode stripping on output, while the other fixes DeepSeek checkpoint-to-tokenizer-class routing during load. They touch different code paths and solve different bugs.",
-        "confidence": 0.97,
+        "summary": "The cluster contains two unrelated pull requests: one fixes CodeLlama decode behavior removing real leading spaces, while the other fixes DeepSeek-Coder/DeepSeek-LLM checkpoint routing to the correct tokenizer class.",
+        "confidence": 0.99,
         "canonical_issue_reason": null,
         "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46574 is the more self-contained representative fix: it addresses a single, concrete decode regression in CodeLlama with a direct code-path change and targeted tests. PR 46580 is a separate auto-tokenizer routing fix for different checkpoints.",
+        "best_pr_reason": null,
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46574",
             "right": "pull_request:46580",
             "accept": false,
-            "reason": "Different underlying bugs and code paths: 46574 changes CodeLlama decode behavior to preserve leading spaces, while 46580 changes AutoTokenizer checkpoint-prefix routing for DeepSeek models to avoid loading the wrong tokenizer class."
+            "reason": "Different bugs and different code paths: 46574 changes CodeLlama tokenizer decoding in `tokenization_code_llama.py`, while 46580 changes auto-tokenizer checkpoint mapping in `tokenization_auto.py` for DeepSeek models. They are not the same underlying fix and would not be merged into one PR."
           }
         ]
       },
@@ -26298,7 +26326,7 @@
       "soft_pairs": [
         "pull_request:46773|pull_request:47087"
       ],
-      "prepared_review_unit_hash": "69873021d9a0bcd35555cfdaaf6a801b5a0b7beffc29c7a392b9df1d26d0c9d8",
+      "prepared_review_unit_hash": "661f1e21191d284ac8230a5f38c2606ea782fba09184f37204de05c3a9a52154",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -26320,18 +26348,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These two PRs are related to padding-mask handling, but they fix different bugs in different paths: one targets batch-size-1 masking in linear-attention model forwards, the other fixes continued-forward/cache/chunked-prefill behavior in recurrent masking utilities. They should not be deduplicated.",
-        "confidence": 0.92,
+        "summary": "These PRs are both about padding-mask handling, but they fix different bugs in different code paths: PR 46773 corrects a batch-size-1 masking guard in linear-attention model code, while PR 47087 changes recurrent masking utility behavior for cache/chunked continued forwards. They are related conceptually but not duplicates.",
+        "confidence": 0.95,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 47087 is the better canonical PR because it fixes the shared masking utility path, has an explicit issue target, and its scope is clearer and more general than the model-specific batch-size-1 fix in PR 46773.",
+        "canonical_pr_reason": "PR 47087 is the better representative of the cluster because it fixes the shared masking utility at the framework level and includes an explicit issue reference, while PR 46773 is a model-specific fast-path condition fix.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 47087 is the strongest representative of the cluster: it addresses a concrete recurrent-mask continuation bug in core masking code, with tests and an explicit tracking issue. PR 46773 is narrower and concerns a different condition and code path.",
+        "best_pr_reason": "PR 47087 is the strongest standalone PR in this set because it addresses a broader reusable masking path, has clearer problem framing, and is more central than the model-specific fix in PR 46773.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46773",
             "right": "pull_request:47087",
             "accept": false,
-            "reason": "Different underlying bugs: 46773 fixes padding-mask application being skipped when batch size is 1 in linear-attention model code, while 47087 fixes recurrent padding-mask skipping during continued forwards/chunked prefill/cache continuation in masking_utils. They are not the same concrete change and would not be merged as one PR."
+            "reason": "Different underlying bugs: 46773 fixes a condition that skipped padding masking for batch size 1 in linear-attention model code, whereas 47087 fixes recurrent-mask creation during continued/cache forwards and single-token decode. They do not appear mergeable into one PR for the same concrete code-path problem."
           }
         ]
       },
@@ -26363,7 +26391,7 @@
         "pull_request:46781|pull_request:46993",
         "pull_request:46993|pull_request:46994"
       ],
-      "prepared_review_unit_hash": "34173075ea7cf6fd5c87b8cbc97de1d161649fa9967a80c63a77d1bc59b0f37d",
+      "prepared_review_unit_hash": "1182ceb2c520f0b0f4134d33ba790ff0fb174ea061fe22ebb34dba23e6b2954d",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
@@ -26385,24 +26413,24 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "All three pull requests are unrelated bug fixes in neighboring generation/cache code, not duplicates of the same underlying change.",
-        "confidence": 0.96,
+        "summary": "The cluster contains three unrelated pull requests: one for QuantizedCache validation, one for prompt lookup decoding length capping, and one for empty-list validation in generation logits processing. They share only broad generation/theme similarity and should not be treated as duplicates.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
         "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46993 is the clearest standalone bug fix in the set: it corrects prompt lookup decoding length capping and includes a targeted test, making it the strongest representative if one must be chosen.",
+        "best_pr_reason": null,
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46781",
             "right": "pull_request:46993",
             "accept": false,
-            "reason": "Different problems and code paths: QuantizedCache layer-type validation vs prompt lookup decoding max_length handling. They only share a broad generation/cache area, not the same bug."
+            "reason": "Different code paths and bugs: cache layer-type validation vs prompt lookup decoding max_length handling. No plausible single merged fix."
           },
           {
             "left": "pull_request:46993",
             "right": "pull_request:46994",
             "accept": false,
-            "reason": "Both are generation-related, but one fixes candidate length capping in prompt lookup decoding while the other tightens validation for empty bad_words_ids/sequence_bias inputs. Separate defects, not mergeable as one PR."
+            "reason": "Both are generation-related, but they fix unrelated validation/logic bugs in different modules (candidate generation vs logits processor validation)."
           }
         ]
       },
@@ -26432,12 +26460,12 @@
       "soft_pairs": [
         "pull_request:47026|pull_request:47046"
       ],
-      "prepared_review_unit_hash": "39cc16d099daa56c20d7a3d8e539deca60ccbb6ab4e1955e94a2744dae2cd9f6",
+      "prepared_review_unit_hash": "f48a065bc422c4ec5e6aa2960062dc82f69bade5fa35e169b23486ad215e600d",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4442,
+        "serialized_chars": 4441,
         "estimated_input_tokens": 1111,
         "estimated_eval_tokens": 2478
       },
@@ -26445,7 +26473,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4442,
+        "serialized_chars": 4441,
         "estimated_input_tokens": 1111,
         "estimated_eval_tokens": 2478
       },
@@ -26454,18 +26482,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs target the same xLSTM train-mode failure in `modeling_xlstm.py` under the shared upstream issue; PR 47046 looks like the broader, more complete fix and likely subsumes the narrower 47026 change.",
-        "confidence": 0.81,
+        "summary": "Two open xLSTM PRs target the same train-mode forward path in `modeling_xlstm.py` and the same tracking issue. PR 47046 appears to be the broader, more complete fix (it explicitly covers both `return_last_states=False` and `use_cache`), so it is the best representative for the cluster.",
+        "confidence": 0.71,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 47046 is the better canonical representative because it addresses the same train-mode crash but also extends the fix to the related `use_cache` path and includes tests, making it the more complete change.",
+        "canonical_pr_reason": "PR 47046 is the more complete and likely superseding fix: it targets the same xLSTM train-mode crash path as 47026 but also covers `use_cache` and includes tests, making it the better canonical PR.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 47046 best represents the cluster because it covers the broader concrete failure mode in xLSTM train mode rather than only the narrower `return_last_states=False` case.",
+        "best_pr_reason": "PR 47046 is the strongest single PR to keep because it addresses the shared xLSTM train-mode failure mode more comprehensively than 47026.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:47026",
             "right": "pull_request:47046",
             "accept": true,
-            "reason": "Same underlying xLSTM train-mode crash in `train_with_padding`; 47046 appears to expand the same forward-path fix to the related `use_cache` case, so these are close enough to treat as duplicates."
+            "reason": "Both PRs fix the same xLSTM train-mode code path in `modeling_xlstm.py` and share the same issue target; 47046 looks like a superset of the 47026 fix rather than a separate change."
           }
         ]
       },
@@ -26503,12 +26531,12 @@
         "pull_request:47060|pull_request:47076",
         "pull_request:47074|pull_request:47076"
       ],
-      "prepared_review_unit_hash": "5726c39ac34c7b39d994520027025ac8f7169b9ca0fa414c3b95699ec2835bea",
+      "prepared_review_unit_hash": "1851f642cbcf838d0dab29dfd1a2d8e2ee8d5dff1a6171bb6d33c9b2dec681b5",
       "component_budget": {
         "node_count": 5,
         "item_count": 5,
         "soft_pair_count": 6,
-        "serialized_chars": 12761,
+        "serialized_chars": 12763,
         "estimated_input_tokens": 3191,
         "estimated_eval_tokens": 6638
       },
@@ -26516,7 +26544,7 @@
         "node_count": 5,
         "item_count": 5,
         "soft_pair_count": 6,
-        "serialized_chars": 12761,
+        "serialized_chars": 12763,
         "estimated_input_tokens": 3191,
         "estimated_eval_tokens": 6638
       },
@@ -26525,48 +26553,48 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These PRs are mostly unrelated multimodal/processor fixes. The only mild overlap is around image_text_to_text generation plumbing, but the concrete bugs differ enough that I would not collapse them as duplicates.",
-        "confidence": 0.87,
+        "summary": "These PRs are thematically related around multimodal pipeline/processor handling, but they fix distinct bugs and changes rather than the same underlying issue. No soft edge is a true duplicate.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 47051 is the broadest and most central multimodal-generation plumbing change, touching the shared pipeline code that also overlaps with 47032 and 47074.",
+        "canonical_pr_reason": "47051 is the broadest umbrella change in the shared multimodal pipeline codepath and most representative of the cluster, but it still addresses a different fix than the other PRs.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 47051 is the best representative cluster member because it addresses the common pipeline-layer generation-kwargs path rather than a model-specific edge case.",
+        "best_pr_reason": "47051 is the best single representative PR because it spans the shared pipeline plumbing touched by nearby PRs, even though the cluster does not contain a true duplicate set.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:47032",
             "right": "pull_request:47051",
             "accept": false,
-            "reason": "Both touch image_text_to_text, but 47032 is a specific stop_sequence-to-stop_strings fix while 47051 is broader generation-kwargs forwarding; related area, not the same concrete bug."
+            "reason": "Same area (image_text_to_text) but different bugs: stop_sequence behavior vs direct generation kwarg forwarding."
           },
           {
             "left": "pull_request:47051",
             "right": "pull_request:47074",
             "accept": false,
-            "reason": "Both involve AnyToAnyPipeline, but one changes kwarg forwarding and the other normalizes input keys; different failure modes and not mergeable as one duplicate fix."
+            "reason": "Both touch AnyToAnyPipeline, but one forwards kwargs while the other normalizes input keys; separate code-path fixes."
           },
           {
             "left": "pull_request:47051",
             "right": "pull_request:47060",
             "accept": false,
-            "reason": "Different code paths and bugs: multimodal generation kwarg plumbing vs Qwen Omni batched decode handling."
+            "reason": "Different subsystems and bugs: multimodal pipeline kwarg plumbing vs Qwen Omni batched post-processing."
           },
           {
             "left": "pull_request:47060",
             "right": "pull_request:47074",
             "accept": false,
-            "reason": "Qwen Omni post-processing and AnyToAny input normalization are unrelated changes despite both being multimodal."
+            "reason": "Unrelated fixes: Qwen batched decoding behavior versus AnyToAnyPipeline input key normalization."
           },
           {
             "left": "pull_request:47060",
             "right": "pull_request:47076",
             "accept": false,
-            "reason": "Different models and distinct processor logic; no shared concrete bug."
+            "reason": "Different model families and failure modes: Qwen post-processing vs Gemma4 audio placeholder validation."
           },
           {
             "left": "pull_request:47074",
             "right": "pull_request:47076",
             "accept": false,
-            "reason": "Different subsystems and behaviors: AnyToAny key normalization vs Gemma4 placeholder validation."
+            "reason": "Both are multimodal, but they fix different problems in different components: pipeline input mapping vs processor placeholder consistency."
           }
         ]
       },
@@ -26596,7 +26624,7 @@
       "soft_pairs": [
         "pull_request:47105|pull_request:47129"
       ],
-      "prepared_review_unit_hash": "1199759cbebde22260767ed0b498af0a074b8b0404f2049cf42c74664d105d93",
+      "prepared_review_unit_hash": "5e5adcdbccf9a97bd2f73884338dbc2150d095a5f5fcc19aaa4d3114bf3f9515",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -26618,18 +26646,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "The two PRs are unrelated: one re-enables a VITS forward test, the other fixes an RT-DETR v2 weight-conversion utility. They do not share a concrete bug or change path, so the soft edge should be rejected.",
-        "confidence": 0.97,
+        "summary": "These two PRs are unrelated: one re-enables and adjusts a VITS model forward test, while the other updates an RT-DETR v2 weight-conversion script for changed model keys.",
+        "confidence": 0.99,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "pull_request:47129 is the better canonical PR for this cluster because it contains the substantive source-code fix in the RT-DETR v2 converter, whereas 47105 is only a test adjustment/re-enable with no shared code path.",
+        "canonical_pr_reason": "No clear canonical PR: the cluster contains two unrelated pull requests touching different models and different code paths.",
         "best_issue_reason": null,
-        "best_pr_reason": "pull_request:47129 is the strongest representative overall since it directly modifies production conversion logic and is more self-contained as a code fix.",
+        "best_pr_reason": "Neither PR is a good cluster representative for deduping because they address separate, non-overlapping changes; if forced, 47129 is slightly more self-contained, but it is not a duplicate of 47105.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:47105",
             "right": "pull_request:47129",
             "accept": false,
-            "reason": "Different models and different artifacts: a VITS test re-enable versus an RT-DETR v2 converter fix. They do not appear to fix the same underlying bug or be mergeable into one PR."
+            "reason": "Different artifacts and code paths: VITS test changes versus RT-DETR v2 converter fixes. They do not appear to be the same underlying bug or change."
           }
         ]
       },
@@ -26662,12 +26690,12 @@
         "pull_request:47122|pull_request:47124",
         "pull_request:47123|pull_request:47124"
       ],
-      "prepared_review_unit_hash": "e4a7d9eb3432cc9c2363f762de3bc8b421b251d4fd2001ee32e5d63834be47c5",
+      "prepared_review_unit_hash": "24d92246f23caefccd79e3d7932d01b4bcbff295a37656d4f67388169c1a40e3",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 6619,
+        "serialized_chars": 6618,
         "estimated_input_tokens": 1655,
         "estimated_eval_tokens": 3566
       },
@@ -26675,7 +26703,7 @@
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 6619,
+        "serialized_chars": 6618,
         "estimated_input_tokens": 1655,
         "estimated_eval_tokens": 3566
       },
@@ -26684,30 +26712,30 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Three open PRs with only superficial similarity: one code fix for EdgeTam runtime kwargs handling and two separate test-only updates for different model suites. They are not duplicates of the same underlying bug/change.",
-        "confidence": 0.93,
+        "summary": "These are three unrelated pull requests fixing separate model-specific integration-test failures: EdgeTam runtime kwargs handling, MiniMax-M3-VL load/offload test setup, and Falcon-Mamba expected output updates. They are not duplicates of one another.",
+        "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 47122 is the strongest canonical PR candidate because it fixes an actual runtime code path in EdgeTam by dropping unsupported `original_sizes` before calling the TimmWrapper backbone; the others are test harness adjustments for unrelated models.",
+        "canonical_pr_reason": "PR 47122 is the best canonical PR because it contains the clearest concrete code-path fix: removing unsupported `original_sizes` before calling the EdgeTam backbone. The other PRs are test-only adjustments for different models and failure modes.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 47122 is the best overall PR: it contains the substantive bug fix, targets the concrete failure mode, and is more merge-worthy than the test-only PRs.",
+        "best_pr_reason": "PR 47122 is the strongest representative of the cluster because it fixes the underlying model implementation bug rather than only changing test harness expectations or loader setup.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:47122",
             "right": "pull_request:47123",
             "accept": false,
-            "reason": "Different models and failure modes: EdgeTam runtime kwargs incompatibility vs MiniMax-M3-VL load/offload setup in tests. Not the same underlying bug or change."
+            "reason": "Different models and different failures: EdgeTam fixes unsupported `original_sizes` passed into the backbone, while MiniMax-M3-VL only adds an offload folder in a test loader. Not the same underlying bug or change."
           },
           {
             "left": "pull_request:47122",
             "right": "pull_request:47124",
             "accept": false,
-            "reason": "Unrelated fixes: EdgeTam code-path argument cleanup vs Falcon-Mamba expected-output updates in integration tests. Too broad to be one duplicate change."
+            "reason": "EdgeTam runtime kwargs handling vs. Falcon-Mamba test expectation updates. They touch different code paths and failure types, so they are not duplicates."
           },
           {
             "left": "pull_request:47123",
             "right": "pull_request:47124",
             "accept": false,
-            "reason": "Both are test-related, but for different models and different issues (`load_error` vs `output_mismatch`). They should not be merged as duplicates."
+            "reason": "Both are test-related, but for different models and different issues: one addresses model loading/offloading, the other adjusts generation expectations. Not the same underlying fix."
           }
         ]
       },
diff --git a/analysis/current/manifest.json b/analysis/current/manifest.json
index de5335b3ec48c319b4540a384d4aaaed23923493..831c64f3b04d0ea8d934535c2c4bdb74b1d83f8d 100644
--- a/analysis/current/manifest.json
+++ b/analysis/current/manifest.json
@@ -1,8 +1,8 @@
 {
-  "analysis_id": "hybrid-model-20260707t180018z",
+  "analysis_id": "hybrid-model-20260708t000615z",
   "archived_artifacts": {
-    "hybrid": "snapshots/20260707T180018Z/analysis-runs/hybrid-model-20260707t180018z/analysis-report-hybrid.json",
-    "hybrid_reviews": "snapshots/20260707T180018Z/analysis-runs/hybrid-model-20260707t180018z/analysis-report-hybrid.llm-reviews.json"
+    "hybrid": "snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.json",
+    "hybrid_reviews": "snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.llm-reviews.json"
   },
   "artifacts": {
     "hybrid": "analysis/current/analysis-report-hybrid.json",
@@ -10,9 +10,9 @@
   },
   "channel": "canonical",
   "model": null,
-  "published_at": "2026-07-07T18:14:26Z",
+  "published_at": "2026-07-08T00:12:28Z",
   "repo": "huggingface/transformers",
   "schema_version": 1,
-  "snapshot_id": "20260707T180018Z",
+  "snapshot_id": "20260708T000615Z",
   "variant": "hybrid"
 }
diff --git a/snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.json b/snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.json
new file mode 100644
index 0000000000000000000000000000000000000000..25898d28b3d77aae20d9f5db6ecb5534f659c2ea
--- /dev/null
+++ b/snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.json
@@ -0,0 +1,2851 @@
+{
+  "schema_version": "1.0",
+  "repo": "huggingface/transformers",
+  "snapshot_id": "20260708T000615Z",
+  "generated_at": "2026-07-08T00:12:27Z",
+  "evidence_quality": "full",
+  "llm_enrichment": true,
+  "meta_bugs": [
+    {
+      "cluster_id": "cluster-43979-11",
+      "summary": "Cluster of 1 issues and 10 PRs centered on issue #43979.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 43979,
+      "canonical_pr_number": 44007,
+      "issue_numbers": [
+        43979
+      ],
+      "pr_numbers": [
+        43996,
+        44007,
+        44013,
+        44044,
+        44066,
+        44072,
+        44085,
+        44129,
+        44154,
+        44722
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target",
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 43996,
+          "right_pr_number": 44007,
+          "code_similarity": 0.179,
+          "size_similarity": 0.576,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.429,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 43996,
+          "right_pr_number": 44013,
+          "code_similarity": 0.122,
+          "size_similarity": 0.318,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.392,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 43996,
+          "right_pr_number": 44044,
+          "code_similarity": 0.245,
+          "size_similarity": 0.864,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.479,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 43996,
+          "right_pr_number": 44066,
+          "code_similarity": 0.225,
+          "size_similarity": 0.818,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.408,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 43996,
+          "right_pr_number": 44072,
+          "code_similarity": 0.14,
+          "size_similarity": 0.303,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.528,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 43996,
+          "right_pr_number": 44085,
+          "code_similarity": 0.216,
+          "size_similarity": 0.783,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.398,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 43996,
+          "right_pr_number": 44129,
+          "code_similarity": 0.163,
+          "size_similarity": 0.643,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.229,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 43996,
+          "right_pr_number": 44154,
+          "code_similarity": 0.153,
+          "size_similarity": 0.535,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.31,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 43996,
+          "right_pr_number": 44722,
+          "code_similarity": 0.225,
+          "size_similarity": 0.848,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.368,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44007,
+          "right_pr_number": 44013,
+          "code_similarity": 0.19,
+          "size_similarity": 0.553,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.531,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44007,
+          "right_pr_number": 44044,
+          "code_similarity": 0.186,
+          "size_similarity": 0.667,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.354,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44007,
+          "right_pr_number": 44066,
+          "code_similarity": 0.188,
+          "size_similarity": 0.704,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.315,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44007,
+          "right_pr_number": 44072,
+          "code_similarity": 0.212,
+          "size_similarity": 0.526,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.708,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44007,
+          "right_pr_number": 44085,
+          "code_similarity": 0.195,
+          "size_similarity": 0.735,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.318,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44007,
+          "right_pr_number": 44129,
+          "code_similarity": 0.103,
+          "size_similarity": 0.37,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.191,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44007,
+          "right_pr_number": 44154,
+          "code_similarity": 0.238,
+          "size_similarity": 0.93,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.344,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44007,
+          "right_pr_number": 44722,
+          "code_similarity": 0.178,
+          "size_similarity": 0.679,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.28,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44013,
+          "right_pr_number": 44044,
+          "code_similarity": 0.126,
+          "size_similarity": 0.368,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.351,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44013,
+          "right_pr_number": 44066,
+          "code_similarity": 0.127,
+          "size_similarity": 0.389,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.325,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44013,
+          "right_pr_number": 44072,
+          "code_similarity": 0.29,
+          "size_similarity": 0.952,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.667,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44013,
+          "right_pr_number": 44085,
+          "code_similarity": 0.131,
+          "size_similarity": 0.406,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.329,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44013,
+          "right_pr_number": 44129,
+          "code_similarity": 0.07,
+          "size_similarity": 0.205,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.192,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44013,
+          "right_pr_number": 44154,
+          "code_similarity": 0.177,
+          "size_similarity": 0.594,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.389,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44013,
+          "right_pr_number": 44722,
+          "code_similarity": 0.118,
+          "size_similarity": 0.375,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.287,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44044,
+          "right_pr_number": 44066,
+          "code_similarity": 0.25,
+          "size_similarity": 0.947,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.404,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44044,
+          "right_pr_number": 44072,
+          "code_similarity": 0.136,
+          "size_similarity": 0.351,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.442,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44044,
+          "right_pr_number": 44085,
+          "code_similarity": 0.24,
+          "size_similarity": 0.906,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.394,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44044,
+          "right_pr_number": 44129,
+          "code_similarity": 0.147,
+          "size_similarity": 0.555,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.243,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44044,
+          "right_pr_number": 44154,
+          "code_similarity": 0.17,
+          "size_similarity": 0.62,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.306,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44044,
+          "right_pr_number": 44722,
+          "code_similarity": 0.257,
+          "size_similarity": 0.982,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.402,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44066,
+          "right_pr_number": 44072,
+          "code_similarity": 0.133,
+          "size_similarity": 0.37,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.393,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44066,
+          "right_pr_number": 44085,
+          "code_similarity": 0.763,
+          "size_similarity": 0.957,
+          "file_overlap": 0.5,
+          "area_overlap": 0.825,
+          "patch_similarity": 0.887,
+          "shared_filenames": [
+            "src/transformers/models/gptj/modeling_gptj.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/gptj/modeling_gptj.py",
+              "left_ranges": [
+                [
+                  33,
+                  41
+                ],
+                [
+                  174,
+                  181
+                ],
+                [
+                  250,
+                  257
+                ],
+                [
+                  398,
+                  419
+                ],
+                [
+                  425,
+                  434
+                ],
+                [
+                  459,
+                  466
+                ],
+                [
+                  471,
+                  488
+                ],
+                [
+                  518,
+                  540
+                ],
+                [
+                  554,
+                  560
+                ],
+                [
+                  566,
+                  575
+                ],
+                [
+                  580,
+                  598
+                ],
+                [
+                  601,
+                  606
+                ],
+                [
+                  634,
+                  640
+                ],
+                [
+                  646,
+                  653
+                ],
+                [
+                  658,
+                  674
+                ],
+                [
+                  717,
+                  722
+                ],
+                [
+                  738,
+                  744
+                ],
+                [
+                  749,
+                  772
+                ],
+                [
+                  790,
+                  795
+                ]
+              ],
+              "right_ranges": [
+                [
+                  33,
+                  40
+                ],
+                [
+                  173,
+                  180
+                ],
+                [
+                  249,
+                  256
+                ],
+                [
+                  397,
+                  405
+                ],
+                [
+                  408,
+                  419
+                ],
+                [
+                  425,
+                  434
+                ],
+                [
+                  459,
+                  465
+                ],
+                [
+                  470,
+                  489
+                ],
+                [
+                  517,
+                  539
+                ],
+                [
+                  553,
+                  559
+                ],
+                [
+                  565,
+                  574
+                ],
+                [
+                  579,
+                  597
+                ],
+                [
+                  600,
+                  611
+                ],
+                [
+                  633,
+                  639
+                ],
+                [
+                  645,
+                  652
+                ],
+                [
+                  657,
+                  673
+                ],
+                [
+                  716,
+                  728
+                ],
+                [
+                  737,
+                  743
+                ],
+                [
+                  748,
+                  771
+                ],
+                [
+                  789,
+                  794
+                ]
+              ]
+            }
+          ]
+        },
+        {
+          "left_pr_number": 44066,
+          "right_pr_number": 44129,
+          "code_similarity": 0.145,
+          "size_similarity": 0.526,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.263,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44066,
+          "right_pr_number": 44154,
+          "code_similarity": 0.174,
+          "size_similarity": 0.654,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.286,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44066,
+          "right_pr_number": 44722,
+          "code_similarity": 0.907,
+          "size_similarity": 0.964,
+          "file_overlap": 1.0,
+          "area_overlap": 0.808,
+          "patch_similarity": 0.874,
+          "shared_filenames": [
+            "src/transformers/models/codegen/modeling_codegen.py",
+            "src/transformers/models/gptj/modeling_gptj.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/codegen/modeling_codegen.py",
+              "left_ranges": [
+                [
+                  245,
+                  266
+                ]
+              ],
+              "right_ranges": [
+                [
+                  228,
+                  234
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/gptj/modeling_gptj.py",
+              "left_ranges": [
+                [
+                  33,
+                  41
+                ],
+                [
+                  174,
+                  181
+                ],
+                [
+                  250,
+                  257
+                ],
+                [
+                  398,
+                  419
+                ],
+                [
+                  425,
+                  434
+                ],
+                [
+                  459,
+                  466
+                ],
+                [
+                  471,
+                  488
+                ],
+                [
+                  518,
+                  540
+                ],
+                [
+                  554,
+                  560
+                ],
+                [
+                  566,
+                  575
+                ],
+                [
+                  580,
+                  598
+                ],
+                [
+                  601,
+                  606
+                ],
+                [
+                  634,
+                  640
+                ],
+                [
+                  646,
+                  653
+                ],
+                [
+                  658,
+                  674
+                ],
+                [
+                  717,
+                  722
+                ],
+                [
+                  738,
+                  744
+                ],
+                [
+                  749,
+                  772
+                ],
+                [
+                  790,
+                  795
+                ]
+              ],
+              "right_ranges": [
+                [
+                  33,
+                  42
+                ],
+                [
+                  175,
+                  182
+                ],
+                [
+                  251,
+                  258
+                ],
+                [
+                  399,
+                  420
+                ],
+                [
+                  426,
+                  435
+                ],
+                [
+                  460,
+                  467
+                ],
+                [
+                  472,
+                  489
+                ],
+                [
+                  519,
+                  540
+                ],
+                [
+                  554,
+                  560
+                ],
+                [
+                  566,
+                  575
+                ],
+                [
+                  580,
+                  598
+                ],
+                [
+                  601,
+                  612
+                ],
+                [
+                  634,
+                  640
+                ],
+                [
+                  646,
+                  653
+                ],
+                [
+                  658,
+                  663
+                ],
+                [
+                  666,
+                  674
+                ],
+                [
+                  717,
+                  722
+                ],
+                [
+                  738,
+                  744
+                ],
+                [
+                  749,
+                  772
+                ],
+                [
+                  790,
+                  795
+                ]
+              ]
+            }
+          ]
+        },
+        {
+          "left_pr_number": 44072,
+          "right_pr_number": 44085,
+          "code_similarity": 0.137,
+          "size_similarity": 0.387,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.398,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44072,
+          "right_pr_number": 44129,
+          "code_similarity": 0.074,
+          "size_similarity": 0.195,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.231,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44072,
+          "right_pr_number": 44154,
+          "code_similarity": 0.175,
+          "size_similarity": 0.566,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.414,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44072,
+          "right_pr_number": 44722,
+          "code_similarity": 0.124,
+          "size_similarity": 0.357,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.347,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44085,
+          "right_pr_number": 44129,
+          "code_similarity": 0.141,
+          "size_similarity": 0.503,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.272,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44085,
+          "right_pr_number": 44154,
+          "code_similarity": 0.18,
+          "size_similarity": 0.684,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.289,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44085,
+          "right_pr_number": 44722,
+          "code_similarity": 0.728,
+          "size_similarity": 0.923,
+          "file_overlap": 0.5,
+          "area_overlap": 0.791,
+          "patch_similarity": 0.78,
+          "shared_filenames": [
+            "src/transformers/models/gptj/modeling_gptj.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/gptj/modeling_gptj.py",
+              "left_ranges": [
+                [
+                  33,
+                  40
+                ],
+                [
+                  173,
+                  180
+                ],
+                [
+                  249,
+                  256
+                ],
+                [
+                  397,
+                  405
+                ],
+                [
+                  408,
+                  419
+                ],
+                [
+                  425,
+                  434
+                ],
+                [
+                  459,
+                  465
+                ],
+                [
+                  470,
+                  489
+                ],
+                [
+                  517,
+                  539
+                ],
+                [
+                  553,
+                  559
+                ],
+                [
+                  565,
+                  574
+                ],
+                [
+                  579,
+                  597
+                ],
+                [
+                  600,
+                  611
+                ],
+                [
+                  633,
+                  639
+                ],
+                [
+                  645,
+                  652
+                ],
+                [
+                  657,
+                  673
+                ],
+                [
+                  716,
+                  728
+                ],
+                [
+                  737,
+                  743
+                ],
+                [
+                  748,
+                  771
+                ],
+                [
+                  789,
+                  794
+                ]
+              ],
+              "right_ranges": [
+                [
+                  33,
+                  42
+                ],
+                [
+                  175,
+                  182
+                ],
+                [
+                  251,
+                  258
+                ],
+                [
+                  399,
+                  420
+                ],
+                [
+                  426,
+                  435
+                ],
+                [
+                  460,
+                  467
+                ],
+                [
+                  472,
+                  489
+                ],
+                [
+                  519,
+                  540
+                ],
+                [
+                  554,
+                  560
+                ],
+                [
+                  566,
+                  575
+                ],
+                [
+                  580,
+                  598
+                ],
+                [
+                  601,
+                  612
+                ],
+                [
+                  634,
+                  640
+                ],
+                [
+                  646,
+                  653
+                ],
+                [
+                  658,
+                  663
+                ],
+                [
+                  666,
+                  674
+                ],
+                [
+                  717,
+                  722
+                ],
+                [
+                  738,
+                  744
+                ],
+                [
+                  749,
+                  772
+                ],
+                [
+                  790,
+                  795
+                ]
+              ]
+            }
+          ]
+        },
+        {
+          "left_pr_number": 44129,
+          "right_pr_number": 44154,
+          "code_similarity": 0.099,
+          "size_similarity": 0.344,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.199,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44129,
+          "right_pr_number": 44722,
+          "code_similarity": 0.146,
+          "size_similarity": 0.545,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.247,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 44154,
+          "right_pr_number": 44722,
+          "code_similarity": 0.164,
+          "size_similarity": 0.631,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.255,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-41211-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #41211.",
+      "status": "open",
+      "confidence": 0.55,
+      "canonical_issue_number": 41211,
+      "canonical_pr_number": 44339,
+      "issue_numbers": [
+        41211
+      ],
+      "pr_numbers": [
+        41356,
+        44339
+      ],
+      "evidence_types": [
+        "closing_reference"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 41356,
+          "right_pr_number": 44339,
+          "code_similarity": 0.155,
+          "size_similarity": 0.094,
+          "file_overlap": 0.273,
+          "area_overlap": 0.078,
+          "patch_similarity": 0.183,
+          "shared_filenames": [
+            "docs/source/en/model_doc/deimv2.md",
+            "src/transformers/models/deimv2/__init__.py",
+            "src/transformers/models/deimv2/configuration_deimv2.py",
+            "src/transformers/models/deimv2/modeling_deimv2.py",
+            "tests/models/deimv2/__init__.py",
+            "tests/models/deimv2/test_modeling_deimv2.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "docs/source/en/model_doc/deimv2.md",
+              "left_ranges": [
+                [
+                  1,
+                  132
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  65
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/deimv2/__init__.py",
+              "left_ranges": [
+                [
+                  1,
+                  15
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  29
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/deimv2/configuration_deimv2.py",
+              "left_ranges": [
+                [
+                  1,
+                  74
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  266
+                ]
+              ]
+            },
+            {
+              "filename": "tests/models/deimv2/test_modeling_deimv2.py",
+              "left_ranges": [
+                [
+                  1,
+                  15
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  1734
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-46196-3",
+      "summary": "Cluster of 3 related pull requests linked by soft_similarity.",
+      "status": "open",
+      "confidence": 0.5,
+      "canonical_issue_number": null,
+      "canonical_pr_number": 46196,
+      "issue_numbers": [],
+      "pr_numbers": [
+        46196,
+        46836,
+        47143
+      ],
+      "evidence_types": [
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 46196,
+          "right_pr_number": 46836,
+          "code_similarity": 0.205,
+          "size_similarity": 0.936,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.12,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 46196,
+          "right_pr_number": 47143,
+          "code_similarity": 0.124,
+          "size_similarity": 0.513,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.145,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        },
+        {
+          "left_pr_number": 46836,
+          "right_pr_number": 47143,
+          "code_similarity": 0.169,
+          "size_similarity": 0.48,
+          "file_overlap": 0.111,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.266,
+          "shared_filenames": [
+            "src/transformers/models/__init__.py",
+            "src/transformers/models/auto/auto_mappings.py",
+            "src/transformers/models/auto/modeling_auto.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/__init__.py",
+              "left_ranges": [
+                [
+                  452,
+                  458
+                ]
+              ],
+              "right_ranges": [
+                [
+                  305,
+                  311
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/auto/auto_mappings.py",
+              "left_ranges": [
+                [
+                  628,
+                  638
+                ],
+                [
+                  905,
+                  914
+                ],
+                [
+                  1112,
+                  1118
+                ],
+                [
+                  1238,
+                  1244
+                ]
+              ],
+              "right_ranges": [
+                [
+                  412,
+                  418
+                ],
+                [
+                  975,
+                  981
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/auto/modeling_auto.py",
+              "left_ranges": [
+                [
+                  501,
+                  507
+                ],
+                [
+                  1107,
+                  1113
+                ]
+              ],
+              "right_ranges": [
+                [
+                  348,
+                  354
+                ],
+                [
+                  1960,
+                  1966
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-46894-5",
+      "summary": "Cluster of 2 issues and 3 PRs centered on issue #46894.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 46894,
+      "canonical_pr_number": 46895,
+      "issue_numbers": [
+        46894,
+        46897
+      ],
+      "pr_numbers": [
+        46895,
+        46958,
+        47090
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target",
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 46895,
+          "right_pr_number": 46958,
+          "code_similarity": 0.779,
+          "size_similarity": 0.915,
+          "file_overlap": 1.0,
+          "area_overlap": 0.5,
+          "patch_similarity": 0.81,
+          "shared_filenames": [
+            "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
+            "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+            "tests/models/cohere_asr/test_modeling_cohere_asr.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
+              "left_ranges": [
+                [
+                  637,
+                  650
+                ]
+              ],
+              "right_ranges": [
+                [
+                  22,
+                  28
+                ],
+                [
+                  638,
+                  645
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+              "left_ranges": [
+                [
+                  500,
+                  513
+                ]
+              ],
+              "right_ranges": [
+                [
+                  16,
+                  22
+                ],
+                [
+                  501,
+                  508
+                ]
+              ]
+            },
+            {
+              "filename": "tests/models/cohere_asr/test_modeling_cohere_asr.py",
+              "left_ranges": [
+                [
+                  35,
+                  41
+                ],
+                [
+                  176,
+                  214
+                ]
+              ],
+              "right_ranges": [
+                [
+                  15,
+                  21
+                ],
+                [
+                  180,
+                  221
+                ]
+              ]
+            }
+          ]
+        },
+        {
+          "left_pr_number": 46895,
+          "right_pr_number": 47090,
+          "code_similarity": 0.819,
+          "size_similarity": 0.932,
+          "file_overlap": 1.0,
+          "area_overlap": 0.783,
+          "patch_similarity": 0.389,
+          "shared_filenames": [
+            "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
+            "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+            "tests/models/cohere_asr/test_modeling_cohere_asr.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
+              "left_ranges": [
+                [
+                  637,
+                  650
+                ]
+              ],
+              "right_ranges": [
+                [
+                  637,
+                  649
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+              "left_ranges": [
+                [
+                  500,
+                  513
+                ]
+              ],
+              "right_ranges": [
+                [
+                  500,
+                  512
+                ]
+              ]
+            },
+            {
+              "filename": "tests/models/cohere_asr/test_modeling_cohere_asr.py",
+              "left_ranges": [
+                [
+                  35,
+                  41
+                ],
+                [
+                  176,
+                  214
+                ]
+              ],
+              "right_ranges": [
+                [
+                  175,
+                  222
+                ]
+              ]
+            }
+          ]
+        },
+        {
+          "left_pr_number": 46958,
+          "right_pr_number": 47090,
+          "code_similarity": 0.742,
+          "size_similarity": 0.852,
+          "file_overlap": 1.0,
+          "area_overlap": 0.611,
+          "patch_similarity": 0.387,
+          "shared_filenames": [
+            "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
+            "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+            "tests/models/cohere_asr/test_modeling_cohere_asr.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/cohere_asr/modeling_cohere_asr.py",
+              "left_ranges": [
+                [
+                  22,
+                  28
+                ],
+                [
+                  638,
+                  645
+                ]
+              ],
+              "right_ranges": [
+                [
+                  637,
+                  649
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/cohere_asr/modular_cohere_asr.py",
+              "left_ranges": [
+                [
+                  16,
+                  22
+                ],
+                [
+                  501,
+                  508
+                ]
+              ],
+              "right_ranges": [
+                [
+                  500,
+                  512
+                ]
+              ]
+            },
+            {
+              "filename": "tests/models/cohere_asr/test_modeling_cohere_asr.py",
+              "left_ranges": [
+                [
+                  15,
+                  21
+                ],
+                [
+                  180,
+                  221
+                ]
+              ],
+              "right_ranges": [
+                [
+                  175,
+                  222
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-43240-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #43240.",
+      "status": "open",
+      "confidence": 0.75,
+      "canonical_issue_number": 43240,
+      "canonical_pr_number": 43251,
+      "issue_numbers": [
+        43240
+      ],
+      "pr_numbers": [
+        43251,
+        43254
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 43251,
+          "right_pr_number": 43254,
+          "code_similarity": 0.794,
+          "size_similarity": 0.64,
+          "file_overlap": 1.0,
+          "area_overlap": 0.667,
+          "patch_similarity": 0.882,
+          "shared_filenames": [
+            "src/transformers/loss/loss_utils.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/loss/loss_utils.py",
+              "left_ranges": [
+                [
+                  30,
+                  50
+                ]
+              ],
+              "right_ranges": [
+                [
+                  30,
+                  43
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-46693-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #46693.",
+      "status": "open",
+      "confidence": 0.55,
+      "canonical_issue_number": 46693,
+      "canonical_pr_number": 47134,
+      "issue_numbers": [
+        46693
+      ],
+      "pr_numbers": [
+        46696,
+        47134
+      ],
+      "evidence_types": [
+        "closing_reference"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 46696,
+          "right_pr_number": 47134,
+          "code_similarity": 0.464,
+          "size_similarity": 0.8,
+          "file_overlap": 1.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.029,
+          "shared_filenames": [
+            "src/transformers/modeling_flash_attention_utils.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/modeling_flash_attention_utils.py",
+              "left_ranges": [
+                [
+                  667,
+                  692
+                ]
+              ],
+              "right_ranges": [
+                [
+                  316,
+                  323
+                ],
+                [
+                  366,
+                  373
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-45647-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #45647.",
+      "status": "open",
+      "confidence": 0.75,
+      "canonical_issue_number": 45647,
+      "canonical_pr_number": 45738,
+      "issue_numbers": [
+        45647
+      ],
+      "pr_numbers": [
+        45737,
+        45738
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 45737,
+          "right_pr_number": 45738,
+          "code_similarity": 0.526,
+          "size_similarity": 0.11,
+          "file_overlap": 0.5,
+          "area_overlap": 1.0,
+          "patch_similarity": 0.024,
+          "shared_filenames": [
+            "src/transformers/models/musicgen_melody/modeling_musicgen_melody.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/musicgen_melody/modeling_musicgen_melody.py",
+              "left_ranges": [
+                [
+                  491,
+                  497
+                ]
+              ],
+              "right_ranges": [
+                [
+                  491,
+                  497
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-46869-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #46869.",
+      "status": "open",
+      "confidence": 0.75,
+      "canonical_issue_number": 46869,
+      "canonical_pr_number": 46886,
+      "issue_numbers": [
+        46869
+      ],
+      "pr_numbers": [
+        46886,
+        46922
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 46886,
+          "right_pr_number": 46922,
+          "code_similarity": 0.711,
+          "size_similarity": 0.857,
+          "file_overlap": 1.0,
+          "area_overlap": 0.5,
+          "patch_similarity": 0.429,
+          "shared_filenames": [
+            "src/transformers/models/yolos/modeling_yolos.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/yolos/modeling_yolos.py",
+              "left_ranges": [
+                [
+                  446,
+                  452
+                ]
+              ],
+              "right_ranges": [
+                [
+                  434,
+                  440
+                ],
+                [
+                  446,
+                  452
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-46159-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #46159.",
+      "status": "open",
+      "confidence": 0.55,
+      "canonical_issue_number": 46159,
+      "canonical_pr_number": 46174,
+      "issue_numbers": [
+        46159
+      ],
+      "pr_numbers": [
+        46174,
+        46780
+      ],
+      "evidence_types": [
+        "closing_reference"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 46174,
+          "right_pr_number": 46780,
+          "code_similarity": 0.841,
+          "size_similarity": 0.87,
+          "file_overlap": 1.0,
+          "area_overlap": 0.767,
+          "patch_similarity": 0.657,
+          "shared_filenames": [
+            "src/transformers/models/aimv2/modeling_aimv2.py",
+            "src/transformers/models/d_fine/modeling_d_fine.py",
+            "src/transformers/models/deimv2/modeling_deimv2.py",
+            "src/transformers/models/pp_doclayout_v2/modeling_pp_doclayout_v2.py",
+            "src/transformers/models/pp_doclayout_v3/modeling_pp_doclayout_v3.py",
+            "src/transformers/models/rt_detr/modeling_rt_detr.py",
+            "src/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py",
+            "src/transformers/models/vit_mae/modeling_vit_mae.py",
+            "src/transformers/models/vit_mae/modular_vit_mae.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/aimv2/modeling_aimv2.py",
+              "left_ranges": [
+                [
+                  134,
+                  141
+                ],
+                [
+                  144,
+                  154
+                ],
+                [
+                  157,
+                  165
+                ]
+              ],
+              "right_ranges": [
+                [
+                  134,
+                  140
+                ],
+                [
+                  143,
+                  153
+                ],
+                [
+                  156,
+                  162
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/d_fine/modeling_d_fine.py",
+              "left_ranges": [
+                [
+                  631,
+                  638
+                ],
+                [
+                  641,
+                  651
+                ],
+                [
+                  654,
+                  662
+                ]
+              ],
+              "right_ranges": [
+                [
+                  631,
+                  637
+                ],
+                [
+                  640,
+                  650
+                ],
+                [
+                  653,
+                  659
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/deimv2/modeling_deimv2.py",
+              "left_ranges": [
+                [
+                  703,
+                  710
+                ],
+                [
+                  713,
+                  723
+                ],
+                [
+                  726,
+                  734
+                ]
+              ],
+              "right_ranges": [
+                [
+                  703,
+                  709
+                ],
+                [
+                  712,
+                  722
+                ],
+                [
+                  725,
+                  731
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/pp_doclayout_v2/modeling_pp_doclayout_v2.py",
+              "left_ranges": [
+                [
+                  1551,
+                  1558
+                ],
+                [
+                  1561,
+                  1571
+                ],
+                [
+                  1574,
+                  1582
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1551,
+                  1557
+                ],
+                [
+                  1560,
+                  1570
+                ],
+                [
+                  1573,
+                  1579
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/pp_doclayout_v3/modeling_pp_doclayout_v3.py",
+              "left_ranges": [
+                [
+                  804,
+                  811
+                ],
+                [
+                  814,
+                  824
+                ],
+                [
+                  827,
+                  835
+                ]
+              ],
+              "right_ranges": [
+                [
+                  804,
+                  810
+                ],
+                [
+                  813,
+                  823
+                ],
+                [
+                  826,
+                  832
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/rt_detr/modeling_rt_detr.py",
+              "left_ranges": [
+                [
+                  847,
+                  854
+                ],
+                [
+                  857,
+                  867
+                ],
+                [
+                  870,
+                  878
+                ]
+              ],
+              "right_ranges": [
+                [
+                  847,
+                  853
+                ],
+                [
+                  856,
+                  866
+                ],
+                [
+                  869,
+                  875
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py",
+              "left_ranges": [
+                [
+                  976,
+                  983
+                ],
+                [
+                  986,
+                  996
+                ],
+                [
+                  999,
+                  1007
+                ]
+              ],
+              "right_ranges": [
+                [
+                  976,
+                  982
+                ],
+                [
+                  985,
+                  995
+                ],
+                [
+                  998,
+                  1004
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/vit_mae/modeling_vit_mae.py",
+              "left_ranges": [
+                [
+                  152,
+                  159
+                ],
+                [
+                  162,
+                  172
+                ],
+                [
+                  175,
+                  183
+                ]
+              ],
+              "right_ranges": [
+                [
+                  152,
+                  158
+                ],
+                [
+                  161,
+                  171
+                ],
+                [
+                  174,
+                  180
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/vit_mae/modular_vit_mae.py",
+              "left_ranges": [
+                [
+                  62,
+                  69
+                ],
+                [
+                  72,
+                  82
+                ],
+                [
+                  85,
+                  93
+                ]
+              ],
+              "right_ranges": [
+                [
+                  62,
+                  68
+                ],
+                [
+                  71,
+                  81
+                ],
+                [
+                  84,
+                  90
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-47013-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #47013.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 47013,
+      "canonical_pr_number": 47046,
+      "issue_numbers": [
+        47013
+      ],
+      "pr_numbers": [
+        47026,
+        47046
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target",
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 47026,
+          "right_pr_number": 47046,
+          "code_similarity": 0.308,
+          "size_similarity": 0.264,
+          "file_overlap": 0.5,
+          "area_overlap": 0.229,
+          "patch_similarity": 0.169,
+          "shared_filenames": [
+            "src/transformers/models/xlstm/modeling_xlstm.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/xlstm/modeling_xlstm.py",
+              "left_ranges": [
+                [
+                  816,
+                  840
+                ],
+                [
+                  843,
+                  851
+                ]
+              ],
+              "right_ranges": [
+                [
+                  25,
+                  39
+                ],
+                [
+                  789,
+                  795
+                ],
+                [
+                  808,
+                  817
+                ],
+                [
+                  823,
+                  846
+                ],
+                [
+                  849,
+                  857
+                ],
+                [
+                  1388,
+                  1394
+                ],
+                [
+                  1399,
+                  1405
+                ],
+                [
+                  1461,
+                  1473
+                ],
+                [
+                  1504,
+                  1522
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-45561-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #45561.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 45561,
+      "canonical_pr_number": 45645,
+      "issue_numbers": [
+        45561
+      ],
+      "pr_numbers": [
+        45639,
+        45645
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target",
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 45639,
+          "right_pr_number": 45645,
+          "code_similarity": 0.721,
+          "size_similarity": 0.973,
+          "file_overlap": 0.75,
+          "area_overlap": 0.563,
+          "patch_similarity": 0.696,
+          "shared_filenames": [
+            "src/transformers/testing_utils.py",
+            "tests/utils/test_testing_utils.py",
+            "utils/notification_service.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/testing_utils.py",
+              "left_ranges": [
+                [
+                  3611,
+                  3618
+                ],
+                [
+                  3835,
+                  3869
+                ]
+              ],
+              "right_ranges": [
+                [
+                  3525,
+                  3558
+                ],
+                [
+                  3782,
+                  3788
+                ]
+              ]
+            },
+            {
+              "filename": "tests/utils/test_testing_utils.py",
+              "left_ranges": [
+                [
+                  1,
+                  144
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  114
+                ]
+              ]
+            },
+            {
+              "filename": "utils/notification_service.py",
+              "left_ranges": [
+                [
+                  935,
+                  967
+                ]
+              ],
+              "right_ranges": [
+                [
+                  935,
+                  967
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-45081-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #45081.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 45081,
+      "canonical_pr_number": 45317,
+      "issue_numbers": [
+        45081
+      ],
+      "pr_numbers": [
+        45086,
+        45317
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target",
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 45086,
+          "right_pr_number": 45317,
+          "code_similarity": 0.589,
+          "size_similarity": 0.34,
+          "file_overlap": 0.5,
+          "area_overlap": 1.0,
+          "patch_similarity": 0.136,
+          "shared_filenames": [
+            "src/transformers/tokenization_utils_tokenizers.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/tokenization_utils_tokenizers.py",
+              "left_ranges": [
+                [
+                  1360,
+                  1370
+                ],
+                [
+                  1374,
+                  1380
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1360,
+                  1370
+                ],
+                [
+                  1374,
+                  1380
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-43698-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #43698.",
+      "status": "open",
+      "confidence": 0.75,
+      "canonical_issue_number": 43698,
+      "canonical_pr_number": 43779,
+      "issue_numbers": [
+        43698
+      ],
+      "pr_numbers": [
+        43779,
+        43816
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 43779,
+          "right_pr_number": 43816,
+          "code_similarity": 0.418,
+          "size_similarity": 0.538,
+          "file_overlap": 1.0,
+          "area_overlap": 0.02,
+          "patch_similarity": 0.02,
+          "shared_filenames": [
+            "src/transformers/integrations/integration_utils.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/integrations/integration_utils.py",
+              "left_ranges": [
+                [
+                  2227,
+                  2241
+                ],
+                [
+                  2303,
+                  2309
+                ]
+              ],
+              "right_ranges": [
+                [
+                  2278,
+                  2291
+                ],
+                [
+                  2309,
+                  2322
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-44001-2",
+      "summary": "Cluster of 2 related pull requests linked by soft_similarity.",
+      "status": "open",
+      "confidence": 0.5,
+      "canonical_issue_number": null,
+      "canonical_pr_number": 44004,
+      "issue_numbers": [],
+      "pr_numbers": [
+        44001,
+        44004
+      ],
+      "evidence_types": [
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 44001,
+          "right_pr_number": 44004,
+          "code_similarity": 0.087,
+          "size_similarity": 0.237,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.262,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-44002-2",
+      "summary": "Cluster of 2 related pull requests linked by soft_similarity.",
+      "status": "open",
+      "confidence": 0.5,
+      "canonical_issue_number": null,
+      "canonical_pr_number": 44002,
+      "issue_numbers": [],
+      "pr_numbers": [
+        44002,
+        44029
+      ],
+      "evidence_types": [
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 44002,
+          "right_pr_number": 44029,
+          "code_similarity": 0.135,
+          "size_similarity": 0.43,
+          "file_overlap": 0.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.329,
+          "shared_filenames": [],
+          "shared_file_areas": []
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-44018-2",
+      "summary": "Cluster of 2 related pull requests linked by soft_similarity.",
+      "status": "open",
+      "confidence": 0.5,
+      "canonical_issue_number": null,
+      "canonical_pr_number": 44068,
+      "issue_numbers": [],
+      "pr_numbers": [
+        44018,
+        44068
+      ],
+      "evidence_types": [
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 44018,
+          "right_pr_number": 44068,
+          "code_similarity": 0.766,
+          "size_similarity": 0.939,
+          "file_overlap": 1.0,
+          "area_overlap": 0.425,
+          "patch_similarity": 0.866,
+          "shared_filenames": [
+            "src/transformers/models/gpt_neo/modeling_gpt_neo.py",
+            "tests/models/gpt_neo/test_modeling_gpt_neo.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/gpt_neo/modeling_gpt_neo.py",
+              "left_ranges": [
+                [
+                  26,
+                  46
+                ],
+                [
+                  138,
+                  143
+                ],
+                [
+                  182,
+                  187
+                ],
+                [
+                  283,
+                  295
+                ],
+                [
+                  327,
+                  341
+                ],
+                [
+                  348,
+                  354
+                ],
+                [
+                  360,
+                  369
+                ],
+                [
+                  399,
+                  406
+                ],
+                [
+                  411,
+                  419
+                ],
+                [
+                  428,
+                  433
+                ],
+                [
+                  472,
+                  492
+                ],
+                [
+                  507,
+                  513
+                ],
+                [
+                  519,
+                  528
+                ],
+                [
+                  541,
+                  559
+                ],
+                [
+                  562,
+                  567
+                ],
+                [
+                  595,
+                  601
+                ],
+                [
+                  607,
+                  614
+                ],
+                [
+                  627,
+                  643
+                ],
+                [
+                  685,
+                  690
+                ],
+                [
+                  708,
+                  714
+                ],
+                [
+                  720,
+                  727
+                ],
+                [
+                  740,
+                  757
+                ],
+                [
+                  761,
+                  766
+                ],
+                [
+                  780,
+                  786
+                ],
+                [
+                  791,
+                  798
+                ],
+                [
+                  807,
+                  822
+                ],
+                [
+                  840,
+                  845
+                ]
+              ],
+              "right_ranges": [
+                [
+                  26,
+                  31
+                ],
+                [
+                  34,
+                  44
+                ],
+                [
+                  136,
+                  141
+                ],
+                [
+                  180,
+                  185
+                ],
+                [
+                  281,
+                  293
+                ],
+                [
+                  325,
+                  339
+                ],
+                [
+                  346,
+                  352
+                ],
+                [
+                  358,
+                  367
+                ],
+                [
+                  397,
+                  404
+                ],
+                [
+                  409,
+                  417
+                ],
+                [
+                  426,
+                  434
+                ],
+                [
+                  463,
+                  483
+                ],
+                [
+                  498,
+                  504
+                ],
+                [
+                  510,
+                  519
+                ],
+                [
+                  532,
+                  550
+                ],
+                [
+                  553,
+                  564
+                ],
+                [
+                  586,
+                  592
+                ],
+                [
+                  598,
+                  605
+                ],
+                [
+                  618,
+                  634
+                ],
+                [
+                  676,
+                  688
+                ],
+                [
+                  699,
+                  705
+                ],
+                [
+                  711,
+                  718
+                ],
+                [
+                  731,
+                  748
+                ],
+                [
+                  752,
+                  762
+                ],
+                [
+                  771,
+                  777
+                ],
+                [
+                  782,
+                  789
+                ],
+                [
+                  798,
+                  813
+                ],
+                [
+                  831,
+                  836
+                ]
+              ]
+            },
+            {
+              "filename": "tests/models/gpt_neo/test_modeling_gpt_neo.py",
+              "left_ranges": [
+                [
+                  458,
+                  464
+                ]
+              ],
+              "right_ranges": [
+                [
+                  458,
+                  464
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    }
+  ],
+  "duplicate_issues": [
+    {
+      "cluster_id": "cluster-46894-5",
+      "canonical_issue_number": 46894,
+      "duplicate_issue_numbers": [
+        46897
+      ],
+      "reason": "Issues in cluster-46894-5 are treated as duplicates because they share closing_reference, shared_issue_target, soft_similarity evidence."
+    }
+  ],
+  "duplicate_prs": [
+    {
+      "cluster_id": "cluster-41211-3",
+      "canonical_pr_number": 44339,
+      "duplicate_pr_numbers": [
+        41356
+      ],
+      "target_issue_number": 41211,
+      "reason": "PRs in cluster-41211-3 are treated as duplicates because they converge on issue #41211 with closing_reference evidence."
+    },
+    {
+      "cluster_id": "cluster-43240-3",
+      "canonical_pr_number": 43251,
+      "duplicate_pr_numbers": [
+        43254
+      ],
+      "target_issue_number": 43240,
+      "reason": "PRs in cluster-43240-3 are treated as duplicates because they converge on issue #43240 with closing_reference, shared_issue_target evidence."
+    },
+    {
+      "cluster_id": "cluster-43698-3",
+      "canonical_pr_number": 43779,
+      "duplicate_pr_numbers": [
+        43816
+      ],
+      "target_issue_number": 43698,
+      "reason": "PRs in cluster-43698-3 are treated as duplicates because they converge on issue #43698 with closing_reference, shared_issue_target evidence."
+    },
+    {
+      "cluster_id": "cluster-43979-11",
+      "canonical_pr_number": 44007,
+      "duplicate_pr_numbers": [
+        43996,
+        44013,
+        44044,
+        44066,
+        44072,
+        44085,
+        44129,
+        44154,
+        44722
+      ],
+      "target_issue_number": 43979,
+      "reason": "PRs in cluster-43979-11 are treated as duplicates because they converge on issue #43979 with closing_reference, shared_issue_target, soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-44001-2",
+      "canonical_pr_number": 44004,
+      "duplicate_pr_numbers": [
+        44001
+      ],
+      "target_issue_number": null,
+      "reason": "PRs in cluster-44001-2 are treated as duplicates because they share soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-44002-2",
+      "canonical_pr_number": 44002,
+      "duplicate_pr_numbers": [
+        44029
+      ],
+      "target_issue_number": null,
+      "reason": "PRs in cluster-44002-2 are treated as duplicates because they share soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-44018-2",
+      "canonical_pr_number": 44068,
+      "duplicate_pr_numbers": [
+        44018
+      ],
+      "target_issue_number": null,
+      "reason": "PRs in cluster-44018-2 are treated as duplicates because they share soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-45081-3",
+      "canonical_pr_number": 45317,
+      "duplicate_pr_numbers": [
+        45086
+      ],
+      "target_issue_number": 45081,
+      "reason": "PRs in cluster-45081-3 are treated as duplicates because they converge on issue #45081 with closing_reference, shared_issue_target, soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-45561-3",
+      "canonical_pr_number": 45645,
+      "duplicate_pr_numbers": [
+        45639
+      ],
+      "target_issue_number": 45561,
+      "reason": "PRs in cluster-45561-3 are treated as duplicates because they converge on issue #45561 with closing_reference, shared_issue_target, soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-45647-3",
+      "canonical_pr_number": 45738,
+      "duplicate_pr_numbers": [
+        45737
+      ],
+      "target_issue_number": 45647,
+      "reason": "PRs in cluster-45647-3 are treated as duplicates because they converge on issue #45647 with closing_reference, shared_issue_target evidence."
+    },
+    {
+      "cluster_id": "cluster-46159-3",
+      "canonical_pr_number": 46174,
+      "duplicate_pr_numbers": [
+        46780
+      ],
+      "target_issue_number": 46159,
+      "reason": "PRs in cluster-46159-3 are treated as duplicates because they converge on issue #46159 with closing_reference evidence."
+    },
+    {
+      "cluster_id": "cluster-46196-3",
+      "canonical_pr_number": 46196,
+      "duplicate_pr_numbers": [
+        46836,
+        47143
+      ],
+      "target_issue_number": null,
+      "reason": "PRs in cluster-46196-3 are treated as duplicates because they share soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-46693-3",
+      "canonical_pr_number": 47134,
+      "duplicate_pr_numbers": [
+        46696
+      ],
+      "target_issue_number": 46693,
+      "reason": "PRs in cluster-46693-3 are treated as duplicates because they converge on issue #46693 with closing_reference evidence."
+    },
+    {
+      "cluster_id": "cluster-46869-3",
+      "canonical_pr_number": 46886,
+      "duplicate_pr_numbers": [
+        46922
+      ],
+      "target_issue_number": 46869,
+      "reason": "PRs in cluster-46869-3 are treated as duplicates because they converge on issue #46869 with closing_reference, shared_issue_target evidence."
+    },
+    {
+      "cluster_id": "cluster-46894-5",
+      "canonical_pr_number": 46895,
+      "duplicate_pr_numbers": [
+        46958,
+        47090
+      ],
+      "target_issue_number": 46894,
+      "reason": "PRs in cluster-46894-5 are treated as duplicates because they converge on issue #46894 with closing_reference, shared_issue_target, soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-47013-3",
+      "canonical_pr_number": 47046,
+      "duplicate_pr_numbers": [
+        47026
+      ],
+      "target_issue_number": 47013,
+      "reason": "PRs in cluster-47013-3 are treated as duplicates because they converge on issue #47013 with closing_reference, shared_issue_target, soft_similarity evidence."
+    }
+  ],
+  "best_issue": {
+    "cluster_id": "cluster-46894-5",
+    "issue_number": 46894,
+    "reason": "Issue #46894 is the strongest global issue candidate because it is open, belongs to a cluster with 5 artifacts, and carries 1 discussion comments plus 9 inbound references.",
+    "score": 161.0
+  },
+  "best_pr": {
+    "cluster_id": "cluster-43979-11",
+    "pr_number": 44007,
+    "reason": "PR #44007 is the strongest global PR candidate because it is open, belongs to a cluster with 11 artifacts, links to 1 issues, and carries 2 review/discussion events.",
+    "score": 238.886
+  }
+}
diff --git a/snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.llm-reviews.json b/snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.llm-reviews.json
new file mode 100644
index 0000000000000000000000000000000000000000..d2e308944cc1a88fef32f692de48f0dddece4c79
--- /dev/null
+++ b/snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.llm-reviews.json
@@ -0,0 +1,26750 @@
+{
+  "schema_version": "1.0",
+  "repo": "huggingface/transformers",
+  "snapshot_id": "20260708T000615Z",
+  "generated_at": "2026-07-08T00:12:27Z",
+  "model": "gpt-5.4-mini?service_tier=flex",
+  "reviews": [
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:24643",
+        "issue:30277",
+        "issue:34634",
+        "issue:34928",
+        "issue:35545",
+        "issue:36010",
+        "issue:36331",
+        "issue:39290",
+        "issue:41720",
+        "issue:41762",
+        "issue:42915",
+        "issue:43039",
+        "issue:43425",
+        "issue:43541",
+        "issue:43596",
+        "issue:43638",
+        "issue:43716",
+        "issue:43827",
+        "issue:43828",
+        "issue:43975",
+        "issue:44292",
+        "issue:44322",
+        "issue:44457",
+        "issue:44509",
+        "issue:44512",
+        "issue:44560",
+        "issue:44805",
+        "issue:44841",
+        "issue:44918",
+        "issue:45081",
+        "issue:45161",
+        "issue:45237",
+        "issue:45464",
+        "issue:45507",
+        "issue:45685",
+        "issue:45800",
+        "issue:45823",
+        "issue:45859",
+        "issue:45874",
+        "issue:45901",
+        "issue:45902",
+        "issue:46015",
+        "issue:46165",
+        "issue:46294",
+        "issue:46468",
+        "issue:46515",
+        "issue:46661",
+        "issue:46869"
+      ],
+      "soft_pairs": [
+        "issue:44322|issue:45464",
+        "issue:45237|issue:45685",
+        "issue:43716|issue:45237",
+        "issue:43828|issue:45237",
+        "issue:45237|issue:46468",
+        "issue:45237|issue:45507",
+        "issue:30277|issue:35545",
+        "issue:35545|issue:42915",
+        "issue:35545|issue:41720",
+        "issue:45901|issue:46015",
+        "issue:35545|issue:36010",
+        "issue:44509|issue:44512",
+        "issue:44918|issue:46165",
+        "issue:46165|issue:46869",
+        "issue:44805|issue:45823",
+        "issue:43638|issue:44805",
+        "issue:43638|issue:46165",
+        "issue:43596|issue:44805",
+        "issue:44805|issue:46294",
+        "issue:44805|issue:45161",
+        "issue:46165|issue:46468",
+        "issue:44560|issue:46165",
+        "issue:34634|issue:35545",
+        "issue:30277|issue:43638",
+        "issue:43596|issue:43638",
+        "issue:43039|issue:45161",
+        "issue:43975|issue:44457",
+        "issue:44292|issue:44457",
+        "issue:44841|issue:46661",
+        "issue:43716|issue:46869",
+        "issue:24643|issue:30277",
+        "issue:34928|issue:44805",
+        "issue:45685|issue:46869",
+        "issue:45859|issue:45902",
+        "issue:34928|issue:36331",
+        "issue:43638|issue:45161",
+        "issue:43596|issue:45161",
+        "issue:43716|issue:45507",
+        "issue:44292|issue:45237",
+        "issue:45874|issue:46294",
+        "issue:39290|issue:45081",
+        "issue:45800|issue:46515",
+        "issue:45823|issue:45874",
+        "issue:34928|issue:41720",
+        "issue:43425|issue:44292",
+        "issue:43828|issue:45161",
+        "issue:43541|issue:43716",
+        "issue:43828|issue:44292",
+        "issue:43716|issue:45685",
+        "issue:45507|issue:45685",
+        "issue:44292|issue:44918",
+        "issue:43716|issue:44292",
+        "issue:44918|issue:45161",
+        "issue:43039|issue:44918",
+        "issue:41762|issue:43638",
+        "issue:43827|issue:44512",
+        "issue:43828|issue:45685",
+        "issue:43039|issue:43828",
+        "issue:30277|issue:36331",
+        "issue:34928|issue:36010",
+        "issue:45237|issue:46015",
+        "issue:43039|issue:44805"
+      ],
+      "prepared_review_unit_hash": "fd41fd4600e534e2737f40f28019b2d87da97e530a1e11b17469eece061125c3",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 62,
+        "serialized_chars": 46282,
+        "estimated_input_tokens": 11571,
+        "estimated_eval_tokens": 23398
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most pairs are not true duplicates; they span unrelated model/runtime issues. The only clear issue duplicates are the two DeepSpeed ZeRO-3 BERT index-out-of-bounds reports, and the torch 2.4.1 compatibility reports are likely the same underlying version mismatch. The docs v5 cleanup items are similar but not close enough to merge as duplicates.",
+        "confidence": 0.79,
+        "canonical_issue_reason": "issue:43596 is the clearest representative of the ZeRO-3/BERT index-out-of-bounds bug and has a strong duplicate in issue:43638.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43596 is the best issue representative in this set because it is a concrete bug report with a clear matching duplicate and a broader title than the variant report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44322",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Different failures: missing config attr in Qwen3_5 vs streaming chat/completions error in Qwen3.5-0.8B."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Different subsystems and failures: AMD GPU inference vs MPS histogram op on Int."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Different models and root causes: image preprocessor dtype mismatch vs AMD GPU runtime failure."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Both are dtype/device issues, but on different models and code paths; not the same bug."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "Different model/example failures with no clear shared code path."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "AMD GPU failure vs invalid method call in GraniteMoEHybrid; unrelated."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:35545",
+            "accept": false,
+            "reason": "DeepSpeed collective mismatch is unrelated to ONNX export error."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "ONNX export problem vs Qwen3Moe FP8 failure; different change surface."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "ONNX export error and CUDA assert in auto device mapping are unrelated."
+          },
+          {
+            "left": "issue:45901",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "Table QA crash and outdated RAG examples are different issues."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Export failure vs GenerationMixin import breakage; different code paths."
+          },
+          {
+            "left": "issue:44509",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are v5 docs cleanup items, but they reference different removed features and likely different edits."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:46165",
+            "accept": false,
+            "reason": "TRL SFT embedding unpacking vs resume-training/remote-code issue are unrelated."
+          },
+          {
+            "left": "issue:46165",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "Resume training with remote code is unrelated to YOLO multi-GPU device-map failure."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Mask shape IndexError and Gemma4 device mismatch are different forward-pass failures."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Both are tensor/shape errors, but the ZeRO-3 Bert index bug is not the same as the mask-shape mismatch."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:46165",
+            "accept": false,
+            "reason": "DeepSpeed/BERT index error vs resume-training remote-code issue; no overlap."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Both are IndexErrors, but the reported shapes and code paths differ; not the same underlying bug."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "Mask/tensor shape mismatch is unrelated to Qwen3.5 video token/feature mismatch."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Mask IndexError vs GPT-OSS MoE TP issue are different problems."
+          },
+          {
+            "left": "issue:46165",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "Remote-code resume problem is unrelated to the Nemotron example failure."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:46165",
+            "accept": false,
+            "reason": "Video StopIteration in Qwen3-vl-embedding is unrelated to training resume/remote code."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:35545",
+            "accept": false,
+            "reason": "BarkProcessor voice_preset and ModernBERT ONNX export are unrelated."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Different DeepSpeed symptoms: collective mismatch vs Bert zero-size index error."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:43638",
+            "accept": true,
+            "reason": "Same ZeRO-3/BERT failure: index 0 out of bounds from size-0 tensors during model init/load."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Liger kernel cross-entropy path is unrelated to GPT-OSS tensor-parallel failure."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:44457",
+            "accept": false,
+            "reason": "Incorrect detokenization and LoRA merge/reload mismatch are different bugs."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44457",
+            "accept": false,
+            "reason": "NVFP4 model runtime error is unrelated to LoRA save/load inconsistency."
+          },
+          {
+            "left": "issue:44841",
+            "right": "issue:46661",
+            "accept": false,
+            "reason": "Processor failure for Voxtral is unrelated to remote-code config_class handling."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "Different models and failures: Mistral-3 dtype mismatch vs YOLO multi-GPU device map."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:30277",
+            "accept": false,
+            "reason": "Both involve DeepSpeed, but one is a weight dimensionality error and the other is a collective mismatch."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "FSDP activation-checkpointing recompute mismatch is unrelated to the mask-shape IndexError."
+          },
+          {
+            "left": "issue:45685",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "MPS histogram op failure and YOLO device-map failure are unrelated."
+          },
+          {
+            "left": "issue:45859",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Same model family, but different bugs: CPU-offload incompatibility vs uninitialized memory after materialization."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "FSDP tensor-size recompute issue is unrelated to CustomTrainer.compute_loss signature breakage."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "ZeRO-3 Bert index error and GPT-OSS tensor-parallel issue do not share a concrete code path."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "BERT ZeRO-3 index error is unrelated to GPT-OSS MoE tensor-parallel failure."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Mistral-3 image dtype mismatch and GraniteMoEHybrid invalid method call are different bugs."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "NVFP4 runtime error is unrelated to AMD GPU failure on GPT-OSS-20B."
+          },
+          {
+            "left": "issue:45874",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "inputs_embeds reverse-embedding cost and video token mismatch are unrelated."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Gemma3 vLLM attribute error and Mistral regex patch crash are unrelated."
+          },
+          {
+            "left": "issue:45800",
+            "right": "issue:46515",
+            "accept": true,
+            "reason": "Both report the same torch 2.4.1 incompatibility; 46515 is a more generic restatement of the same version breakage."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:45874",
+            "accept": false,
+            "reason": "Device-map forward mismatch and expensive reverse embedding lookup are different Gemma4 issues."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "FSDP activation-checkpointing mismatch is unrelated to Qwen3 auto device-map CUDA assert."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Generic torch 2.10 incompatibility and Qwen-3-8B-NVFP4 runtime error are not the same bug."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Phi-tiny-MoE autocast dtype mismatch is unrelated to GPT-OSS TP failure."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:43716",
+            "accept": false,
+            "reason": "Mixtral torch dynamo tracing error and Mistral-3 preprocessor dtype mismatch are different code paths."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Different models and failures: autocast dtype mismatch vs NVFP4 runtime error."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Mistral-3 image dtype mismatch is unrelated to MPS histogram Int failure."
+          },
+          {
+            "left": "issue:45507",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Invalid method call in GraniteMoEHybrid and MPS histogram failure are unrelated."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Qwen NVFP4 runtime error and TRL SFT embedding unpacking are different issues."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Different models and failure modes; no shared bug."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "TRL SFT embedding unpacking and GPT-OSS TP failure are unrelated."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Liger kernel cross-entropy path and Qwen3.5 TRL embedding unpacking are unrelated."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Gemma3 ZeRO-3 load failure and Bert ZeRO-3 load failure are similar in shape, but not the same concrete bug."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are v5 docs regressions, but one is pipeline-task references and the other is the transformers run command; not one duplicate change."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Autocast dtype mismatch and MPS Int histogram op failure are different issues."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Liger kernel cross-entropy call and Phi-tiny-MoE autocast dtype mismatch are unrelated."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "DeepSpeed collective mismatch and custom trainer signature breakage are different failures."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "FSDP activation-checkpointing mismatch is unrelated to GenerationMixin import failure."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "AMD GPU inference failure and outdated RAG examples are unrelated."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Liger kernel cross-entropy issue is unrelated to mask shape IndexError."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded overall, and the soft-edge calls stay conservative. The two accepted pairs are well-supported by the titles, and the summary does not overclaim beyond the packet."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 2,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:36010",
+        "issue:39290",
+        "issue:41093",
+        "issue:41720",
+        "issue:41762",
+        "issue:43039",
+        "issue:43296",
+        "issue:43404",
+        "issue:43425",
+        "issue:43531",
+        "issue:43541",
+        "issue:43572",
+        "issue:43596",
+        "issue:43638",
+        "issue:43792",
+        "issue:43828",
+        "issue:43854",
+        "issue:43866",
+        "issue:43872",
+        "issue:43901",
+        "issue:43975",
+        "issue:44292",
+        "issue:44512",
+        "issue:44560",
+        "issue:44589",
+        "issue:44841",
+        "issue:44863",
+        "issue:45020",
+        "issue:45070",
+        "issue:45084",
+        "issue:45237",
+        "issue:45362",
+        "issue:45399",
+        "issue:45464",
+        "issue:45507",
+        "issue:45685",
+        "issue:45800",
+        "issue:45812",
+        "issue:45823",
+        "issue:45901",
+        "issue:45920",
+        "issue:46015",
+        "issue:46294",
+        "issue:46304",
+        "issue:46402",
+        "issue:46661",
+        "issue:46772",
+        "issue:46869"
+      ],
+      "soft_pairs": [
+        "issue:45800|issue:46304",
+        "issue:43296|issue:46304",
+        "issue:43866|issue:46402",
+        "issue:41762|issue:43596",
+        "issue:39290|issue:41720",
+        "issue:43901|issue:44512",
+        "issue:44589|issue:46304",
+        "issue:45237|issue:45901",
+        "issue:43039|issue:44560",
+        "issue:45823|issue:46294",
+        "issue:43975|issue:46015",
+        "issue:44841|issue:45823",
+        "issue:43866|issue:46772",
+        "issue:44863|issue:46015",
+        "issue:45020|issue:46304",
+        "issue:45812|issue:45920",
+        "issue:43872|issue:46304",
+        "issue:44292|issue:46015",
+        "issue:43866|issue:44863",
+        "issue:39290|issue:43531",
+        "issue:41093|issue:43638",
+        "issue:43792|issue:45901",
+        "issue:43854|issue:43866",
+        "issue:43425|issue:45070",
+        "issue:43975|issue:45901",
+        "issue:43792|issue:45070",
+        "issue:45362|issue:45464",
+        "issue:43596|issue:45800",
+        "issue:43541|issue:46869",
+        "issue:44292|issue:45901",
+        "issue:39290|issue:45823",
+        "issue:45464|issue:45685",
+        "issue:36010|issue:39290",
+        "issue:43541|issue:45685",
+        "issue:43541|issue:43828",
+        "issue:44841|issue:45084",
+        "issue:45507|issue:46869",
+        "issue:43404|issue:43866",
+        "issue:41093|issue:41762",
+        "issue:45399|issue:46661",
+        "issue:43572|issue:45070",
+        "issue:45800|issue:45823"
+      ],
+      "prepared_review_unit_hash": "709757637a15ba306d521fc03ccf75ff5b631899f4bad6b832461072c22edaed",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 42,
+        "serialized_chars": 40046,
+        "estimated_input_tokens": 10012,
+        "estimated_eval_tokens": 20280
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a heterogeneous mix of Transformers bug reports; only a couple of pairs look like true duplicates. The clearest representative is the broad AutoTokenizer v4\u2192v5 regression report.",
+        "confidence": 0.62,
+        "canonical_issue_reason": "issue:45920 is the broadest and most central report: it describes the same wrong-token-ID regression across multiple model families, and it subsumes the narrower Granite-only report.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45920 best represents the main recurring bug class in this set because it is broad, concrete, and directly overlaps with at least one narrower duplicate candidate.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45800",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "Both are version-compatibility complaints, but they concern different failures and different breakages; not the same underlying bug."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "PaddleOCR-VL loading in vLLM and USE_DISTRIBUTED=0 support are unrelated."
+          },
+          {
+            "left": "issue:43866",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Corrupted checkpoint vs missing loaded weights are different model-loading failures."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:43596",
+            "accept": true,
+            "reason": "Both report the same DeepSpeed ZeRO-3 init pattern: index 0 out of bounds from an empty/size-0 tensor during model loading."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Different models and different failures; one is a missing config attribute, the other is a CUDA assert with auto device mapping."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are docs issues, but they cover different documentation problems and are not the same change."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "Storage-type lookup failure is unrelated to the USE_DISTRIBUTED=0 compatibility regression."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "Different task/model failures with no concrete shared bug path."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Liger Kernel cross-entropy behavior and Qwen3-vl video processing are unrelated."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "Device-map forward mismatch and chat-template video token mismatch are different code paths."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "Tokenizer detokenization bug vs outdated RAG docs are unrelated."
+          },
+          {
+            "left": "issue:44841",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Different model families and different failure modes."
+          },
+          {
+            "left": "issue:43866",
+            "right": "issue:46772",
+            "accept": false,
+            "reason": "Checkpoint corruption and NVFP4 compatibility are not the same underlying issue."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "NemotronH checkpoint loading and RAG docs are unrelated."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "Both are broad regressions, but they affect different subsystems and different failures."
+          },
+          {
+            "left": "issue:45812",
+            "right": "issue:45920",
+            "accept": true,
+            "reason": "Same bug class: silent v4\u2192v5 AutoTokenizer wrong token IDs; 45920 is the broader superset."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "bitsandbytes constructor incompatibility is unrelated to the torch/distributed support regression."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "NVFP4 execution failure and RAG docs are unrelated."
+          },
+          {
+            "left": "issue:43866",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Different model-loading problems with different models and symptoms."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Both mention sliding_window, but they are different model-specific bugs and not clearly the same code-path issue."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Mask shape mismatch and ZeRO-3 index-out-of-bounds are different errors."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "Whisper run failure and table-question-answering crash are unrelated."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "Different models and different load failures."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Torch version incompatibility and PretrainedConfig/pydantic regression are distinct issues."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "Tokenizer detokenization bug and table QA crash are unrelated."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Whisper model failure and config-schema regression are unrelated."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Both involve Qwen3.5 chat paths, but the concrete failures differ and are not clearly the same bug."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Both are versioned regressions, but one is a Zero-3 loading bug and the other is a torch/transformers compatibility issue."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "Grouped_mm tracing failure and YOLO multi-GPU device-map failure are unrelated."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "NVFP4 model execution and table QA are unrelated."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Missing sliding_window_pattern and Gemma4 device mismatch are different bugs."
+          },
+          {
+            "left": "issue:45464",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Streaming chat API failure and MPS histogram kernel error are unrelated."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:39290",
+            "accept": false,
+            "reason": "ImportError for GenerationMixin and Gemma3 config attribute error are unrelated."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Different MoE-related failures on different backends."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Both concern model execution, but the tracing/grouped_mm error and autocast dtype mismatch are not the same bug."
+          },
+          {
+            "left": "issue:44841",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Processor failure and template-node compilation error are unrelated."
+          },
+          {
+            "left": "issue:45507",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "Invalid method call in GraniteMoEHybrid and YOLO device-map failure are unrelated."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "Untied lm_head weights and corrupted checkpoint are different issues."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:41762",
+            "accept": false,
+            "reason": "Different index-related failures with different shapes and different loading paths."
+          },
+          {
+            "left": "issue:45399",
+            "right": "issue:46661",
+            "accept": false,
+            "reason": "flash-attn fallback logic and remote-code config_class handling are unrelated."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "A missing StableLmConfig field and a pydantic PretrainedConfig regression are different bugs."
+          },
+          {
+            "left": "issue:45800",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Torch/transformers compatibility and Gemma4 device mismatch are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet, and the soft-edge judgments are conservative overall. The two accepted duplicates are well-supported by the issue titles and scopes, and the rejects avoid overclaiming equivalence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 3,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:39186",
+        "issue:40990",
+        "issue:41093",
+        "issue:42915",
+        "issue:43054",
+        "issue:43296",
+        "issue:43366",
+        "issue:43404",
+        "issue:43531",
+        "issue:43541",
+        "issue:43596",
+        "issue:43645",
+        "issue:43716",
+        "issue:43828",
+        "issue:43873",
+        "issue:43994",
+        "issue:44291",
+        "issue:44360",
+        "issue:44387",
+        "issue:44479",
+        "issue:44492",
+        "issue:44512",
+        "issue:44530",
+        "issue:44805",
+        "issue:44863",
+        "issue:44918",
+        "issue:44960",
+        "issue:45005",
+        "issue:45020",
+        "issue:45070",
+        "issue:45161",
+        "issue:45183",
+        "issue:45399",
+        "issue:45464",
+        "issue:45507",
+        "issue:45632",
+        "issue:45698",
+        "issue:45759",
+        "issue:45800",
+        "issue:45902",
+        "issue:46165",
+        "issue:46294",
+        "issue:46463",
+        "issue:46468",
+        "issue:46640",
+        "issue:46661",
+        "issue:46869"
+      ],
+      "soft_pairs": [
+        "issue:43296|issue:43366",
+        "issue:44805|issue:44918",
+        "issue:43541|issue:45161",
+        "issue:43828|issue:45464",
+        "issue:43366|issue:45464",
+        "issue:44387|issue:45005",
+        "issue:44291|issue:45070",
+        "issue:43531|issue:45070",
+        "issue:41093|issue:43596",
+        "issue:43716|issue:46294",
+        "issue:45020|issue:45800",
+        "issue:44479|issue:46294",
+        "issue:43054|issue:43994",
+        "issue:46468|issue:46869",
+        "issue:45759|issue:45902",
+        "issue:43596|issue:46165",
+        "issue:46640|issue:46661",
+        "issue:43716|issue:44918",
+        "issue:39186|issue:44918",
+        "issue:43645|issue:45399",
+        "issue:40990|issue:44960",
+        "issue:45632|issue:45698",
+        "issue:42915|issue:45399",
+        "issue:43404|issue:45005",
+        "issue:44360|issue:44512",
+        "issue:43873|issue:45005",
+        "issue:44492|issue:44512",
+        "issue:44530|issue:45005",
+        "issue:44918|issue:45507",
+        "issue:45183|issue:46463",
+        "issue:44863|issue:45005"
+      ],
+      "prepared_review_unit_hash": "13b83c2385d110f37da2bc8d25c660b30d20474829d53ee7bf77b46672f4784a",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36240,
+        "estimated_input_tokens": 9060,
+        "estimated_eval_tokens": 18376
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most items are unrelated bug reports. I only see two likely true duplicate pairs: the Qwen video-input regression reports and the remote-code/custom-module loading/cache-collision reports.",
+        "confidence": 0.84,
+        "canonical_issue_reason": "No single canonical issue cleanly spans this mostly unrelated set; the best root-cause anchor is the remote-code cache-path collision report because it gives the clearest concrete failure mode and matches the wrong-module-loading symptom.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45632 is the strongest standalone issue here: it is concrete, reproducible, and names a specific root cause that plausibly explains the related wrong custom-module loading report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43296",
+            "right": "issue:43366",
+            "accept": false,
+            "reason": "Different models and failure modes: PaddleOCR-VL load failure vs GPT-OSS GGUF support request."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Both mention indexing/shape problems, but the concrete paths differ: generic mask mismatch vs Qwen3.5 input-embedding unpacking in TRL."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Different MoE-related problems: torch Dynamo/grouped_mm tracing error vs GPT-OSS tensor-parallel support."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Different code paths and symptoms: autocast dtype mismatch vs streaming chat/completions failure."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Unrelated: GPT-OSS GGUF support vs Qwen3.5 streaming inference."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Different issues: int4 quantization memory regression vs tied-weights behavior in translation models."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Different root causes: init_empty_weights argument handling vs pydantic PretrainedConfig field breakage."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Different topics: Qwen3-MoE sliding window behavior vs pydantic config field regression."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:43596",
+            "accept": false,
+            "reason": "Both are IndexErrors, but the specific causes differ: mask-length mismatch vs ZeRO-3/BertModel zero-sized tensor indexing."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "Both are multimodal dtype/feature mismatch reports, but one is image preprocessing for Mistral-3 and the other is video chat-template handling for Qwen3.5."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Different failures: remote_code breakage in newer transformers vs torch/transformers version incompatibility."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:46294",
+            "accept": true,
+            "reason": "These describe the same Qwen video-input regression: video features/tokens mismatch across Qwen3.5/Qwen3 VL variants."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Same model family, but the reports are not clearly the same bug: degraded text embeddings vs nonsensical AutoModel/pipeline outputs."
+          },
+          {
+            "left": "issue:46468",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "Unrelated: broken Nemotron example vs YOLO multi-GPU device-map auto failure."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Different Qwen3.5 issues: config unwrapping in from_config vs uninitialized rotary embedding state after to_empty()."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:46165",
+            "accept": false,
+            "reason": "Different remote-code/training problems: zero-sized tensor index error during ZeRO-3 init vs resume-training with local trust_remote_code models."
+          },
+          {
+            "left": "issue:46640",
+            "right": "issue:46661",
+            "accept": false,
+            "reason": "Different problems: Flash Attention 2 issue vs remote-code model_class/config_class AttributeError."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Different modalities and paths: image preprocessor dtype mismatch vs Qwen3.5 embedding unpacking in TRL."
+          },
+          {
+            "left": "issue:39186",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Different failures: FSDP 2-D weight requirement vs Qwen3.5 embedding unpacking error."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Unrelated: Jupyter custom-model initialization regression vs FA2 fallback blocked by other checks."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "Different models and symptoms: GPT-OSS perplexity issue vs GLM5 issue."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:45698",
+            "accept": true,
+            "reason": "Both point to the same remote-code/custom-module loading bug, with a cache/path collision causing the wrong custom module to be loaded."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Different areas: Qwen3-MoE FineGrainedFP8 failure vs FlashAttention2 fallback gating."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Both involve tied weights, but the concrete bugs differ: missing lm_head tying in Mistral3 multimodal generation vs translation-model v5 tied-weight regressions."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Unrelated: DSA indexer behavior vs documentation typo."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Different issues: offloading with quantization vs tied-weight behavior in translation models."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are low-level content fixes, but one is a cache-strategy typo and the other is stale docs about a removed command."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Different concrete bugs: Qwen3.5 PagedAttentionCache crash vs tied-weight regressions."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Unrelated: Qwen3.5 embedding unpacking vs GraniteMoEHybrid invalid-method call."
+          },
+          {
+            "left": "issue:45183",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Both are validation/error-message issues, but they are different validation paths: input_ids/inputs_embeds XOR vs T5 decoder input validation."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Unrelated: NemotronH checkpoint loading vs tied-weight regressions on translation models."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is appropriately cautious and stays close to the packet. The two accepted soft pairs are plausibly duplicate reports, and the remaining pairs are conservatively rejected without overclaiming. No mergeability issue applies here."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 4,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:24643",
+        "issue:28530",
+        "issue:36331",
+        "issue:38175",
+        "issue:39692",
+        "issue:41553",
+        "issue:41762",
+        "issue:42915",
+        "issue:42947",
+        "issue:43054",
+        "issue:43493",
+        "issue:43643",
+        "issue:43644",
+        "issue:43782",
+        "issue:43824",
+        "issue:43828",
+        "issue:43856",
+        "issue:43872",
+        "issue:43883",
+        "issue:43950",
+        "issue:43975",
+        "issue:44292",
+        "issue:44387",
+        "issue:44403",
+        "issue:44451",
+        "issue:44488",
+        "issue:44534",
+        "issue:44589",
+        "issue:44910",
+        "issue:44928",
+        "issue:44960",
+        "issue:44991",
+        "issue:45020",
+        "issue:45399",
+        "issue:45520",
+        "issue:45663",
+        "issue:45685",
+        "issue:45701",
+        "issue:45823",
+        "issue:45859",
+        "issue:45902",
+        "issue:45920",
+        "issue:45987",
+        "issue:46489",
+        "issue:46491",
+        "issue:46858",
+        "issue:46899",
+        "issue:46932"
+      ],
+      "soft_pairs": [
+        "issue:43975|issue:44292",
+        "issue:46489|issue:46491",
+        "issue:39692|issue:43054",
+        "issue:43950|issue:44534",
+        "issue:36331|issue:41762",
+        "issue:43644|issue:43950",
+        "issue:43872|issue:44589",
+        "issue:46899|issue:46932",
+        "issue:43828|issue:44928",
+        "issue:43054|issue:45701",
+        "issue:45920|issue:46489",
+        "issue:38175|issue:43054",
+        "issue:45987|issue:46858",
+        "issue:43950|issue:45902",
+        "issue:43054|issue:43493",
+        "issue:44910|issue:45859",
+        "issue:42947|issue:44387",
+        "issue:45663|issue:45902",
+        "issue:44403|issue:44451",
+        "issue:41553|issue:43883",
+        "issue:24643|issue:28530",
+        "issue:43975|issue:45520",
+        "issue:44403|issue:44488",
+        "issue:43643|issue:45020",
+        "issue:42947|issue:43856",
+        "issue:41762|issue:42915",
+        "issue:44991|issue:45399",
+        "issue:43824|issue:43872",
+        "issue:45685|issue:45823",
+        "issue:43782|issue:44960"
+      ],
+      "prepared_review_unit_hash": "f446e0e1631e47b6a44db0efc8648234539b4d5f3f4ec52b1f8c24cfe9c607eb",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36904,
+        "estimated_input_tokens": 9226,
+        "estimated_eval_tokens": 18708
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly unrelated issues overall, but there is a clear duplicate cluster around the v5 non-persistent-buffer corruption regression: issue:43644 is the earliest report, and issue:43950 and issue:44534 are closely matching follow-ups of the same bug.",
+        "confidence": 0.73,
+        "canonical_issue_reason": "issue:43644 is the earliest report in the non-persistent-buffer junk/corruption regression cluster, making it the best canonical anchor for that bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43950 is the strongest representative issue overall for this cluster because it states the regression most explicitly and concretely, even though it is not the earliest report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43975",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Tokenizer detokenization regression and Qwen3 NVFP4 load/runtime error are different failures in different paths."
+          },
+          {
+            "left": "issue:46489",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "Both are tokenizer bugs, but they target different models and concrete failure modes; not the same defect."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43054",
+            "accept": false,
+            "reason": "Doc/example errors plus quantization failure are not the same bug as a SigLIP2 text-embedding quality regression."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:44534",
+            "accept": true,
+            "reason": "Same underlying v5 non-persistent-buffer corruption regression, described with slightly different wording."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41762",
+            "accept": false,
+            "reason": "Trainer API signature mismatch and DeepSpeed ZeRO-3 load-time index error are unrelated code paths."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:43950",
+            "accept": false,
+            "reason": "Remote-code AutoConfig field loss is a different bug from non-persistent-buffer corruption."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Both are quantization/storage related, but they are different incompatibility bugs."
+          },
+          {
+            "left": "issue:46899",
+            "right": "issue:46932",
+            "accept": true,
+            "reason": "Same Gemma4 quantized vision/audio embedder dtype-cast bug; 46932 is an explicit fix for the same issue family."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:44928",
+            "accept": false,
+            "reason": "Different models and symptoms: Phi MoE autocast dtype mismatch vs Qwen3.5 position_ids/SDPA fallback and NaN issue."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "SigLIP2 embedding-quality regression is not the same as a broad tokenization-change report."
+          },
+          {
+            "left": "issue:45920",
+            "right": "issue:46489",
+            "accept": false,
+            "reason": "Broad multi-model tokenizer regression and a DeepSeek-Coder v1-specific round-trip bug are not the same concrete defect."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43054",
+            "accept": false,
+            "reason": "Same model family, but zero probabilities and worse text embeddings are different reported failures."
+          },
+          {
+            "left": "issue:45987",
+            "right": "issue:46858",
+            "accept": true,
+            "reason": "Same cache.get_seq_length() return-type regression; one report emphasizes the contract mismatch, the other downstream slice arithmetic breakage."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Both mention buffers, but one is a load corruption regression and the other is Qwen3_5Moe rotary embedding initialization after to_empty()."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Related SigLIP2 concerns, but one is a performance gap and the other is a HF-vs-JAX discrepancy; not clearly the same bug."
+          },
+          {
+            "left": "issue:44910",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "FlashAttention crash from 3D position_ids is unrelated to CPU-offload incompatibility in rotary embedding."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "LoRA/gradient-checkpointing issue and int4 quantization OOM are different training problems."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Different model paths and failures: FSDP2 shared_kv_states rebuild bug vs uninitialized rotary embedding memory."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Generic loading noise is not the same as a specific model load failure."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "AutoTokenizer error for Voxtral and missing all_tied_weights_keys on Molmo are unrelated."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:28530",
+            "accept": false,
+            "reason": "DeepSpeed weight-dimension runtime error and early-stopping metric handling are unrelated training issues."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Tokenizer detokenization bug and flash_attn import KeyError are unrelated."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Generic loading noise is not the same as the specific sleng-bert loading failure."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Both involve remote_code, but one is a missing-fields config return bug while the other is a broader compatibility regression."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "Different problems: ineffective gradient checkpointing vs excessive memory usage in Qwen3 MoE training."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO-3 load-time index error and Qwen3Moe FineGrainedFP8Config failure are different code paths."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Tokenizer load failure for a specific model and flash-attn2 fallback blocking are unrelated."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43872",
+            "accept": false,
+            "reason": "Missing Qwen2_5_VL class export and bitsandbytes Int8Params kwarg mismatch are different compatibility issues."
+          },
+          {
+            "left": "issue:45685",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "MPS MoE histogram limitation and Gemma4 PLE device mismatch are different backend/model bugs."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "Qwen3VL weight_only load error and a generic GLM5 issue are not the same underlying bug."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43950",
+            "accept": true,
+            "reason": "Same non-persistent-buffer regression; 43644 is the earlier report and 43950 is a later, more explicit framing of the same bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The clustering is mostly conservative, but the canonical/best-issue claim is not fully grounded: issue:43644 is an earlier and nearly identical report of the same non-persistent-buffer regression, so issue:43950 cannot be called the earliest statement of that bug. The summary should also acknowledge that 43644 belongs in the same cluster. The rest of the soft-edge verdicts look mostly cautious, but this factual overstatement is enough to reject."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 5,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:24643",
+        "issue:30064",
+        "issue:36331",
+        "issue:41093",
+        "issue:41553",
+        "issue:42491",
+        "issue:42947",
+        "issue:43257",
+        "issue:43404",
+        "issue:43792",
+        "issue:43827",
+        "issue:43828",
+        "issue:43854",
+        "issue:43872",
+        "issue:43881",
+        "issue:43883",
+        "issue:43940",
+        "issue:43950",
+        "issue:44220",
+        "issue:44291",
+        "issue:44368",
+        "issue:44479",
+        "issue:44509",
+        "issue:44560",
+        "issue:44661",
+        "issue:44805",
+        "issue:44863",
+        "issue:44928",
+        "issue:44991",
+        "issue:45020",
+        "issue:45081",
+        "issue:45305",
+        "issue:45313",
+        "issue:45356",
+        "issue:45399",
+        "issue:45488",
+        "issue:45588",
+        "issue:45715",
+        "issue:45759",
+        "issue:45823",
+        "issue:45923",
+        "issue:46315",
+        "issue:46489",
+        "issue:46661",
+        "issue:46869",
+        "issue:46894",
+        "issue:46897"
+      ],
+      "soft_pairs": [
+        "issue:45081|issue:45399",
+        "issue:41553|issue:43881",
+        "issue:45020|issue:45823",
+        "issue:44661|issue:45020",
+        "issue:43828|issue:45823",
+        "issue:45020|issue:45356",
+        "issue:42491|issue:44368",
+        "issue:43881|issue:43883",
+        "issue:45715|issue:46315",
+        "issue:43872|issue:46869",
+        "issue:45759|issue:46661",
+        "issue:43872|issue:44291",
+        "issue:43827|issue:44509",
+        "issue:45488|issue:46489",
+        "issue:45020|issue:46661",
+        "issue:43404|issue:43950",
+        "issue:46894|issue:46897",
+        "issue:43792|issue:44220",
+        "issue:43257|issue:45305",
+        "issue:42947|issue:45923",
+        "issue:36331|issue:41093",
+        "issue:43940|issue:45313",
+        "issue:43854|issue:44863",
+        "issue:44928|issue:45588",
+        "issue:24643|issue:30064",
+        "issue:44479|issue:44560",
+        "issue:41093|issue:44805",
+        "issue:41553|issue:44991"
+      ],
+      "prepared_review_unit_hash": "10eacfcf5e54d0f330b7e6d70387c71df3b04d34fa2150b862541969a269f5c3",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 28,
+        "serialized_chars": 36182,
+        "estimated_input_tokens": 9046,
+        "estimated_eval_tokens": 18348
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a false-positive similarity cluster spanning unrelated tokenizer, model-loading, DeepSpeed, and training-loss bugs. The only clearly plausible duplicate-like pairs are the mask-shape mismatch reports (41093/44805) and the training-loss double-shift reports (46894/46897), both of which look like near-duplicates with the same error pattern rather than proven exact matches.",
+        "confidence": 0.8,
+        "canonical_issue_reason": "No single canonical issue fits the whole cluster because it spans multiple unrelated bugs. If a representative anchor is needed, issue 41093 is the best mask-shape representative, with 44805 as a near-duplicate showing the same error pattern.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 41093 is the strongest issue anchor because it is the clearest representative of the shared mask-shape mismatch problem, and it has a close near-duplicate in 44805.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45081",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Different bug classes: Mistral regex patching vs flash-attn fallback gating."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "Both are loading failures, but for different models and different causes."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Remote-code regression and device-map mismatch are unrelated code paths."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "One is tokenizer-mapping utility behavior; the other is a broad remote-code loading regression."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Autocast dtype mismatch and device-map forwarding are different problems."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Tokenizer codec/regression issue is not the same as remote-code breakage."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Both mention Qwen/tie-word-embeddings, but one is a version compatibility break and the other is just a config warning."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "GLM loading failure and missing all_tied_weights_keys are different model bugs."
+          },
+          {
+            "left": "issue:45715",
+            "right": "issue:46315",
+            "accept": false,
+            "reason": "Both are tokenizer performance regressions, but in different conversion APIs and likely different causes."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "Bitsandbytes parameter-construction incompatibility is unrelated to YOLO multi-GPU device-map failure."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:46661",
+            "accept": false,
+            "reason": "Both involve auto-model loading, but one is composite config unwrapping and the other is remote-code config_class handling."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:44291",
+            "accept": true,
+            "reason": "Same unexpected `_is_hf_initialized` kwarg during weight/model initialization; likely one regression across loading paths."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "Both are docs/pipeline references, but they are separate documentation cleanup issues."
+          },
+          {
+            "left": "issue:45488",
+            "right": "issue:46489",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but for different DeepSeek tokenization problems and code paths."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:46661",
+            "accept": false,
+            "reason": "Remote-code breakage and remote-code model_class/config_class errors are related in theme but not the same bug."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43950",
+            "accept": false,
+            "reason": "lm_head tying and non-persistent buffer corruption are distinct state-loading bugs."
+          },
+          {
+            "left": "issue:46894",
+            "right": "issue:46897",
+            "accept": true,
+            "reason": "Same double-shift training-loss bug pattern across two model reports; plausible duplicate-like reports of the same underlying issue."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Whisper model failure and `_torch_extract_fbank_features()` are too broad to treat as the same underlying bug from the titles alone."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45305",
+            "accept": false,
+            "reason": "DeepSpeed weight-loading failure and gradient averaging under GAS are different DeepSpeed bugs."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Gradient checkpointing ineffectiveness in PEFT LoRA is not the same as a missing supports_gradient_checkpointing flag."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41093",
+            "accept": false,
+            "reason": "Unexpected compute_loss kwarg and mask/tensor shape mismatch are unrelated failures."
+          },
+          {
+            "left": "issue:43940",
+            "right": "issue:45313",
+            "accept": false,
+            "reason": "Both are Qwen3 DeepSpeed load failures, but the model variants and failure modes are not clearly the same concrete bug."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "GLM-4.7-Flash unit-test loading and NemotronH checkpoint loading are separate model integration issues."
+          },
+          {
+            "left": "issue:44928",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "RLHF gradient explosion and flash_attention s_aux handling are unrelated."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "DeepSpeed weight shape error and void segmentation-map processing are completely different."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Both involve video, but the regression and StopIteration failures do not look like the same bug."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:44805",
+            "accept": true,
+            "reason": "Same mask/tensor shape mismatch error pattern; near-duplicate reports of the same underlying bug, though not enough evidence to call it an exact duplicate."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Tokenizer loading errors for Voxtral and EMBEDDIA/est-roberta are different model-specific failures."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The cluster-level summary is broadly plausible, but it overstates the 41093/44805 relationship. From the packet we can only see highly similar mask/tensor shape mismatch titles, not an exact duplicate match, so phrases like \"exact matching duplicate\" / \"exact-duplicate subpair\" are too strong without body evidence. Please downgrade that to conservative language such as \"near-duplicate with the same error pattern.\""
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 6,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:36010",
+        "issue:38175",
+        "issue:41553",
+        "issue:42491",
+        "issue:42617",
+        "issue:42915",
+        "issue:43366",
+        "issue:43531",
+        "issue:43550",
+        "issue:43596",
+        "issue:43638",
+        "issue:43645",
+        "issue:43646",
+        "issue:43716",
+        "issue:43720",
+        "issue:43824",
+        "issue:43828",
+        "issue:43950",
+        "issue:43957",
+        "issue:43994",
+        "issue:44155",
+        "issue:44488",
+        "issue:44530",
+        "issue:44560",
+        "issue:44589",
+        "issue:44661",
+        "issue:44683",
+        "issue:44805",
+        "issue:44863",
+        "issue:44877",
+        "issue:44910",
+        "issue:44918",
+        "issue:45030",
+        "issue:45325",
+        "issue:45584",
+        "issue:45663",
+        "issue:45672",
+        "issue:45685",
+        "issue:45812",
+        "issue:45823",
+        "issue:45834",
+        "issue:45902",
+        "issue:46402",
+        "issue:46468",
+        "issue:46489",
+        "issue:46515",
+        "issue:46858",
+        "issue:46899"
+      ],
+      "soft_pairs": [
+        "issue:43716|issue:46468",
+        "issue:44530|issue:44863",
+        "issue:44863|issue:46402",
+        "issue:43366|issue:43828",
+        "issue:43366|issue:43531",
+        "issue:43596|issue:46515",
+        "issue:45325|issue:45902",
+        "issue:41553|issue:44488",
+        "issue:43366|issue:44589",
+        "issue:43638|issue:46515",
+        "issue:44155|issue:45584",
+        "issue:45663|issue:45823",
+        "issue:42915|issue:43957",
+        "issue:43824|issue:43957",
+        "issue:43716|issue:44560",
+        "issue:43950|issue:45834",
+        "issue:42617|issue:43366",
+        "issue:43645|issue:43646",
+        "issue:44560|issue:45685",
+        "issue:44910|issue:46858",
+        "issue:36010|issue:42915",
+        "issue:43550|issue:44910",
+        "issue:44560|issue:44918",
+        "issue:44683|issue:46515",
+        "issue:45812|issue:46489",
+        "issue:45823|issue:46515",
+        "issue:44661|issue:45823",
+        "issue:44560|issue:44805",
+        "issue:38175|issue:43994",
+        "issue:44877|issue:45030",
+        "issue:45672|issue:46899",
+        "issue:43638|issue:44661",
+        "issue:43596|issue:44661",
+        "issue:42491|issue:43720"
+      ],
+      "prepared_review_unit_hash": "355c1b19e3d7a495cd724985027f9bc67ab9eb1d875802ecf67ba97ba3a4ecdb",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 38043,
+        "estimated_input_tokens": 9511,
+        "estimated_eval_tokens": 19278
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly heterogeneous cluster with one clear duplicate pair around Transformers 5.0.0 custom-model initialization, plus one likely duplicate SigLIP2 pair. The rest of the title-only soft edges look like related subsystems or similar error classes, but not the same concrete bug.",
+        "confidence": 0.67,
+        "canonical_issue_reason": "43646 is the broadest representative of the custom-model-initialization regression; 43645 looks like the notebook-specific manifestation of the same issue.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43646 is the best issue anchor because it is the most general statement of the custom model initialization regression and cleanly covers the 43645 variant.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43716",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "Different model families and failure modes: multimodal dtype mismatch vs Nemotron example loading."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Both are model-specific failures, but the concrete bugs and targets differ."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Both are loading-related, but for different Nemotron/Sam3 paths and symptoms."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Unrelated bugs: GGUF support for gpt-oss vs Phi-tiny-MoE autocast dtype mismatch."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Same broad Qwen/MoE area, but not the same concrete regression."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "A specific ZeRO3+BERT index error is not the same as a generic torch-version incompatibility."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Both involve position or rotary state, but the model families and failure paths differ."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both are loading failures, but they appear to affect different models for different reasons."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "No concrete overlap beyond general runtime failure; these look unrelated."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "The BERT ZeRO3 index error and the torch 2.4.1 incompatibility report are not the same bug."
+          },
+          {
+            "left": "issue:44155",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Different subsystems and symptoms: AudioFlamingo batched inference leak vs Whisper empty-transcription failure."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Same model family, but training/FSDP2 shared_kv_states is different from device_map forward mismatch."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 failure and meta-device loading breakage are separate problems."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "ImportError for one class is not the same as broader meta-device loading failure."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Different models and failure mechanisms; they do not look like the same bug."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Non-persistent buffer corruption and long-input index error are unrelated regressions."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:43366",
+            "accept": false,
+            "reason": "A script startup failure and GGUF model support are not the same underlying issue."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:43646",
+            "accept": true,
+            "reason": "Likely the same Transformers 5.0.0 custom-model-initialization regression; 43645 just adds the Jupyter-notebook context."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Different runtime errors in different code paths and model setups."
+          },
+          {
+            "left": "issue:44910",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Both touch cache/attention internals, but the concrete bugs differ."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "ImportError for GenerationMixin and Qwen3Moe FP8 failure are unrelated."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44910",
+            "accept": false,
+            "reason": "torch.compile+SDPA on Bamba and Flash Attention illegal access on Qwen3.5 are different code-path problems."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Different models and different failure shapes; not the same bug."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "torch>=2.9 compiled flex_attention failure is not the same as a torch 2.4.1 incompatibility report."
+          },
+          {
+            "left": "issue:45812",
+            "right": "issue:46489",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but for different model families and likely different fixes."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "Gemma4 device-map mismatch is a model-forward bug, not a generic torch-version incompatibility."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Tokenizer-mapping handling and Gemma4 device placement are unrelated issues."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Duplicate of the earlier verdict: same generic runtime-error pattern, but not the same concrete bug."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43994",
+            "accept": true,
+            "reason": "Likely the same SigLIP2 regression on the same model; the titles differ mainly in symptom wording."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:45030",
+            "accept": false,
+            "reason": "Different config-validation/loading problems affecting different models."
+          },
+          {
+            "left": "issue:45672",
+            "right": "issue:46899",
+            "accept": false,
+            "reason": "Both are Gemma4 4-bit issues, but the concrete dtype-cast failures appear to be different code paths."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "ZeRO3+BERT index error and tokenizer-mapping-name failure are unrelated."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "A BERT ZeRO3 index error and a BitNet packed-weight loading issue are different bugs."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "Mostly conservative, but a few claims go beyond the packet. In particular, \"detailed repro\" for 43645 is not supported by the metadata provided here, and calling the ZeRO3+BERT index-error items a \"clear duplicate pocket\" is stronger than the evidence in the packet. The 38175/43994 accept also reads a bit too definitive; from titles alone, \"likely same issue\" would be safer than \"same underlying regression.\""
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 7,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:38175",
+        "issue:39692",
+        "issue:41720",
+        "issue:42491",
+        "issue:42915",
+        "issue:43257",
+        "issue:43317",
+        "issue:43632",
+        "issue:43644",
+        "issue:43645",
+        "issue:43708",
+        "issue:43854",
+        "issue:43856",
+        "issue:43873",
+        "issue:43931",
+        "issue:43994",
+        "issue:44075",
+        "issue:44164",
+        "issue:44291",
+        "issue:44387",
+        "issue:44451",
+        "issue:44488",
+        "issue:44534",
+        "issue:44661",
+        "issue:44756",
+        "issue:44805",
+        "issue:44863",
+        "issue:44960",
+        "issue:44977",
+        "issue:44991",
+        "issue:45276",
+        "issue:45292",
+        "issue:45313",
+        "issue:45335",
+        "issue:45357",
+        "issue:45362",
+        "issue:45406",
+        "issue:45488",
+        "issue:45507",
+        "issue:45812",
+        "issue:45834",
+        "issue:46050",
+        "issue:46459",
+        "issue:46468",
+        "issue:46512",
+        "issue:46846",
+        "issue:46897",
+        "issue:46901"
+      ],
+      "soft_pairs": [
+        "issue:43632|issue:45834",
+        "issue:44661|issue:44805",
+        "issue:43317|issue:44756",
+        "issue:38175|issue:39692",
+        "issue:43644|issue:44534",
+        "issue:41720|issue:42915",
+        "issue:42915|issue:44756",
+        "issue:42491|issue:45276",
+        "issue:44451|issue:44488",
+        "issue:44960|issue:44977",
+        "issue:44863|issue:46846",
+        "issue:46459|issue:46512",
+        "issue:45313|issue:45357",
+        "issue:45507|issue:46468",
+        "issue:43644|issue:43645",
+        "issue:39692|issue:43994",
+        "issue:44488|issue:44991",
+        "issue:43257|issue:43854",
+        "issue:43856|issue:43873",
+        "issue:42491|issue:45292",
+        "issue:43856|issue:44387",
+        "issue:42491|issue:45335",
+        "issue:43854|issue:43931",
+        "issue:43708|issue:44075",
+        "issue:44960|issue:46050",
+        "issue:44164|issue:44291",
+        "issue:46897|issue:46901",
+        "issue:45276|issue:45335",
+        "issue:45488|issue:45812",
+        "issue:45362|issue:45406"
+      ],
+      "prepared_review_unit_hash": "7d320a20d8630bd2e14cb3bf022a6d41866e95c7b13a479770e2b83a6b5a0876",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36517,
+        "estimated_input_tokens": 9130,
+        "estimated_eval_tokens": 18516
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "There are a few clear near-duplicates, but most pairs are only loosely related by subsystem or symptom. The strongest duplicate is the v5 non-persistent buffer corruption report; most other soft pairs should stay separate.",
+        "confidence": 0.74,
+        "canonical_issue_reason": "issue:43644 is the cleanest canonical report: it describes the generic v5 non-persistent-buffer corruption regression, while issue:44534 is just a later rephrasing of the same bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43644 is the best representative issue overall because it is broad, concrete, and regression-focused, making it the most reusable target for a fix.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43632",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Different bugs: one is an `_is_hf_initialized` regression, the other is an OCR long-input indexing error."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Unrelated failure modes: tokenizer-mapping registration vs a mask/tensor shape mismatch."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Both involve loading/memory, but one is quantized offload load failure and the other is Strix Halo mmap OOM; not the same bug."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:39692",
+            "accept": false,
+            "reason": "Both mention SigLIP2, but one is zero probabilities at inference and the other is a doc example/model-processor mismatch plus quantization failure."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44534",
+            "accept": true,
+            "reason": "Same regression: v5 fills non-persistent buffers with junk; the later issue is just a paraphrase of the same bug."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Different Qwen3 failure paths: auto device mapping CUDA assert vs FineGrainedFP8Config load failure."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Qwen3 MoE FP8 loading is unrelated to Strix Halo mmap OOM mitigation."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "One is a LoRA compatibility break between hf4.x and hf5.x; the other is a resize_token_embeddings bug in Gemma4."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both are model-loading complaints, but they concern different checkpoints and do not describe the same concrete bug."
+          },
+          {
+            "left": "issue:44960",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Unrelated models and symptoms: GLM5 is not the same as a Qwen3.5 flash-attention generation bug."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:46846",
+            "accept": false,
+            "reason": "NemotronH checkpoint loading and Qwen3.5 TP-plan omissions are different code paths and failures."
+          },
+          {
+            "left": "issue:46459",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "Both concern dtype handling, but one is ignored dtype on composite checkpoint load and the other is stale config.dtype after from_config."
+          },
+          {
+            "left": "issue:45313",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Both are Qwen3.5/Gemma-related regressions, but one is ZeRO-3 weight loading and the other is wrong visual-encoder keys on save_pretrained."
+          },
+          {
+            "left": "issue:45507",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "Invalid method call in GraniteMoEHybrid is not the same as the broken official Nemotron example."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Related to Transformers v5 initialization, but one is buffer corruption and the other is custom-model initialization in notebooks."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Both involve SigLIP2 docs/loading, but one is a documentation example error and the other is nonsensical inference results."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are tokenizer/model-loading failures, but they affect different models and do not point to the same underlying defect."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Accelerate+deepspeed MoE conversion failure and GLM-4.7 flash test loading are unrelated."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Both mention quantization or memory, but one is training memory inefficiency and the other is offloading not working as expected."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:45292",
+            "accept": false,
+            "reason": "Different issues: LoRA compatibility across versions vs resize_token_embeddings not updating output embeddings."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Both are memory-related, but one is Qwen3 MoE training memory and the other is int4 quantization increasing CUDA reserved memory."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "The hf4.x\u2192hf5.x LoRA compatibility problem is not the same as t5gemma embedding-resize propagation."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different model-load errors: GLM-4.7 Flash unit test loading vs Qwen3-VL weight-shape mismatch."
+          },
+          {
+            "left": "issue:43708",
+            "right": "issue:44075",
+            "accept": false,
+            "reason": "Trainer checkpoint-resume step counting and SGD argument handling are unrelated bugs."
+          },
+          {
+            "left": "issue:44960",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "GLM5 is unrelated to a save_pretrained/from_pretrained breakage in quantized sinq models."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Both involve save/load initialization behavior, but they are distinct bugs: extra_state handling vs unexpected `_is_hf_initialized` args."
+          },
+          {
+            "left": "issue:46897",
+            "right": "issue:46901",
+            "accept": false,
+            "reason": "Both describe training-loss double-shift symptoms, but they target different model implementations and there is no evidence of one shared code path."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "Closely related symptom, but the affected models and embedding targets differ; not enough evidence they are the same concrete bug."
+          },
+          {
+            "left": "issue:45488",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are tokenizer regressions in v5, but one is LlamaTokenizer pre-tokenizer behavior and the other is Granite AutoTokenizer ID generation."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Different crashes: Qwen3.5 chat generation vs Gemma4Processor missing `_tokenizer` in transformers serve."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet, and the one accepted soft pair (43644/44534) is strongly supported by the nearly identical regression descriptions. The remaining soft-edge verdicts are conservative and do not overclaim duplication."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 8,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:36010",
+        "issue:39404",
+        "issue:39692",
+        "issue:41720",
+        "issue:42175",
+        "issue:42796",
+        "issue:42947",
+        "issue:43335",
+        "issue:43381",
+        "issue:43404",
+        "issue:43425",
+        "issue:43454",
+        "issue:43645",
+        "issue:43854",
+        "issue:43873",
+        "issue:43957",
+        "issue:44079",
+        "issue:44155",
+        "issue:44164",
+        "issue:44291",
+        "issue:44292",
+        "issue:44387",
+        "issue:44402",
+        "issue:44479",
+        "issue:44488",
+        "issue:44589",
+        "issue:44756",
+        "issue:44863",
+        "issue:44877",
+        "issue:44910",
+        "issue:44912",
+        "issue:44928",
+        "issue:44960",
+        "issue:45216",
+        "issue:45276",
+        "issue:45292",
+        "issue:45335",
+        "issue:45588",
+        "issue:45663",
+        "issue:45684",
+        "issue:45698",
+        "issue:45704",
+        "issue:45759",
+        "issue:45823",
+        "issue:45859",
+        "issue:46326",
+        "issue:46468",
+        "issue:46829"
+      ],
+      "soft_pairs": [
+        "issue:44910|issue:45588",
+        "issue:43425|issue:43957",
+        "issue:44402|issue:44488",
+        "issue:39692|issue:44960",
+        "issue:41720|issue:44155",
+        "issue:44291|issue:44756",
+        "issue:45663|issue:45759",
+        "issue:42947|issue:45704",
+        "issue:43404|issue:43454",
+        "issue:43873|issue:44387",
+        "issue:44079|issue:46829",
+        "issue:45663|issue:45859",
+        "issue:44292|issue:44912",
+        "issue:42947|issue:43381",
+        "issue:36010|issue:39404",
+        "issue:43957|issue:44756",
+        "issue:44155|issue:45823",
+        "issue:44387|issue:44756",
+        "issue:43335|issue:45663",
+        "issue:44863|issue:46468",
+        "issue:45276|issue:45292",
+        "issue:44479|issue:44877",
+        "issue:43381|issue:44928",
+        "issue:44164|issue:44756",
+        "issue:42796|issue:46326",
+        "issue:45588|issue:45663",
+        "issue:45684|issue:45698",
+        "issue:43854|issue:45216",
+        "issue:45292|issue:45335",
+        "issue:43645|issue:44164",
+        "issue:42175|issue:43957",
+        "issue:44589|issue:44756"
+      ],
+      "prepared_review_unit_hash": "ee181d71d2e2f123ba5ce3e69020fa945b6ccf3673a214deafcbc729774f1456",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37385,
+        "estimated_input_tokens": 9347,
+        "estimated_eval_tokens": 18950
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The set is mostly heterogeneous; only the resize_token_embeddings/output_embeddings reports form a true duplicate family. I would keep 45292 as the representative issue for that small cluster and reject the rest as separate bugs or feature regressions.",
+        "confidence": 0.89,
+        "canonical_issue_reason": "Issue 45292 is the best canonical representative of the only clear duplicate family here: the generic resize_token_embeddings/output_embeddings regression, which the Gemma4 and t5gemma reports specialize.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "45292 is the broadest, model-agnostic version of the resize_token_embeddings bug, so it is the best global issue target among these items.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44910",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Different flash-attention failures: 3D position_ids misparse vs s_aux=None crash in sink-less models."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Version-compatibility claim vs meta-device loading bug; different root causes and code paths."
+          },
+          {
+            "left": "issue:44402",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Different model-specific tokenizer/load issues; no shared underlying bug is evident."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "Unrelated models and symptoms: SigLIP2 docs example errors vs GLM5 issue."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44155",
+            "accept": false,
+            "reason": "Distinct bugs in different models: CUDA assert on auto device mapping vs batched inference token leakage."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Unexpected init_empty_weights argument vs mmap/OOM behavior; separate loading problems."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Different model families and paths: Gemma4 FSDP2 attention state vs Qwen multimodal config unwrapping."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Training checkpointing/LoRA behavior is unrelated to T5 apex.FusedRMSNorm memory leak."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43454",
+            "accept": false,
+            "reason": "Both mention untied lm_head weights, but they are model-specific reports without evidence of the same shared code-path bug."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Both are quantization/memory complaints, but one is offloading behavior and the other is reserved-memory OOM under int4."
+          },
+          {
+            "left": "issue:44079",
+            "right": "issue:46829",
+            "accept": false,
+            "reason": "Generic ModelOutput key handling is not the same bug as VideoPrism's hidden_states misassignment."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Unrelated model-specific bugs: Gemma4 FSDP2 state rebuild vs Qwen3_5Moe CPU offload incompatibility."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Different models and quantization formats; NVFP4 load failure is not the same as MXFP4 bf16 fallback."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "Both mention gradient checkpointing, but one is ineffective LoRA behavior and the other is eval-mode prohibition."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:39404",
+            "accept": false,
+            "reason": "ImportError for GenerationMixin is unrelated to Whisper return_language pipeline regression."
+          },
+          {
+            "left": "issue:43957",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Meta-device loading bug and Strix Halo mmap/OOM issue are separate problems."
+          },
+          {
+            "left": "issue:44155",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Different model paths: AudioFlamingo3 batch leakage vs Gemma4 device-map mismatch."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Int4 quantization memory regression is not the same as disabling mmap to avoid Strix Halo OOM."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "Single-layer SwitchTransformers config bug is unrelated to Gemma4 FSDP2 shared_kv_states rebuild."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "NemotronH checkpoint loading and the official example being broken may be related by model family, but they are not clearly the same concrete bug."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45292",
+            "accept": true,
+            "reason": "Same resize_token_embeddings/output_embeddings regression; 45276 is a model-specific variant of the broader bug in 45292."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Video-input regression for Qwen models is unrelated to Granite speech config strictness."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:44928",
+            "accept": false,
+            "reason": "Gradient checkpointing eval-mode issue is not the same as the RLHF NaN/SDPA fallback regression."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "save/from_pretrained extra_state handling is unrelated to Strix Halo mmap/OOM behavior."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Both involve mistral_common tokenizers, but one is BPE marker handling and the other is continue_final_message logic."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "flash_attention s_aux=None crash is a different code path from Gemma4 FSDP2 shared_kv_states rebuilding."
+          },
+          {
+            "left": "issue:45684",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Custom-model save permission propagation and wrong-module loading are different save/load bugs."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "GLM-4.7-Flash load/test failure and Qwen3.5 save_pretrained regression are unrelated models and failure modes."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45335",
+            "accept": true,
+            "reason": "Both are the same resize_token_embeddings family; 45335 is a t5gemma-specific manifestation of the generic output_embeddings bug."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:44164",
+            "accept": false,
+            "reason": "Custom model initialization in Jupyter is unrelated to extra_state save/load failure."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Packaging/backend dependency issue is unrelated to meta-device model loading."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Missing Float8 storage type and Strix Halo mmap/OOM are different runtime issues."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: 45292/45276/45335 clearly form the only obvious duplicate family from their titles, and the remaining pairwise rejections stay conservative given the distinct model names and failure modes. No overstatement stands out."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 9,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:36032",
+        "issue:39404",
+        "issue:43425",
+        "issue:43452",
+        "issue:43611",
+        "issue:43645",
+        "issue:43646",
+        "issue:43708",
+        "issue:43716",
+        "issue:43828",
+        "issue:43856",
+        "issue:43901",
+        "issue:43906",
+        "issue:43994",
+        "issue:44079",
+        "issue:44164",
+        "issue:44265",
+        "issue:44402",
+        "issue:44451",
+        "issue:44479",
+        "issue:44509",
+        "issue:44530",
+        "issue:44589",
+        "issue:44671",
+        "issue:44683",
+        "issue:44756",
+        "issue:44805",
+        "issue:44898",
+        "issue:44912",
+        "issue:44991",
+        "issue:45081",
+        "issue:45084",
+        "issue:45479",
+        "issue:45701",
+        "issue:45759",
+        "issue:45812",
+        "issue:45823",
+        "issue:46050",
+        "issue:46315",
+        "issue:46326",
+        "issue:46491",
+        "issue:46665",
+        "issue:46821",
+        "issue:46869",
+        "issue:46894",
+        "issue:46899",
+        "issue:46901"
+      ],
+      "soft_pairs": [
+        "issue:44402|issue:44451",
+        "issue:43646|issue:43708",
+        "issue:46894|issue:46901",
+        "issue:43425|issue:43828",
+        "issue:46315|issue:46491",
+        "issue:45812|issue:46315",
+        "issue:39404|issue:45701",
+        "issue:44530|issue:45084",
+        "issue:43611|issue:43645",
+        "issue:43425|issue:46869",
+        "issue:44683|issue:45823",
+        "issue:39404|issue:46665",
+        "issue:43425|issue:43716",
+        "issue:43425|issue:45823",
+        "issue:44589|issue:45084",
+        "issue:44479|issue:44805",
+        "issue:43901|issue:44509",
+        "issue:43994|issue:44079",
+        "issue:43906|issue:44530",
+        "issue:44164|issue:45759",
+        "issue:36032|issue:43452",
+        "issue:44265|issue:44898",
+        "issue:46050|issue:46326",
+        "issue:43906|issue:46326",
+        "issue:39404|issue:44912",
+        "issue:44991|issue:45081",
+        "issue:44671|issue:44991",
+        "issue:45823|issue:46899",
+        "issue:45479|issue:46894",
+        "issue:43856|issue:44756",
+        "issue:44164|issue:46821"
+      ],
+      "prepared_review_unit_hash": "fa5641339c872aae55ef90794da625db6c0c94732ca3b7411fd250d8485448a9",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36640,
+        "estimated_input_tokens": 9160,
+        "estimated_eval_tokens": 18576
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly unrelated issue set; the soft pairs share broad themes (tokenizers, quantization, multimodal, docs) but not the same concrete defect, so none are duplicate-worthy.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "Issue 45812 is the strongest canonical anchor: it is a broad, clearly scoped v4\u2192v5 regression affecting all Granite tokenization and has a precise symptom, making it the most useful representative issue in this cluster.",
+        "canonical_pr_reason": "No pull requests are present in the provided items.",
+        "best_issue_reason": "Issue 45812 has the clearest high-impact regression framing and the broadest affected scope among the provided issues.",
+        "best_pr_reason": "No pull requests are present in the provided items.",
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44402",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Both are tokenizer-load failures, but one is a vocab-size inconsistency and the other is a specific model load failure; not the same bug."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:43708",
+            "accept": false,
+            "reason": "Custom model initialization in notebooks and Trainer resume max_steps are different code paths and failure modes."
+          },
+          {
+            "left": "issue:46894",
+            "right": "issue:46901",
+            "accept": false,
+            "reason": "Both mention a training-loss double-shift pattern, but they concern different model implementations and are not shown to share one concrete fix."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Both involve torch version/dtype issues, but one is generic Torch 2.10 incompatibility and the other is an autocast mismatch for a specific model."
+          },
+          {
+            "left": "issue:46315",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "One is a tokenizer performance regression, the other is an encode/decode whitespace regression; different defects."
+          },
+          {
+            "left": "issue:45812",
+            "right": "issue:46315",
+            "accept": false,
+            "reason": "Granite token IDs being wrong is a correctness bug; convert_tokens_to_ids slowdown is a performance regression."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Whisper pipeline return_language and general tokenization changes are related only at a very broad level, not the same issue."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "PagedAttentionCache linear_attention crash and compile-time template-node error are unrelated failure classes."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Base_model_prefix loading breakage and notebook custom-model initialization are distinct custom-model problems."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "Torch incompatibility and YOLO multi-GPU device-map failure are different bugs in different paths."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Compiled flex_attention on torch>=2.9 and Gemma4 device-map mismatch are unrelated issues."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:46665",
+            "accept": false,
+            "reason": "Whisper pipeline behavior and from_pretrained inside Python classes are different regressions."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43716",
+            "accept": false,
+            "reason": "Generic torch compatibility and Mistral-3 dtype mismatch are not the same underlying defect."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "A Torch version incompatibility is not the same as a Gemma4 forward device mismatch."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Float8 storage lookup failure and template-node compilation error are unrelated."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Video-input regression for Qwen variants and a mask-shape IndexError are different bugs."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "Both are documentation-related, but they cover different deprecated pipeline/task mentions and should not be merged as duplicates."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Bad SigLIP2 outputs and ModelOutput key assignment issues are different code-path problems."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "An isolated reproduction of an older issue and a Qwen3.5 cache crash are not the same defect."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Extra_state save/load handling and multimodal config unwrapping are unrelated serialization/config bugs."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "T5Tokenizer special-token method conflict and gguf_file load breakage are different tokenizer/model-loading failures."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "torch.export/exportability and Perceiver interpolation at non-default resolution are unrelated model-export vs vision-input bugs."
+          },
+          {
+            "left": "issue:46050",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Quantized save/load corruption and mistral_common continue_final_message behavior are separate issues."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "These describe different reproduction/behavioral problems and do not point to one shared code defect."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Whisper return_language and MXFP4 quantization fallback are unrelated regressions."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both affect tokenizers, but one is a tokenizer load regression and the other is a Mistral regex patch crash; different causes."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "CamemBERT masked-LM prediction errors and est-roberta tokenizer loading failures are not the same bug."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46899",
+            "accept": false,
+            "reason": "Both are Gemma4-related, but one is a device-map forward mismatch and the other is unsafe dtype casting in multimodal embedders under 4-bit quantization."
+          },
+          {
+            "left": "issue:45479",
+            "right": "issue:46894",
+            "accept": false,
+            "reason": "Both concern loss computation, but they are different model families and the bugs are described differently enough that they should not be merged."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Qwen3 MoE memory usage and disabling mmap on Strix Halo address different OOM sources."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "save/from_pretrained extra_state handling and a TimesFM window_size AttributeError are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 10,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:32090",
+        "issue:34567",
+        "issue:36032",
+        "issue:39404",
+        "issue:40444",
+        "issue:40990",
+        "issue:42175",
+        "issue:42796",
+        "issue:42915",
+        "issue:43299",
+        "issue:43404",
+        "issue:43421",
+        "issue:43452",
+        "issue:43482",
+        "issue:43541",
+        "issue:43582",
+        "issue:43596",
+        "issue:43611",
+        "issue:43632",
+        "issue:43643",
+        "issue:43644",
+        "issue:43645",
+        "issue:43646",
+        "issue:43716",
+        "issue:43931",
+        "issue:44206",
+        "issue:44220",
+        "issue:44451",
+        "issue:44462",
+        "issue:44488",
+        "issue:44530",
+        "issue:44661",
+        "issue:44683",
+        "issue:44737",
+        "issue:44811",
+        "issue:44987",
+        "issue:45081",
+        "issue:45084",
+        "issue:45479",
+        "issue:45584",
+        "issue:45859",
+        "issue:46178",
+        "issue:46285",
+        "issue:46326",
+        "issue:46402",
+        "issue:46489",
+        "issue:46897",
+        "issue:47050"
+      ],
+      "soft_pairs": [
+        "issue:43299|issue:43931",
+        "issue:43596|issue:44683",
+        "issue:43632|issue:47050",
+        "issue:45479|issue:46897",
+        "issue:44488|issue:44987",
+        "issue:44488|issue:45081",
+        "issue:43611|issue:43644",
+        "issue:36032|issue:42915",
+        "issue:42796|issue:43404",
+        "issue:44462|issue:46489",
+        "issue:43716|issue:45084",
+        "issue:46285|issue:46402",
+        "issue:44661|issue:44683",
+        "issue:32090|issue:39404",
+        "issue:43643|issue:44530",
+        "issue:44451|issue:44987",
+        "issue:43541|issue:45084",
+        "issue:42915|issue:44530",
+        "issue:44737|issue:45859",
+        "issue:40990|issue:43421",
+        "issue:34567|issue:40444",
+        "issue:43644|issue:43646",
+        "issue:42175|issue:46178",
+        "issue:40444|issue:46326",
+        "issue:43582|issue:43632",
+        "issue:36032|issue:43645",
+        "issue:44206|issue:44220",
+        "issue:44811|issue:45584",
+        "issue:43452|issue:43482"
+      ],
+      "prepared_review_unit_hash": "ff7bf742deba396a88d159f10ba18af1f314507c6886fd612d77664882b16c8a",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36550,
+        "estimated_input_tokens": 9138,
+        "estimated_eval_tokens": 18532
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous: it mixes unrelated regressions across Trainer, Whisper, tokenizers, model loading, and backend/package issues. Only one soft pair looks like the same underlying GGUF-loading regression; the rest should stay separate.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "Issue 43452 is the broadest representative of the only plausible duplicate subcluster: it describes the general GGUF loading breakage, while 43482 is the Qwen2.5-specific instance of that same failure path.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43452 is the best issue representative because it captures the shared GGUF loading regression more generally than the model-specific report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43299",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Both mention Qwen/VL loading, but one is a v5 MoE regression and the other is a checkpoint weight-shape mismatch; not the same bug."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Unrelated failures: Deepspeed ZeRO3/BERT init versus torch>=2.9 compiled flex_attention."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:47050",
+            "accept": false,
+            "reason": "Different code paths: `_is_hf_initialized` flag regression versus Canine ONNX export dtype mismatch."
+          },
+          {
+            "left": "issue:45479",
+            "right": "issue:46897",
+            "accept": false,
+            "reason": "Both concern loss behavior, but one is zero-loss in sequence classification and the other is a Florence2 double-shift training-loss bug."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Both are model-load complaints, but they target different models and likely different tokenizer/config paths."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Different tokenizer failures: loading a specific BERT model versus `_patch_mistral_regex` crashing on Mistral tokenizers."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43644",
+            "accept": false,
+            "reason": "Both are v5 regressions, but one is base_model_prefix loading and the other is junk non-persistent buffers; distinct issues."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Tokenizer method conflict versus Qwen3Moe FP8 configuration failure; no shared bug."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43404",
+            "accept": false,
+            "reason": "Tokenizer BPE-marker handling and Mistral3 lm_head weight tying are unrelated."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:46489",
+            "accept": false,
+            "reason": "Both involve tokenizers, but one ignores repository tokenizer.json and the other changes DeepSeek-Coder output; different failures."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Mistral image-preprocessor dtype mismatch versus a template-node compilation error; unrelated."
+          },
+          {
+            "left": "issue:46285",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "CLIP structure inconsistency and Sam3 text-encoder weights not loading are different model-loading bugs."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Model-registration utility failure versus compiled flex_attention breakage; no shared code path."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:39404",
+            "accept": false,
+            "reason": "Trainer GPU broadcast NoneType error is unrelated to Whisper pipeline return_language behavior."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "AutoConfig trust_remote_code field loss and Qwen3.5 PagedAttentionCache invalid group type are separate regressions."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Different affected models and symptoms; same broad area of loading, but not the same underlying bug."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Torch dynamo grouped_mm tracing failure is unrelated to the non-template-node compilation error."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "Both are Qwen-family issues, but one is FP8 config support and the other is cache/group-type handling in Qwen3.5."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Different device-placement bugs in different models: XLNet CPU arange versus Qwen3_5Moe CPU offload."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:43421",
+            "accept": false,
+            "reason": "Perplexity regression on GPT-OSS is unrelated to runtime post-processor updates in TokenizersBackend."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "TrainerState token counting and Qwen2.5-VL iterable-dataset multi-image failures are unrelated."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43646",
+            "accept": false,
+            "reason": "Both are v5 initialization regressions, but non-persistent buffers being junk and custom model initialization are different failure modes."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "One is a missing tensorflow extra in the torch install, the other is an undefined tensorflow_text backend mapping."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Qwen2.5-VL multi-image dataset failure and mistral_common continue_final_message behavior are unrelated."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:43632",
+            "accept": false,
+            "reason": "Apple Silicon allocator warmup TypeError and `_is_hf_initialized` are unrelated regressions."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Tokenizer special-token conflict and Jupyter custom-model initialization are separate problems."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Both are audio feature extraction issues, but one is an unsupported `center` argument and the other is a lower-level `_torch_extract_fbank_features()` problem."
+          },
+          {
+            "left": "issue:44811",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Both are Whisper-related, but batch_decode skip_special_tokens and empty-transcription handling after align_special_tokens are different bugs."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43482",
+            "accept": true,
+            "reason": "Same GGUF loading path: 43452 reports the general AutoTokenizer/AutoModel breakage, and 43482 is the Qwen2.5-GGUF instance of that same v5 regression."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The only accepted soft pair, 43452/43482, is a plausible GGUF-loading duplicate; the rest are reasonably separated without overclaiming evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 11,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28530",
+        "issue:34567",
+        "issue:36032",
+        "issue:42915",
+        "issue:43208",
+        "issue:43262",
+        "issue:43317",
+        "issue:43377",
+        "issue:43404",
+        "issue:43482",
+        "issue:43531",
+        "issue:43632",
+        "issue:43645",
+        "issue:43646",
+        "issue:43742",
+        "issue:43746",
+        "issue:43792",
+        "issue:43819",
+        "issue:43883",
+        "issue:43994",
+        "issue:44008",
+        "issue:44062",
+        "issue:44265",
+        "issue:44291",
+        "issue:44403",
+        "issue:44451",
+        "issue:44464",
+        "issue:44466",
+        "issue:44488",
+        "issue:44534",
+        "issue:44589",
+        "issue:44610",
+        "issue:44849",
+        "issue:44877",
+        "issue:44987",
+        "issue:44991",
+        "issue:45081",
+        "issue:45084",
+        "issue:45325",
+        "issue:45507",
+        "issue:45987",
+        "issue:46326",
+        "issue:46661",
+        "issue:46821",
+        "issue:46869",
+        "issue:46899",
+        "issue:47013"
+      ],
+      "soft_pairs": [
+        "issue:44987|issue:44991",
+        "issue:45084|issue:45507",
+        "issue:43317|issue:43482",
+        "issue:43746|issue:46326",
+        "issue:43262|issue:43377",
+        "issue:43208|issue:47013",
+        "issue:44403|issue:45081",
+        "issue:43792|issue:46869",
+        "issue:42915|issue:45084",
+        "issue:43645|issue:46661",
+        "issue:43646|issue:44534",
+        "issue:43377|issue:43819",
+        "issue:44849|issue:45325",
+        "issue:43883|issue:44488",
+        "issue:44987|issue:45081",
+        "issue:28530|issue:34567",
+        "issue:43404|issue:44466",
+        "issue:42915|issue:44291",
+        "issue:43994|issue:44610",
+        "issue:36032|issue:43742",
+        "issue:44451|issue:45081",
+        "issue:44464|issue:45987",
+        "issue:44062|issue:44987",
+        "issue:44008|issue:46899",
+        "issue:42915|issue:44589",
+        "issue:44265|issue:44610",
+        "issue:43632|issue:46821",
+        "issue:43531|issue:44877"
+      ],
+      "prepared_review_unit_hash": "05602bc4a51c16421b4e1769e62fb49ba3c036c3cd9b58785595ca61d84587e6",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 28,
+        "serialized_chars": 35712,
+        "estimated_input_tokens": 8928,
+        "estimated_eval_tokens": 18112
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a loose semantic bucket of unrelated Transformers v5 regression reports. The titles often overlap on broad themes like model loading, custom models, quantization, or tokenizer handling, but the concrete bugs differ, so none of the soft pairs should be merged as duplicates.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43646 is the broadest representative of the dominant theme in this cluster: custom-model initialization regressions under Transformers v5. It is the clearest umbrella issue, though it is not a duplicate of most other items here.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44987",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are v5 loading failures, but they affect different models and different loader/tokenizer paths; not the same concrete bug."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "One is a template-compilation error, the other is an invalid-method call in GraniteMoEHybrid; unrelated failures."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "GPU/CPU offload of a dequantized model and Qwen2.5-GGUF loading are separate quantization/loading problems."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "PEFT local-checkpoint loading and continue_final_message behavior are different code paths with different symptoms."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43377",
+            "accept": false,
+            "reason": "Audio processor sampling-rate defaults and MIMI batched-vs-single output mismatch are distinct audio bugs."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:47013",
+            "accept": false,
+            "reason": "Same model family, but the reported failures are different: a bundle of three xLSTM bugs versus a train-mode unpacking crash."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Unnecessary loading noise and a Mistral tokenizer regex crash are not the same bug."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "Whisper runtime failure and YOLO multi-GPU device-map failure are unrelated model-loading issues."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "FineGrainedFP8Config failure in Qwen3-MoE and a non-template-node compile error do not share a concrete code path."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:46661",
+            "accept": false,
+            "reason": "Both involve model initialization/loading, but one is custom models in notebooks and the other is remote-code models missing config_class; different root causes."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Custom model initialization breakage and junk-filled non-persistent buffers are separate regressions."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "MIMI padding-mask divergence and DAC.from_latents missing STE are different model behaviors and fixes."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Qwen3.5 hidden-states bug and Qwen2.5-VL rope-index scaling bug are model-specific and unrelated."
+          },
+          {
+            "left": "issue:43883",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Missing all_tied_weights_keys in Molmo and failing to load ScandiBERT are different loading issues."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "A model load failure for physical-intelligence/fast and a Mistral regex-patching crash are unrelated."
+          },
+          {
+            "left": "issue:28530",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Early stopping metric selection and num_input_tokens_seen updating are separate trainer-state issues."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Weight-tying in Mistral3ForConditionalGeneration and device-dependent lm_head serialization are related only at a very high level, not the same bug."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Qwen3-MoE FP8 config failure and _is_hf_initialized init_empty_weights regression are different compatibility problems."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Siglip2 bad outputs and OmDet-Turbo processor size mismatch are unrelated model/preprocessor issues."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "T5 tokenizer special-token conflict and MobileLLM key-error loading are different tokenizer/model failures."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both are loading issues, but ScandiBERT loading and Mistral tokenizer regex patching are not the same defect."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45987",
+            "accept": false,
+            "reason": "Chunked generation inconsistency with compiled forward and StaticCache seq-length type mismatch are separate generation/cache bugs."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "AddedToken argument conflict and a model loading failure for physical-intelligence/fast are unrelated."
+          },
+          {
+            "left": "issue:44008",
+            "right": "issue:46899",
+            "accept": false,
+            "reason": "Gemma3n variable-name collision and Gemma4 4-bit quantization dtype-cast bug affect different model versions and code paths."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "FineGrainedFP8Config and missing Float8 storage are different quantization/runtime issues."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "torch.export/export compatibility and processor input-size mismatch are unrelated."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "_is_hf_initialized flag regression and TimesFM window_size AttributeError do not describe the same failure."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior and strict granite_speech config loading are distinct bugs."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 12,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:33666",
+        "issue:35707",
+        "issue:36032",
+        "issue:40444",
+        "issue:40990",
+        "issue:41720",
+        "issue:42796",
+        "issue:43089",
+        "issue:43296",
+        "issue:43322",
+        "issue:43381",
+        "issue:43421",
+        "issue:43425",
+        "issue:43482",
+        "issue:43525",
+        "issue:43531",
+        "issue:43540",
+        "issue:43611",
+        "issue:43632",
+        "issue:43716",
+        "issue:43746",
+        "issue:43792",
+        "issue:43883",
+        "issue:44079",
+        "issue:44220",
+        "issue:44291",
+        "issue:44403",
+        "issue:44479",
+        "issue:44534",
+        "issue:44568",
+        "issue:44589",
+        "issue:44683",
+        "issue:44756",
+        "issue:44811",
+        "issue:44991",
+        "issue:45081",
+        "issue:45216",
+        "issue:45325",
+        "issue:45357",
+        "issue:45663",
+        "issue:45812",
+        "issue:45923",
+        "issue:45987",
+        "issue:46315",
+        "issue:46491",
+        "issue:46710",
+        "issue:46899"
+      ],
+      "soft_pairs": [
+        "issue:43322|issue:43540",
+        "issue:36032|issue:41720",
+        "issue:44479|issue:45325",
+        "issue:43089|issue:45987",
+        "issue:43425|issue:44403",
+        "issue:43296|issue:43531",
+        "issue:45216|issue:45357",
+        "issue:40444|issue:45325",
+        "issue:43611|issue:44534",
+        "issue:44991|issue:46315",
+        "issue:44220|issue:44683",
+        "issue:43883|issue:44991",
+        "issue:43716|issue:46899",
+        "issue:44403|issue:44991",
+        "issue:43381|issue:45923",
+        "issue:42796|issue:43746",
+        "issue:45812|issue:46491",
+        "issue:43632|issue:44756",
+        "issue:43792|issue:44589",
+        "issue:33666|issue:40444",
+        "issue:44291|issue:44589",
+        "issue:42796|issue:46710",
+        "issue:40990|issue:44811",
+        "issue:43482|issue:43525",
+        "issue:35707|issue:40444",
+        "issue:44079|issue:45663",
+        "issue:43531|issue:44291",
+        "issue:43421|issue:44568",
+        "issue:41720|issue:45081"
+      ],
+      "prepared_review_unit_hash": "01cdfa21a5c34f23529b161c8d376d0bed629ed5b05cd3f6c55d8b50a88f1a03",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 35919,
+        "estimated_input_tokens": 8980,
+        "estimated_eval_tokens": 18216
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a heterogeneous mix of unrelated bug reports. The only strong duplicate-like pair is the Qwen3.5 checkpoint/save_pretrained regression; the rest differ by model, subsystem, or failure mode.",
+        "confidence": 0.87,
+        "canonical_issue_reason": "issue:45216 is earlier and broader, describing the core Qwen3.5 save_pretrained checkpoint corruption/regression; issue:45357 looks like a later, narrower follow-up about the same save-path bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45216 is the best global issue candidate because it captures the underlying Qwen3.5 save_pretrained regression at the right abstraction level and predates the narrower follow-up.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43322",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Different models and different failures: Llava Next loading segfault vs Qwen3OmniMoe video ValueError."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Unrelated subsystems: T5 tokenizer special-token method conflict vs Qwen3 auto device-map CUDA assert."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Both involve Qwen VL, but one is a broad video regression and the other is a still-image rope-index scaling bug; not the same concrete defect."
+          },
+          {
+            "left": "issue:43089",
+            "right": "issue:45987",
+            "accept": false,
+            "reason": "Generation performance/sync overhead is not the same bug as StaticCache.get_seq_length() returning the wrong type."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44403",
+            "accept": false,
+            "reason": "Torch version incompatibility and loading noise are distinct issues."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "PaddleOCR-VL load failure in vLLM is unrelated to Qwen3-MoE sliding_window behavior."
+          },
+          {
+            "left": "issue:45216",
+            "right": "issue:45357",
+            "accept": true,
+            "reason": "Same underlying Qwen3.5 save_pretrained regression; 45357 is a narrower follow-up about incorrect visual encoder keys."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "IterableDataset multi-image finetuning failure is different from still-image temporal position scaling."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Different v5 regressions: base_model_prefix loading vs non-persistent buffers being filled with junk."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:46315",
+            "accept": false,
+            "reason": "Tokenizer load failure for a specific model is not the same as slow convert_tokens_to_ids performance regression."
+          },
+          {
+            "left": "issue:44220",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Audio fbank extraction issue and compiled flex_attention torch-compat issue are unrelated."
+          },
+          {
+            "left": "issue:43883",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Missing all_tied_weights_keys in Molmo is unrelated to tokenizer loading for EMBEDDIA/est-roberta."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:46899",
+            "accept": false,
+            "reason": "Both are multimodal dtype problems, but they affect different models and code paths; one is image preprocessor/model mismatch, the other 4-bit embedders cast."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Loading noise vs tokenizer loading failure are different symptoms and likely different fixes."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Eval-mode gradient checkpointing restriction is not the same as a missing supports_gradient_checkpointing flag on a remote-code variant."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Tokenizer BPE marker handling and PEFT/local-checkpoint loading for GraniteSpeech are unrelated."
+          },
+          {
+            "left": "issue:45812",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they concern different models and different behaviors (wrong IDs vs space-stripping round-trip)."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Transformers v5 initialization flag breakage is unrelated to Strix Halo mmap OOM behavior."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Whisper model runtime failure and Float8 storage lookup error are not the same code-path problem."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "Both are Qwen VL training-related, but one is generic multi-GPU training and the other is multi-image IterableDataset finetuning; too broad to be duplicates."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "init_empty_weights/_is_hf_initialized constructor incompatibility is unrelated to missing Float8 storage type."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:46710",
+            "accept": false,
+            "reason": "Tokenizer BPE marker bug and DeepSeek accuracy regression are different issues."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "High perplexity on a GPT-OSS benchmark and Whisper batch_decode special-token handling are unrelated."
+          },
+          {
+            "left": "issue:43482",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Qwen2.5-GGUF loading failure and Llama4Config missing pad_token_id are different model/load bugs."
+          },
+          {
+            "left": "issue:35707",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "Progressive generation with inputs_embeds/past_key_values is unrelated to Qwen2.5-VL multi-image dataset finetuning."
+          },
+          {
+            "left": "issue:44079",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "ModelOutput key assignment bug and Gemma4 FSDP2 shared_kv_states KeyError are unrelated."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior is unrelated to the init_empty_weights constructor regression."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Runtime special-token post-processor updates and mdeberta add_special_tokens BOS/EOS regression are different tokenizer issues."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "CUDA assert in auto device mapping and Mistral regex patch backend_tokenizer crash are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The selected canonical issue pair (45216/45357) is plausibly the same Qwen3.5 save_pretrained regression, while the other pair judgments mostly reflect clear differences in model, subsystem, or failure mode without overstating duplicate evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 13,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:40444",
+        "issue:40990",
+        "issue:41553",
+        "issue:42175",
+        "issue:42796",
+        "issue:42915",
+        "issue:43296",
+        "issue:43317",
+        "issue:43377",
+        "issue:43454",
+        "issue:43482",
+        "issue:43493",
+        "issue:43582",
+        "issue:43606",
+        "issue:43611",
+        "issue:43644",
+        "issue:43646",
+        "issue:43701",
+        "issue:43818",
+        "issue:43856",
+        "issue:43873",
+        "issue:43994",
+        "issue:44079",
+        "issue:44265",
+        "issue:44291",
+        "issue:44462",
+        "issue:44466",
+        "issue:44479",
+        "issue:44509",
+        "issue:44534",
+        "issue:44829",
+        "issue:44877",
+        "issue:45072",
+        "issue:45325",
+        "issue:45357",
+        "issue:45381",
+        "issue:45479",
+        "issue:45491",
+        "issue:45538",
+        "issue:45542",
+        "issue:45584",
+        "issue:45701",
+        "issue:45812",
+        "issue:45823",
+        "issue:45834",
+        "issue:46402",
+        "issue:46821",
+        "issue:46909"
+      ],
+      "soft_pairs": [
+        "issue:42175|issue:45542",
+        "issue:43482|issue:45538",
+        "issue:43296|issue:46821",
+        "issue:44079|issue:44534",
+        "issue:44462|issue:45812",
+        "issue:42175|issue:43296",
+        "issue:43317|issue:43856",
+        "issue:44509|issue:46909",
+        "issue:43611|issue:43646",
+        "issue:43377|issue:45491",
+        "issue:43644|issue:44079",
+        "issue:43818|issue:46402",
+        "issue:40444|issue:45381",
+        "issue:44265|issue:45072",
+        "issue:43818|issue:45357",
+        "issue:45325|issue:45834",
+        "issue:43493|issue:43994",
+        "issue:45325|issue:45381",
+        "issue:44829|issue:45479",
+        "issue:43454|issue:44466",
+        "issue:44291|issue:44877",
+        "issue:42915|issue:43482",
+        "issue:43701|issue:45325",
+        "issue:43873|issue:45701",
+        "issue:42796|issue:45381",
+        "issue:43606|issue:44265",
+        "issue:41553|issue:43582",
+        "issue:40990|issue:45584",
+        "issue:44479|issue:45823"
+      ],
+      "prepared_review_unit_hash": "b12e3456d184be1a58c9d8af674026145382e8a98917815f747ed1d260ffe372",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36481,
+        "estimated_input_tokens": 9121,
+        "estimated_eval_tokens": 18498
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is a mixed bag of unrelated issues, with a few likely duplicate subclusters around transformers v5 regressions and Qwen2.5-VL video/position-id handling. I\u2019d anchor the cluster on the clearest concrete Qwen2.5-VL bug report.",
+        "confidence": 0.47,
+        "canonical_issue_reason": "issue:45325 is the most explicit root-cause report in the set: it names the exact function, version, and failure mode, and it cleanly represents the Qwen2.5-VL position-id regression subcluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45325 is the best representative issue because it is specific, actionable, and directly describes the concrete bug path rather than a broader symptom.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42175",
+            "right": "issue:45542",
+            "accept": true,
+            "reason": "Both report an incomplete TensorFlow backend install / packaging extra problem that leaves the tf backend unavailable at runtime."
+          },
+          {
+            "left": "issue:43482",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "These are unrelated: one is a Qwen2.5-GGUF v5 loading regression, the other is a CLIPTokenizer max-length constant issue."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "PaddleOCR-VL loading in vLLM and TimesFM window_size AttributeError are different models and different failure paths."
+          },
+          {
+            "left": "issue:44079",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "ModelOutput key bookkeeping and non-persistent buffer initialization are separate internal bugs."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "AutoTokenizer ignoring tokenizer.json and Granite token-ID regressions are both tokenizer-related but not the same underlying bug."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43296",
+            "accept": false,
+            "reason": "Package-extras/TensorFlow backend installation is unrelated to a PaddleOCR-VL model-load failure."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "gpu+cpu offload load failure and Qwen3 MoE training memory inefficiency are different problems."
+          },
+          {
+            "left": "issue:44509",
+            "right": "issue:46909",
+            "accept": false,
+            "reason": "Both are docs issues, but they concern different documentation content and are not the same change."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43646",
+            "accept": true,
+            "reason": "Both describe the same transformers v5 model-loading/custom-initialization regression affecting custom models and base_model_prefix handling."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Both involve batching/padding symptoms, but they hit different models and different mechanisms."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "A non-persistent buffer serialization bug is not the same as ModelOutput key assignment when a key was previously None."
+          },
+          {
+            "left": "issue:43818",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Missing temporal attention / shared weights in Video-LLaVA is unrelated to Sam3Model text-encoder weights not loading."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Multiple-images-per-prompt finetuning failure is a different multimodal path from the Qwen2.5-VL video position-id regression."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "torch.export compatibility with torch_compilable_check is unrelated to dtype mismatch during bfloat16 inference."
+          },
+          {
+            "left": "issue:43818",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Video-LLaVA temporal-attention/weight-sharing and Qwen3.5 save_pretrained visual-key serialization are different bugs."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Qwen2.5-VL temporal position-id scaling and Kosmos2.5 long-OCR index errors are different models and likely different code paths."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43994",
+            "accept": true,
+            "reason": "Both point to the same SigLIP2 correctness regression: HF outputs diverge from expected behavior and produce bad/nonsensical results for the same model family."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45381",
+            "accept": true,
+            "reason": "These are the same Qwen2.5-VL video/position-id bug in transformers 5.3.0, just described from slightly different angles."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Degenerate training from flash_attention_3 is a different issue than zero-loss classification behavior with num_labels=1."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Tying lm_head weights for garbage text generation and inconsistent lm_head.weight serialization are related to weights, but not the same concrete bug."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Unexpected _is_hf_initialized in init_empty_weights and strict Granite config rejection are different loading failures."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Qwen3MoE FineGrainedFP8Config failure and Qwen2.5-GGUF v5 loading failure are unrelated model/quantization issues."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Checkpoint resume key mismatch is not the same code path as the Qwen2.5-VL position-id regression."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Quantization/offloading behavior and tokenization changing across transformers versions are separate issues."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Ministral BPE-marker handling and Qwen2.5-VL vision_position_ids are unrelated tokenizer vs multimodal-position bugs."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44265",
+            "accept": false,
+            "reason": "CPU-offload device mismatch and torch.export / torch_compilable_check are different failure modes."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "Bad AutoTokenizer error messaging for Voxtral and Apple Silicon caching_allocator_warmup TypeError are unrelated."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "High perplexity on gpt-oss-20b and Whisper empty-transcription failure after align_special_tokens are different issues."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "The Qwen video-input regression and Gemma4 device-map mismatch are different model families and different bugs."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays appropriately conservative about the cluster being mostly unrelated issues with a few likely duplicates. The chosen canonical issue is well supported by the evidence, and the soft-edge acceptances are plausible without overstating certainty."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 14,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:34634",
+        "issue:35707",
+        "issue:36331",
+        "issue:39692",
+        "issue:40444",
+        "issue:40990",
+        "issue:41720",
+        "issue:42617",
+        "issue:42796",
+        "issue:42915",
+        "issue:42947",
+        "issue:43257",
+        "issue:43262",
+        "issue:43317",
+        "issue:43322",
+        "issue:43377",
+        "issue:43454",
+        "issue:43550",
+        "issue:43572",
+        "issue:43606",
+        "issue:43720",
+        "issue:43746",
+        "issue:43749",
+        "issue:43819",
+        "issue:43828",
+        "issue:43873",
+        "issue:44186",
+        "issue:44265",
+        "issue:44387",
+        "issue:44534",
+        "issue:44568",
+        "issue:44610",
+        "issue:44877",
+        "issue:44898",
+        "issue:44977",
+        "issue:44991",
+        "issue:45072",
+        "issue:45701",
+        "issue:45704",
+        "issue:45759",
+        "issue:45823",
+        "issue:45920",
+        "issue:45923",
+        "issue:46050",
+        "issue:46209",
+        "issue:46305",
+        "issue:46315",
+        "issue:46821"
+      ],
+      "soft_pairs": [
+        "issue:43262|issue:43746",
+        "issue:43828|issue:46821",
+        "issue:35707|issue:41720",
+        "issue:34634|issue:35707",
+        "issue:39692|issue:40444",
+        "issue:41720|issue:45823",
+        "issue:44186|issue:44265",
+        "issue:43606|issue:43819",
+        "issue:43454|issue:43550",
+        "issue:43257|issue:45923",
+        "issue:42796|issue:46209",
+        "issue:45823|issue:46821",
+        "issue:44568|issue:45920",
+        "issue:43377|issue:44610",
+        "issue:43454|issue:45072",
+        "issue:42617|issue:42915",
+        "issue:40990|issue:42947",
+        "issue:46305|issue:46315",
+        "issue:43572|issue:45759",
+        "issue:43377|issue:44898",
+        "issue:43454|issue:43720",
+        "issue:43749|issue:44534",
+        "issue:43317|issue:43873",
+        "issue:43322|issue:46050",
+        "issue:36331|issue:42617",
+        "issue:41720|issue:43322",
+        "issue:36331|issue:42915",
+        "issue:41720|issue:44977",
+        "issue:44877|issue:44991",
+        "issue:44387|issue:45704",
+        "issue:40990|issue:45701",
+        "issue:41720|issue:42617"
+      ],
+      "prepared_review_unit_hash": "505a76d8e6b3b66d53f1b282f43aca88e3975da4646749c18ac6310b191aa197",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37388,
+        "estimated_input_tokens": 9347,
+        "estimated_eval_tokens": 18950
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are mostly unrelated issue reports spanning tokenization, multimodal loading, training, quantization/offloading, and docs. None of the soft pairs look like safe duplicates or the same concrete fix path.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "issue:43873 is the broadest open report and the closest reusable target for offloading/quantization regressions, but the overall set is too heterogeneous for a true cluster canonical.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43873 is the best issue candidate only in the limited sense that it is open, broad, and reusable for offloading/quantization-related follow-ups; it still does not unify the rest of the cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43262",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Audio processor chat-template sampling-rate bug vs GraniteSpeech PEFT checkpoint loading; different components and failure modes."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Autocast dtype mismatch in Phi-tiny-MoE vs TimesFM window_size AttributeError; unrelated code paths."
+          },
+          {
+            "left": "issue:35707",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "inputs_embeds/past_key_values generation regression vs Qwen3 auto device-map CUDA assert; not the same bug."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "BarkProcessor voice_preset issue vs progressive generation with inputs_embeds/past_key_values; unrelated."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "SigLIP2 docs example errors vs Qwen2.5-VL IterableDataset multi-image finetuning; different surfaces and fixes."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Both are device-mapping failures, but on different models and likely different code paths; not mergeable as one bug."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44265",
+            "accept": false,
+            "reason": "LayoutLMv2Tokenizer NER/batching crash vs torch.export export failure with torch_compilable_check; unrelated."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Bark-small CPU offload device mismatch vs DAC.from_latents forward mismatch; different subsystems."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "AyaVision lm_head tying bug vs Bamba-9B-v2 torch.compile+SDPA failure; unrelated."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Qwen3 MoE weights conversion during accelerate+deepspeed vs missing supports_gradient_checkpointing flag; different problems."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Tokenizer BPE marker handling vs FP8 inference on B200; unrelated."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Gemma4 device mismatch vs TimesFM window_size AttributeError; different models and bugs."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45920",
+            "accept": false,
+            "reason": "mDeBERTa add_special_tokens regression vs broad wrong-token-id regression across models; both tokenization-related but not the same concrete defect."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "MIMI padding-mask batching bug vs OmDet-Turbo processor/model resolution mismatch; unrelated."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Weight tying issue vs bfloat16 dtype mismatch in inference; different code paths."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "3d_parallel.py runtime problem vs Qwen3MoE FineGrainedFP8Config failure; unrelated."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:42947",
+            "accept": false,
+            "reason": "High perplexity on gpt-oss-20b vs gradient-checkpointing ineffective with PEFT LoRA; unrelated."
+          },
+          {
+            "left": "issue:46305",
+            "right": "issue:46315",
+            "accept": false,
+            "reason": "MoonshineStreaming off-by-one mask bug vs slow convert_tokens_to_ids regression; unrelated."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Missing pad_token_idx in StableLmConfig vs AutoModelForCausalLM.from_config text_config unwrapping; different config bugs."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "MIMI padding-mask issue vs Perceiver non-default resolution interpolation failure; unrelated."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "AyaVision weight tying bug vs BitNet packed-weights unpacking bug; unrelated."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "FSDP CPU RAM efficient loading regression vs non-persistent buffers filled with junk; both loading-related but not the same concrete issue."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Specific device_map auto load/offload failure vs broad quantization offloading issue; too broad to treat as one duplicate."
+          },
+          {
+            "left": "issue:43322",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Llava Next loading segfault vs sinq save/load quantization regression; unrelated."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "CustomTrainer unexpected kwarg error vs 3d_parallel.py execution failure; unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43322",
+            "accept": false,
+            "reason": "Qwen3 device-map CUDA assert vs Llava Next loading segfault; different models and failure modes."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Trainer compute_loss signature regression vs Qwen3MoE FP8 config failure; unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Qwen3 device-map CUDA assert vs Qwen3.5 flash-attention generation issue; same family name but different concrete bugs."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "granite_speech strict-config loading bug vs est-roberta tokenizer load failure; unrelated."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "int4 quantization OOM regression vs T5 apex RMSNorm memory leak; both memory-related, but different layers and fixes."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "High perplexity on a model eval vs version-dependent tokenization change; not the same bug."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "Qwen3 auto device-map assert vs 3d_parallel.py runtime failure; unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 15,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:26413",
+        "issue:28530",
+        "issue:33666",
+        "issue:36010",
+        "issue:41720",
+        "issue:42617",
+        "issue:42915",
+        "issue:43065",
+        "issue:43295",
+        "issue:43317",
+        "issue:43404",
+        "issue:43450",
+        "issue:43531",
+        "issue:43550",
+        "issue:43572",
+        "issue:43577",
+        "issue:43606",
+        "issue:43646",
+        "issue:43708",
+        "issue:43720",
+        "issue:43873",
+        "issue:43881",
+        "issue:44008",
+        "issue:44186",
+        "issue:44423",
+        "issue:44451",
+        "issue:44466",
+        "issue:44488",
+        "issue:44610",
+        "issue:44734",
+        "issue:44756",
+        "issue:44898",
+        "issue:44991",
+        "issue:45072",
+        "issue:45161",
+        "issue:45357",
+        "issue:45406",
+        "issue:45685",
+        "issue:45701",
+        "issue:45823",
+        "issue:46402",
+        "issue:46612",
+        "issue:46665",
+        "issue:46821",
+        "issue:46858",
+        "issue:46869",
+        "issue:47020"
+      ],
+      "soft_pairs": [
+        "issue:43881|issue:44488",
+        "issue:43450|issue:45701",
+        "issue:44466|issue:45357",
+        "issue:43065|issue:46402",
+        "issue:26413|issue:28530",
+        "issue:43873|issue:44756",
+        "issue:44186|issue:45072",
+        "issue:46612|issue:46858",
+        "issue:45685|issue:46821",
+        "issue:43295|issue:45701",
+        "issue:43577|issue:43708",
+        "issue:43550|issue:43720",
+        "issue:43646|issue:46665",
+        "issue:36010|issue:42617",
+        "issue:44451|issue:47020",
+        "issue:44186|issue:44610",
+        "issue:43550|issue:43606",
+        "issue:42915|issue:43317",
+        "issue:43404|issue:45357",
+        "issue:44610|issue:44898",
+        "issue:46821|issue:46869",
+        "issue:28530|issue:33666",
+        "issue:41720|issue:45161",
+        "issue:44734|issue:46858",
+        "issue:44008|issue:45823",
+        "issue:43881|issue:44991",
+        "issue:44423|issue:45406",
+        "issue:43531|issue:43572"
+      ],
+      "prepared_review_unit_hash": "739bcadce1f2f7617250563163a855d5fd5d0e2457379e8474c5f119488dcbb5",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 28,
+        "serialized_chars": 35561,
+        "estimated_input_tokens": 8891,
+        "estimated_eval_tokens": 18038
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "No soft edges look like true duplicates: most pairs share only a broad subsystem (trainer, tokenizer, serving, cache, offload, multimodal), but the concrete failure modes and affected code paths differ. I\u2019d keep the open quantization/offload issue as the best canonical issue candidate.",
+        "confidence": 0.89,
+        "canonical_issue_reason": "issue:43873 is open, has multiple inbound references, and describes a concrete, reusable offloading/quantization failure pattern that is well-suited as a canonical tracker.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43873 has the strongest triage signal for a canonical issue: open state, broader user impact, inbound references, and a clearly described runtime failure.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43881",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both are loading failures, but they involve different models and likely different tokenizer/model-conversion bugs."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "One is a batched video processor shape bug; the other is a version-dependent tokenization change. Different code paths."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Both concern save/load serialization, but the broken fields and model families differ (lm_head weight vs visual encoder keys)."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Same Sam3 family, but one is about a dummy Conv2d in the pixel decoder and the other about missing text encoder weights. Distinct bugs."
+          },
+          {
+            "left": "issue:26413",
+            "right": "issue:28530",
+            "accept": false,
+            "reason": "Trainer checkpoint resume vs early stopping metric resolution are separate training-control issues."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Quantization/offload failure and Strix Halo mmap OOM are different root causes and mitigations."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Tokenizer crash on NER/batched padding is unrelated to dtype mismatches in model inference."
+          },
+          {
+            "left": "issue:46612",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Both touch cache logic, but one is missing beam-search cache reorder and the other is a static-cache seq-length tensor arithmetic bug."
+          },
+          {
+            "left": "issue:45685",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "MPS histogram on Int and TimesFM window_size AttributeError are unrelated model/backend failures."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Custom processor/tokenizer regression is different from tokenization behavior changes across versions."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43708",
+            "accept": false,
+            "reason": "One is a dtype-loading bug for Blip2; the other is a Trainer max_steps miscalculation on resume."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "torch.compile+SDPA on Bamba and packed-weight unpacking in BitNet are different implementation bugs."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:46665",
+            "accept": false,
+            "reason": "General custom model initialization breakage is not the same as inconsistent from_pretrained behavior in Python classes."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "GenerationMixin import failure and inability to run 3d_parallel.py are unrelated."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:47020",
+            "accept": false,
+            "reason": "Different tokenizer loading failures with different root causes and model/tokenizer stacks."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Tokenizer crash for NER/padding is unrelated to OmDet-Turbo processor image-size mismatch."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43606",
+            "accept": false,
+            "reason": "SDPA/torch.compile failure on Bamba and CPU offload device mismatch are different bugs."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43317",
+            "accept": false,
+            "reason": "FP8 configuration failure on Qwen3Moe and dequantized-model offload failure are different loading paths."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Both involve multimodal model weight handling, but one is untied lm_head weights and the other is incorrect visual encoder key serialization."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Processor output size mismatch and Perceiver non-default-resolution failure are distinct vision preprocessing/model issues."
+          },
+          {
+            "left": "issue:46821",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "TimesFM window_size AttributeError and YOLO multi-GPU device_map failure are unrelated."
+          },
+          {
+            "left": "issue:28530",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "Early stopping metric lookup and Qwen2-VL multi-GPU training are not the same bug."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Qwen3 auto device-map cudaErrorAssert and GPT-OSS tensor-parallel failure are different distributed/inference problems."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Serve KV-cache continuation indexing and static-cache seq-length device-tensor arithmetic are both cache-related but not the same concrete bug."
+          },
+          {
+            "left": "issue:44008",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Variable name collision in Gemma3n forward and Gemma4 PLE device mismatch are unrelated runtime errors."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Different tokenizer/model loading failures for different repositories and failure modes."
+          },
+          {
+            "left": "issue:44423",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both affect serving multimodal models, but one is a .to() on str bug and the other is a missing _tokenizer attribute. Different root causes."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Sliding-window handling in Qwen3-MoE and missing pad_token_idx in StableLmConfig are unrelated issues."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 16,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28530",
+        "issue:29159",
+        "issue:30064",
+        "issue:33357",
+        "issue:34567",
+        "issue:34634",
+        "issue:35707",
+        "issue:36032",
+        "issue:36331",
+        "issue:41720",
+        "issue:42490",
+        "issue:42915",
+        "issue:43316",
+        "issue:43322",
+        "issue:43329",
+        "issue:43425",
+        "issue:43452",
+        "issue:43526",
+        "issue:43531",
+        "issue:43540",
+        "issue:43550",
+        "issue:43606",
+        "issue:43653",
+        "issue:43720",
+        "issue:43746",
+        "issue:43782",
+        "issue:43873",
+        "issue:43883",
+        "issue:44186",
+        "issue:44361",
+        "issue:44464",
+        "issue:44492",
+        "issue:44509",
+        "issue:44534",
+        "issue:44671",
+        "issue:44743",
+        "issue:44913",
+        "issue:45632",
+        "issue:45684",
+        "issue:45701",
+        "issue:45704",
+        "issue:45759",
+        "issue:46291",
+        "issue:46665",
+        "issue:46821",
+        "issue:46858",
+        "issue:47020",
+        "issue:47145"
+      ],
+      "soft_pairs": [
+        "issue:43526|issue:43883",
+        "issue:42915|issue:46291",
+        "issue:36331|issue:41720",
+        "issue:44464|issue:46858",
+        "issue:45632|issue:45684",
+        "issue:44671|issue:44743",
+        "issue:44186|issue:44361",
+        "issue:43425|issue:46821",
+        "issue:28530|issue:30064",
+        "issue:43316|issue:44913",
+        "issue:45701|issue:47145",
+        "issue:42915|issue:43322",
+        "issue:33357|issue:35707",
+        "issue:29159|issue:36032",
+        "issue:30064|issue:34634",
+        "issue:43873|issue:45704",
+        "issue:43540|issue:43782",
+        "issue:43531|issue:45759",
+        "issue:44492|issue:44509",
+        "issue:43550|issue:43653",
+        "issue:44534|issue:46665",
+        "issue:43452|issue:47020",
+        "issue:41720|issue:43746",
+        "issue:43606|issue:43720",
+        "issue:42490|issue:43329",
+        "issue:34567|issue:35707",
+        "issue:43531|issue:46821"
+      ],
+      "prepared_review_unit_hash": "c3b354b3ba526c1d458be9ae2f5b250261f0d5325eaf5a32f7623c4fade6642f",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 35842,
+        "estimated_input_tokens": 8961,
+        "estimated_eval_tokens": 18178
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous: the soft pairs share broad vocabulary or subsystem overlap, but they describe different bugs, configs, or code paths. I do not see any true duplicate pairs, and there are no PRs in the set.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43526",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "Different problems: BeitImageProcessorFast label reduction vs Molmo tied-weights missing attribute."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Both are loading-related, but one is Qwen3Moe FP8 failure and the other is a kernels import/version requirement."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Trainer API keyword mismatch is unrelated to Qwen3 auto device mapping cuda assert."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Both involve generation/cache behavior, but compiled chunked generation inconsistency is a different bug from static-cache seq_length device tensors."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:45684",
+            "accept": false,
+            "reason": "Both are path/save/load issues, but cache-path collisions and read-only permissions are distinct failure modes."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "CamemBERT masked-LM wrong outputs and Qwen3 recurrent-state reset are different model-specific bugs."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Tokenizer crashes in NER/padding for LayoutLMv2 vs MLuke task AttributeError are not the same bug."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Torch 2.10 compatibility and TimesFM window_size AttributeError are unrelated compatibility/config issues."
+          },
+          {
+            "left": "issue:28530",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "Early-stopping metric lookup and image processor void segmentation maps are different subsystems and symptoms."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Gemma3 text config API mismatch is not the same as GPTNeoX rotary_pct persisting incorrectly."
+          },
+          {
+            "left": "issue:45701",
+            "right": "issue:47145",
+            "accept": false,
+            "reason": "Tokenization version drift and FP8 weight_scale corruption are unrelated issues."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43322",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 loading failure is unrelated to Llava Next segmentation faults."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "MacOS CLIP bus error and inputs_embeds/past_key_values generation issue do not share a concrete code path."
+          },
+          {
+            "left": "issue:29159",
+            "right": "issue:36032",
+            "accept": false,
+            "reason": "Slow/fast tokenizer inconsistency is a different tokenizer bug than T5 add_special_tokens loading conflict."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:34634",
+            "accept": false,
+            "reason": "Void segmentation map processing and BarkProcessor voice_preset handling are unrelated processor bugs."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Quantization offloading behavior and T5 apex RMSNorm memory leak are different root causes."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:43782",
+            "accept": false,
+            "reason": "Video-input processing error in Qwen3OmniMoe is unrelated to Qwen3VL weight_only loading failure."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Sliding-window behavior in Qwen3-MoE and composite config unwrapping in AutoModelForCausalLM are distinct bugs."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "Cache strategy typo and docs mentioning removed pipeline tasks are unrelated documentation issues."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "torch.compile+SDPA failure in Bamba-9B-v2 and BigBirdTokenizer special-token registration are not the same bug."
+          },
+          {
+            "left": "issue:44534",
+            "right": "issue:46665",
+            "accept": false,
+            "reason": "Non-persistent buffer corruption in v5 is unrelated to from_pretrained behavior inside Python classes."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:47020",
+            "accept": false,
+            "reason": "gguf_file from_pretrained breakage and XLMRobertaTokenizer Unigram vocab type mismatch are separate loading/tokenizer issues."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping cuda assert and GraniteSpeech PEFT local-checkpoint loading are different failure modes."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "CPU offload device mismatch for bark-small and BitNet packed-weights unpacking are different accelerate-loading bugs."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Trailing-slash local model import warning and multimodal video branch undefined variables are unrelated."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "TrainerState token counter not updating is unrelated to progressive generation with inputs_embeds/past_key_values."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding-window handling and TimesFM window_size AttributeError are separate model-specific issues."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 17,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:34634",
+        "issue:39692",
+        "issue:41720",
+        "issue:42886",
+        "issue:43381",
+        "issue:43388",
+        "issue:43421",
+        "issue:43526",
+        "issue:43531",
+        "issue:43540",
+        "issue:43606",
+        "issue:43644",
+        "issue:43653",
+        "issue:43746",
+        "issue:43844",
+        "issue:43873",
+        "issue:43881",
+        "issue:43927",
+        "issue:44062",
+        "issue:44077",
+        "issue:44368",
+        "issue:44451",
+        "issue:44589",
+        "issue:44610",
+        "issue:44625",
+        "issue:44843",
+        "issue:44898",
+        "issue:44918",
+        "issue:44936",
+        "issue:44987",
+        "issue:44991",
+        "issue:45042",
+        "issue:45059",
+        "issue:45072",
+        "issue:45357",
+        "issue:45406",
+        "issue:45464",
+        "issue:45479",
+        "issue:45704",
+        "issue:45759",
+        "issue:45859",
+        "issue:45923",
+        "issue:46050",
+        "issue:46402",
+        "issue:46655",
+        "issue:46665",
+        "issue:46693",
+        "issue:46858"
+      ],
+      "soft_pairs": [
+        "issue:43388|issue:45059",
+        "issue:34634|issue:41720",
+        "issue:45406|issue:45464",
+        "issue:43606|issue:43653",
+        "issue:43531|issue:44589",
+        "issue:45759|issue:45859",
+        "issue:43540|issue:46858",
+        "issue:44451|issue:44991",
+        "issue:44077|issue:44625",
+        "issue:45357|issue:46402",
+        "issue:43644|issue:46665",
+        "issue:44987|issue:45042",
+        "issue:43526|issue:43881",
+        "issue:44062|issue:44451",
+        "issue:44898|issue:45072",
+        "issue:44625|issue:45479",
+        "issue:43381|issue:43844",
+        "issue:39692|issue:44610",
+        "issue:41720|issue:45704",
+        "issue:43746|issue:46050",
+        "issue:44368|issue:45357",
+        "issue:43381|issue:45704",
+        "issue:46655|issue:46693",
+        "issue:44918|issue:44936",
+        "issue:43421|issue:43927",
+        "issue:45704|issue:45923",
+        "issue:45759|issue:45923",
+        "issue:43746|issue:43873",
+        "issue:42886|issue:44843"
+      ],
+      "prepared_review_unit_hash": "1d0fd6a86632e3b1456cbdf6ccf996b9cbc624cc8e1741847127c0ed6224a323",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36648,
+        "estimated_input_tokens": 9162,
+        "estimated_eval_tokens": 18580
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most pairs are clearly different bugs touching different subsystems. Only the two HF_HUB_OFFLINE tokenizer-loading reports look like the same underlying offline regression in AutoTokenizer/_patch_mistral_regex.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "issue:44843 is the clearest root-cause report: it names the unconditional model_info() call in _patch_mistral_regex and the HF_HUB_OFFLINE breakage directly.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44843 is the best representative issue because it is the most specific and actionable description of the shared offline tokenizer bug.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43388",
+            "right": "issue:45059",
+            "accept": false,
+            "reason": "Different failure modes in different areas: distributed metric gathering vs a multimodal model/processor mismatch."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Bark voice_preset and Qwen3 auto device mapping/cuda assert are unrelated bugs."
+          },
+          {
+            "left": "issue:45406",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Gemma4Processor _tokenizer crash and Qwen3.5 streaming API failure do not share a concrete code path."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "CPU offload device mismatch for bark-small is unrelated to BigBirdTokenizer special-token decode behavior."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window and a Float8 storage lookup error are distinct runtime problems."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Both involve Qwen 3.x, but one is config unwrapping and the other is CPU-offload rotary embedding behavior."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video-input validation and static-cache slice arithmetic on non-CUDA backends are different bugs."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are tokenizer-loading regressions, but for different models and no clear shared root cause is shown."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44625",
+            "accept": false,
+            "reason": "PatchTSMixer post_init policy and Qwen3.5 num_labels propagation are unrelated config issues."
+          },
+          {
+            "left": "issue:45357",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Incorrect visual encoder keys on save_pretrained and SAM3 text-encoder weights not loading are separate multimodal loading bugs."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:46665",
+            "accept": false,
+            "reason": "Non-persistent buffer junk initialization and inconsistent from_pretrained behavior inside Python classes are not the same defect."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "A model-loading regression for physical-intelligence/fast is unrelated to PIL backend torchvision requirements."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction and glm-4v-9b loading failure are different issues."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "AddedToken constructor duplication and ScandiBERT loading failure may both involve tokenizers, but the underlying bugs are not established as the same."
+          },
+          {
+            "left": "issue:44898",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Perceiver resize failure and bfloat16 dtype mismatches in SwitchTransformers/TimmWrapperModel are unrelated."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "num_labels propagation is a config wiring issue, while zero loss with num_labels=1 is a loss-function semantics bug."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode and ZeRO-3 gradient growth are unrelated training issues."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "SigLIP2 doc errors and OmDet-Turbo processor resolution mismatch are different model-specific problems."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Qwen3 device-mapping cuda assert and T5 apex RMSNorm memory leak are unrelated."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "PEFT local-checkpoint loading and sinq save/load breakage are both persistence bugs but not the same concrete code-path."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "A warning about tie_word_embeddings and incorrect visual encoder keys address different Qwen3.5 behaviors."
+          },
+          {
+            "left": "issue:46655",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "SmolVLM video padding logic and FA2 long-context performance regression are not the same bug."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "TRL SFT embedding unpacking failure and train-to-evaluate failure are distinct trainer issues."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Runtime special-token post-processing and DiaConfig custom-token persistence are different tokenization/config bugs."
+          },
+          {
+            "left": "issue:45704",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "T5 RMSNorm memory leak and missing supports_gradient_checkpointing are unrelated."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Composite Qwen config unwrapping and missing gradient-checkpointing support are different code paths."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Checkpoint loading with PEFT adapters is not the same as offloading/quantization behavior."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:44843",
+            "accept": true,
+            "reason": "Both report HF_HUB_OFFLINE failures when loading tokenizers, and 44843 names the likely shared root cause: an unconditional model_info() call in _patch_mistral_regex."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The only accepted soft pair, issue:42886 with issue:44843, is strongly supported by the titles: both describe HF_HUB_OFFLINE tokenizer-loading failures, and 44843 explicitly names the unconditional model_info() call in _patch_mistral_regex as the cause. The other pairs are reasonably rejected as distinct bugs, and the representative/canonical issue choice is justified by the more specific root-cause description in 44843."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 18,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:30064",
+        "issue:36010",
+        "issue:36331",
+        "issue:39692",
+        "issue:43295",
+        "issue:43421",
+        "issue:43493",
+        "issue:43531",
+        "issue:43540",
+        "issue:43550",
+        "issue:43596",
+        "issue:43611",
+        "issue:43643",
+        "issue:43653",
+        "issue:43720",
+        "issue:43782",
+        "issue:43819",
+        "issue:43873",
+        "issue:43927",
+        "issue:43994",
+        "issue:44062",
+        "issue:44186",
+        "issue:44220",
+        "issue:44361",
+        "issue:44464",
+        "issue:44488",
+        "issue:44514",
+        "issue:44545",
+        "issue:44625",
+        "issue:44671",
+        "issue:44857",
+        "issue:44995",
+        "issue:45081",
+        "issue:45356",
+        "issue:45381",
+        "issue:45834",
+        "issue:45907",
+        "issue:45987",
+        "issue:46032",
+        "issue:46294",
+        "issue:46463",
+        "issue:46665",
+        "issue:46693",
+        "issue:46829",
+        "issue:46832",
+        "issue:46858",
+        "issue:47050"
+      ],
+      "soft_pairs": [
+        "issue:44514|issue:44545",
+        "issue:30064|issue:36010",
+        "issue:39692|issue:43493",
+        "issue:43421|issue:43653",
+        "issue:43994|issue:44671",
+        "issue:44464|issue:47050",
+        "issue:43782|issue:45381",
+        "issue:46032|issue:46463",
+        "issue:44857|issue:47050",
+        "issue:45081|issue:45356",
+        "issue:43421|issue:44186",
+        "issue:45356|issue:45907",
+        "issue:44995|issue:46858",
+        "issue:39692|issue:43873",
+        "issue:39692|issue:47050",
+        "issue:43720|issue:43819",
+        "issue:43295|issue:45834",
+        "issue:43653|issue:43720",
+        "issue:46829|issue:46832",
+        "issue:43550|issue:43927",
+        "issue:43550|issue:44361",
+        "issue:43596|issue:44220",
+        "issue:44625|issue:47050",
+        "issue:43540|issue:46294",
+        "issue:43531|issue:43643",
+        "issue:44062|issue:44488",
+        "issue:45381|issue:46693",
+        "issue:45987|issue:46829",
+        "issue:43611|issue:46665",
+        "issue:30064|issue:36331"
+      ],
+      "prepared_review_unit_hash": "827011fee78503d8ea4f4b385bf64bbfd604e379c4f95ee3f52684e755b6c607",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 30,
+        "serialized_chars": 36417,
+        "estimated_input_tokens": 9105,
+        "estimated_eval_tokens": 18466
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a set of unrelated HF bug reports. Only one soft edge looks like a true duplicate: the Qwen2_5_VLProcessor batched-input padding=False crash (44514/44545).",
+        "confidence": 0.91,
+        "canonical_issue_reason": "No single canonical issue fits the whole set; issue 44514 is the best anchor only because it is the earliest, clearest report of the one confirmed duplicate pair.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44514 is the best representative for triage because it exactly matches 44545 and describes the concrete failure most clearly.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44514",
+            "right": "issue:44545",
+            "accept": true,
+            "reason": "Same Qwen2_5_VLProcessor.apply_chat_template batched-input crash with padding=False; clearly the same bug."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Different components and failures: image segmentation map processing vs GenerationMixin import error."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Both involve SigLIP2, but one is a docs example error and the other is an implementation discrepancy; not the same bug."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Both touch tokenizer special-token handling, but the concrete failures differ (runtime post-processor update vs mask token decode output)."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Different models and different failure modes; no shared concrete bug."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:47050",
+            "accept": false,
+            "reason": "Unrelated code paths: compiled generation inconsistency vs ONNX fp16 attention dtype mismatch."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both are Qwen vision-language issues, but the errors concern different models and different inputs."
+          },
+          {
+            "left": "issue:46032",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Different subsystems: cache/use_cache correctness vs T5 decoder input validation message."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:47050",
+            "accept": false,
+            "reason": "Different export/runtime problems in different models; not the same underlying bug."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Both mention tokenizer regex fixes, but one is a crash in _patch_mistral_regex and the other is a Kimi regression with codec handling; too different to merge."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44186",
+            "accept": false,
+            "reason": "Shared tokenizer area only broadly; one is special-token propagation, the other is LayoutLMv2 NER/padding crashes."
+          },
+          {
+            "left": "issue:45356",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Tokenizer regression vs unrelated Hub API argument forwarding bug."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Both mention cache, but they affect different cache mechanisms and different failure points."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "SigLIP2 doc/model issue is unrelated to offloading with quantization."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:47050",
+            "accept": false,
+            "reason": "Different model families and unrelated bugs."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Different problems: packed-weight loading vs DAC from_latents / STE mismatch."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Both are multimodal processor issues, but one is a processor/tokenizer regression and the other is an OCR length index error."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Tokenizer special-token decoding bug vs BitNet packed-weight loading bug; unrelated."
+          },
+          {
+            "left": "issue:46829",
+            "right": "issue:46832",
+            "accept": false,
+            "reason": "Different model heads and different forward-path errors."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure in Bamba vs DiaConfig save/load token ID loss; unrelated."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Different models and different failure types."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "BertModel zero3 init IndexError vs audio feature extraction issue; no concrete overlap."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:47050",
+            "accept": false,
+            "reason": "Qwen3.5 config propagation bug vs unrelated ONNX dtype mismatch."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "Both are Qwen video-input bugs, but they concern different models and different validation paths."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "sliding_window MoE bug vs AutoConfig trust_remote_code missing fields; unrelated."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Generic tokenizer constructor error vs a specific model load failure; not enough evidence of the same bug."
+          },
+          {
+            "left": "issue:45381",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "Qwen2.5-VL video position IDs bug vs FA2 long-context performance regression; different issues."
+          },
+          {
+            "left": "issue:45987",
+            "right": "issue:46829",
+            "accept": false,
+            "reason": "StaticCache return-type bug vs VideoPrism output field mismatch; unrelated."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:46665",
+            "accept": false,
+            "reason": "Both involve loading/from_pretrained behavior, but the concrete regressions are different."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Image processor void-segmentation failure vs trainer loss-signature TypeError; unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The only accepted soft edge, 44514/44545, is well supported by the near-identical titles and matching failure description. The remaining pairwise rejections do not overclaim duplicate evidence based on the packet."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 19,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:34634",
+        "issue:36010",
+        "issue:39692",
+        "issue:41720",
+        "issue:43322",
+        "issue:43335",
+        "issue:43421",
+        "issue:43540",
+        "issue:43550",
+        "issue:43606",
+        "issue:43618",
+        "issue:43643",
+        "issue:43644",
+        "issue:43653",
+        "issue:43697",
+        "issue:43749",
+        "issue:43756",
+        "issue:43819",
+        "issue:44008",
+        "issue:44186",
+        "issue:44220",
+        "issue:44442",
+        "issue:44448",
+        "issue:44488",
+        "issue:44561",
+        "issue:44661",
+        "issue:44671",
+        "issue:44811",
+        "issue:44821",
+        "issue:44843",
+        "issue:44877",
+        "issue:44898",
+        "issue:44991",
+        "issue:45020",
+        "issue:45072",
+        "issue:45081",
+        "issue:45084",
+        "issue:45183",
+        "issue:45479",
+        "issue:45584",
+        "issue:45685",
+        "issue:45834",
+        "issue:45987",
+        "issue:46291",
+        "issue:46858",
+        "issue:46932",
+        "issue:47020"
+      ],
+      "soft_pairs": [
+        "issue:41720|issue:44991",
+        "issue:44220|issue:44661",
+        "issue:44186|issue:44898",
+        "issue:43653|issue:43819",
+        "issue:44488|issue:47020",
+        "issue:45987|issue:46932",
+        "issue:43550|issue:43819",
+        "issue:43335|issue:43756",
+        "issue:43618|issue:43697",
+        "issue:43643|issue:44877",
+        "issue:44008|issue:46858",
+        "issue:44821|issue:45084",
+        "issue:34634|issue:36010",
+        "issue:44843|issue:45081",
+        "issue:39692|issue:45834",
+        "issue:43540|issue:44008",
+        "issue:44186|issue:44442",
+        "issue:44448|issue:44671",
+        "issue:45584|issue:45834",
+        "issue:45183|issue:45479",
+        "issue:44991|issue:47020",
+        "issue:43322|issue:45685",
+        "issue:43644|issue:43749",
+        "issue:44991|issue:46291",
+        "issue:43606|issue:45072",
+        "issue:44561|issue:45020",
+        "issue:43421|issue:44811",
+        "issue:43550|issue:45072"
+      ],
+      "prepared_review_unit_hash": "4e2fab4e7a26243af86f6b42956c8bbdf767b09ca9b5942810c4f4e1f50b5ac3",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 28,
+        "serialized_chars": 35696,
+        "estimated_input_tokens": 8924,
+        "estimated_eval_tokens": 18104
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "All soft pairs look like false positives: they span different models, APIs, and failure modes, and none appear to be the same concrete bug or change.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "Issue #45020 is the broadest, most umbrella-like report in this set (remote_code regressions across recent transformers versions), so it is the best anchor if one issue must represent the cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #45020 is the most generally scoped and well-framed regression among mostly narrow, model-specific reports; it is the strongest issue candidate to keep as the representative.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41720",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Different failures: Qwen3 auto device mapping CUDA assert vs tokenizer loading for EMBEDDIA/est-roberta."
+          },
+          {
+            "left": "issue:44220",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "Unrelated code paths: audio fbank extraction vs add-new-model-like/tokenizer mapping logic."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "LayoutLMv2 tokenizer NER/padding crash vs Perceiver image-classification resolution handling."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration bug vs DAC.from_latents/forward mismatch; different subsystems."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:47020",
+            "accept": false,
+            "reason": "Both are tokenizer-loading problems, but one is a specific model load failure and the other is a Unigram constructor type mismatch; not the same bug."
+          },
+          {
+            "left": "issue:45987",
+            "right": "issue:46932",
+            "accept": false,
+            "reason": "StaticCache seq-length return type bug vs Gemma4 quantized embedder dtype casting; unrelated."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure in Bamba-9B-v2 vs DAC latent reconstruction mismatch; different models and paths."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43756",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer config bug vs Smollm3 RoPE-layer count issue; not the same defect."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "Missing CLIPOutput attentions assignment vs RTDetrV2 output drift in v5; different regressions."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "AutoConfig trust_remote_code missing fields vs strict config blocking granite_speech loading; both config-related but distinct failures."
+          },
+          {
+            "left": "issue:44008",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Gemma3n variable-name collision in forward vs cache.get_seq_length device-tensor issue; unrelated bugs."
+          },
+          {
+            "left": "issue:44821",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "AutoImageProcessor URL loading vs template-node compilation error; different APIs."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "BarkProcessor voice_preset bug vs GenerationMixin import error; not the same issue."
+          },
+          {
+            "left": "issue:44843",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "One is an offline-breaking model_info call, the other an AttributeError in _patch_mistral_regex; same helper, different concrete bugs."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "SigLIP2 doc example errors vs Kosmos2.5 long-OCR index error; unrelated models and contexts."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44008",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video-input ValueError vs Gemma3n audio mask AttributeError; different multimodal paths."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "Tokenizer crash for LayoutLMv2 NER/padding vs AutoTokenizer failing on FastSpeech2ConformerTokenizer; different tokenizer bugs."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Pegasus v4/v5 output drift vs CamemBERT masked-LM prediction regression; different model-specific changes."
+          },
+          {
+            "left": "issue:45584",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Whisper empty-transcription generation failure vs Kosmos2.5 OCR index error; unrelated model behaviors."
+          },
+          {
+            "left": "issue:45183",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Input validation XOR error messaging vs degenerate zero loss in sequence classification; different logic."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:47020",
+            "accept": false,
+            "reason": "Both tokenizer-loading regressions, but they involve different tokenizer classes and failure mechanisms; not the same concrete bug."
+          },
+          {
+            "left": "issue:43322",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Llava Next segmentation fault vs MPS histogram-not-implemented error in moe code; unrelated."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Non-persistent buffer initialization junk vs FSDP CPU RAM efficient loading; different symptoms and code paths."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Tokenizer load failure for one model vs kernels import/version requirement during pipelines import; unrelated."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "CPU-offload device mismatch in bark-small vs bfloat16 dtype mismatch in SwitchTransformers/TimmWrapperModel; different bugs."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "trust_remote_code breakage from removing is_torch_fx_available vs broader remote_code regressions; related area but not the same concrete code-path problem."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Runtime special-token/post-processor update bug vs Whisper batch_decode skip_special_tokens handling; different layers and fixes."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "torch.compile + SDPA failure vs bf16 dtype mismatch; different model/back-end regressions."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 20,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28218",
+        "issue:29159",
+        "issue:34634",
+        "issue:36331",
+        "issue:42175",
+        "issue:43262",
+        "issue:43388",
+        "issue:43540",
+        "issue:43577",
+        "issue:43582",
+        "issue:43645",
+        "issue:43716",
+        "issue:43720",
+        "issue:43749",
+        "issue:43782",
+        "issue:43818",
+        "issue:43819",
+        "issue:43825",
+        "issue:43827",
+        "issue:43931",
+        "issue:43994",
+        "issue:44008",
+        "issue:44291",
+        "issue:44479",
+        "issue:44610",
+        "issue:44671",
+        "issue:44805",
+        "issue:44857",
+        "issue:44936",
+        "issue:44977",
+        "issue:45003",
+        "issue:45072",
+        "issue:45081",
+        "issue:45325",
+        "issue:45381",
+        "issue:45701",
+        "issue:45834",
+        "issue:46032",
+        "issue:46175",
+        "issue:46285",
+        "issue:46294",
+        "issue:46463",
+        "issue:46512",
+        "issue:46665",
+        "issue:46697",
+        "issue:46728",
+        "issue:46846",
+        "issue:46932"
+      ],
+      "soft_pairs": [
+        "issue:43577|issue:43818",
+        "issue:44479|issue:45834",
+        "issue:43582|issue:46512",
+        "issue:43782|issue:44977",
+        "issue:44805|issue:44936",
+        "issue:43825|issue:43994",
+        "issue:43388|issue:46032",
+        "issue:46294|issue:46463",
+        "issue:43716|issue:44008",
+        "issue:43931|issue:46846",
+        "issue:43540|issue:45381",
+        "issue:28218|issue:29159",
+        "issue:44610|issue:45072",
+        "issue:43749|issue:44291",
+        "issue:43720|issue:45072",
+        "issue:43645|issue:46665",
+        "issue:43825|issue:43827",
+        "issue:34634|issue:36331",
+        "issue:42175|issue:45081",
+        "issue:43577|issue:43582",
+        "issue:44671|issue:45701",
+        "issue:45003|issue:46285",
+        "issue:46175|issue:46932",
+        "issue:43262|issue:43819",
+        "issue:46697|issue:46728",
+        "issue:43782|issue:45325",
+        "issue:43582|issue:44857"
+      ],
+      "prepared_review_unit_hash": "8629f61903064f499c168b13520c06e590a04c646a4783be80a03455f0c0487e",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 36359,
+        "estimated_input_tokens": 9090,
+        "estimated_eval_tokens": 18436
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most pairs are unrelated and should be kept separate. The only strong duplicate-looking pair is the tokenizer spacing/inconsistency reports (#28218 and #29159).",
+        "confidence": 0.88,
+        "canonical_issue_reason": "#28218 is the clearest, most specific report of the shared tokenizer-added-token spacing bug and is open, making it the best canonical issue for that duplicate set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#28218 is the most actionable issue in the only true duplicate pair: it states the concrete symptom directly and can anchor the cluster cleanly.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43577",
+            "right": "issue:43818",
+            "accept": false,
+            "reason": "Both are vision/model-loading problems, but one is a dtype-loading mismatch in BLIP2 and the other is a Video-LLaVA architecture issue; not the same bug."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Different models and failure modes: a Qwen video regression versus a Kosmos2.5 OCR index error."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "Apple Silicon allocator TypeError is unrelated to from_config leaving config.dtype stale."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "A weight_only loading error and a flash-attention generation bug are different code paths and symptoms."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Mask-shape IndexError is not the same as trainer.evaluate() failing after training."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "One is a wrong error message about pipeline support; the other is a model producing bad outputs via pipeline, which is a different defect."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:46032",
+            "accept": false,
+            "reason": "Metric gathering truncation and Mamba2 cache corruption are unrelated training/inference bugs."
+          },
+          {
+            "left": "issue:46294",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Video feature/token mismatch for Qwen and T5 decoder-input validation messaging are unrelated."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44008",
+            "accept": false,
+            "reason": "Both mention dtype/attribute issues, but they affect different models and different forward paths."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:46846",
+            "accept": false,
+            "reason": "Qwen3-VL weight-shape loading mismatch and Qwen3.5 TP-plan omissions are distinct loading/model-parallel issues."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Different Qwen video bugs on different model families and different tensors/position-id logic."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:29159",
+            "accept": true,
+            "reason": "Both describe the same tokenizer behavior regression around added tokens/slow-vs-fast inconsistency producing an extra space."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Processor image-size mismatch and dtype mismatches in bf16 inference are unrelated."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "These are both loading-related, but one is FSDP CPU RAM-efficient loading and the other is init_empty_weights argument handling."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Packed-weight unpacking during accelerate loading is a different problem from bf16 dtype mismatches."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:46665",
+            "accept": false,
+            "reason": "Jupyter custom-model initialization and from_pretrained inside Python classes are not the same concrete bug."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "Both are about v5 pipeline removal, but one is runtime error messaging and the other is documentation lag; not the same underlying defect."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Bark voice_preset and CustomTrainer.compute_loss signature mismatch are unrelated."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Backend packaging for TensorFlow and Mistral regex patching are unrelated."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "Both involve loading/precision issues, but one is BLIP2 dtype mismatch and the other is an Apple Silicon allocator TypeError."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "CamemBERT masked-LM prediction regression and generic tokenization changes are too broad to treat as the same bug."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "Unsafe sys.modules access is unrelated to CLIP model structure differences."
+          },
+          {
+            "left": "issue:46175",
+            "right": "issue:46932",
+            "accept": false,
+            "reason": "A huge one-hot allocation issue and quantized embedder dtype casting are different Gemma4 bugs."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Chat-template sampling-rate defaulting and DAC latent/forward mismatch are unrelated audio bugs."
+          },
+          {
+            "left": "issue:46697",
+            "right": "issue:46728",
+            "accept": false,
+            "reason": "Over-padding divisible images and wrong patch counts for odd multiples are similar in theme but not the same defect."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Qwen3VL weight_only loading error and Qwen2.5-VL rope-index scaling are different model-specific issues."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Allocator warmup TypeError and LwDetr float16 AMP crash are unrelated runtime failures."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary stays grounded in the packet and the soft-edge calls are appropriately conservative. The only accepted pair, #28218/#29159, is plausibly the same tokenizer regression from the titles alone, while the rest are reasonably rejected as different bugs or different code paths. No PR-pair mergeability issue applies here."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 21,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:36010",
+        "issue:36331",
+        "issue:38175",
+        "issue:39692",
+        "issue:42175",
+        "issue:42490",
+        "issue:43257",
+        "issue:43262",
+        "issue:43329",
+        "issue:43452",
+        "issue:43526",
+        "issue:43531",
+        "issue:43550",
+        "issue:43638",
+        "issue:43643",
+        "issue:43645",
+        "issue:43653",
+        "issue:43697",
+        "issue:43742",
+        "issue:43746",
+        "issue:43819",
+        "issue:43854",
+        "issue:43906",
+        "issue:43927",
+        "issue:44060",
+        "issue:44368",
+        "issue:44554",
+        "issue:44589",
+        "issue:44610",
+        "issue:44811",
+        "issue:44912",
+        "issue:44936",
+        "issue:44977",
+        "issue:44991",
+        "issue:45020",
+        "issue:45042",
+        "issue:45081",
+        "issue:45084",
+        "issue:45372",
+        "issue:45406",
+        "issue:45538",
+        "issue:45600",
+        "issue:45759",
+        "issue:45859",
+        "issue:46209",
+        "issue:46291",
+        "issue:46326",
+        "issue:47050"
+      ],
+      "soft_pairs": [
+        "issue:43653|issue:44912",
+        "issue:44589|issue:45042",
+        "issue:43742|issue:45020",
+        "issue:45372|issue:46326",
+        "issue:43550|issue:44912",
+        "issue:36010|issue:43645",
+        "issue:36010|issue:36331",
+        "issue:43257|issue:43329",
+        "issue:43906|issue:46209",
+        "issue:43531|issue:45042",
+        "issue:43526|issue:43697",
+        "issue:43452|issue:44991",
+        "issue:42175|issue:46291",
+        "issue:43638|issue:44936",
+        "issue:43819|issue:43927",
+        "issue:45081|issue:45406",
+        "issue:44060|issue:44368",
+        "issue:43329|issue:45600",
+        "issue:45084|issue:45538",
+        "issue:43526|issue:44554",
+        "issue:42490|issue:43257",
+        "issue:43746|issue:44811",
+        "issue:45859|issue:47050",
+        "issue:43854|issue:44977",
+        "issue:43643|issue:45759",
+        "issue:38175|issue:44610",
+        "issue:43262|issue:44811",
+        "issue:39692|issue:44912"
+      ],
+      "prepared_review_unit_hash": "fa3570422ebf3a1e2887da9d472a27036e0933397ed0a60815cb283d4f2cd3b0",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36560,
+        "estimated_input_tokens": 9140,
+        "estimated_eval_tokens": 18536
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly unrelated closed issues spanning loading/import regressions, tokenizer/processor bugs, training failures, quantization, and model-specific correctness problems. No soft edges should be accepted here; the Qwen3 tied-weights pair is only superficially related and differs in model/version and failure mode.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "issue:45020 is the broadest umbrella in the set (recent-version / remote_code loading regressions) and is the most useful cluster anchor, though most other items are still distinct bugs.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45020 is the best representative issue because it is the most general and central to the loading/config-regression subset of this cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43653",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Different problems: BigBirdTokenizer special-token decoding vs MXFP4 quantization fallback."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Unrelated subsystems: float8 storage lookup vs PIL image processor torchvision dependency."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Both involve loading regressions, but one is a model-specific KeyError and the other is a broad remote_code breakage report."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Both involve mistral_common, but the concrete failures differ: missing ReasoningEffort import vs continue_final_message behavior."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Different code paths: torch.compile/SDPA attention bug vs quantization loading fallback."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Import path regression vs custom-model initialization breakage in notebooks; not the same bug."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "GenerationMixin import failure is unrelated to CustomTrainer.compute_loss signature breakage."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Qwen3 MoE loading/deepspeed issue is unrelated to multimodal token counting video-branch bug."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Different model/problem: isolated reproduction of an old issue vs FP8 inference failure on B200."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior is unrelated to PIL backend torchvision requirement."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "Different models and failures: BEiT label reduction vs RTDetrV2 output drift in v5."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Tokenizer/model loading failures, but one is gguf_file handling and the other is a specific tokenizer compatibility regression."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Packaging backend omission is unrelated to kernels v0.15 import-time requirements."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Zero3/non-pretrained Bert index error is not the same as evaluate() failing after train()."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Different model internals: DAC latent/STE mismatch vs DiaConfig save/load custom token IDs."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both are tokenizer/processor loading errors, but they affect different classes and different failure points."
+          },
+          {
+            "left": "issue:44060",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Related only at a high level: Qwen3-Next tied-weights warning vs Qwen3.5 LoRA warning about tie_word_embeddings=False; different models and concrete failure modes."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Undefined video-patch helpers in multimodal token counting is unrelated to removed config references in auto_mappings.py."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Template compilation error vs tokenizer max_length bug; different behavior and code paths."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "Different models and failures: BEiT label reduction vs MPS attention correctness issue."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43257",
+            "accept": false,
+            "reason": "Trailing-slash local path warning is unrelated to Qwen3 MoE weight conversion under accelerate/deepspeed."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "PEFT local checkpoint loading vs Whisper batch_decode skip_special_tokens; different components."
+          },
+          {
+            "left": "issue:45859",
+            "right": "issue:47050",
+            "accept": false,
+            "reason": "CPU offload incompatibility in Qwen3 MoE rotary embedding is unrelated to Canine fp16 ONNX export dtype mismatch."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Different model-specific regressions: GLM-4.7-Flash loading vs Qwen3.5 flash-attention generation."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both are config-related, but one is trust_remote_code field loss and the other is composite multimodal config unwrapping."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "SigLIP2 zero-probability behavior is unrelated to OmDet-Turbo processor image-size mismatch."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Audio sampling-rate defaulting is unrelated to Whisper batch_decode token skipping."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Documentation/model-processor mismatch plus quantization failure is not the same as a separate MXFP4 loading fallback bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The overall summary is cautious enough, but one soft-edge verdict is too strong: accepting issue:44060 \u2194 issue:44368 is not well-grounded from the packet alone. Their titles suggest related themes around tied-weights warnings, but they appear to describe different models and different failure modes, so treating them as the same bugfix overstates the evidence. Please downgrade that pair to conservative false unless you have stronger linkage."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 22,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28282",
+        "issue:29159",
+        "issue:29870",
+        "issue:33357",
+        "issue:34567",
+        "issue:34689",
+        "issue:39379",
+        "issue:40990",
+        "issue:42175",
+        "issue:42915",
+        "issue:43257",
+        "issue:43329",
+        "issue:43425",
+        "issue:43450",
+        "issue:43454",
+        "issue:43525",
+        "issue:43540",
+        "issue:43577",
+        "issue:43650",
+        "issue:43696",
+        "issue:43697",
+        "issue:43742",
+        "issue:43749",
+        "issue:43824",
+        "issue:43906",
+        "issue:43931",
+        "issue:44291",
+        "issue:44387",
+        "issue:44479",
+        "issue:44554",
+        "issue:44568",
+        "issue:44617",
+        "issue:44756",
+        "issue:44821",
+        "issue:44912",
+        "issue:44987",
+        "issue:45042",
+        "issue:45356",
+        "issue:45357",
+        "issue:45459",
+        "issue:45538",
+        "issue:45803",
+        "issue:46291",
+        "issue:46421",
+        "issue:46424",
+        "issue:46701",
+        "issue:47145"
+      ],
+      "soft_pairs": [
+        "issue:43257|issue:43696",
+        "issue:43329|issue:43450",
+        "issue:43329|issue:44479",
+        "issue:42915|issue:43742",
+        "issue:43257|issue:43931",
+        "issue:44387|issue:44617",
+        "issue:44912|issue:47145",
+        "issue:44291|issue:45042",
+        "issue:34567|issue:40990",
+        "issue:42915|issue:44821",
+        "issue:29159|issue:34567",
+        "issue:45459|issue:45803",
+        "issue:46421|issue:46424",
+        "issue:43650|issue:43906",
+        "issue:43525|issue:46291",
+        "issue:28282|issue:29870",
+        "issue:43749|issue:44756",
+        "issue:44987|issue:45538",
+        "issue:28282|issue:34689",
+        "issue:43425|issue:46291",
+        "issue:43540|issue:46701",
+        "issue:43454|issue:43577",
+        "issue:44568|issue:45356",
+        "issue:33357|issue:34567",
+        "issue:39379|issue:44554",
+        "issue:42175|issue:43824",
+        "issue:43697|issue:45357"
+      ],
+      "prepared_review_unit_hash": "390c0c088f0a8e7c6a663647c03409c4f13d615264141f5b4410a9c54e56588c",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 27,
+        "serialized_chars": 35248,
+        "estimated_input_tokens": 8812,
+        "estimated_eval_tokens": 17880
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most items are unrelated one-off bugs; the soft links are mostly superficial. The only plausible duplicate family is the Qwen video-input regression around multimodal token handling.",
+        "confidence": 0.76,
+        "canonical_issue_reason": "Issue 43329 is the clearest root-cause report in the only plausible duplicate family here: the Qwen video-path multimodal token counting bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44479 is the strongest standalone issue: it clearly states a regression, lists the affected model families, and describes a user-visible failure mode.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43257",
+            "right": "issue:43696",
+            "accept": false,
+            "reason": "Both are model-loading failures, but one is Qwen3 MoE weight conversion and the other is GPT-oss OOM; different code paths and fixes."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Both mention video inputs, but one is a missing/undefined helper issue in token counting while the other is a batched-shape bug in video processors."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:44479",
+            "accept": true,
+            "reason": "Both point to the same Qwen video-input regression path; 43329 names the broken video token-counting code, and 44479 reports the resulting regression across Qwen VL variants."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Different failures: FineGrainedFP8Config on Qwen3Moe vs a key error loading MobileLLM-125M."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Both are loading issues, but one is accelerate/deepspeed MoE conversion and the other is a shape mismatch for a different Qwen3-VL model."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "One is an int4 CUDA memory regression, the other is a separate SAM3 video OOM; same symptom, different bug."
+          },
+          {
+            "left": "issue:44912",
+            "right": "issue:47145",
+            "accept": false,
+            "reason": "Different quantization failures: MXFP4 fallback behavior versus FP8 frozen checkpoint weight_scale not being applied."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Init-empty-weights argument handling and PIL backend torchvision requirements are unrelated."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:40990",
+            "accept": false,
+            "reason": "Trainer state token counting and GPT-oss perplexity are unrelated bugs."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config failure and AutoImageProcessor-from-URL loading are unrelated."
+          },
+          {
+            "left": "issue:29159",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Tokenizer slow/fast inconsistency and TrainerState token accounting are different subsystems and failure modes."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Both hide errors, but one is protobuf decode handling in tokenizers and the other is a bare-except in Fuyu tensor conversion."
+          },
+          {
+            "left": "issue:46421",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Both involve cache behavior, but one is lazy init causing recompilation and the other is max_cache_length being honored only for quantized cache."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "A generic data request is not the same as an isolated reproduction of an existing issue."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Llama4Config pad_token_id and pipelines import/kernel version requirements are unrelated."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:29870",
+            "accept": false,
+            "reason": "ImportError for missing PyTorch and Bart generate() decoder_attention_mask assignment are different bugs."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Both mention OOM, but one is FSDP CPU RAM efficient loading and the other is mmap on Strix Halo."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Loading physical-intelligence/fast and CLIPTokenizer model_max_length are unrelated."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Missing PyTorch import versus Llama 3.2 Vision loading regression are different failures."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Torch 2.10 incompatibility and kernels import/version requirements are not the same underlying bug."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:46701",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video processing ValueError and DiffusionGemmaGenerationOutput pipeline failure are unrelated."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "AyaVision lm_head tying bug and BLIP2 dtype propagation bug affect different model-loading paths."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Tokenizer special-token insertion regression and Kimi-K2.5 codec handling regression are distinct tokenizer issues."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "A MacOS CLIP bus error and TrainerState token counting are unrelated."
+          },
+          {
+            "left": "issue:39379",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "Incorrect full_text_row_masked_out_mask handling in mllama and an MPS attention correctness issue are different code paths."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Backend packaging for torch-only install and missing Qwen2_5_VL import are unrelated."
+          },
+          {
+            "left": "issue:43697",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "RTDetrV2 output differences and Qwen3.5 save_pretrained key regression are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded overall and appropriately conservative. The accepted 43329 \u2194 44479 link is well supported by the issue titles, and the remaining soft pairs are reasonably rejected as distinct bugs. Minor note: the best-issue phrasing for 44479 leans a bit on inference about user impact, but it stays within the packet evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 23,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:24643",
+        "issue:27957",
+        "issue:28282",
+        "issue:29127",
+        "issue:29159",
+        "issue:34567",
+        "issue:36331",
+        "issue:38175",
+        "issue:41720",
+        "issue:42175",
+        "issue:42915",
+        "issue:43404",
+        "issue:43452",
+        "issue:43540",
+        "issue:43577",
+        "issue:43582",
+        "issue:43742",
+        "issue:43749",
+        "issue:44062",
+        "issue:44291",
+        "issue:44479",
+        "issue:44488",
+        "issue:44560",
+        "issue:44568",
+        "issue:44792",
+        "issue:44821",
+        "issue:44849",
+        "issue:44857",
+        "issue:44898",
+        "issue:44936",
+        "issue:44964",
+        "issue:44987",
+        "issue:44991",
+        "issue:45081",
+        "issue:45127",
+        "issue:45137",
+        "issue:45381",
+        "issue:45405",
+        "issue:45488",
+        "issue:45538",
+        "issue:45701",
+        "issue:45907",
+        "issue:46512",
+        "issue:46693",
+        "issue:46821",
+        "issue:46832",
+        "issue:46869",
+        "issue:47020"
+      ],
+      "soft_pairs": [
+        "issue:44857|issue:46832",
+        "issue:45137|issue:45405",
+        "issue:43404|issue:45127",
+        "issue:28282|issue:29127",
+        "issue:45081|issue:47020",
+        "issue:43742|issue:44291",
+        "issue:44936|issue:46821",
+        "issue:44479|issue:46693",
+        "issue:42915|issue:45538",
+        "issue:45538|issue:46869",
+        "issue:44062|issue:44991",
+        "issue:24643|issue:29159",
+        "issue:43452|issue:43582",
+        "issue:27957|issue:28282",
+        "issue:42915|issue:43749",
+        "issue:43577|issue:46512",
+        "issue:41720|issue:43540",
+        "issue:44849|issue:45701",
+        "issue:42175|issue:44488",
+        "issue:43540|issue:44560",
+        "issue:43540|issue:44479",
+        "issue:44987|issue:45907",
+        "issue:44291|issue:44821",
+        "issue:44479|issue:45381",
+        "issue:38175|issue:44898",
+        "issue:34567|issue:36331",
+        "issue:44792|issue:44964",
+        "issue:44568|issue:45488",
+        "issue:44062|issue:45081"
+      ],
+      "prepared_review_unit_hash": "b6bb4f4687445670b830a9543edd9a835fdf360f74ba21911efa3c05821783c8",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36411,
+        "estimated_input_tokens": 9103,
+        "estimated_eval_tokens": 18462
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most candidates are distinct bug reports across different model families and code paths. I only see one plausible duplicate-style match: the Qwen video-input regression pair around transformers v5.3.0.",
+        "confidence": 0.79,
+        "canonical_issue_reason": "issue:44479 is the broadest, most canonical description of the Qwen video-input regression, and it can subsume the more specific qwen2.5-vl vision_position_ids report.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44479 is the best anchor because it states the versioned regression clearly and covers multiple affected Qwen VL variants, making it the strongest canonical issue candidate in this set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44857",
+            "right": "issue:46832",
+            "accept": false,
+            "reason": "Both are vision/object-detection related, but they describe different models and different failure modes (AMP crash vs feature-level mismatch)."
+          },
+          {
+            "left": "issue:45137",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO3 deque failure and PEFT version pinning are unrelated."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45127",
+            "accept": false,
+            "reason": "Both mention tied weights, but one is a Mistral3 loading bug and the other is a LoRA merge collapse with extended vocab; not the same concrete bug."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "ImportError about missing PyTorch is unrelated to a LayoutLMv3 error-message clarity issue."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:47020",
+            "accept": false,
+            "reason": "Different tokenizer classes and different code paths; one is a Mistral regex patch crash, the other is an XLMRoberta/Unigram constructor type error."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Both are model-loading failures, but they point to different root causes: a MobileLLM key error vs an init_empty_weights argument mismatch."
+          },
+          {
+            "left": "issue:44936",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Trainer evaluation sequencing and TimesFM window_size handling are unrelated."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "Both are regressions, but one is a video-input bug and the other is a long-context FA2 performance regression; different bugs."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 loading and CLIPTokenizer model_max_length are unrelated."
+          },
+          {
+            "left": "issue:45538",
+            "right": "issue:46869",
+            "accept": false,
+            "reason": "CLIPTokenizer max length behavior is unrelated to YOLO multi-GPU device map failure."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "A duplicated AddedToken keyword error may underlie tokenizer loading failures, but the specific reports and models are not clearly the same bug."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:29159",
+            "accept": false,
+            "reason": "DeepSpeed training weight-shape error and tokenizer slow/fast inconsistency are unrelated."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "gguf_file tokenizer/model loading and AppleSilicon allocator warmup are unrelated."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Flash Attention 2 with XLMRoberta is unrelated to missing PyTorch import handling."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config failure and FSDP CPU RAM efficient loading are different training/config issues."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "Both are dtype-related, but one is BLIP2 model component dtype loading and the other is from_config config-state staleness."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Different Qwen3 problems and different failure points; auto device mapping vs video-input processing."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "A hidden-states bug in Qwen3.5 is unrelated to a tokenizer-version behavior change."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Backend packaging/install issue and a model-loading failure are unrelated."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Both are Qwen video-input failures, but the model families and symptoms differ enough that the same concrete bug is not clear."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Same broad area, but one is Qwen3OmniMoe-specific and the other is a broader Qwen2.5/3/3.5 VL regression; not clearly one bug."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "A model-loading failure and a repo-files revision-forwarding bug are different code paths."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Unexpected _is_hf_initialized handling and loading AutoImageProcessor from URL are unrelated."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45381",
+            "accept": true,
+            "reason": "Both describe the same Qwen2.5-VL video-input regression in v5.3.0; the latter is a narrower symptom of the broader video-input bug."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Unexpected zero probabilities in SigLIP2 and Perceiver image-resolution failure are unrelated."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "TrainerState token counting and custom compute_loss signature mismatch are different trainer API issues."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different multimodal models and different failure modes; one test failure is not the same as a model-loading regression."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45488",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different tokenizer implementations and different semantics."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Qwen3.5 hidden-states bug and a version-wide tokenization change are not the same underlying issue."
+          },
+          {
+            "left": "issue:46512",
+            "right": "issue:45488",
+            "accept": false,
+            "reason": "from_config dtype/config staleness and LlamaTokenizer pre-tokenizer override are unrelated."
+          },
+          {
+            "left": "issue:46869",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "YOLO multi-GPU device mapping and CLIPTokenizer max-length behavior are unrelated."
+          },
+          {
+            "left": "issue:47020",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "XLMRoberta/Unigram constructor type error and Mistral regex patch crash are different tokenizer bugs."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The canonical anchor choice of issue:44479 is supported by its broad, versioned Qwen VL video-input regression title, and the single accepted soft match to issue:45381 is plausible from the titles alone. The other pair verdicts read appropriately cautious rather than overstated."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 24,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:24643",
+        "issue:28282",
+        "issue:29870",
+        "issue:32090",
+        "issue:43257",
+        "issue:43452",
+        "issue:43526",
+        "issue:43531",
+        "issue:43540",
+        "issue:43577",
+        "issue:43582",
+        "issue:43697",
+        "issue:43716",
+        "issue:43742",
+        "issue:43756",
+        "issue:43906",
+        "issue:44488",
+        "issue:44568",
+        "issue:44589",
+        "issue:44625",
+        "issue:44737",
+        "issue:44821",
+        "issue:44877",
+        "issue:44913",
+        "issue:45042",
+        "issue:45072",
+        "issue:45081",
+        "issue:45216",
+        "issue:45325",
+        "issue:45405",
+        "issue:45479",
+        "issue:45588",
+        "issue:45600",
+        "issue:45663",
+        "issue:45672",
+        "issue:45674",
+        "issue:45823",
+        "issue:46121",
+        "issue:46399",
+        "issue:46463",
+        "issue:46468",
+        "issue:46491",
+        "issue:46693",
+        "issue:46701",
+        "issue:46821",
+        "issue:46829",
+        "issue:46846",
+        "issue:46899"
+      ],
+      "soft_pairs": [
+        "issue:45042|issue:45081",
+        "issue:43257|issue:43756",
+        "issue:43582|issue:46821",
+        "issue:43742|issue:44589",
+        "issue:43531|issue:44821",
+        "issue:43756|issue:46399",
+        "issue:43756|issue:45325",
+        "issue:45405|issue:45600",
+        "issue:44737|issue:46463",
+        "issue:43716|issue:46701",
+        "issue:43577|issue:45072",
+        "issue:45823|issue:46463",
+        "issue:43906|issue:46693",
+        "issue:44568|issue:46491",
+        "issue:24643|issue:29870",
+        "issue:45672|issue:45674",
+        "issue:44625|issue:44877",
+        "issue:43540|issue:46693",
+        "issue:43452|issue:44488",
+        "issue:43906|issue:46468",
+        "issue:28282|issue:32090",
+        "issue:43697|issue:46829",
+        "issue:44913|issue:46121",
+        "issue:45663|issue:46846",
+        "issue:44737|issue:45588",
+        "issue:43526|issue:45479",
+        "issue:46463|issue:46899",
+        "issue:43697|issue:45216"
+      ],
+      "prepared_review_unit_hash": "40cac19a3d5903c698352ec888e441defaee313780ade8b4f36e817141f37eef",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36612,
+        "estimated_input_tokens": 9153,
+        "estimated_eval_tokens": 18562
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous: the paired issues cover unrelated bugs across loading, tokenizers, generation, quantization, config, and model-specific regressions. The soft-similarity scores are not enough to treat any pair as the same underlying change or bug, so all candidate edges should be rejected.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "No single canonical issue stands out because the set is not a duplicate cluster; it is a grab bag of distinct bugs with different models, code paths, and failure modes.",
+        "canonical_pr_reason": "No PR is present in the cluster.",
+        "best_issue_reason": "None of the issues is a good global representative for the whole cluster, since the topics are too diverse to normalize to one underlying bug.",
+        "best_pr_reason": "No PR is present in the cluster.",
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45042",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both involve image/tokenizer-related loading behavior, but one is about PIL backend torchvision requirements and the other is a Mistral regex patch crash; different code paths and failures."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43756",
+            "accept": false,
+            "reason": "One is a Qwen3 MoE weight conversion problem with accelerate/deepspeed, the other is about Smollm3 RoPE layer dropping; unrelated model internals."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Different functions and platforms: Apple Silicon caching allocator warmup versus TimesFM window_size handling."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Both are load-time errors, but one is a MobileLLM key lookup issue and the other is a Float8 storage type error; no shared underlying bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window handling is unrelated to AutoImageProcessor URL loading."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Both mention Qwen3.5/Qwen families, but one is RoPE layer composition and the other is kernel registration for apply_rotary_pos_emb; distinct defects."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Smollm3 RoPE layer count and Qwen2.5-VL rope indexing are different model families and different logic."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "PEFT version bump and removed Sam3LiteText auto-mapping references are unrelated maintenance issues."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "XLNet CPU device placement bug and T5 decoder-input validation message bug are unrelated."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:46701",
+            "accept": false,
+            "reason": "Both concern multimodal models, but one is a dtype mismatch in Mistral-3 preprocessing and the other is a DiffusionGemma generation-output pipeline failure."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "BLIP2 dtype propagation and SwitchTransformers/TimmWrapperModel bfloat16 mismatches are different model-specific dtype issues."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Gemma4 device-map placement in forward and T5 decoder validation are unrelated bugs."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "A reproduction of an earlier issue is not the same as a FA2 long-context performance regression."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "Tokenizer special-token behavior for mdeberta-v3 is unrelated to CodeLlama leading-space round-trip regression."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:29870",
+            "accept": false,
+            "reason": "DeepSpeed training weight-shape error and BART generate decoder_attention_mask scoping bug are different failure modes."
+          },
+          {
+            "left": "issue:45672",
+            "right": "issue:45674",
+            "accept": false,
+            "reason": "Both touch quantized inference, but one is unsafe uint8 casting in Gemma4 audio/vision embedders and the other is BitsAndBytes skip-module exclusion logic."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Qwen3.5 num_labels propagation and granite_speech strict config loading are unrelated config issues."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video-input processing and FA2 long-context performance regression are distinct areas and symptoms."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "GGUF file handling for AutoTokenizer/AutoModel loading and sleng-bert load failure may both be loading-related, but the model-specific causes are different."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "An isolated reproduction of an existing issue is not a distinct duplicate target for an official Nemotron example breakage."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Missing PyTorch import and Trainer _gpu_broadcast_one NoneType error are completely different startup/runtime failures."
+          },
+          {
+            "left": "issue:43697",
+            "right": "issue:46829",
+            "accept": false,
+            "reason": "RTDetrV2 output divergence in v5 and VideoPrism hidden_states mislabeling are unrelated model-output bugs."
+          },
+          {
+            "left": "issue:44913",
+            "right": "issue:46121",
+            "accept": false,
+            "reason": "GPTNeoX rotary_pct reload loss and JSON-loaded rope-validation list type error are separate config serialization problems."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:46846",
+            "accept": false,
+            "reason": "Gemma4 FSDP2 shared_kv_states rebuilding and Qwen3.5 tensor-parallel plan omissions are different distributed-training bugs."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "XLNet CPU allocation and flash_attention sink-less-model crash have no shared code path beyond both being runtime bugs."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "BeitImageProcessorFast reduce_labels truncation and sequence-classification zero-loss behavior are unrelated preprocessing vs loss-calculation issues."
+          },
+          {
+            "left": "issue:46463",
+            "right": "issue:46899",
+            "accept": false,
+            "reason": "T5 decoder input validation messaging and Gemma4 multimodal 4-bit dtype casting are unrelated."
+          },
+          {
+            "left": "issue:43697",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "RTDetrV2 output mismatch and Qwen3.5 save_pretrained regression are different subsystems and bugs."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 25,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:29159",
+        "issue:33357",
+        "issue:38175",
+        "issue:41762",
+        "issue:43295",
+        "issue:43452",
+        "issue:43525",
+        "issue:43531",
+        "issue:43643",
+        "issue:43646",
+        "issue:43653",
+        "issue:43688",
+        "issue:43720",
+        "issue:43742",
+        "issue:43749",
+        "issue:43761",
+        "issue:43819",
+        "issue:43994",
+        "issue:44361",
+        "issue:44442",
+        "issue:44466",
+        "issue:44493",
+        "issue:44568",
+        "issue:44811",
+        "issue:44843",
+        "issue:44857",
+        "issue:44898",
+        "issue:44936",
+        "issue:45005",
+        "issue:45200",
+        "issue:45356",
+        "issue:45357",
+        "issue:45372",
+        "issue:45375",
+        "issue:45542",
+        "issue:45701",
+        "issue:45704",
+        "issue:45920",
+        "issue:46050",
+        "issue:46170",
+        "issue:46178",
+        "issue:46463",
+        "issue:46701",
+        "issue:46829",
+        "issue:46832",
+        "issue:46932",
+        "issue:47000"
+      ],
+      "soft_pairs": [
+        "issue:45357|issue:45920",
+        "issue:44466|issue:45005",
+        "issue:43720|issue:44898",
+        "issue:41762|issue:44936",
+        "issue:45200|issue:46463",
+        "issue:44493|issue:46050",
+        "issue:43643|issue:45375",
+        "issue:43688|issue:43749",
+        "issue:38175|issue:43295",
+        "issue:44361|issue:44442",
+        "issue:43653|issue:44361",
+        "issue:43452|issue:44843",
+        "issue:43819|issue:44811",
+        "issue:43994|issue:44898",
+        "issue:43531|issue:43742",
+        "issue:44857|issue:47000",
+        "issue:45372|issue:46170",
+        "issue:43761|issue:46829",
+        "issue:29159|issue:33357",
+        "issue:43525|issue:46701",
+        "issue:45542|issue:46178",
+        "issue:46832|issue:46932",
+        "issue:43688|issue:45704",
+        "issue:43295|issue:45356",
+        "issue:43688|issue:45701",
+        "issue:43653|issue:44568",
+        "issue:43646|issue:43749"
+      ],
+      "prepared_review_unit_hash": "33569943ba7444b8a88e64a1f2610ffa5b5465f91bac0804b732bdf40ebd378f",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 27,
+        "serialized_chars": 35558,
+        "estimated_input_tokens": 8890,
+        "estimated_eval_tokens": 18036
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a set of unrelated transformer/tokenizer/model-loading regressions that were likely grouped by broad lexical similarity. Only one pair looks like the same concrete bug: tied-weights serialization in v5. No pull requests are present.",
+        "confidence": 0.47,
+        "canonical_issue_reason": "issue:45920 is the broadest representative tokenizer-regression report, affecting multiple model families and matching the cluster\u2019s dominant tokenizer theme.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45920 is the best single anchor for this cluster because it is broad, regression-focused, and central to the tokenizer-related subset, though it does not cover the whole cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45357",
+            "right": "issue:45920",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but one is Kimi-K2.5 codec/warning handling and the other is wrong token IDs across multiple models; not the same bug."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45005",
+            "accept": true,
+            "reason": "Both report v5 tied-weights serialization problems, including device-dependent lm_head.weight behavior; these look like the same underlying bug."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Different model paths and failures: BitNet packed-weight loading vs Perceiver interpolation/classification."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Gemma3 ZeRO-3 loading IndexError and trainer.evaluate() failure after train() are unrelated issues."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Gemma 4 token-type defaults and T5 decoder input error messaging are separate problems."
+          },
+          {
+            "left": "issue:44493",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Unexpected position-id keys and sinq quantized save/load breakage are different code paths."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both concern config fields, but one is AutoConfig/trust_remote_code missing fields and the other is a specific model config field being dropped; not the same bug."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Aux-loss normalization and FSDP CPU RAM-efficient loading are unrelated."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "SigLIP2 zero-probability output and processor.tokenizer/image-passing regression are different failures."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "Different tokenizer classes and symptoms: MLuke AttributeError vs FastSpeech2ConformerTokenizer load failure."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "BigBird special-token decode bug and MLuke task AttributeError are unrelated."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44843",
+            "accept": false,
+            "reason": "gguf_file support breakage and unconditional model_info() offline breakage are separate AutoTokenizer issues."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "DAC latent reconstruction mismatch and Whisper skip_special_tokens handling are unrelated."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Both are vision-model complaints, but the concrete bugs differ: SigLIP2 nonsense outputs vs Perceiver resolution/interpolation failure."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window and MobileLLM-125M key error are unrelated."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "Both are detection-loss crashes, but they involve different models and causes; not mergeable as one fix."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Gemma4 processor import failure and Windows ROCm distributed dependency errors are unrelated."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:46829",
+            "accept": false,
+            "reason": "Both mention hidden_states, but one omits them and the other misassigns them; different model-specific bugs."
+          },
+          {
+            "left": "issue:29159",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Slow/fast tokenizer inconsistency and a MacOS bus error in CLIP are unrelated."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:46701",
+            "accept": false,
+            "reason": "Missing pad_token_id in Llama4Config and DiffusionGemma pipeline failure are unrelated."
+          },
+          {
+            "left": "issue:45542",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "Both involve TF/backend registration, but one is an undefined backend error from tensorboard-only installs and the other is a missing BACKENDS_MAPPING entry for tensorflow_text; not the same bug."
+          },
+          {
+            "left": "issue:46832",
+            "right": "issue:46932",
+            "accept": false,
+            "reason": "RT-DETR feature-level crash and Gemma4 quantized embedder dtype casting are unrelated."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Aux-loss normalization and T5 apex.FusedRMSNorm memory leak are unrelated."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they concern different model families and different root causes."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Aux-loss normalization does not match a general tokenization-version regression."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Different special-token issues on different tokenizers; not the same underlying change."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Custom model initialization regression and FSDP loading breakage are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative about the soft edges. The one accepted pair (44466/45005) is reasonably supported as the same tied-weights serialization bug, while the other pair judgments avoid overclaiming. No PRs are present, so there is nothing to merge-check there."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 26,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:32090",
+        "issue:33666",
+        "issue:36010",
+        "issue:36331",
+        "issue:41720",
+        "issue:42898",
+        "issue:43262",
+        "issue:43278",
+        "issue:43295",
+        "issue:43452",
+        "issue:43504",
+        "issue:43550",
+        "issue:43577",
+        "issue:43606",
+        "issue:43611",
+        "issue:43653",
+        "issue:43688",
+        "issue:43720",
+        "issue:43756",
+        "issue:43761",
+        "issue:43782",
+        "issue:43927",
+        "issue:44464",
+        "issue:44466",
+        "issue:44479",
+        "issue:44484",
+        "issue:44514",
+        "issue:44554",
+        "issue:44625",
+        "issue:44704",
+        "issue:44734",
+        "issue:44743",
+        "issue:44821",
+        "issue:44843",
+        "issue:44977",
+        "issue:44991",
+        "issue:45290",
+        "issue:45440",
+        "issue:45491",
+        "issue:46002",
+        "issue:46050",
+        "issue:46402",
+        "issue:46491",
+        "issue:46612",
+        "issue:46693",
+        "issue:46821",
+        "issue:46846",
+        "issue:47013"
+      ],
+      "soft_pairs": [
+        "issue:43756|issue:46402",
+        "issue:44464|issue:44743",
+        "issue:43452|issue:43577",
+        "issue:43653|issue:43927",
+        "issue:44821|issue:46821",
+        "issue:42898|issue:46491",
+        "issue:44843|issue:44991",
+        "issue:44479|issue:44625",
+        "issue:43761|issue:47013",
+        "issue:33666|issue:36010",
+        "issue:32090|issue:33666",
+        "issue:44484|issue:44704",
+        "issue:44466|issue:46050",
+        "issue:41720|issue:44464",
+        "issue:43688|issue:45440",
+        "issue:44977|issue:46693",
+        "issue:44464|issue:44977",
+        "issue:44734|issue:46612",
+        "issue:43504|issue:43550",
+        "issue:33666|issue:36331",
+        "issue:43262|issue:43295",
+        "issue:43278|issue:43688",
+        "issue:43504|issue:43720",
+        "issue:43782|issue:46693",
+        "issue:45491|issue:46002",
+        "issue:43720|issue:43927",
+        "issue:43577|issue:44554",
+        "issue:43611|issue:46846",
+        "issue:43606|issue:43927",
+        "issue:44514|issue:45290"
+      ],
+      "prepared_review_unit_hash": "7a1822442a029be06151f9a6ff0f89c5fe3d9930c07e5aac60d76f6a72a6468e",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 37020,
+        "estimated_input_tokens": 9255,
+        "estimated_eval_tokens": 18766
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is highly heterogeneous: the soft pairs mostly share only broad subsystem words (generation, tokenizer, save/load, attention, trainer, MoE) but not the same concrete bug or code path. I would not merge any of the suggested pairs.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43756",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Different models and failures: Smollm3 RoPE-layer omission vs Sam3Model text-encoder weights not loading."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Both involve generation state, but one is chunked generation with compiled forward and the other is recurrent-state reset in modular_qwen3_5; not the same bug."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "GGUF loading/tokenizer path issue vs BLIP2 dtype propagation issue; different code paths and symptoms."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Both mention token IDs/tokens, but one is BigBird special-token registration and the other is DiaConfig save/load of custom IDs."
+          },
+          {
+            "left": "issue:44821",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Unrelated failures: loading AutoImageProcessor from URL vs TimesFM window_size AttributeError."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "One is a v5 behavior change discussion, the other is a CodeLlama round-trip regression; not the same defect."
+          },
+          {
+            "left": "issue:44843",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Offline model_info() call in mistral regex patch vs tokenizer load failure for a specific model; different root causes."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44625",
+            "accept": false,
+            "reason": "Video-input regression for Qwen VL models vs num_labels propagation in Qwen3.5 config."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:47013",
+            "accept": false,
+            "reason": "CLIPVision hidden_states omission and xLSTM train-mode unpacking crash are unrelated."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Multi-GPU training issue vs GenerationMixin import error; no shared code path."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "Trainer broadcast TypeError and Qwen2-VL multi-GPU training are different problems."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "save_pretrained shard-size question vs AutoProcessor kwargs not forwarded to cached_file; unrelated."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Both concern save/load, but one is device-dependent lm_head serialization and the other is sinq quantization persistence."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44464",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping cudaErrorAssert is unrelated to chunked generation with compiled forward."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Aux-loss normalization bug in OLMoE/GPT OSS vs DeepSeekV3MoE implementation divergence; different defects."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "Flash-attention correctness regression vs FA2 long-context performance regression; not the same underlying bug."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Compiled-forward chunked generation inconsistency is a different issue from Qwen3.5 flash-attention generation failure."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:46612",
+            "accept": false,
+            "reason": "Both involve cache/generation, but one is KV-cache continuation indexing in serve and the other is cache reorder missing for several models."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "BEiT legacy-field loading bug vs Bamba torch.compile+SDPA failure; unrelated."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Multi-GPU training issue vs CustomTrainer.compute_loss signature break; different failure modes."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate default bug vs processor.tokenizer/image-passing regression; not the same bug."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43688",
+            "accept": false,
+            "reason": "Training/evaluate dtype mismatch vs auxiliary-loss normalization; no shared code path."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Legacy field handling in model loading vs packed-weights unpacking during accelerate loading."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "Qwen3VL weight_only load error is unrelated to the FA2 long-context performance regression."
+          },
+          {
+            "left": "issue:45491",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Both mention NaNs, but one is Gemma3 sliding-window attention on GPU and the other is classification-head NaNs after checkout/init."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Packed-weight loading bug vs config custom-token-ID persistence bug; different artifacts and root causes."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "BLIP2 dtype propagation and MPS attention correctness are unrelated."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:46846",
+            "accept": false,
+            "reason": "base_model_prefix loading regression vs tensor-parallel plan missing linear_attn entries; different problems."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "CPU offload device mismatch vs DiaConfig custom token IDs; no overlap beyond generic loading."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both are apply_chat_template crashes, but the triggering inputs and failure paths differ: batched padding=False vs assistant tool-call messages with no content."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 27,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28218",
+        "issue:33357",
+        "issue:33666",
+        "issue:39692",
+        "issue:42796",
+        "issue:43278",
+        "issue:43295",
+        "issue:43452",
+        "issue:43493",
+        "issue:43504",
+        "issue:43525",
+        "issue:43572",
+        "issue:43575",
+        "issue:43643",
+        "issue:43646",
+        "issue:43701",
+        "issue:43720",
+        "issue:43825",
+        "issue:43844",
+        "issue:43927",
+        "issue:43994",
+        "issue:44291",
+        "issue:44361",
+        "issue:44451",
+        "issue:44466",
+        "issue:44488",
+        "issue:44554",
+        "issue:44561",
+        "issue:44568",
+        "issue:44589",
+        "issue:44610",
+        "issue:44625",
+        "issue:44792",
+        "issue:44843",
+        "issue:44849",
+        "issue:44964",
+        "issue:44977",
+        "issue:45405",
+        "issue:45520",
+        "issue:45584",
+        "issue:45812",
+        "issue:45823",
+        "issue:46050",
+        "issue:46192",
+        "issue:46209",
+        "issue:46365",
+        "issue:46489",
+        "issue:46710"
+      ],
+      "soft_pairs": [
+        "issue:43504|issue:44610",
+        "issue:43525|issue:43572",
+        "issue:43643|issue:44561",
+        "issue:43720|issue:44361",
+        "issue:44964|issue:45405",
+        "issue:44625|issue:45584",
+        "issue:43493|issue:43825",
+        "issue:44977|issue:46209",
+        "issue:44554|issue:46365",
+        "issue:28218|issue:33357",
+        "issue:43452|issue:44451",
+        "issue:45823|issue:46050",
+        "issue:39692|issue:43295",
+        "issue:43994|issue:46192",
+        "issue:42796|issue:44977",
+        "issue:43493|issue:43844",
+        "issue:43278|issue:43844",
+        "issue:43646|issue:44466",
+        "issue:43278|issue:43701",
+        "issue:46209|issue:46710",
+        "issue:44568|issue:46489",
+        "issue:33357|issue:33666",
+        "issue:43927|issue:44361",
+        "issue:43575|issue:45405",
+        "issue:44792|issue:45405",
+        "issue:44568|issue:45812",
+        "issue:44849|issue:46209",
+        "issue:44488|issue:44843",
+        "issue:43572|issue:44291",
+        "issue:44589|issue:45520"
+      ],
+      "prepared_review_unit_hash": "266421d04084fe0e17aa51516725a41519faf20f74f20e6509f6c58830eed585",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36762,
+        "estimated_input_tokens": 9191,
+        "estimated_eval_tokens": 18638
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This looks like a heterogeneous grab-bag of unrelated regressions, mostly tokenizer/model-loading failures. None of the soft pairs are true duplicates.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No single canonical issue: the soft pairs span different models, symptoms, and code paths.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45812 is the closest representative because it is a broad tokenizer regression with multiple inbound references, but the overall set is too heterogeneous to anchor on it.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43504",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Different model families and different failures: legacy field load error vs. processor resolution/size mismatch."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Both are missing-padding-field regressions, but they affect different config classes and attribute names; not the same concrete bug."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "Both involve trust_remote_code breakage, but one is missing config fields and the other is removal of a utility symbol; different root causes."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Unrelated areas: packed-weight loading in BitNet vs. a tokenizer AttributeError in MLuke."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Model-loading failure vs. dependency-version bump problem; no shared underlying bug."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Different subsystems: Qwen3.5 label propagation vs. Whisper empty-transcription generation after token alignment."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "SigLIP2 implementation discrepancy is unrelated to a pipeline error-message bug about translation support."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Both are runtime failures, but one is Qwen3.5 flash-attention generation and the other is FP8/B200 inference; different code paths."
+          },
+          {
+            "left": "issue:44554",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Completely different problems: MPS attention correctness vs. importing Trainer being blocked by datasets."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Tokenizer spacing bug and a MacOS bus error with CLIP are unrelated."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Both are load-time failures, but they concern different model types and different failure mechanisms."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Device-map forward mismatch vs. quantized save/load breakage; not the same bug."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Documentation example errors are not the same issue as a regression in processor.tokenizer access and image handling."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Nonsensical model output and return_tensors being ignored are unrelated tokenizer/API behaviors."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Tokenizer BPE-marker handling and Qwen3.5 flash-attention generation are distinct issues."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "SigLIP2 HF/JAX mismatch is unrelated to a gradient explosion in ZeRO-3 training."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "Embedding dtype drift in evaluate is unrelated to the ZeRO-3 gradient regression."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Custom model initialization breakage and lm_head serialization inconsistency are different bugs."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Different problems: BF16/FP32 dtype mismatch vs. resume_from_checkpoint key mismatch."
+          },
+          {
+            "left": "issue:46209",
+            "right": "issue:46710",
+            "accept": false,
+            "reason": "Hardware-specific FP8 inference failure vs. accuracy regression after upgrade; not the same code-path problem."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:46489",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different models and different tokenizer behaviors; too weak to merge."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "MacOS bus error in CLIP and Qwen2-VL multi-GPU training are unrelated."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "DiaConfig save/load token-ID loss is unrelated to MLukeTokenizer task AttributeErrors."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "OOM on loading a specific model is unrelated to a PEFT version-pin release problem."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "A failed janus image-generation test is not the same issue as a dependency-version bump."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are tokenizer-related, but they are different regressions affecting different models and different tokenization paths."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Qwen3.5 hidden-state output bug and FP8/B200 inference failure are unrelated."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44843",
+            "accept": false,
+            "reason": "Model-loading failure for a specific BERT model is unrelated to an offline-mode tokenizer patch bug."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Missing pad_token_idx in one config and an init_empty_weights TypeError are different breakages."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Different low-level import/storage errors affecting different optional backends; not the same bug."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 28,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28218",
+        "issue:29159",
+        "issue:29942",
+        "issue:30333",
+        "issue:32090",
+        "issue:34567",
+        "issue:42490",
+        "issue:43257",
+        "issue:43299",
+        "issue:43334",
+        "issue:43504",
+        "issue:43572",
+        "issue:43606",
+        "issue:43653",
+        "issue:43696",
+        "issue:43824",
+        "issue:43844",
+        "issue:43873",
+        "issue:43937",
+        "issue:43994",
+        "issue:44112",
+        "issue:44464",
+        "issue:44493",
+        "issue:44589",
+        "issue:44849",
+        "issue:44869",
+        "issue:44964",
+        "issue:45072",
+        "issue:45084",
+        "issue:45200",
+        "issue:45229",
+        "issue:45245",
+        "issue:45357",
+        "issue:45520",
+        "issue:45584",
+        "issue:45588",
+        "issue:45759",
+        "issue:45823",
+        "issue:46050",
+        "issue:46285",
+        "issue:46399",
+        "issue:46701",
+        "issue:46710",
+        "issue:46752",
+        "issue:46821",
+        "issue:46832",
+        "issue:46846",
+        "issue:47145"
+      ],
+      "soft_pairs": [
+        "issue:46821|issue:46832",
+        "issue:43334|issue:45759",
+        "issue:43824|issue:44589",
+        "issue:44849|issue:45357",
+        "issue:43504|issue:43606",
+        "issue:45245|issue:46710",
+        "issue:43334|issue:43572",
+        "issue:44493|issue:46285",
+        "issue:43257|issue:43299",
+        "issue:43504|issue:43653",
+        "issue:43606|issue:44112",
+        "issue:45072|issue:46832",
+        "issue:46399|issue:46846",
+        "issue:44849|issue:47145",
+        "issue:44464|issue:45584",
+        "issue:43873|issue:46050",
+        "issue:43994|issue:44493",
+        "issue:43844|issue:44964",
+        "issue:29159|issue:32090",
+        "issue:44869|issue:46752",
+        "issue:45200|issue:45229",
+        "issue:45588|issue:46050",
+        "issue:43937|issue:46701",
+        "issue:45229|issue:45823",
+        "issue:42490|issue:43696",
+        "issue:45084|issue:45520",
+        "issue:29942|issue:30333",
+        "issue:28218|issue:34567"
+      ],
+      "prepared_review_unit_hash": "15547b9ce092115eb86780b6b64edacc854b21c2ff4f338950658b888a344a77",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36153,
+        "estimated_input_tokens": 9039,
+        "estimated_eval_tokens": 18334
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous and does not contain a convincing duplicate set; the items span unrelated tokenizer, model-loading, quantization/offload, generation, and CI regressions across different model families.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "No single canonical issue fits: the cluster is a grab-bag of distinct bugs with no exact duplicate pair or shared concrete code path.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "If forced to pick a representative, issue 43873 is the broadest/high-activity report, but it is still too generic to serve as a true canonical for the cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46821",
+            "right": "issue:46832",
+            "accept": false,
+            "reason": "Different model/component and failure mode: TimesFM window_size attribute error vs RT-DETR feature-level forward mismatch."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both involve Qwen multimodal config loading, but the concrete bugs differ: missing pad_token_id vs not unwrapping text_config in from_config."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Unrelated: missing Qwen2.5-VL import symbol vs missing Float8 storage type."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different Qwen3.5 bugs: output_hidden_states handling vs save_pretrained writing incorrect visual encoder keys."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43606",
+            "accept": false,
+            "reason": "Different subsystems and symptoms: Beit legacy field loading vs Bark CPU-offload device mismatch."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:46710",
+            "accept": false,
+            "reason": "No shared bug: category-count runtime limit vs accuracy regression after upgrade."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Both are missing config attributes, but they affect different models and fields (pad_token_id vs pad_token_idx) with different load paths."
+          },
+          {
+            "left": "issue:44493",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "Both touch model structure changes, but the reported bugs are different and not the same code-path failure."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43299",
+            "accept": false,
+            "reason": "Both mention Qwen3 MoE loading, but one is accelerate/deepspeed weight conversion and the other is a transformers dev load regression; too different to merge."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Different areas: pretrained model preset legacy-field loading vs tokenizer special-token registration/decoding."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Different failures: CPU offload device mismatch in Bark vs a stale GraniteSpeech CI test."
+          },
+          {
+            "left": "issue:45072",
+            "right": "issue:46832",
+            "accept": false,
+            "reason": "Different models and bugs: dtype mismatch in inference vs RT-DETR forward feature-level logic."
+          },
+          {
+            "left": "issue:46399",
+            "right": "issue:46846",
+            "accept": false,
+            "reason": "Both are Qwen3.5-related, but one is kernel registration/decorator failure and the other is tensor-parallel plan coverage; not the same bug."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:47145",
+            "accept": false,
+            "reason": "Different problems: Qwen3.5 hidden-state output bug vs FP8 frozen checkpoint loading without weight_scale."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Both are generation issues, but they affect different pipelines and failure modes; no shared underlying bug."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Both are quantization/offload-adjacent, but one is generic offloading behavior and the other is sinq save/load corruption; not the same concrete issue."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Different model families and symptoms: SigLIP2 auto-model/pipeline output vs positional-id key warnings."
+          },
+          {
+            "left": "issue:43844",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different regressions: ZeRO-3 gradient growth vs Phi-4 multimodal model loading."
+          },
+          {
+            "left": "issue:29159",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Tokenizer inconsistency vs Trainer/_gpu_broadcast_one NoneType error are unrelated."
+          },
+          {
+            "left": "issue:44869",
+            "right": "issue:46752",
+            "accept": false,
+            "reason": "Both are edge-case crashes, but one is Whisper timestamp decoding and the other is chat template handling; unrelated code paths."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Different Gemma 4 issues: missing default mm_token_type_ids vs multi-GPU inference OOM."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Different bugs: flash_attention integration crashing on s_aux=None vs quantized sinq save/load corruption."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:46701",
+            "accept": false,
+            "reason": "Different generation/pipeline failures in different models; no evidence of the same underlying bug."
+          },
+          {
+            "left": "issue:45229",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Same model family but different failures: OOM during inference vs device mismatch with device_map=\"auto\"."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43696",
+            "accept": false,
+            "reason": "Unrelated: local-path import warning vs CUDA OOM for GPT-oss-20b."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Different import/runtime failures: non-template-node compilation vs flash_attn import_utils KeyError on Python 3.13."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:30333",
+            "accept": false,
+            "reason": "Failing Flash Attention 2 tests and MLFlow reporting after job failure are unrelated."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Tokenizer added-space behavior and TrainerState token counting are distinct bugs with different subsystems."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 29,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:15354",
+        "issue:29870",
+        "issue:29942",
+        "issue:37428",
+        "issue:38175",
+        "issue:39692",
+        "issue:41720",
+        "issue:42490",
+        "issue:42915",
+        "issue:43065",
+        "issue:43295",
+        "issue:43450",
+        "issue:43493",
+        "issue:43540",
+        "issue:43550",
+        "issue:43572",
+        "issue:43575",
+        "issue:43701",
+        "issue:43824",
+        "issue:43844",
+        "issue:43873",
+        "issue:43957",
+        "issue:44186",
+        "issue:44423",
+        "issue:44521",
+        "issue:44554",
+        "issue:44610",
+        "issue:44734",
+        "issue:44977",
+        "issue:44987",
+        "issue:45072",
+        "issue:45198",
+        "issue:45200",
+        "issue:45290",
+        "issue:45406",
+        "issue:45446",
+        "issue:45520",
+        "issue:45588",
+        "issue:45759",
+        "issue:45823",
+        "issue:46121",
+        "issue:46178",
+        "issue:46192",
+        "issue:46369",
+        "issue:46701",
+        "issue:46710",
+        "issue:46821"
+      ],
+      "soft_pairs": [
+        "issue:43065|issue:46369",
+        "issue:15354|issue:38175",
+        "issue:44554|issue:45446",
+        "issue:43493|issue:43701",
+        "issue:41720|issue:46710",
+        "issue:44977|issue:46701",
+        "issue:44423|issue:44734",
+        "issue:37428|issue:43957",
+        "issue:43295|issue:43450",
+        "issue:44977|issue:46710",
+        "issue:45072|issue:45198",
+        "issue:45406|issue:46121",
+        "issue:45588|issue:45759",
+        "issue:43540|issue:45200",
+        "issue:43295|issue:43873",
+        "issue:43572|issue:46821",
+        "issue:43295|issue:44610",
+        "issue:44186|issue:46192",
+        "issue:45823|issue:46192",
+        "issue:43550|issue:43844",
+        "issue:43450|issue:43575",
+        "issue:29870|issue:29942",
+        "issue:43572|issue:44987",
+        "issue:45520|issue:46178",
+        "issue:37428|issue:43824",
+        "issue:45200|issue:45823",
+        "issue:42490|issue:43493",
+        "issue:39692|issue:46192",
+        "issue:43295|issue:43493",
+        "issue:44521|issue:45290",
+        "issue:42915|issue:43824"
+      ],
+      "prepared_review_unit_hash": "6e35e1ec62a071dd791524c8877130822e572777ca950e2f896a10b39af67774",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36578,
+        "estimated_input_tokens": 9145,
+        "estimated_eval_tokens": 18546
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items do not form a single duplicate cluster. They span many unrelated bugs across model loading, processors/tokenizers, serving, flash-attention, quantization, and training, so the soft links are mostly superficial keyword overlaps.",
+        "confidence": 0.96,
+        "canonical_issue_reason": "No single canonical issue is justified here; the candidate issues cover unrelated root causes and code paths, so picking one as the representative duplicate target would be misleading.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No best issue in a duplicate-triage sense: none of these issues is a plausible umbrella for the others, and the similarities are too shallow to merge the cluster around one report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43065",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Same broad SAM3 family, but one is about a dummy Conv2d and the other about memory overflow; different failure modes and likely different fixes."
+          },
+          {
+            "left": "issue:15354",
+            "right": "issue:38175",
+            "accept": false,
+            "reason": "Different models and unrelated symptoms: TorchScript GeneratorExp export issue vs SigLIP2 zero probabilities."
+          },
+          {
+            "left": "issue:44554",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Both touch flex attention, but one is an MPS attention correctness bug and the other is a PyTorch version-check/import problem."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "SigLIP2 implementation discrepancy and resume_from_checkpoint key mismatch are unrelated bugs."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:46710",
+            "accept": false,
+            "reason": "Different models and different problems: CUDA assert with device mapping vs accuracy regression after upgrade."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:46701",
+            "accept": false,
+            "reason": "Both involve generation, but they affect different models and code paths; no shared underlying bug is evident."
+          },
+          {
+            "left": "issue:44423",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Both are serve crashes, but one is a multimodal continuous-batching attribute error and the other is KV-cache continuation indexing."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "An import error for a flash-attention helper and a meta-device loading regression are unrelated."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Both involve processors, but one is a processor.tokenizer regression and the other is an incorrect batched video shape."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:46710",
+            "accept": false,
+            "reason": "Different issue class and model: flash-attention generation failure vs post-upgrade accuracy regression."
+          },
+          {
+            "left": "issue:45072",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Dtype mismatches in inference and Wav2Vec2 save/tokenization failures are unrelated."
+          },
+          {
+            "left": "issue:45406",
+            "right": "issue:46121",
+            "accept": false,
+            "reason": "Gemma4Processor _tokenizer crash and rope-parameter JSON list TypeError do not share the same code path."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Flash-attention integration crash with s_aux=None is unrelated to AutoModelForCausalLM text_config unwrapping."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Both are multimodal-related, but one is a Qwen3OmniMoe video-processing error and the other is a Gemma4 fine-tuning token-type defaulting issue."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Processor/tokenizer regression and quantization offloading behavior are different bugs."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "A missing config field after update and a TimesFM window_size AttributeError are unrelated."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Both involve processors, but the reported failures are different: processor.tokenizer access vs wrong image size for OmDet-Turbo."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Tokenizer crashes on NER/batched padding versus return_tensors being ignored are distinct API bugs."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Gemma4 device-map mismatch and return_tensors handling are unrelated."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "torch.compile with SDPA on Bamba and gradient growth with ZeRO-3 are different training/runtime problems."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Incorrect batched video shape and tensor-parallel OOM are unrelated."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:29942",
+            "accept": false,
+            "reason": "Bart generate() UnboundLocalError and Flash Attention 2 test failures do not indicate the same underlying bug."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "StableLmConfig pad_token_idx regression and loading physical-intelligence/fast failing are not the same issue."
+          },
+          {
+            "left": "issue:45520",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "Both are import/backend mapping issues, but they involve different missing entries and different failure contexts."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Different import errors for different symbols/modules; no shared root cause is shown."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Gemma4 token-type defaults and Gemma4 PLE device mismatch are separate forward-path bugs."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "A local-path trailing-slash import warning and a SigLIP2 implementation discrepancy are unrelated."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "SigLIP2 documentation errors and return_tensors being ignored are different problems."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Processor.tokenizer regression and SigLIP2 HF-vs-JAX discrepancy are not the same bug."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both are apply_chat_template issues, but one is zero assistant masks for multimodal inputs and the other is a crash with tool-call assistant messages."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 failure and Qwen2.5-VL import error are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 30,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28218",
+        "issue:35141",
+        "issue:41720",
+        "issue:43262",
+        "issue:43278",
+        "issue:43334",
+        "issue:43335",
+        "issue:43388",
+        "issue:43450",
+        "issue:43452",
+        "issue:43525",
+        "issue:43526",
+        "issue:43550",
+        "issue:43606",
+        "issue:43653",
+        "issue:43819",
+        "issue:43824",
+        "issue:43827",
+        "issue:43856",
+        "issue:43901",
+        "issue:43950",
+        "issue:43975",
+        "issue:44112",
+        "issue:44442",
+        "issue:44466",
+        "issue:44561",
+        "issue:44610",
+        "issue:44737",
+        "issue:44792",
+        "issue:44857",
+        "issue:44898",
+        "issue:45083",
+        "issue:45127",
+        "issue:45198",
+        "issue:45200",
+        "issue:45325",
+        "issue:45381",
+        "issue:45412",
+        "issue:45563",
+        "issue:45593",
+        "issue:46192",
+        "issue:46209",
+        "issue:46234",
+        "issue:46285",
+        "issue:46512",
+        "issue:46894",
+        "issue:46899",
+        "issue:46932"
+      ],
+      "soft_pairs": [
+        "issue:45381|issue:46209",
+        "issue:28218|issue:35141",
+        "issue:44466|issue:45127",
+        "issue:43278|issue:43856",
+        "issue:44737|issue:46894",
+        "issue:43452|issue:46512",
+        "issue:43653|issue:44112",
+        "issue:43262|issue:43450",
+        "issue:43824|issue:43975",
+        "issue:43950|issue:46285",
+        "issue:43334|issue:43525",
+        "issue:44442|issue:45198",
+        "issue:43526|issue:46192",
+        "issue:45200|issue:46899",
+        "issue:45200|issue:46932",
+        "issue:44898|issue:45412",
+        "issue:44792|issue:45563",
+        "issue:43819|issue:44112",
+        "issue:45200|issue:45325",
+        "issue:43335|issue:43550",
+        "issue:41720|issue:46209",
+        "issue:44561|issue:46234",
+        "issue:44610|issue:44857",
+        "issue:43827|issue:43901",
+        "issue:43819|issue:45083",
+        "issue:44857|issue:45593",
+        "issue:43388|issue:43526",
+        "issue:43606|issue:45083"
+      ],
+      "prepared_review_unit_hash": "da96d2dc974d004187949b25c995d68b84e94f3da598b330008162b49271ee58",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36574,
+        "estimated_input_tokens": 9144,
+        "estimated_eval_tokens": 18544
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is heterogeneous: mostly unrelated bug reports across tokenization, multimodal processing, model config, serialization, and training/runtime edge cases. None of the soft pairs look like true duplicates of the same underlying bug/change. If forced to nominate a representative, the broadest core regression is the `from_pretrained()` non-persistent-buffer corruption issue.",
+        "confidence": 0.31,
+        "canonical_issue_reason": "issue:43950 is the broadest, most central transformers 5.x regression in the set: it affects core model loading/state handling and has wide applicability beyond one model family.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43950 is the best representative issue because it is a clear, high-impact framework-level regression rather than a model-specific or docs-only problem.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45381",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Different failures: Qwen2.5-VL position-id logic vs FP8 inference on B200. Shared hardware/runtime theme is too broad."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Tokenizer spacing after added tokens is unrelated to embedding resizing/reinitialization."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45127",
+            "accept": false,
+            "reason": "Both involve tied weights, but one is serialization/device-dependent and the other is LoRA merge collapse. Not the same bug."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "BF16-to-FP32 eval dtype drift is unrelated to Qwen3 MoE memory usage."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:46894",
+            "accept": false,
+            "reason": "XLNet CPU arange placement and CohereASR double-shift training loss are different code paths and symptoms."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "gguf_file loading failure is unrelated to stale `config.dtype` after `from_config()`."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration and a GraniteSpeech CI device override test are not the same issue."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate default and batched video output shape are unrelated processor bugs."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43975",
+            "accept": false,
+            "reason": "Qwen2.5-VL import/export issue versus DeepSeek detokenization regression; different models and failures."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "Both are v5 regressions, but one is buffer loading corruption and the other is CLIP module structure inconsistency."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Same missing-`pad_token_id` symptom, but on different configs/models and likely separate fixes."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Tokenizer loading failure and save/tokenization failure in a different speech model are related only at a very high level."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Label reduction bug and `return_tensors` being ignored are unrelated processor behaviors."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:46899",
+            "accept": false,
+            "reason": "`mm_token_type_ids` defaulting to zeros is unrelated to unsafe dtype casting in quantized embedders."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:46932",
+            "accept": false,
+            "reason": "These are different Gemma4 problems: multimodal token-type defaults vs quantized embedder dtype casting."
+          },
+          {
+            "left": "issue:44898",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Perceiver positional interpolation and RT-DETR memory release are different subsystems and bug classes."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "A failing image-generation test for Janus is unrelated to a stale paged-generate warning."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "DAC latent conversion mismatch and a GraniteSpeech CI test issue do not describe the same underlying bug."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Gemma4 multimodal token-type defaults and Qwen2.5-VL rope-index scaling are different multimodal bugs."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Sparse-layer creation in SwitchTransformers and Bamba SDPA compile failure are unrelated model-specific issues."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Both mention GPU/runtime failures, but one is auto device mapping assert and the other is FP8 inference on B200."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "Removal of `is_torch_fx_available` breaking remote-code models is unrelated to a missing import_utils helper."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "OmDet-Turbo input-size mismatch and LwDetrImageLoss AMP crash are different vision-model bugs."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Both are docs issues, but one is about removed `pipeline()` references and the other about `return_all_scores` wording."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "DAC forward/latents mismatch is unrelated to the Qwen3 Omni MoE feature-extraction helper behavior."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Mixed-precision loss crash and missing auxiliary losses when denoising is off are separate D-FINE problems."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Both touch labels, but one drops tuple elements in metric gathering while the other truncates image labels; different layers and failure modes."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "CPU offload device mismatch in Bark and a Qwen3 Omni MoE helper-length issue are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 31,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28282",
+        "issue:29127",
+        "issue:29942",
+        "issue:32090",
+        "issue:33357",
+        "issue:34689",
+        "issue:35141",
+        "issue:37428",
+        "issue:38175",
+        "issue:39692",
+        "issue:42490",
+        "issue:43278",
+        "issue:43335",
+        "issue:43425",
+        "issue:43504",
+        "issue:43526",
+        "issue:43550",
+        "issue:43611",
+        "issue:43653",
+        "issue:43819",
+        "issue:43825",
+        "issue:43854",
+        "issue:44112",
+        "issue:44206",
+        "issue:44351",
+        "issue:44448",
+        "issue:44466",
+        "issue:44493",
+        "issue:44514",
+        "issue:44521",
+        "issue:44743",
+        "issue:44855",
+        "issue:45083",
+        "issue:45290",
+        "issue:45412",
+        "issue:45520",
+        "issue:45584",
+        "issue:45588",
+        "issue:45593",
+        "issue:45759",
+        "issue:45859",
+        "issue:46170",
+        "issue:46285",
+        "issue:46326",
+        "issue:46424",
+        "issue:46512",
+        "issue:46821",
+        "issue:46858"
+      ],
+      "soft_pairs": [
+        "issue:46512|issue:46821",
+        "issue:44206|issue:45584",
+        "issue:37428|issue:43425",
+        "issue:43335|issue:43653",
+        "issue:45759|issue:46285",
+        "issue:38175|issue:45412",
+        "issue:46424|issue:46858",
+        "issue:29942|issue:32090",
+        "issue:28282|issue:33357",
+        "issue:43611|issue:44466",
+        "issue:43335|issue:43504",
+        "issue:44855|issue:45083",
+        "issue:43550|issue:44112",
+        "issue:45520|issue:46170",
+        "issue:37428|issue:44351",
+        "issue:44448|issue:44743",
+        "issue:43335|issue:45083",
+        "issue:43335|issue:44112",
+        "issue:45290|issue:46326",
+        "issue:43335|issue:45593",
+        "issue:44514|issue:44521",
+        "issue:43278|issue:43825",
+        "issue:34689|issue:35141",
+        "issue:45588|issue:45859",
+        "issue:42490|issue:43854",
+        "issue:29127|issue:32090",
+        "issue:43819|issue:44493",
+        "issue:39692|issue:43526"
+      ],
+      "prepared_review_unit_hash": "c08abf1d356e0f6cc27e20782b7f8ecd8ebb4a323b54396cb5781baf874a05d3",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36517,
+        "estimated_input_tokens": 9130,
+        "estimated_eval_tokens": 18516
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is a heterogeneous collection of unrelated issue reports; none of the proposed soft pairs look like true duplicates or the same code change. The strongest standalone issue is the open apply_chat_template multimodal assistant_masks bug.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "issue:44521 is the clearest, most actionable report: it has a focused repro, active discussion, and a concrete failure in apply_chat_template for multimodal inputs.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44521 is the best representative issue here because it is specific, reproducible, and currently open, making it the most useful anchor for triage.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46512",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Different bugs: stale config dtype after from_config versus window_size causing AttributeError in TimesFM 2.5."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Unrelated failure modes: unsupported center arg in LasrFeatureExtractor versus Whisper empty-transcription generation after align_special_tokens."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43425",
+            "accept": false,
+            "reason": "Different subsystem and symptom: missing _flash_supports_window_size import versus Torch 2.10 compatibility."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Unrelated issues: SwitchTransformers sparse layer creation versus BigBirdTokenizer special-token registration."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "Different model families and code paths: AutoModelForCausalLM config unwrapping versus CLIP class structure inconsistency."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "No shared underlying bug: zero probabilities in SigLIP2 versus RT-DETR memory not being released."
+          },
+          {
+            "left": "issue:46424",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Both touch cache behavior, but the concrete problems differ: cache length config handling versus device tensor returned from get_seq_length."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Different test/runtime failures: Flash Attention 2 test breakage versus Trainer broadcast rejecting NoneType."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Completely different errors: missing PyTorch ImportError versus MacOS bus error loading a CLIP model."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Both involve model loading/serialization, but they are different bugs with different failure points and fixes."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43504",
+            "accept": false,
+            "reason": "Different models and failures: SwitchTransformers sparse-layer config bug versus Beit legacy field load failure."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "No shared underlying issue: Python 3.13 IndentationError in DebertaV2 versus a Qwen3 Omni helper-function behavior bug."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Different scope: torch.compile/SDPA failure in Bamba versus a stale CI test in GraniteSpeech."
+          },
+          {
+            "left": "issue:45520",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Different platform dependency problems: flash_attn import key error on Python 3.13 versus Windows ROCm torch.distributed import/type issues."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Unrelated import errors: flash-support helper missing versus HybridCache import failure."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Different model behaviors: Pegasus output drift across v4/v5 versus Qwen3_5 recurrent state reset with cache."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Different components: SwitchTransformers config bug versus qwen3_omni_moe feature-length helper behavior."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "No duplicate evidence; one is a model-creation bug, the other is a CI test regression."
+          },
+          {
+            "left": "issue:45290",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Both concern chat-template code, but they describe different edge cases and different tokenizer stacks; not clearly the same bug."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Different models and symptoms: SwitchTransformers sparse-layer creation versus D-FINE auxiliary-loss handling."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Related area but not the same bug: one is a batched padding=False crash, the other is all-zero assistant masks on multimodal inputs."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Different issues entirely: dtype drift between train/eval versus an incorrect pipeline() error message."
+          },
+          {
+            "left": "issue:34689",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Different problems: model loading regression for Llama 3.2 Vision versus embedding resize reinitializing output embeddings."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Different code paths: flash_attention.py handling of s_aux=None versus Qwen3_5MoeTextRotaryEmbedding CPU-offload compatibility."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Different symptoms and contexts: local-path trailing-slash import warning versus GLM-4.7-Flash unit-test load failure."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Different failure domains: LayoutLMv3 box-content error message clarity versus Trainer GPU broadcast TypeError."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Different model bugs: DAC latent/forward mismatch versus unexpected position-id key warnings across many models."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Different areas: SigLIP2 docs example problems versus BeitImageProcessorFast reduce_labels returning the wrong labels."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 32,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:10000",
+        "issue:15354",
+        "issue:26413",
+        "issue:27957",
+        "issue:28218",
+        "issue:29870",
+        "issue:29942",
+        "issue:31501",
+        "issue:32090",
+        "issue:33357",
+        "issue:34689",
+        "issue:35141",
+        "issue:38175",
+        "issue:39692",
+        "issue:42490",
+        "issue:43278",
+        "issue:43329",
+        "issue:43550",
+        "issue:43575",
+        "issue:43606",
+        "issue:43701",
+        "issue:43720",
+        "issue:44112",
+        "issue:44206",
+        "issue:44448",
+        "issue:44464",
+        "issue:44493",
+        "issue:44617",
+        "issue:44779",
+        "issue:44857",
+        "issue:45072",
+        "issue:45083",
+        "issue:45127",
+        "issue:45245",
+        "issue:45290",
+        "issue:45310",
+        "issue:45381",
+        "issue:45459",
+        "issue:45478",
+        "issue:45593",
+        "issue:45823",
+        "issue:46234",
+        "issue:46424",
+        "issue:46431",
+        "issue:46566",
+        "issue:46697",
+        "issue:47000"
+      ],
+      "soft_pairs": [
+        "issue:46431|issue:46566",
+        "issue:43606|issue:44857",
+        "issue:43720|issue:44112",
+        "issue:39692|issue:44493",
+        "issue:28218|issue:32090",
+        "issue:45310|issue:45478",
+        "issue:44448|issue:44779",
+        "issue:44206|issue:45823",
+        "issue:43701|issue:45083",
+        "issue:44617|issue:46697",
+        "issue:26413|issue:31501",
+        "issue:45127|issue:45245",
+        "issue:45459|issue:46234",
+        "issue:32090|issue:33357",
+        "issue:44857|issue:45072",
+        "issue:42490|issue:43575",
+        "issue:45593|issue:47000",
+        "issue:43550|issue:45083",
+        "issue:27957|issue:34689",
+        "issue:38175|issue:44493",
+        "issue:44464|issue:46424",
+        "issue:29942|issue:35141",
+        "issue:29870|issue:32090",
+        "issue:45290|issue:45381",
+        "issue:43278|issue:43329",
+        "issue:10000|issue:15354"
+      ],
+      "prepared_review_unit_hash": "686332fb5f3e0767f901c85ca1feada5e867ba526cf7b4a37db166c33e5213bb",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 26,
+        "serialized_chars": 34966,
+        "estimated_input_tokens": 8742,
+        "estimated_eval_tokens": 17740
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most soft links are superficial topical matches, not true duplicates. Only the two Qwen3.5 MoE from_pretrained reports appear to describe the same regression and should be merged as duplicates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "issue:45310 is the better canonical issue: it\u2019s the earlier report, has the same Qwen3.5 MoE from_pretrained failure, and provides the fuller initial context.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45310 is the best global representative because it most clearly captures the shared bug and predates the later duplicate.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46431",
+            "right": "issue:46566",
+            "accept": false,
+            "reason": "Different failure modes: missing chat template vs device_map/meta-device loading bug."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Both are runtime errors, but they affect different models and code paths; not the same bug."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "One is an accelerate loading/unpacking problem; the other is a stale CI device override test."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Unrelated problems: documentation/model mismatch versus position-id/key handling across models."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Tokenizer spacing issue and Trainer broadcast error are unrelated."
+          },
+          {
+            "left": "issue:45310",
+            "right": "issue:45478",
+            "accept": true,
+            "reason": "Same concrete regression: Qwen3.5 MoE from_pretrained fails on transformers>=5.4.0."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both mention v5 regressions, but they concern different models and different behaviors."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Unsupported processor argument vs device-mapping mismatch; no shared code-path bug."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Resume-from-checkpoint key mismatch is unrelated to qwen3_omni_moe feature-length logic."
+          },
+          {
+            "left": "issue:44617",
+            "right": "issue:46697",
+            "accept": false,
+            "reason": "CUDA OOM in a video model is not the same as image over-padding."
+          },
+          {
+            "left": "issue:26413",
+            "right": "issue:31501",
+            "accept": false,
+            "reason": "Both are training-start/resume complaints, but the reported errors and likely root causes differ."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "Extended-vocab LoRA collapse is unrelated to category-count limits."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "Tokenizer error masking and missing import_utils symbol are different issues."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "GPU broadcast NoneType error and MacOS bus error on CLIP are unrelated."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Both are dtype-related, but they affect different models and different inference paths."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Path-trailing-slash import warning is unrelated to tensor-parallel OOM on Qwen2-57B."
+          },
+          {
+            "left": "issue:45593",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "Different detection-loss failures with different symptoms and likely causes."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "torch.compile+SDPA failure in Bamba-9B-v2 is unrelated to qwen3_omni_moe length calculation."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Flash Attention 2 support issue and Llama 3.2 vision loading regression are different problems."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Zero probabilities in SigLIP2 is not the same as unexpected position-id keys."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Inconsistent compiled chunked generation and quantized-cache max length handling are distinct bugs."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Failing FA2 tests and post-init reinitialization after resizing embeddings are unrelated."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "BART decoder_attention_mask bug is unrelated to Trainer GPU broadcast NoneType."
+          },
+          {
+            "left": "issue:45290",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Chat-template tool-call crash and Qwen2.5-VL video position-id bug are different code paths."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "BF16-to-FP32 evaluation dtype drift is unrelated to multimodal token counting video-branch undefined variables."
+          },
+          {
+            "left": "issue:10000",
+            "right": "issue:15354",
+            "accept": false,
+            "reason": "DistilBertModel issue and torch.jit.script GeneratorExp export failure are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: the only clearly mergeable duplicate pair here is issue:45310 and issue:45478, which share the same title/regression and are ordered correctly with the earlier issue as canonical. The other soft-edge rejections are conservative and do not overclaim equivalence based on the limited evidence in the packet."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 33,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:25251",
+        "issue:29127",
+        "issue:29870",
+        "issue:32090",
+        "issue:33290",
+        "issue:33357",
+        "issue:34567",
+        "issue:37428",
+        "issue:39692",
+        "issue:40444",
+        "issue:41720",
+        "issue:42175",
+        "issue:42490",
+        "issue:43278",
+        "issue:43381",
+        "issue:43526",
+        "issue:43575",
+        "issue:43701",
+        "issue:43873",
+        "issue:43927",
+        "issue:43976",
+        "issue:44186",
+        "issue:44360",
+        "issue:44361",
+        "issue:44448",
+        "issue:44466",
+        "issue:44561",
+        "issue:44568",
+        "issue:44779",
+        "issue:44855",
+        "issue:44871",
+        "issue:44977",
+        "issue:45198",
+        "issue:45229",
+        "issue:45290",
+        "issue:45440",
+        "issue:45491",
+        "issue:45563",
+        "issue:45588",
+        "issue:45698",
+        "issue:45803",
+        "issue:45823",
+        "issue:45920",
+        "issue:46192",
+        "issue:46399",
+        "issue:46421",
+        "issue:46655",
+        "issue:46858"
+      ],
+      "soft_pairs": [
+        "issue:37428|issue:42175",
+        "issue:45290|issue:46655",
+        "issue:29870|issue:33290",
+        "issue:44448|issue:44977",
+        "issue:44360|issue:44561",
+        "issue:43278|issue:43381",
+        "issue:44855|issue:46399",
+        "issue:40444|issue:43526",
+        "issue:45920|issue:46192",
+        "issue:32090|issue:33290",
+        "issue:44977|issue:45229",
+        "issue:43927|issue:45698",
+        "issue:46421|issue:46858",
+        "issue:42490|issue:43873",
+        "issue:43873|issue:45229",
+        "issue:25251|issue:32090",
+        "issue:44448|issue:44466",
+        "issue:43381|issue:43701",
+        "issue:43976|issue:45440",
+        "issue:44871|issue:44977",
+        "issue:29127|issue:34567",
+        "issue:33357|issue:39692",
+        "issue:43278|issue:43575",
+        "issue:44561|issue:44779",
+        "issue:45588|issue:45823",
+        "issue:44568|issue:45491",
+        "issue:41720|issue:45229",
+        "issue:44186|issue:45198",
+        "issue:44361|issue:45198",
+        "issue:45198|issue:45803",
+        "issue:44977|issue:45563"
+      ],
+      "prepared_review_unit_hash": "64d9da7c97aeefcc89a12bc3007910b3f61f8c8d9c9d5cdb4de99754c6a2b061",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 37315,
+        "estimated_input_tokens": 9329,
+        "estimated_eval_tokens": 18914
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a grab-bag of unrelated issues across tokenization, generation, evaluation, quantization, and packaging. None of the soft pairs look like the same underlying bug or change, so I rejected all pairings.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "issue:45920 is the strongest representative issue here: it describes a clear regression, affects multiple models, and has broad impact with a concrete tokenizer failure mode.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45920 is the best single issue to anchor the cluster because it is specific, well-scoped, and broad enough to be useful as a representative bug report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:37428",
+            "right": "issue:42175",
+            "accept": false,
+            "reason": "Flash-attention import failure and TensorFlow installation/backend packaging are unrelated."
+          },
+          {
+            "left": "issue:45290",
+            "right": "issue:46655",
+            "accept": false,
+            "reason": "Chat template tool-call crash and video padding calculation bug are different code paths."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "BART generate() UnboundLocalError and DeepSpeed Adafactor OOM are unrelated failures."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Pegasus output regression and Qwen3.5 flash-attention generation crash are different model-specific issues."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "DSA ReLU discussion and removal of is_torch_fx_available breaking trust_remote_code are not the same bug."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "Eval dtype drift and gradient-checkpointing-in-eval are separate evaluation-mode problems."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Python 3.13 JIT parsing error and Qwen3.5 decorator registration error are distinct decorator-related failures."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Multi-image IterableDataset finetuning and BeitImageProcessorFast label reduction are unrelated preprocessing bugs."
+          },
+          {
+            "left": "issue:45920",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Wrong token IDs in AutoTokenizer and silently ignored return_tensors are different tokenizer API bugs."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "NoneType GPU broadcast error and Adafactor/DeepSpeed OOM are different training failures."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Flash-attention generation bug and multi-GPU CUDA OOM are not the same concrete issue."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Config losing custom token IDs and loading the wrong custom module after save_pretrained are separate save/load problems."
+          },
+          {
+            "left": "issue:46421",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Spurious recompilation from lazy cache init and device-tensor seq_length arithmetic failure are different cache/compile bugs."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Trailing-slash import warning and quantization offloading not working are unrelated loading/runtime issues."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Quantization offloading malfunction and Gemma4 multi-GPU OOM are different memory-management problems."
+          },
+          {
+            "left": "issue:25251",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Pipeline top_k nested-list behavior and GPU broadcast NoneType error are unrelated."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Different generation output vs. device-dependent serialization of lm_head.weight are not the same bug."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode and resume_from_checkpoint key mismatch are unrelated trainer issues."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Python version compatibility and DeepSeekV3MoE implementation divergence are unrelated."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Gemma eos_token_id mismatch and Qwen3.5 flash-attention generation failure are different model-specific bugs."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "LayoutLMv3 error messaging and TrainerState token counting are unrelated."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:39692",
+            "accept": false,
+            "reason": "MacOS bus error for CLIP and SigLIP2 documentation/model mismatch are different issues."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Eval dtype drift and Qwen2-57B tensor-parallel OOM are not the same underlying problem."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "trust_remote_code breakage from removed helper and DeepSeek tokenizer regression are distinct v5 issues."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "flash_attention.py crash on s_aux=None and Gemma4 PLE device mismatch are different forward-pass failures."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Tokenizer special-token omission and Gemma3 NaN embeddings with sliding-window attention are unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping assert and Gemma4 multi-GPU OOM are different multi-device failures."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "LayoutLMv2Tokenizer crash on NER/padding and Wav2Vec2 save/tokenization failure are unrelated tokenization bugs."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "MLukeTokenizer AttributeError and Wav2Vec2 save/tokenization failure are different tokenizer/model issues."
+          },
+          {
+            "left": "issue:45198",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Wav2Vec2 save/tokenization failure and Fuyu bare-except exception handling are unrelated."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Qwen3.5 flash-attention generation failure and stale paged-generate warning are not the same code-path."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 34,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:29127",
+        "issue:29870",
+        "issue:29942",
+        "issue:30064",
+        "issue:33357",
+        "issue:34689",
+        "issue:35141",
+        "issue:42490",
+        "issue:42915",
+        "issue:42994",
+        "issue:43065",
+        "issue:43066",
+        "issue:43381",
+        "issue:43412",
+        "issue:43421",
+        "issue:43526",
+        "issue:43575",
+        "issue:43611",
+        "issue:43653",
+        "issue:43701",
+        "issue:43756",
+        "issue:43994",
+        "issue:44206",
+        "issue:44265",
+        "issue:44387",
+        "issue:44442",
+        "issue:44462",
+        "issue:44479",
+        "issue:44743",
+        "issue:45072",
+        "issue:45198",
+        "issue:45245",
+        "issue:45372",
+        "issue:45406",
+        "issue:45412",
+        "issue:45440",
+        "issue:45459",
+        "issue:45491",
+        "issue:45563",
+        "issue:45588",
+        "issue:45593",
+        "issue:45712",
+        "issue:45834",
+        "issue:46121",
+        "issue:46369",
+        "issue:46693",
+        "issue:46701"
+      ],
+      "soft_pairs": [
+        "issue:43611|issue:46121",
+        "issue:29870|issue:34689",
+        "issue:29942|issue:30064",
+        "issue:43653|issue:44743",
+        "issue:44206|issue:44479",
+        "issue:42915|issue:42994",
+        "issue:43756|issue:45440",
+        "issue:43653|issue:44442",
+        "issue:43065|issue:43994",
+        "issue:43065|issue:45412",
+        "issue:44265|issue:45198",
+        "issue:43575|issue:43701",
+        "issue:29127|issue:33357",
+        "issue:43412|issue:45588",
+        "issue:45491|issue:46693",
+        "issue:45372|issue:45588",
+        "issue:42490|issue:43381",
+        "issue:29127|issue:35141",
+        "issue:45245|issue:45563",
+        "issue:45459|issue:45712",
+        "issue:45440|issue:45593",
+        "issue:43994|issue:45412",
+        "issue:45372|issue:45406",
+        "issue:45406|issue:46701",
+        "issue:43575|issue:44387",
+        "issue:43421|issue:45491",
+        "issue:45412|issue:46369",
+        "issue:43066|issue:44462",
+        "issue:45372|issue:45459",
+        "issue:44442|issue:45072",
+        "issue:43526|issue:45834"
+      ],
+      "prepared_review_unit_hash": "ccaa664e4ed9b881d51b29193f332eee0c0e7f8ca5731198381bd9e1d83b5fac",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36515,
+        "estimated_input_tokens": 9129,
+        "estimated_eval_tokens": 18514
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster does not look like a true duplicate set; the issues span many unrelated models, tokenizer/loading paths, export bugs, quantization, and performance regressions. No soft edge is strong enough to merge.",
+        "confidence": 0.95,
+        "canonical_issue_reason": "issue:46693 is the most active and concrete runtime-regression report in the set, but it should only serve as a loose representative because the cluster is heterogeneous.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:46693 has the richest discussion and is a clear user-facing regression report, making it the best single issue representative of the set despite weak cluster coherence.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43611",
+            "right": "issue:46121",
+            "accept": false,
+            "reason": "Unrelated bugs: base_model_prefix loading vs rope-validation type handling."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Different failures in unrelated code paths: BART generate() vs Llama 3.2 Vision model loading."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "Flash Attention test failures and void segmentation-map processing are different problems."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Tokenizer special-token registration vs Qwen recurrent-state cache reset are unrelated."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Feature extractor center-arg regression is unrelated to Qwen video-input regression."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:42994",
+            "accept": false,
+            "reason": "Both involve quantization, but one is FP8 config failure and the other is quantized model saving; different bugs."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Both are model-implementation divergence reports, but for different architectures and symptoms."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "BigBirdTokenizer decode issue and FastSpeech2ConformerTokenizer load failure are distinct tokenizer bugs."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "SAM3 pixel-decoder dummy layer issue is unrelated to SigLIP2 AutoModel/pipeline output quality."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Dummy Conv2d placeholder and RT-DETR memory leak are different defects."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "torch.export failures from torch_compilable_check are unrelated to Wav2Vec2 save/tokenization issues."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Tensor-parallel OOM and resume_from_checkpoint key mismatch are unrelated."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "LayoutLMv3 box-validation error and CLIP bus error on macOS are unrelated."
+          },
+          {
+            "left": "issue:43412",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Gemma3n export problems and flash_attention s_aux=None crash are different code paths."
+          },
+          {
+            "left": "issue:45491",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "Both mention FA2/long-context behavior, but one is NaN embeddings in Gemma3 and the other is a performance regression after 5.2.0."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Gemma4 processor import failure and flash_attention integration crash are unrelated."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "Local-path import warning and gradient-checkpointing eval-mode restriction are unrelated."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Poor LayoutLMv3 error messaging and embedding reinitialization on resize are unrelated."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Category-cardinality runtime error and paged generate warning are different issues."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Tokenizer error masking and leftover dummy classes are unrelated maintenance/runtime problems."
+          },
+          {
+            "left": "issue:45440",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "DeepSeek implementation divergence and D-FINE auxiliary-loss behavior are unrelated."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "SigLIP2 model output issues and RT-DETR memory retention are different bugs."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both involve Gemma4, but one is an import dependency failure and the other is a missing processor attribute."
+          },
+          {
+            "left": "issue:45406",
+            "right": "issue:46701",
+            "accept": false,
+            "reason": "Gemma4 serve crash and DiffusionGemma image-text-to-text pipeline failure are different failures."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "OOM from loading a specific Qwen model and int4 reserved-memory regression are related only broadly by memory pressure."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Tokenizer post-processor update behavior and Gemma3 NaN embeddings are unrelated."
+          },
+          {
+            "left": "issue:45412",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "RT-DETR cleanup/memory leak and Sam3-litetext memory overflow are different memory issues."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:44462",
+            "accept": false,
+            "reason": "Wrong tokenizer decoder type in v5 and ignoring tokenizer.json are distinct tokenizer-loading bugs."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "Gemma4 processor import dependency issue and protobuf error masking are unrelated."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "FastSpeech2ConformerTokenizer loading and bf16 dtype mismatches are unrelated."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction bug and Kosmos2.5 OCR index error are unrelated image-processing bugs."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 35,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:27957",
+        "issue:29127",
+        "issue:29870",
+        "issue:33290",
+        "issue:34689",
+        "issue:35141",
+        "issue:42175",
+        "issue:42490",
+        "issue:42915",
+        "issue:43262",
+        "issue:43316",
+        "issue:43334",
+        "issue:43381",
+        "issue:43525",
+        "issue:43526",
+        "issue:43531",
+        "issue:43646",
+        "issue:43720",
+        "issue:43761",
+        "issue:43824",
+        "issue:43873",
+        "issue:43931",
+        "issue:43994",
+        "issue:44077",
+        "issue:44186",
+        "issue:44206",
+        "issue:44279",
+        "issue:44442",
+        "issue:44514",
+        "issue:44743",
+        "issue:44811",
+        "issue:44857",
+        "issue:44898",
+        "issue:44977",
+        "issue:45003",
+        "issue:45245",
+        "issue:45250",
+        "issue:45381",
+        "issue:45440",
+        "issue:45588",
+        "issue:45859",
+        "issue:46121",
+        "issue:46234",
+        "issue:46369",
+        "issue:46399",
+        "issue:46640",
+        "issue:46693",
+        "issue:46821"
+      ],
+      "soft_pairs": [
+        "issue:27957|issue:33290",
+        "issue:27957|issue:29127",
+        "issue:43526|issue:46693",
+        "issue:43824|issue:44279",
+        "issue:33290|issue:34689",
+        "issue:33290|issue:35141",
+        "issue:44857|issue:45588",
+        "issue:43994|issue:46369",
+        "issue:42915|issue:43931",
+        "issue:43526|issue:44186",
+        "issue:43334|issue:43531",
+        "issue:45250|issue:46640",
+        "issue:43381|issue:43873",
+        "issue:45588|issue:46821",
+        "issue:43646|issue:44077",
+        "issue:46121|issue:46399",
+        "issue:44514|issue:45381",
+        "issue:44977|issue:45245",
+        "issue:43761|issue:45859",
+        "issue:44743|issue:44811",
+        "issue:43262|issue:44206",
+        "issue:44442|issue:44898",
+        "issue:42490|issue:44743",
+        "issue:43720|issue:44442",
+        "issue:27957|issue:29870",
+        "issue:42175|issue:46234",
+        "issue:43316|issue:43525",
+        "issue:45003|issue:45440"
+      ],
+      "prepared_review_unit_hash": "25117b36f95d7b4f1c80b4cde3ab76ceb555957f8826b058b530bbb1a2967fa9",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 35947,
+        "estimated_input_tokens": 8987,
+        "estimated_eval_tokens": 18230
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a heterogeneous set of issues with a few weak theme overlaps around Flash Attention, Qwen-family models, and loading/config regressions. I would not merge any of the soft pairs as duplicates; they do not describe the same concrete bug/change closely enough.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "issue:46640 is the clearest umbrella Flash Attention 2 report in the set and the most plausible representative anchor for that small theme, but it still doesn\u2019t match the other items closely enough to be their duplicate.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:46640 is the best global issue anchor here because it is the most general FA2 ticket and has the strongest chance of collecting related discussion, even though the surrounding items are not true duplicates.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:27957",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "Different failures: XLMRoberta/Flash Attention 2 support vs. Adafactor+DeepSpeed OOM."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "Unrelated subsystems: Flash Attention 2 on XLMRoberta vs. LayoutLMv3 box-validation error messaging."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "Different problems: BeitImageProcessorFast label reduction bug vs. FA2 long-context performance regression."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Import error for a specific Qwen2.5 VL class vs. a generic dependency issue; no same code-path evidence."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "OOM with Adafactor/DeepSpeed is unrelated to Llama 3.2 Vision model loading breakage."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Optimizer memory use vs. embedding reinitialization during post_init are different bugs."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Different models and failures: LW-DETR AMP/CUDA crash vs. flash_attention sink-less AttributeError."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Different symptom and scope: SigLIP2 bad outputs vs. SAM3-LiteText memory overflow."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different Qwen model/config issues: FineGrainedFP8Config failure vs. Qwen3-VL weight-shape mismatch."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44186",
+            "accept": false,
+            "reason": "Different image-processor/tokenizer bugs; no shared underlying code-path."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Qwen3-VL missing pad_token_id vs. Qwen3-MoE sliding_window behavior are separate issues."
+          },
+          {
+            "left": "issue:45250",
+            "right": "issue:46640",
+            "accept": false,
+            "reason": "Both mention Flash Attention 2, but the reports are too generic to prove the same concrete bug/change."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode is unrelated to quantization offloading behavior."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Flash-attention sink handling vs. TimesFM window_size AttributeError are different code paths."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "Generic custom-model-init regression vs. patchtsmixer post_init policy change; not the same bug."
+          },
+          {
+            "left": "issue:46121",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Rope-validation TypeError and Qwen3.5 kernel registration failure are separate problems."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Batched apply_chat_template padding crash vs. video vision_position_ids bug are different Qwen2.5-VL issues."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "FA2 generation problem vs. category-count limit error; no overlap beyond broad model/runtime context."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "CLIPVision hidden_states regression and Qwen3_5Moe CPU offload incompatibility are unrelated."
+          },
+          {
+            "left": "issue:44743",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Recurrent-state cache reset bug vs. Whisper batch_decode skip_special_tokens handling are different behaviors."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44206",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate default issue vs. LasrFeatureExtractor center-arg crash are unrelated."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Tokenizer loading failure and Perceiver image-classification resolution bug are different models and paths."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Local-path import warning is unrelated to recurrent-state reset with cache."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "BitNet packed-weight loading bug vs. tokenizer load failure for FastSpeech2Conformer are unrelated."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:29870",
+            "accept": false,
+            "reason": "XLMRoberta Flash Attention 2 support issue is not the same as BART decoder_attention_mask UnboundLocalError."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "Packaging/backend dependency complaint vs. missing import_utils symbol are different issues."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "API discrepancy in Gemma3TextConfig is unrelated to Llama4Config missing pad_token_id."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "sys.modules access safety bug vs. DeepseekV3MoE divergence from remote implementation are different code concerns."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 36,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:16998",
+        "issue:22355",
+        "issue:25251",
+        "issue:27957",
+        "issue:28282",
+        "issue:32090",
+        "issue:42617",
+        "issue:42907",
+        "issue:43232",
+        "issue:43316",
+        "issue:43388",
+        "issue:43493",
+        "issue:43550",
+        "issue:43575",
+        "issue:43606",
+        "issue:43643",
+        "issue:43701",
+        "issue:43742",
+        "issue:43749",
+        "issue:43761",
+        "issue:43827",
+        "issue:43873",
+        "issue:43931",
+        "issue:44062",
+        "issue:44077",
+        "issue:44079",
+        "issue:44188",
+        "issue:44190",
+        "issue:44442",
+        "issue:44479",
+        "issue:44496",
+        "issue:44855",
+        "issue:44871",
+        "issue:44913",
+        "issue:44987",
+        "issue:45137",
+        "issue:45216",
+        "issue:45325",
+        "issue:45698",
+        "issue:45759",
+        "issue:45803",
+        "issue:45907",
+        "issue:46178",
+        "issue:46506",
+        "issue:46697",
+        "issue:46829",
+        "issue:46932"
+      ],
+      "soft_pairs": [
+        "issue:44077|issue:44479",
+        "issue:43493|issue:43827",
+        "issue:43873|issue:44871",
+        "issue:45325|issue:46697",
+        "issue:44062|issue:45907",
+        "issue:43643|issue:44913",
+        "issue:44855|issue:45137",
+        "issue:43316|issue:45759",
+        "issue:43749|issue:43931",
+        "issue:44987|issue:46178",
+        "issue:44190|issue:44496",
+        "issue:43931|issue:45216",
+        "issue:42617|issue:43575",
+        "issue:43701|issue:46506",
+        "issue:44442|issue:45803",
+        "issue:43232|issue:43388",
+        "issue:43550|issue:43701",
+        "issue:43575|issue:43742",
+        "issue:43761|issue:44079",
+        "issue:27957|issue:32090",
+        "issue:43761|issue:44188",
+        "issue:43606|issue:43701",
+        "issue:25251|issue:28282",
+        "issue:46829|issue:46932",
+        "issue:42907|issue:45698",
+        "issue:16998|issue:22355"
+      ],
+      "prepared_review_unit_hash": "f7261b8e0ae2a35e2947119162186201a995ed8edca4e15397696dd0596424e7",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 26,
+        "serialized_chars": 34972,
+        "estimated_input_tokens": 8743,
+        "estimated_eval_tokens": 17742
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This is not a coherent duplicate cluster; the items span many unrelated bugs, regressions, and docs issues across different models and subsystems. All soft pairs are only topical similarities, not the same underlying bug/change, so they should be rejected.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44077",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Different models and failure modes: patchtsmixer post_init policy vs Qwen video input regression."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "One is a SigLIP2 implementation discrepancy; the other is a docs update after pipeline removals."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Quantization/offloading behavior and Gemma eos_token_id mismatch are unrelated code paths."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:46697",
+            "accept": false,
+            "reason": "Both are vision-related preprocessing issues, but one is rope/position_ids logic and the other is padding behavior."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Tokenizer AddedToken construction bug is unrelated to Hub list_repo_files revision forwarding."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Both involve config handling, but one is missing fields with trust_remote_code and the other is rotary_pct persistence on reload."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Python 3.13 import parsing error and DeepSpeed ZeRO3 deque IndexError are different failures."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both touch composite configs, but they concern different APIs and different broken behaviors."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "FSDP CPU RAM efficient loading and Qwen3-VL weight-shape mismatch are separate loading bugs."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "A model-loading regression and a backend mapping error are not the same change or code path."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Local dataset loading in a training script is unrelated to unrecognized model/config loading."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "One is a load-time weight mismatch; the other is a save_pretrained regression for Qwen3.5 checkpoints."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "A failing parallel script and an OOM during tensor-parallel model load are not the same bug."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "Checkpoint resume key mismatch is unrelated to CPU/CUDA result divergence."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Tokenizer loading failure and a bare-except tensor conversion bug are unrelated."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:43388",
+            "accept": false,
+            "reason": "Both affect distributed workflows, but generation kwargs sync and gather_for_metrics label dropping are different paths."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "torch.compile + SDPA divergence is unrelated to resume_from_checkpoint key mismatch."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "OOM on tensor-parallel load and a key error while loading MobileLLM are different loading issues."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "hidden_states missing from a model forward pass is not the same as ModelOutput key assignment with None values."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Flash Attention model support and a Trainer broadcast TypeError are unrelated."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44188",
+            "accept": false,
+            "reason": "A CLIPVisionModel output regression and attention-kernel divergence under torch.compile are different bugs."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "CPU offload device mismatch is unrelated to checkpoint resume key mismatches."
+          },
+          {
+            "left": "issue:25251",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Pipeline top_k nesting behavior and missing PyTorch dependency are unrelated."
+          },
+          {
+            "left": "issue:46829",
+            "right": "issue:46932",
+            "accept": false,
+            "reason": "VideoPrism output field mix-up and Gemma4 quantized embedder dtype casting are different issues."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Save failure for dequantized models and wrong custom-module loading after save_pretrained are not the same concrete bug."
+          },
+          {
+            "left": "issue:16998",
+            "right": "issue:22355",
+            "accept": false,
+            "reason": "DeBERTa model_max_length behavior and a missing transformers.onnx module are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 37,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:42175",
+        "issue:43065",
+        "issue:43295",
+        "issue:43316",
+        "issue:43352",
+        "issue:43404",
+        "issue:43441",
+        "issue:43452",
+        "issue:43504",
+        "issue:43540",
+        "issue:43572",
+        "issue:43646",
+        "issue:43650",
+        "issue:43716",
+        "issue:43825",
+        "issue:44077",
+        "issue:44162",
+        "issue:44242",
+        "issue:44464",
+        "issue:44496",
+        "issue:44514",
+        "issue:44534",
+        "issue:44610",
+        "issue:44617",
+        "issue:44792",
+        "issue:44871",
+        "issue:44913",
+        "issue:44977",
+        "issue:45127",
+        "issue:45200",
+        "issue:45542",
+        "issue:45584",
+        "issue:45600",
+        "issue:45632",
+        "issue:45698",
+        "issue:45803",
+        "issue:46050",
+        "issue:46170",
+        "issue:46178",
+        "issue:46202",
+        "issue:46234",
+        "issue:46291",
+        "issue:46365",
+        "issue:46369",
+        "issue:46421",
+        "issue:46424",
+        "issue:46640",
+        "issue:46894"
+      ],
+      "soft_pairs": [
+        "issue:43295|issue:43825",
+        "issue:43295|issue:44871",
+        "issue:44871|issue:45584",
+        "issue:46234|issue:46291",
+        "issue:44077|issue:44534",
+        "issue:43316|issue:43716",
+        "issue:42175|issue:43316",
+        "issue:44464|issue:46421",
+        "issue:43646|issue:44913",
+        "issue:43452|issue:44496",
+        "issue:44871|issue:46050",
+        "issue:45803|issue:46365",
+        "issue:44162|issue:44792",
+        "issue:43404|issue:43441",
+        "issue:43065|issue:44617",
+        "issue:44242|issue:46894",
+        "issue:43540|issue:44514",
+        "issue:44610|issue:46369",
+        "issue:45632|issue:46424",
+        "issue:43572|issue:46178",
+        "issue:45542|issue:46234",
+        "issue:45803|issue:46170",
+        "issue:44792|issue:45600",
+        "issue:44792|issue:44977",
+        "issue:43504|issue:45698",
+        "issue:43650|issue:46202",
+        "issue:43352|issue:46640",
+        "issue:45127|issue:45200"
+      ],
+      "prepared_review_unit_hash": "f392a381d7a45dee8e9d8e4c668604693da1eb7a45b8550f573cbcb3dd0e9cb3",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 35943,
+        "estimated_input_tokens": 8986,
+        "estimated_eval_tokens": 18228
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is heterogeneous: the paired issues share only broad subsystem or keyword overlap, not the same underlying bug/change. I would not collapse any of the soft pairs; treat them as separate tickets. If a representative issue is needed, issue 45698 is the strongest general candidate because it has the highest discussion/inbound activity and describes a concrete, reusable serialization/loading bug.",
+        "confidence": 0.9,
+        "canonical_issue_reason": "Issue 45698 is the most substantive/high-signal issue in the set, with substantial discussion and inbound references, and it describes a concrete core loading bug affecting custom modules.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45698 is the best global representative because it is concrete, broadly relevant to model loading, and has the strongest evidence of impact/triage interest.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43295",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Both mention regressions, but one is a tokenizer/processor API break and the other is an error-message/task-support issue; different code paths."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "One is about processor.tokenizer behavior; the other is an eos_token_id config mismatch. No shared underlying bug."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Gemma eos_token configuration and Whisper empty-transcription handling are unrelated failures."
+          },
+          {
+            "left": "issue:46234",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Missing availability helper vs pipelines import failing on kernels dependency requirements; different problems."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "PatchTSMixer post_init validation and non-persistent buffer junk are unrelated v5 behaviors."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:43716",
+            "accept": false,
+            "reason": "Gemma3TextConfig API discrepancy vs Mistral-3 dtype mismatch; different models and bugs."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43316",
+            "accept": false,
+            "reason": "TensorFlow backend installation issue vs Gemma3TextConfig API mismatch; no same bug."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:46421",
+            "accept": false,
+            "reason": "Both touch chunked generation/prefill, but one is output inconsistency under compiled forward and the other is cache initialization/recompilation behavior."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Generic custom model initialization breakage vs GPTNeoX rotary_pct reload regression; not the same concrete issue."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "gguf_file/from_pretrained loading breakage and unrecognized model config are different model-loading failures."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "eos_token_id inconsistency and quantized-model save/load corruption are distinct serialization/config bugs."
+          },
+          {
+            "left": "issue:45803",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Bare except swallowing errors in tensor conversion vs Trainer import conflict with datasets; unrelated."
+          },
+          {
+            "left": "issue:44162",
+            "right": "issue:44792",
+            "accept": false,
+            "reason": "ESM2 model breakage and a janus test failure are unrelated model-specific issues."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Weight tying in Mistral3 and FlashAttention support in Ministral-3 are different code-path problems."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Dummy Conv2d placeholder in Sam3PixelDecoder vs Sam3Video OOM; not the same bug."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:46894",
+            "accept": false,
+            "reason": "Different MoE training-loss bugs in different models; too broad to treat as one change."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "Video-input processing in Qwen3OmniMoe and batched chat template padding in Qwen2.5-VL are separate processor bugs."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Input-size mismatch in OmDet-Turbo vs memory overflow in sam3-litetext; no common root cause."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Cache-path collision for trust_remote_code and cache length handling for quantized cache are different caching issues."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "StableLmConfig pad_token_idx regression is unrelated to the tensorflow_text backend mapping error."
+          },
+          {
+            "left": "issue:45542",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "TensorBoard/TensorFlow backend installation problem and missing import_utils helper are different dependency bugs."
+          },
+          {
+            "left": "issue:45803",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Exception handling in Fuyu batch conversion vs Windows/ROCm torch.distributed import failure; unrelated."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Janus image generation test failure and Sam3LiteText auto_mappings CI break are different failures."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Janus test failure and Qwen3.5 flash-attention generation issue are unrelated."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Legacy-field pretrained preset loading issue vs wrong custom-module loading after save_pretrained; different failure modes."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "Non-actionable placeholder issue vs Gemma pipeline roundtrip tokenization bug; not duplicates."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:46640",
+            "accept": false,
+            "reason": "Specific model lacking FlashAttention 2 support vs a broader flash-attention issue; insufficient evidence of same code-path."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "LoRA merge/tied-embeddings collapse and Gemma 4 mm_token_type_ids defaults are unrelated model behaviors."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 38,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:25251",
+        "issue:27957",
+        "issue:29870",
+        "issue:35141",
+        "issue:41720",
+        "issue:41950",
+        "issue:42490",
+        "issue:43335",
+        "issue:43388",
+        "issue:43475",
+        "issue:43525",
+        "issue:43531",
+        "issue:43644",
+        "issue:43673",
+        "issue:43688",
+        "issue:43827",
+        "issue:43931",
+        "issue:43994",
+        "issue:44060",
+        "issue:44077",
+        "issue:44190",
+        "issue:44242",
+        "issue:44534",
+        "issue:44617",
+        "issue:44734",
+        "issue:44743",
+        "issue:44855",
+        "issue:44869",
+        "issue:44877",
+        "issue:44913",
+        "issue:44964",
+        "issue:45072",
+        "issue:45245",
+        "issue:45381",
+        "issue:45406",
+        "issue:45563",
+        "issue:45584",
+        "issue:45588",
+        "issue:45632",
+        "issue:46082",
+        "issue:46369",
+        "issue:46421",
+        "issue:46617",
+        "issue:46697",
+        "issue:46832",
+        "issue:46858",
+        "issue:46897",
+        "issue:46932"
+      ],
+      "soft_pairs": [
+        "issue:43525|issue:43531",
+        "issue:44534|issue:44913",
+        "issue:35141|issue:42490",
+        "issue:41950|issue:43475",
+        "issue:43673|issue:44743",
+        "issue:43388|issue:43688",
+        "issue:44190|issue:44964",
+        "issue:44855|issue:45588",
+        "issue:29870|issue:35141",
+        "issue:45072|issue:46932",
+        "issue:45632|issue:46617",
+        "issue:44869|issue:46832",
+        "issue:45245|issue:46858",
+        "issue:44734|issue:45406",
+        "issue:46082|issue:46697",
+        "issue:43827|issue:43994",
+        "issue:44617|issue:45584",
+        "issue:43644|issue:44913",
+        "issue:25251|issue:27957",
+        "issue:43931|issue:44060",
+        "issue:43335|issue:44855",
+        "issue:44242|issue:46897",
+        "issue:44077|issue:44877",
+        "issue:43673|issue:46421",
+        "issue:45381|issue:46369",
+        "issue:45563|issue:45584",
+        "issue:41720|issue:45245"
+      ],
+      "prepared_review_unit_hash": "4398d102cd9f2cd56b5dee7be3a7eee9a1ecc5f5b873b2e7f327610ed018543b",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 36027,
+        "estimated_input_tokens": 9007,
+        "estimated_eval_tokens": 18270
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a grab-bag of unrelated issue reports that only share broad Transformers vocabulary. I don\u2019t see any soft pair that is safely the same underlying bug or change, so all soft edges are rejected.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "issue:43644 is the broadest and most central report in the cache/runtime-regression subset of this cluster, with external references and a title that generalizes beyond one model or endpoint.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43644 is the best representative issue because it describes a generalized v5 regression and is less model-specific than most of the other reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43525",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Different bugs: missing pad_token_id on Llama4Config vs a sliding_window issue in Qwen3-MoE."
+          },
+          {
+            "left": "issue:44534",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Both are config-related, but one is non-persistent buffers getting junk data and the other is rotary_pct not persisting on reload; different symptoms and code paths."
+          },
+          {
+            "left": "issue:35141",
+            "right": "issue:42490",
+            "accept": false,
+            "reason": "Resizing embeddings reinitialization is unrelated to import warnings from local paths ending in a slash."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Video-classification pipeline processor lookup is unrelated to a SAM3 vision encoder missing attribute."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Both involve cache behavior, but one is cache missing during chunked prefill and the other is recurrent states resetting in Qwen3 modular code; not the same concrete bug."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43688",
+            "accept": false,
+            "reason": "Label truncation in gather_for_metrics is unrelated to auxiliary-loss normalization in MoE models."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Local dataset loading in a training script is unrelated to a model weight-shape mismatch when loading Phi-4 multimodal."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Python 3.13 decorator parsing failure is unrelated to a flash-attention crash on s_aux=None."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "BART generate() UnboundLocalError is a different bug from output embeddings being reinitialized after resizing."
+          },
+          {
+            "left": "issue:45072",
+            "right": "issue:46932",
+            "accept": false,
+            "reason": "Both mention dtype issues, but they affect different models and different embedders; not the same fix."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "A cache path collision for trust_remote_code models is different from a FileNotFoundError when loading from cache."
+          },
+          {
+            "left": "issue:44869",
+            "right": "issue:46832",
+            "accept": false,
+            "reason": "Whisper token decode crash and RT-DETR forward TypeError are unrelated failures."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Category-count limit error and device-tensor slice arithmetic in static cache are unrelated numeric/runtime issues."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "KV-cache continuation tensor indexing in serve is unrelated to Gemma4Processor missing _tokenizer."
+          },
+          {
+            "left": "issue:46082",
+            "right": "issue:46697",
+            "accept": false,
+            "reason": "LlamaConfig head_dim validation is unrelated to Swin2SR over-padding."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Documentation still referencing pipeline() is not the same as SigLIP2 producing wrong outputs with AutoModel/pipeline."
+          },
+          {
+            "left": "issue:44617",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "CUDA OOM in Sam3Video is unrelated to Whisper generation failing on empty transcription."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Non-persistent buffer corruption in v5 and rotary_pct reset on reload are different configuration/regression bugs."
+          },
+          {
+            "left": "issue:25251",
+            "right": "issue:27957",
+            "accept": false,
+            "reason": "Pipeline top_k output nesting and Flash Attention 2 support for XLMRoberta are unrelated."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44060",
+            "accept": false,
+            "reason": "Model loading shape mismatch and incorrect tied-weights warning are different issues in different model families."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "Sparse-layer creation in SwitchTransformersConfig is unrelated to a Python 3.13 IndentationError from torch.jit.script parsing."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:46897",
+            "accept": false,
+            "reason": "Missing load-balancing loss when output_router_logits=False is different from a training-loss double-shift bug."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "patchtsmixer post_init policy and granite_speech strict-config loading are unrelated config-handling issues."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:46421",
+            "accept": false,
+            "reason": "Both mention cache/chunked prefill, but one is missing cache initialization and the other is spurious recompilation from lazy init; not enough to treat as the same bug."
+          },
+          {
+            "left": "issue:45381",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Qwen2.5-VL vision_position_ids are unrelated to SAM3 litetext memory overflow."
+          },
+          {
+            "left": "issue:45563",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "A stale warning for num_return_sequences is not the same as Whisper failing on empty transcription."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping cudaErrorAssert and category-count overflow are unrelated failures."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 39,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:25251",
+        "issue:29870",
+        "issue:39692",
+        "issue:43065",
+        "issue:43295",
+        "issue:43335",
+        "issue:43526",
+        "issue:43540",
+        "issue:43550",
+        "issue:43643",
+        "issue:43673",
+        "issue:43854",
+        "issue:44112",
+        "issue:44190",
+        "issue:44246",
+        "issue:44462",
+        "issue:44464",
+        "issue:44479",
+        "issue:44496",
+        "issue:44521",
+        "issue:44610",
+        "issue:44617",
+        "issue:44734",
+        "issue:44779",
+        "issue:44792",
+        "issue:44933",
+        "issue:44938",
+        "issue:44964",
+        "issue:45003",
+        "issue:45127",
+        "issue:45276",
+        "issue:45290",
+        "issue:45357",
+        "issue:45362",
+        "issue:45381",
+        "issue:45542",
+        "issue:45563",
+        "issue:45584",
+        "issue:45588",
+        "issue:45600",
+        "issue:45907",
+        "issue:46369",
+        "issue:46489",
+        "issue:46655",
+        "issue:46697",
+        "issue:46752",
+        "issue:46858",
+        "issue:47013"
+      ],
+      "soft_pairs": [
+        "issue:43550|issue:44190",
+        "issue:44933|issue:45362",
+        "issue:43065|issue:44610",
+        "issue:43335|issue:44190",
+        "issue:45588|issue:47013",
+        "issue:44479|issue:44734",
+        "issue:44617|issue:46369",
+        "issue:45381|issue:46655",
+        "issue:43854|issue:44792",
+        "issue:39692|issue:44190",
+        "issue:45127|issue:45357",
+        "issue:43295|issue:45276",
+        "issue:44464|issue:45563",
+        "issue:44779|issue:46489",
+        "issue:45276|issue:45584",
+        "issue:43540|issue:44734",
+        "issue:45290|issue:46752",
+        "issue:44112|issue:45600",
+        "issue:44521|issue:45381",
+        "issue:25251|issue:29870",
+        "issue:44462|issue:45907",
+        "issue:43643|issue:44496",
+        "issue:43526|issue:46697",
+        "issue:44938|issue:45542",
+        "issue:43854|issue:44964",
+        "issue:44246|issue:45003",
+        "issue:43673|issue:46858"
+      ],
+      "prepared_review_unit_hash": "9576d34f0b4a8674d98a76677b625cadff95d88ae705a236256fbc98da750ffd",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 35944,
+        "estimated_input_tokens": 8986,
+        "estimated_eval_tokens": 18228
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Only one soft duplicate pair is a credible match: the two DeepSeek tokenizer regression reports. All other soft pairs appear to describe different bugs, models, or code paths and should stay separate.",
+        "confidence": 0.96,
+        "canonical_issue_reason": "issue:46489 is the most specific report: it names the exact DeepSeek-Coder v1 regression on transformers v5+ and points to the gap in the prior fix, making it the best canonical issue for this duplicate pair.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:46489 is the strongest representative because it is the most concrete and version-specific description of the shared tokenizer regression.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43550",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Different models and failures: SDPA/torch.compile vs local dataset loading."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "Unrelated problems: missing import from image_utils vs Qwen3.5 chat crash."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Both involve vision models, but they are different models and different bugs."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Completely unrelated config bug vs dataset-loading issue."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:47013",
+            "accept": false,
+            "reason": "Different code paths and symptoms: flash_attention s_aux handling vs xLSTM train-mode unpacking."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Both are generation-related regressions, but they affect different subsystems and failure modes."
+          },
+          {
+            "left": "issue:44617",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Same family, but different models and likely different causes of memory usage."
+          },
+          {
+            "left": "issue:45381",
+            "right": "issue:46655",
+            "accept": false,
+            "reason": "Both are video preprocessing issues, but for different models and distinct bugs."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44792",
+            "accept": false,
+            "reason": "Unrelated model-loading test failure vs janus image-generation test failure."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Doc example problems are unrelated to local dataset loading."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different save/load behaviors and different model components."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "Both touch model/processor APIs, but the concrete bugs are unrelated."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Different generation issues: output inconsistency with compile vs a stale warning."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:46489",
+            "accept": true,
+            "reason": "Same underlying DeepSeek tokenizer regression in v5+: 46489 is the more specific follow-up to the broader 44779 report."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Different features and failures: resize_token_embeddings vs Whisper empty-transcription handling."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Different model families and unrelated runtime errors."
+          },
+          {
+            "left": "issue:45290",
+            "right": "issue:46752",
+            "accept": false,
+            "reason": "Both are apply_chat_template edge cases, but one is tool-call assistant messages and the other is an empty list; not clearly the same bug."
+          },
+          {
+            "left": "issue:44112",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Both are CI-related, but they are separate stale-reference problems in different areas."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Different multimodal bugs: assistant mask computation vs video vision_position_ids."
+          },
+          {
+            "left": "issue:25251",
+            "right": "issue:29870",
+            "accept": false,
+            "reason": "Completely unrelated pipeline output-shape issue vs BART generate() UnboundLocalError."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Tokenizer.json precedence and list_repo_files revision forwarding are separate repository-loading issues."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Both concern config/model loading, but one is missing fields from remote code and the other is an unrecognized model_type."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:46697",
+            "accept": false,
+            "reason": "Different image-processor bugs on different models."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Different import/runtime environment problems."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different model-loading failures for different models."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Import slowness and unsafe sys.modules access are not the same bug."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "Both involve cache/prefill, but the concrete failure modes and root causes differ."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The accepted pair is well-grounded in the packet: 44779 is the broader DeepSeek tokenizer regression, and 46489 is a more specific v5+ follow-up that explicitly describes the remaining gap in the prior fix. The other soft-edge rejections are appropriately conservative and do not overclaim shared root cause from the titles alone."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 40,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:27957",
+        "issue:32090",
+        "issue:35141",
+        "issue:36246",
+        "issue:38617",
+        "issue:41628",
+        "issue:41950",
+        "issue:42175",
+        "issue:42796",
+        "issue:42907",
+        "issue:43295",
+        "issue:43381",
+        "issue:43404",
+        "issue:43441",
+        "issue:43526",
+        "issue:43531",
+        "issue:43582",
+        "issue:43673",
+        "issue:43901",
+        "issue:43976",
+        "issue:44016",
+        "issue:44112",
+        "issue:44220",
+        "issue:44279",
+        "issue:44295",
+        "issue:44336",
+        "issue:44496",
+        "issue:44521",
+        "issue:44683",
+        "issue:44737",
+        "issue:44964",
+        "issue:44987",
+        "issue:45042",
+        "issue:45083",
+        "issue:45216",
+        "issue:45335",
+        "issue:45446",
+        "issue:45593",
+        "issue:45600",
+        "issue:45632",
+        "issue:45759",
+        "issue:45803",
+        "issue:46082",
+        "issue:46121",
+        "issue:46234",
+        "issue:46424",
+        "issue:46506",
+        "issue:46932"
+      ],
+      "soft_pairs": [
+        "issue:27957|issue:35141",
+        "issue:44737|issue:45803",
+        "issue:44964|issue:45600",
+        "issue:44112|issue:44964",
+        "issue:42175|issue:44279",
+        "issue:45632|issue:46234",
+        "issue:42796|issue:44521",
+        "issue:32090|issue:35141",
+        "issue:36246|issue:41628",
+        "issue:36246|issue:38617",
+        "issue:43404|issue:45216",
+        "issue:43531|issue:45600",
+        "issue:44295|issue:46506",
+        "issue:46082|issue:46121",
+        "issue:41950|issue:44016",
+        "issue:43673|issue:46424",
+        "issue:43295|issue:45335",
+        "issue:43582|issue:46932",
+        "issue:45083|issue:46506",
+        "issue:44496|issue:45042",
+        "issue:43976|issue:44987",
+        "issue:43526|issue:43901",
+        "issue:43901|issue:44496",
+        "issue:43673|issue:44336",
+        "issue:44496|issue:45759",
+        "issue:44683|issue:45446",
+        "issue:44016|issue:44220",
+        "issue:43381|issue:45593",
+        "issue:42907|issue:43441"
+      ],
+      "prepared_review_unit_hash": "b2c0af4296b4a445a0c1b3aa35f52ac06ee0fc9369d7b0ea1b1b7c13653c6eb3",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36533,
+        "estimated_input_tokens": 9134,
+        "estimated_eval_tokens": 18524
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The soft pairs are mostly superficial matches on shared subsystem/model names, but they describe different bugs or regressions. I would not merge any of them as duplicates. The most prominent issue in the set is 44521, but the overall set is not a single coherent duplicate cluster.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No single true canonical: the items span unrelated bugs; 44521 is the most active/open issue with inbound references, so it is the best anchor if one must be chosen.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "44521 has the strongest signal (open, detailed report, inbound refs) and is the most plausible representative of the set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:27957",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Different problems: Flash Attention support for XLM-Roberta vs token embedding resize/reinit behavior."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Unrelated bugs: XLNet CPU device placement vs a bare-except in Fuyu tensor conversion."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Different code paths and symptoms: multimodal model loading failure vs stale auto_mappings references in CI."
+          },
+          {
+            "left": "issue:44112",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "CI test failure in GraniteSpeech vs model loading failure for Phi-4 multimodal; not the same bug."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Backend/dependency installation issue vs a generic transformers dependency complaint; no shared concrete defect."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "Cache-path collision for trust_remote_code vs missing import utility symbol; unrelated."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Tokenizer BPE marker handling vs multimodal assistant mask generation; different functionality."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Trainer GPU broadcast TypeError vs embedding resize/post_init reinitialization bug."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and likely different fixes."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Different missing imports in different modules; not the same underlying issue."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Weight tying bug in Mistral3 vs checkpoint save regression in Qwen3.5; related theme but not one bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Sliding-window behavior in Qwen3-MoE vs auto_mappings config reference cleanup; unrelated."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "Position_ids buffer read issue vs CPU/CUDA output divergence in PPDocLayoutV3; different defects."
+          },
+          {
+            "left": "issue:46082",
+            "right": "issue:46121",
+            "accept": false,
+            "reason": "LlamaConfig head_dim validation vs rope-parameter list handling; different validation paths."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:44016",
+            "accept": false,
+            "reason": "Video-classification pipeline processor lookup vs a notebook syntax error; unrelated."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Cache missing in chunked prefill vs max_cache_length only honored for quantized cache; not the same bug."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "Processor/tokenizer API regression vs resize_token_embeddings not updating decoder.embed_tokens; different areas."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:46932",
+            "accept": false,
+            "reason": "Apple Silicon warmup TypeError vs Gemma4 dtype casting in quantized embedders; unrelated."
+          },
+          {
+            "left": "issue:45083",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "Helper output-length calculation in qwen3_omni_moe vs CPU/CUDA result mismatch in PPDocLayoutV3."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Model config import failure vs PIL image-processor torchvision dependency regression; different code paths."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Python version compatibility issue vs loading a specific model on transformers>=5.1.0; not the same underlying bug."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Image processor label reduction bug vs documentation mismatch for return_all_scores."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Docs wording issue vs unrecognized model/config import error; unrelated."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "Generation cache behavior vs ANSI codes in loading_report; different subsystems."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Config/model-type import failure vs AutoModelForCausalLM text_config unwrapping bug; different failures."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Both mention flex_attention and versioning, but one is a torch>=2.9 failure and the other is an incorrect version check; not the same concrete fix."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Notebook syntax error vs feature extraction helper issue; unrelated."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Gradient checkpointing eval-mode restriction vs D-FINE auxiliary-loss behavior; different bugs."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Saving dequantized Ministral models vs FlashAttention failure for Ministral-3; different symptoms and fixes."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 41,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:39692",
+        "issue:41950",
+        "issue:42796",
+        "issue:43295",
+        "issue:43299",
+        "issue:43381",
+        "issue:43441",
+        "issue:43479",
+        "issue:43577",
+        "issue:43650",
+        "issue:43824",
+        "issue:43901",
+        "issue:43976",
+        "issue:44016",
+        "issue:44062",
+        "issue:44075",
+        "issue:44220",
+        "issue:44279",
+        "issue:44327",
+        "issue:44351",
+        "issue:44368",
+        "issue:44466",
+        "issue:44521",
+        "issue:44779",
+        "issue:44792",
+        "issue:44933",
+        "issue:45020",
+        "issue:45216",
+        "issue:45276",
+        "issue:45278",
+        "issue:45335",
+        "issue:45357",
+        "issue:45375",
+        "issue:45405",
+        "issue:45584",
+        "issue:45588",
+        "issue:45812",
+        "issue:46082",
+        "issue:46123",
+        "issue:46369",
+        "issue:46399",
+        "issue:46424",
+        "issue:46515",
+        "issue:46614",
+        "issue:46693",
+        "issue:46752",
+        "issue:47000",
+        "issue:47085"
+      ],
+      "soft_pairs": [
+        "issue:43650|issue:44220",
+        "issue:44016|issue:44062",
+        "issue:43976|issue:44279",
+        "issue:44792|issue:46424",
+        "issue:43441|issue:45588",
+        "issue:44521|issue:46693",
+        "issue:43824|issue:44933",
+        "issue:44279|issue:46515",
+        "issue:43479|issue:45375",
+        "issue:45335|issue:45584",
+        "issue:43299|issue:45216",
+        "issue:44327|issue:46752",
+        "issue:46123|issue:46369",
+        "issue:46614|issue:47085",
+        "issue:44279|issue:45405",
+        "issue:43295|issue:44779",
+        "issue:39692|issue:43901",
+        "issue:42796|issue:43650",
+        "issue:44075|issue:44368",
+        "issue:43577|issue:45216",
+        "issue:44351|issue:45278",
+        "issue:41950|issue:45020",
+        "issue:43976|issue:45405",
+        "issue:44779|issue:45812",
+        "issue:43381|issue:47000",
+        "issue:46082|issue:46399",
+        "issue:44466|issue:45276",
+        "issue:43299|issue:45357"
+      ],
+      "prepared_review_unit_hash": "993d3b79949d83be98a0b6d6125860cc3de6eb09af9a7234aa84021456dbf4b6",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36070,
+        "estimated_input_tokens": 9018,
+        "estimated_eval_tokens": 18292
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "No soft pair looks like the same underlying bug/change with enough specificity to merge; most are unrelated across different models, pipelines, or regressions. The broadest representative issue in the set is the 5.5.0 import-regression report.",
+        "confidence": 0.74,
+        "canonical_issue_reason": "issue:45278 is the broadest, most umbrella-style regression report in the set; it can plausibly cover specific import failures after the 4.57.0\u21925.5.0 upgrade.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the strongest representative issue because it captures a high-impact upgrade regression affecting imports across the library, with clear version context.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43650",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Unrelated topics: a vague data request vs. an audio feature-extraction bug."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44062",
+            "accept": false,
+            "reason": "Notebook syntax error and tokenizer AddedToken argument collision are different failures."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Generic dependency complaint is too broad to match the Python-version-specific failure."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Different subsystems: janus image-generation test failure vs cache-length configuration behavior."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Both mention flash attention, but the models, call paths, and crash causes differ."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "Chat-template mask bug and FA2 performance regression are unrelated."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Both are import errors, but they concern different symbols and different code paths."
+          },
+          {
+            "left": "issue:44279",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "Dependency issue vs torch compatibility problem; no shared concrete bug."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Different model families and different config-handling bugs."
+          },
+          {
+            "left": "issue:45335",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Embedding resize propagation and Whisper empty-transcription generation are not the same issue."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both are Qwen-related regressions, but one is loading failure and the other is save_pretrained serialization."
+          },
+          {
+            "left": "issue:44327",
+            "right": "issue:46752",
+            "accept": false,
+            "reason": "Separate edge cases in unrelated pipelines: QA span decoding vs empty chat-template input."
+          },
+          {
+            "left": "issue:46123",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Partial-batch mask handling and model memory overflow are unrelated."
+          },
+          {
+            "left": "issue:46614",
+            "right": "issue:47085",
+            "accept": false,
+            "reason": "Tokenizer backend language-state bug vs tokenizer thread-safety regression."
+          },
+          {
+            "left": "issue:44279",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Package dependency/version pinning issue is not the same as a torch incompatibility report."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Different tokenizer regressions affecting different APIs and models."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Both are documentation-related, but the affected examples and mismatch types are different."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43650",
+            "accept": false,
+            "reason": "Tokenizer BPE-marker bug is unrelated to the vague data request."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Optimizer argument handling and a tied-weights warning in fine-tuning are not the same bug."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "BLIP2 dtype loading bug and Qwen3.5 save_pretrained regression affect different models and code paths."
+          },
+          {
+            "left": "issue:44351",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "HybridCache import failure could be part of a broader regression, but the broad issue is too general to treat them as the same concrete bug."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Video-classification processor mix-up and remote_code breakage are distinct problems."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Python support regression and PEFT version pinning are unrelated."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Different tokenizer regressions in different model families; not a single code-path bug."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "Gradient checkpointing eval-mode behavior and RTDetr loss infeasibility are unrelated."
+          },
+          {
+            "left": "issue:46082",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Llama config validation and Qwen3.5 rotary-kernel registration failure are different issues."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "Both touch embeddings/tied weights, but one is serialization inconsistency and the other is resize propagation."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Loading failure vs save_pretrained visual-key regression for related but different Qwen3.5 paths."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 42,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:33290",
+        "issue:41950",
+        "issue:42175",
+        "issue:42994",
+        "issue:43299",
+        "issue:43317",
+        "issue:43441",
+        "issue:43479",
+        "issue:43525",
+        "issue:43644",
+        "issue:43650",
+        "issue:43653",
+        "issue:43854",
+        "issue:43873",
+        "issue:43901",
+        "issue:43931",
+        "issue:43976",
+        "issue:44016",
+        "issue:44242",
+        "issue:44246",
+        "issue:44261",
+        "issue:44263",
+        "issue:44291",
+        "issue:44336",
+        "issue:44462",
+        "issue:44623",
+        "issue:44683",
+        "issue:44749",
+        "issue:45081",
+        "issue:45103",
+        "issue:45200",
+        "issue:45372",
+        "issue:45397",
+        "issue:45542",
+        "issue:45588",
+        "issue:45593",
+        "issue:45656",
+        "issue:45823",
+        "issue:45834",
+        "issue:45859",
+        "issue:45907",
+        "issue:46082",
+        "issue:46209",
+        "issue:46512",
+        "issue:46851",
+        "issue:47000",
+        "issue:47145"
+      ],
+      "soft_pairs": [
+        "issue:46082|issue:46512",
+        "issue:43653|issue:45593",
+        "issue:41950|issue:45823",
+        "issue:43650|issue:46209",
+        "issue:44242|issue:45200",
+        "issue:43644|issue:47145",
+        "issue:43299|issue:43854",
+        "issue:43931|issue:47145",
+        "issue:44462|issue:45081",
+        "issue:45588|issue:47000",
+        "issue:44683|issue:45542",
+        "issue:44016|issue:44291",
+        "issue:43479|issue:43525",
+        "issue:45081|issue:45907",
+        "issue:33290|issue:45656",
+        "issue:42994|issue:43317",
+        "issue:43441|issue:45834",
+        "issue:44336|issue:46851",
+        "issue:44623|issue:44749",
+        "issue:43873|issue:47145",
+        "issue:45103|issue:45372",
+        "issue:45859|issue:46851",
+        "issue:44261|issue:44263",
+        "issue:44623|issue:45397",
+        "issue:42175|issue:43976",
+        "issue:43901|issue:44246"
+      ],
+      "prepared_review_unit_hash": "974ec9a2f3fdf67784d7bbd42dc2c6960d47d714962c2c8aa693fa5197eeb4eb",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 26,
+        "serialized_chars": 34848,
+        "estimated_input_tokens": 8712,
+        "estimated_eval_tokens": 17680
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The candidate pairs are mostly unrelated bugs across different subsystems (quantization/offload, tokenizer loading, model configs, docs, CI, etc.). None of the soft edges looks like the same underlying defect, so all should be rejected.",
+        "confidence": 0.89,
+        "canonical_issue_reason": "Issue 43873 is the broadest and most actively discussed item in the set, but this cluster does not actually form a duplicate set around one bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43873 is the best available anchor because it is the most general and active report, though it is not a true canonical duplicate target for the rest of the items.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46082",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "Different bugs: LlamaConfig validation for head_dim vs from_config(dtype) leaving config.dtype stale."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Unrelated: BigBird tokenizer special-token registration vs D-FINE auxiliary-loss behavior."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Different areas and failures: video pipeline processor lookup vs Gemma4 device-mismatch during forward."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "No shared code-path: placeholder data request vs FP8 inference failure on B200."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Different mechanisms: MoE load-balancing loss gating vs Gemma 4 mm_token_type_ids defaulting."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:47145",
+            "accept": false,
+            "reason": "Both involve model loading/weights, but the defects differ: junk non-persistent buffers vs missing FP8 weight_scale application."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Distinct model-specific load failures (Qwen3VL MoE vs GLM-4.7-Flash unit-test loading)."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:47145",
+            "accept": false,
+            "reason": "Weight-shape mismatch for a specific Qwen3-VL model is not the same as FP8 frozen-checkpoint scaling omission."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "AutoTokenizer ignoring tokenizer.json is unrelated to the Mistral regex patch crash."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "FlashAttention AttributeError on s_aux=None is unrelated to RTDetrLoss infeasible cost-matrix failure."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Compiled flex_attention on torch>=2.9 is unrelated to missing TensorFlow backend installation."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Notebook syntax error and init_empty_weights argument handling are separate issues."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Phi4MultimodalConfig default initialization bug is unrelated to Llama4Config missing pad_token_id."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Tokenizer regex crash and list_repo_files revision forwarding are unrelated code paths."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Both mention DeepSpeed/optimizer, but one is OOM with Adafactor while the other is double optimizer stepping."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43317",
+            "accept": false,
+            "reason": "Quantized saving failure and dequantized offload loading failure are different bugs in adjacent areas."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "FlashAttention failure for Ministral-3 is unrelated to Kosmos2.5 OCR index errors."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "ANSI code emission in loading_report is unrelated to variable-length position_ids in Qwen3.5."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "processor.save_pretrained missing files and a Chinese-reported slowdown after upgrade are not the same defect."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:47145",
+            "accept": false,
+            "reason": "Both concern quantization/offloading, but one is general offloading behavior and the other is a specific FP8 checkpoint weight_scale omission."
+          },
+          {
+            "left": "issue:45103",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "auto_docstring annotations crash is unrelated to Gemma 4 processor loading failing on ReasoningEffort import."
+          },
+          {
+            "left": "issue:45859",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "CPU offload incompatibility in a rotary embedding is not the same bug as variable-length position_ids handling."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "MLA q_a_layernorm eps precision issue and GlmMoeDsaIndexer split-return handling are different code paths."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "processor.save_pretrained file omission is unrelated to gemma-4 zero3 from_pretrained."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "TensorFlow backend packaging is unrelated to a Python version compatibility regression."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "TextClassificationPipeline docs mismatch is unrelated to slow intermittent import times."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 43,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:41628",
+        "issue:42796",
+        "issue:43441",
+        "issue:43525",
+        "issue:43526",
+        "issue:43540",
+        "issue:43575",
+        "issue:43643",
+        "issue:43673",
+        "issue:43723",
+        "issue:43761",
+        "issue:43784",
+        "issue:43824",
+        "issue:43881",
+        "issue:43976",
+        "issue:44016",
+        "issue:44162",
+        "issue:44186",
+        "issue:44188",
+        "issue:44246",
+        "issue:44297",
+        "issue:44336",
+        "issue:44373",
+        "issue:44496",
+        "issue:44589",
+        "issue:44671",
+        "issue:44779",
+        "issue:44871",
+        "issue:44987",
+        "issue:45003",
+        "issue:45083",
+        "issue:45325",
+        "issue:45362",
+        "issue:45397",
+        "issue:45542",
+        "issue:45561",
+        "issue:45563",
+        "issue:45600",
+        "issue:45712",
+        "issue:45823",
+        "issue:46291",
+        "issue:46399",
+        "issue:46655",
+        "issue:46726",
+        "issue:46752",
+        "issue:46851",
+        "issue:47000",
+        "issue:47145"
+      ],
+      "soft_pairs": [
+        "issue:45600|issue:45712",
+        "issue:44671|issue:44779",
+        "issue:43673|issue:43761",
+        "issue:43525|issue:43881",
+        "issue:43575|issue:47145",
+        "issue:44016|issue:45823",
+        "issue:44186|issue:47000",
+        "issue:44188|issue:44336",
+        "issue:44987|issue:45542",
+        "issue:43824|issue:43976",
+        "issue:43525|issue:43643",
+        "issue:42796|issue:43441",
+        "issue:43723|issue:44162",
+        "issue:44336|issue:44373",
+        "issue:44589|issue:45362",
+        "issue:45325|issue:46851",
+        "issue:44246|issue:44496",
+        "issue:46291|issue:46399",
+        "issue:44871|issue:47145",
+        "issue:44297|issue:45397",
+        "issue:45083|issue:46851",
+        "issue:46655|issue:46726",
+        "issue:41628|issue:43976",
+        "issue:43784|issue:45542",
+        "issue:46399|issue:46851",
+        "issue:43526|issue:45003",
+        "issue:45561|issue:45563",
+        "issue:43540|issue:46752"
+      ],
+      "prepared_review_unit_hash": "07651e02d0f601fdad351045754d0c110400cda6d5e9b09968af476ce3c921ff",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36042,
+        "estimated_input_tokens": 9011,
+        "estimated_eval_tokens": 18278
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "No soft-edge pair looks like a true duplicate or mergeable same-change PR; the set is heterogeneous across unrelated bugs, regressions, and docs/testing issues.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "No single canonical issue stands out: the items span unrelated model, tokenizer, processor, config, and tooling problems rather than one underlying defect.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No global best issue for the cluster; the closest matches are only superficially similar and not the same concrete bug.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45600",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Both touch repo-maintenance/generated-artifact areas, but one is about stale Sam3LiteText auto-mappings and the other about leftover dummy PT classes; different failures and fixes."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both are tokenizer/model-output regressions in v5, but they concern different model families and different tokenization paths; not the same bug."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:43761",
+            "accept": false,
+            "reason": "Both are v5 regressions, but one is GenerationMixin cache behavior during chunked prefill and the other is CLIPVisionModel hidden_states handling; distinct code paths."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "A missing config field for Llama4 and a loading failure for glm-4v-9b are separate model-specific issues."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:47145",
+            "accept": false,
+            "reason": "One is an OOM when loading a Qwen model with tensor parallelism; the other is a checkpoint corruption bug from missing weight_scale application. Different defects."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Notebook syntax error and Gemma4 device-mapping mismatch are unrelated."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "LayoutLMv2 tokenizer crashes on NER/padding; RTDetrLoss has an infeasible cost matrix. Different subsystems and failure modes."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "torch.compile attention-kernel divergence is unrelated to ANSI output formatting in loading_report."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "A model-loading regression with physical-intelligence/fast is unrelated to the TensorFlow/tensorboard backend import error."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "ImportError for a missing Qwen2.5-VL class and a Python-version compatibility issue are not the same underlying bug."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "Both mention missing fields, but one is a model config attribute bug and the other is incomplete fields returned from AutoConfig with trust_remote_code; not the same issue."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Same model family, but tokenizer BPE-marker handling and FlashAttention failure are separate problems."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "AutoTokenizer loading in v5 and ESM2 breakage are unrelated model-loading issues."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "ANSI-code emission and a wrong docstring are unrelated."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "A missing Float8 storage type and a Qwen chat crash do not describe the same code path."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "Both involve position_ids, but they affect different Qwen model families and different logic (still-image scaling vs variable-length inputs)."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Import-time slowness and an unrecognized model/config.json error are unrelated."
+          },
+          {
+            "left": "issue:46291",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Both mention kernels, but one is an import-time dependency/version requirement and the other is a missing decorator for a specific registration path; not mergeable as one fix."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:47145",
+            "accept": false,
+            "reason": "Gemma eos_token_id inconsistency and FP8 weight_scale corruption are unrelated."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "Tokenizer class metadata mismatch and gemma-4 zero3 loading are different bugs."
+          },
+          {
+            "left": "issue:45083",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "Different Qwen3 Omni/Moe helper behavior versus Qwen3.5 variable-length position_ids; not the same defect."
+          },
+          {
+            "left": "issue:46655",
+            "right": "issue:46726",
+            "accept": false,
+            "reason": "Video padding calculation and missing-vision-backend image-processor reporting are distinct processor issues."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "AutoImageProcessor import failure and Python 3.9/3.10 compatibility are unrelated."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Sentence-transformers import NameError and TensorFlow backend detection error are separate import-time issues."
+          },
+          {
+            "left": "issue:46399",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "Qwen3.5 kernel registration crash and Qwen3.5 variable-length position_ids bug are different concrete failures."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "BeitImageProcessorFast reduce_labels bug and unsafe sys.modules access are unrelated."
+          },
+          {
+            "left": "issue:45561",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "pytest-xdist captured_info race and a stale generate() warning are different behaviors in different code paths."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:46752",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video-processing ValueError and empty-chat-template IndexError are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 44,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:36246",
+        "issue:41950",
+        "issue:42491",
+        "issue:42757",
+        "issue:42886",
+        "issue:43257",
+        "issue:43262",
+        "issue:43299",
+        "issue:43531",
+        "issue:43540",
+        "issue:43550",
+        "issue:43901",
+        "issue:44220",
+        "issue:44242",
+        "issue:44246",
+        "issue:44295",
+        "issue:44336",
+        "issue:44351",
+        "issue:44451",
+        "issue:44479",
+        "issue:44485",
+        "issue:44496",
+        "issue:44749",
+        "issue:44779",
+        "issue:44871",
+        "issue:45003",
+        "issue:45072",
+        "issue:45103",
+        "issue:45137",
+        "issue:45292",
+        "issue:45362",
+        "issue:45375",
+        "issue:45397",
+        "issue:45561",
+        "issue:45588",
+        "issue:45593",
+        "issue:45759",
+        "issue:46132",
+        "issue:46399",
+        "issue:46506",
+        "issue:46614",
+        "issue:46655",
+        "issue:46665",
+        "issue:46752",
+        "issue:46821",
+        "issue:46829",
+        "issue:46851"
+      ],
+      "soft_pairs": [
+        "issue:45103|issue:46399",
+        "issue:45072|issue:46829",
+        "issue:44246|issue:44871",
+        "issue:44295|issue:45397",
+        "issue:43531|issue:45362",
+        "issue:46614|issue:46665",
+        "issue:45588|issue:46851",
+        "issue:42491|issue:43299",
+        "issue:44336|issue:44485",
+        "issue:45003|issue:46821",
+        "issue:44242|issue:45593",
+        "issue:43550|issue:46506",
+        "issue:44779|issue:45292",
+        "issue:42886|issue:46132",
+        "issue:36246|issue:42757",
+        "issue:36246|issue:44351",
+        "issue:43901|issue:45003",
+        "issue:43257|issue:45137",
+        "issue:44496|issue:45003",
+        "issue:43540|issue:46655",
+        "issue:41950|issue:44220",
+        "issue:43262|issue:44246",
+        "issue:44451|issue:46752",
+        "issue:44479|issue:46655",
+        "issue:45375|issue:45759",
+        "issue:44295|issue:44749",
+        "issue:45003|issue:45561"
+      ],
+      "prepared_review_unit_hash": "baacc1cc5f30173ff5c8e9efb8e4cdf990013c597ceeaf13d02640b502215b61",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 27,
+        "serialized_chars": 35295,
+        "estimated_input_tokens": 8824,
+        "estimated_eval_tokens": 17904
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a grab bag of unrelated Transformers bugs and regressions, with a loose concentration around Qwen/model-loading and video-processing issues but no true duplicate pair that clearly shares the same concrete fix.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "issue:44479 is the broadest, most representative report in the set: it describes a clear regression in v5.3.0 affecting multiple Qwen video/VL model families, making it a better anchor than narrower or more incidental failures.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44479 is the best standalone issue to keep because it is specific enough to be actionable yet broad enough to represent a real family of related regressions.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45103",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Both are crashes, but they hit different code paths: auto-docstring annotation handling vs Qwen3.5 rotary-kernel registration."
+          },
+          {
+            "left": "issue:45072",
+            "right": "issue:46829",
+            "accept": false,
+            "reason": "Both involve vision/video models, but one is a bfloat16 dtype mismatch and the other is an output-field mixup."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Import slowdown and eos-token config inconsistency are unrelated bugs."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "One is a position_ids buffer access error; the other is a DeepSpeed zero3 from_pretrained problem."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "Different failures in different Qwen models: sliding_window behavior vs chat-loading crash."
+          },
+          {
+            "left": "issue:46614",
+            "right": "issue:46665",
+            "accept": false,
+            "reason": "Stale phonemizer backend state is not the same as from_pretrained behavior inside Python classes."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "Flash-attention sink handling and variable-length position_ids are different bugs."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43299",
+            "accept": false,
+            "reason": "Both mention Qwen MoE loading, but the triggers and failure modes differ enough that they are not the same bug."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "ANSI-output cleanup and GLM-5 RoPE implementation are unrelated."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Unsafe sys.modules access is not the same as a TimesFM window_size AttributeError."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Both concern auxiliary losses, but for different models and different gating conditions."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure and CPU-vs-CUDA divergence are different code-path problems."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:45292",
+            "accept": false,
+            "reason": "Tokenizer correctness regression and resize_token_embeddings output sync are unrelated."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Offline cache loading and AttentionInterface.register behavior are not the same issue."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are import errors, but they involve different missing symbols and different dependencies."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Different missing imports: Qwen2_5_VLImageProcessor vs HybridCache."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Docs mentioning return_all_scores and sys.modules access are unrelated."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "DeepSpeed-related, but one is weight conversion with accelerate while the other is a ZeRO3 deque underflow."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Unrecognized model config and unsafe sys.modules access are different bugs."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:46655",
+            "accept": false,
+            "reason": "Both are video-related, but they affect different models and likely different processor logic."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Video-classification processor lookup and torch fbank feature extraction are separate subsystems."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "Audio chat-template sampling rate and import slowness are unrelated."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:46752",
+            "accept": false,
+            "reason": "Model loading failure and empty-list chat-template IndexError are different failure modes."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:46655",
+            "accept": false,
+            "reason": "Both are video-input regressions, but they target different model families and different underlying processor bugs."
+          },
+          {
+            "left": "issue:45375",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both touch Qwen multimodal config handling, but one drops a config field while the other fails to unwrap text_config."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "A position_ids buffer error is not the same as a broad performance regression after version upgrade."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "sys.modules access in modeling_utils and a pytest-xdist file race are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 45,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:29870",
+        "issue:30333",
+        "issue:36246",
+        "issue:42199",
+        "issue:43065",
+        "issue:43295",
+        "issue:43352",
+        "issue:43408",
+        "issue:43475",
+        "issue:43526",
+        "issue:43540",
+        "issue:43596",
+        "issue:43653",
+        "issue:44038",
+        "issue:44060",
+        "issue:44077",
+        "issue:44162",
+        "issue:44242",
+        "issue:44368",
+        "issue:44466",
+        "issue:44479",
+        "issue:44623",
+        "issue:44857",
+        "issue:44877",
+        "issue:44977",
+        "issue:44987",
+        "issue:45137",
+        "issue:45200",
+        "issue:45250",
+        "issue:45292",
+        "issue:45335",
+        "issue:45375",
+        "issue:45397",
+        "issue:45412",
+        "issue:45459",
+        "issue:45600",
+        "issue:45656",
+        "issue:45859",
+        "issue:46123",
+        "issue:46170",
+        "issue:46369",
+        "issue:46399",
+        "issue:46506",
+        "issue:46726",
+        "issue:46851",
+        "issue:46894",
+        "issue:46897",
+        "issue:47000"
+      ],
+      "soft_pairs": [
+        "issue:44162|issue:45200",
+        "issue:45292|issue:47000",
+        "issue:45859|issue:46399",
+        "issue:45137|issue:45600",
+        "issue:43295|issue:45292",
+        "issue:36246|issue:43596",
+        "issue:43475|issue:46369",
+        "issue:44466|issue:45335",
+        "issue:44623|issue:46726",
+        "issue:44242|issue:44857",
+        "issue:44038|issue:46506",
+        "issue:44877|issue:45375",
+        "issue:43526|issue:46897",
+        "issue:44987|issue:45459",
+        "issue:29870|issue:30333",
+        "issue:43352|issue:45250",
+        "issue:43526|issue:46894",
+        "issue:44479|issue:46170",
+        "issue:42199|issue:46851",
+        "issue:44060|issue:44077",
+        "issue:44368|issue:44977",
+        "issue:43653|issue:44162",
+        "issue:45397|issue:45656",
+        "issue:45412|issue:46123",
+        "issue:43065|issue:43408",
+        "issue:44877|issue:46170",
+        "issue:43475|issue:43540"
+      ],
+      "prepared_review_unit_hash": "df406a38ebec16452e9731c37716df591091c0dffed678c71727e11d57e28367",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 35932,
+        "estimated_input_tokens": 8983,
+        "estimated_eval_tokens": 18222
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is highly heterogeneous: the issues cover unrelated bugs across different models, processors, configs, and training/runtime paths. All soft-edge pairs look like false positives rather than the same underlying defect.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44162",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Different model families and failures: ESM2 is broadly broken, while Gemma 4 needs default mm_token_type_ids for text-only fine-tuning."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "Unrelated code paths: token embedding resize propagation vs RT-DETR loss infeasible cost matrix."
+          },
+          {
+            "left": "issue:45859",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Both involve Qwen3.5, but the bugs differ: CPU-offload rotary embedding compatibility vs a use_kernels registration/decorator crash."
+          },
+          {
+            "left": "issue:45137",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO3 deque pop failure is unrelated to auto_mappings referencing removed Sam3LiteText configs."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45292",
+            "accept": false,
+            "reason": "Different regressions: processor.tokenizer/images handling vs resize_token_embeddings not updating output embeddings."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:43596",
+            "accept": false,
+            "reason": "Completely different failures: Qwen2_5_VL import error vs DeepSpeed ZeRO3/BertModel index error."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Same broad SAM3 area, but distinct bugs: missing fpn_position_embeddings vs memory overflow in sam3-litetext."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "Both touch tied/embedding weights, but one is a serialization inconsistency and the other is resize propagation in t5gemma."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:46726",
+            "accept": false,
+            "reason": "Processor save_pretrained file omission is not the same as AutoProcessor failing when a vision backend is missing."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Different subsystems: MoE router loss omission vs LwDetrImageLoss AMP/CUDA crash."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe/transformers 5.0 issue is unrelated to PPDocLayoutV3 CPU-vs-CUDA divergence."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both involve config validation, but for different models and different missing-field behaviors."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:46897",
+            "accept": false,
+            "reason": "BEiT label reduction bug is unrelated to Florence2 training-loss double-shift behavior."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "Both are load/parse failures, but one is a model-loading regression and the other is tokenizer error masking without protobuf."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:30333",
+            "accept": false,
+            "reason": "Unrelated issues: BART generate() decoder_attention_mask bug vs MLflow run status not being updated on training failure."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:45250",
+            "accept": false,
+            "reason": "Generic Flash Attention topic, but one is a model support error and the other is just the feature request/topic."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:46894",
+            "accept": false,
+            "reason": "Same pattern of label/count bug is not established; BEiT reduce_labels and CohereASR double-shift are different model-specific failures."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Different regressions: Qwen2_5/qwen3_vl video-input behavior vs Windows ROCm import/type errors from torch.distributed dependency."
+          },
+          {
+            "left": "issue:42199",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "DETR cardinality/background-class logic is unrelated to Qwen3.5 variable-length position_ids handling."
+          },
+          {
+            "left": "issue:44060",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "Different issues in different models: tied-weights warning in Qwen3-Next vs patchtsmixer post_init policy."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Both mention Qwen3.5 and warnings/flash-attn, but one is tied-embeddings warnings during LoRA fine-tuning, the other is a generation regression."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration bug is unrelated to the broad ESM2 breakage report."
+          },
+          {
+            "left": "issue:45397",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Different DeepSpeed-related problems: Gemma-4 zero3 from_pretrained vs optimizer.step being called twice."
+          },
+          {
+            "left": "issue:45412",
+            "right": "issue:46123",
+            "accept": false,
+            "reason": "RT-DETR memory leak on deletion is unrelated to MaskGenerationPipeline dropping final partial batch results."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:43408",
+            "accept": false,
+            "reason": "SAM3 dummy Conv2d in pixel decoder and sam3_video vs sam3_tracker warning are distinct issues."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Config strictness bug for granite_speech is unrelated to Windows ROCm/torch.distributed import failures."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Both are model-specific, but SAM3 vision encoder output missing a field is not the same as Qwen3OmniMoe video input validation errors."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 46,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:42886",
+        "issue:42994",
+        "issue:43065",
+        "issue:43482",
+        "issue:43646",
+        "issue:43701",
+        "issue:43723",
+        "issue:43906",
+        "issue:43931",
+        "issue:43994",
+        "issue:44162",
+        "issue:44188",
+        "issue:44297",
+        "issue:44361",
+        "issue:44368",
+        "issue:44483",
+        "issue:44623",
+        "issue:44749",
+        "issue:44871",
+        "issue:44964",
+        "issue:44995",
+        "issue:45103",
+        "issue:45137",
+        "issue:45200",
+        "issue:45229",
+        "issue:45375",
+        "issue:45478",
+        "issue:45563",
+        "issue:45588",
+        "issue:45593",
+        "issue:45656",
+        "issue:45685",
+        "issue:45987",
+        "issue:46082",
+        "issue:46123",
+        "issue:46132",
+        "issue:46170",
+        "issue:46202",
+        "issue:46421",
+        "issue:46431",
+        "issue:46710",
+        "issue:46726",
+        "issue:46752",
+        "issue:46832",
+        "issue:46858",
+        "issue:47020",
+        "issue:47068",
+        "issue:47145"
+      ],
+      "soft_pairs": [
+        "issue:46170|issue:46832",
+        "issue:42886|issue:45229",
+        "issue:43723|issue:46710",
+        "issue:43931|issue:44368",
+        "issue:44188|issue:46202",
+        "issue:44995|issue:45200",
+        "issue:44368|issue:45563",
+        "issue:46132|issue:47145",
+        "issue:44483|issue:46202",
+        "issue:43906|issue:46202",
+        "issue:44188|issue:46421",
+        "issue:44995|issue:45588",
+        "issue:45593|issue:47068",
+        "issue:45685|issue:46431",
+        "issue:45987|issue:46082",
+        "issue:44361|issue:44749",
+        "issue:45103|issue:46858",
+        "issue:44297|issue:44623",
+        "issue:43065|issue:46123",
+        "issue:43994|issue:46726",
+        "issue:44749|issue:45656",
+        "issue:43701|issue:44749",
+        "issue:43646|issue:44162",
+        "issue:44623|issue:45656",
+        "issue:46752|issue:47020",
+        "issue:44871|issue:46132",
+        "issue:44964|issue:45137",
+        "issue:45375|issue:45478",
+        "issue:42994|issue:43482"
+      ],
+      "prepared_review_unit_hash": "4eea9cb64ac6a5eb432e18f3593aa51049bd2382a63d0ecc892b6aa42246fc07",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36552,
+        "estimated_input_tokens": 9138,
+        "estimated_eval_tokens": 18532
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most items are unrelated individual bug reports across tokenizers, model loading, training, and inference. The only clear duplicate-like pair is the Qwen3.5-Moe config/loading issue (45375/45478), where the specific missing field report appears to explain the broader from_pretrained failure.",
+        "confidence": 0.72,
+        "canonical_issue_reason": "issue:45375 is the most specific root-cause report: it names the missing `deepstack_visual_indexes` field and the `@strict` drop behavior, which plausibly explains the broader loading failure in 45478.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45375 is the best representative issue in this cluster because it is the clearest actionable description of the underlying bug, not just the downstream symptom.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46170",
+            "right": "issue:46832",
+            "accept": false,
+            "reason": "Different failures: Windows/ROCm import dependency vs RT-DETR feature-level shape mismatch."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Tokenizer offline cache loading and Gemma4 multi-GPU CUDA OOM are unrelated."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:46710",
+            "accept": false,
+            "reason": "Tokenizer loading regression vs model accuracy regression after upgrade; different code paths and symptoms."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Qwen3-VL weight-shape mismatch and a tie_word_embeddings warning are not the same bug."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "Attention kernel divergence under torch.compile is unrelated to Gemma pipeline roundtrip behavior."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Stale cache/indexer crash and missing mm_token_type_ids default are different issues."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Both are warnings, but they concern different settings and code paths."
+          },
+          {
+            "left": "issue:46132",
+            "right": "issue:47145",
+            "accept": false,
+            "reason": "Registered attention function behavior change is unrelated to FP8 checkpoint loading corruption."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "Chat completions request handling and Gemma pipeline roundtrip failure are not the same underlying bug."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "An isolated reproduction of another issue is not the same as the Gemma pipeline roundtrip bug."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:46421",
+            "accept": false,
+            "reason": "torch.compile attention branching divergence and cache lazy initialization recompilation are different problems."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "A stale second-forward cache bug is unrelated to the flash-attention `s_aux=None` crash."
+          },
+          {
+            "left": "issue:45593",
+            "right": "issue:47068",
+            "accept": false,
+            "reason": "D-FINE auxiliary-loss behavior and a DeepSpeed Python-conditional host sync concern are unrelated."
+          },
+          {
+            "left": "issue:45685",
+            "right": "issue:46431",
+            "accept": false,
+            "reason": "MPS histogram implementation error and missing chat template handling are different bugs."
+          },
+          {
+            "left": "issue:45987",
+            "right": "issue:46082",
+            "accept": false,
+            "reason": "Cache sequence-length return-type mismatch and LlamaConfig head_dim validation are unrelated."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "MLukeTokenizer AttributeError and a Transformers upgrade slowdown are not the same code-path failure."
+          },
+          {
+            "left": "issue:45103",
+            "right": "issue:46858",
+            "accept": false,
+            "reason": "auto_docstring annotation handling and static-cache sequence-length arithmetic are different issues."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Tokenizer save-pretrained metadata mismatch and processor save-pretrained missing files are related only broadly, not the same concrete bug."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:46123",
+            "accept": false,
+            "reason": "Dummy Conv2d in Sam3PixelDecoder and MaskGenerationPipeline partial-batch dropping are unrelated."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:46726",
+            "accept": false,
+            "reason": "Siglip2 bad outputs and missing vision-backend image-processor handling are different issues."
+          },
+          {
+            "left": "issue:44749",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Performance regression after upgrade and deepspeed optimizer-step double-calling are unrelated."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "resume_from_checkpoint key mismatch and upgrade-related slowdown do not share a concrete failure mode."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "Both are broad regressions, but they affect different models and different initialization paths."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Processor save-pretrained file emission and DeepSpeed optimizer-step duplication are unrelated."
+          },
+          {
+            "left": "issue:46752",
+            "right": "issue:47020",
+            "accept": false,
+            "reason": "Empty-list chat template crash and XLMRobertaTokenizer Unigram initialization error are different tokenizer bugs."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Gemma eos_token_id mismatch and AttentionInterface.register behavior change are unrelated."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Phi-4 multimodal loading and DeepSpeed ZeRO3 deque underflow are unrelated."
+          },
+          {
+            "left": "issue:45375",
+            "right": "issue:45478",
+            "accept": true,
+            "reason": "Both point to the same Qwen3.5-Moe v5.4+ loading problem; the missing `deepstack_visual_indexes` field and its strict-drop behavior plausibly cause the broader from_pretrained failure."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Quantized model saving failure and Qwen2.5-GGUF loading failure are distinct model-format problems."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The only plausible duplicate-like pair is 45375/45478, and the rationale is appropriately tentative. The rejected soft pairs also look cautious rather than overstated. No mergeability issue since these are issue pairs, not PRs."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 47,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28282",
+        "issue:30990",
+        "issue:36010",
+        "issue:39692",
+        "issue:41628",
+        "issue:42371",
+        "issue:42994",
+        "issue:43232",
+        "issue:43475",
+        "issue:43525",
+        "issue:43643",
+        "issue:43673",
+        "issue:43723",
+        "issue:43784",
+        "issue:43824",
+        "issue:43873",
+        "issue:43986",
+        "issue:44038",
+        "issue:44188",
+        "issue:44351",
+        "issue:44373",
+        "issue:44623",
+        "issue:44857",
+        "issue:44869",
+        "issue:44933",
+        "issue:44945",
+        "issue:45081",
+        "issue:45092",
+        "issue:45103",
+        "issue:45137",
+        "issue:45230",
+        "issue:45310",
+        "issue:45356",
+        "issue:45375",
+        "issue:45397",
+        "issue:45561",
+        "issue:45593",
+        "issue:45753",
+        "issue:46326",
+        "issue:46399",
+        "issue:46463",
+        "issue:46726",
+        "issue:46821",
+        "issue:46851",
+        "issue:46877",
+        "issue:47000",
+        "issue:47059"
+      ],
+      "soft_pairs": [
+        "issue:39692|issue:46726",
+        "issue:43475|issue:46821",
+        "issue:43475|issue:43643",
+        "issue:44188|issue:46851",
+        "issue:43232|issue:43673",
+        "issue:45230|issue:45397",
+        "issue:43986|issue:46726",
+        "issue:28282|issue:30990",
+        "issue:46877|issue:47000",
+        "issue:45310|issue:45375",
+        "issue:44373|issue:46851",
+        "issue:44038|issue:44623",
+        "issue:46463|issue:46877",
+        "issue:41628|issue:44933",
+        "issue:42994|issue:43525",
+        "issue:44869|issue:45356",
+        "issue:43723|issue:46326",
+        "issue:42371|issue:45092",
+        "issue:43873|issue:44945",
+        "issue:43784|issue:46399",
+        "issue:45137|issue:47000",
+        "issue:41628|issue:44351",
+        "issue:36010|issue:41628",
+        "issue:44945|issue:45753",
+        "issue:42994|issue:43824",
+        "issue:44857|issue:46877",
+        "issue:45561|issue:47059",
+        "issue:44038|issue:45310",
+        "issue:45081|issue:45103",
+        "issue:45593|issue:46877"
+      ],
+      "prepared_review_unit_hash": "2bab67524f9e95b54492585d345a98d7fc09b0dfc24a9e00f26e0b1bb2f1bd87",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 30,
+        "serialized_chars": 35937,
+        "estimated_input_tokens": 8985,
+        "estimated_eval_tokens": 18226
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is dominated by assorted transformers API regressions, especially missing imports/attributes, but the soft-linked items mostly describe different concrete failures. No PRs are present.",
+        "confidence": 0.74,
+        "canonical_issue_reason": "issue:41628 is the clearest representative of the cluster: a core missing-symbol import regression in transformers, and it has the strongest soft links to other import/export failures.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:41628 is the most central and broadly representative issue here, with multiple related missing-import soft links and the most generic user-facing API breakage.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:39692",
+            "right": "issue:46726",
+            "accept": false,
+            "reason": "Different problems: a SigLIP2 doc example mismatch/quantization failure vs AutoProcessor behavior when vision backend is missing."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "Different model-specific attribute errors on unrelated code paths."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "Sam3 vision encoder output attribute bug is unrelated to AutoConfig losing fields with trust_remote_code."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "Attention-kernel branching under torch.compile is unrelated to Qwen3.5 position_ids handling."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "Both are generation-related, but they describe different failures: sync_gpus model-kwargs handling vs missing cache in chunked prefill."
+          },
+          {
+            "left": "issue:45230",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "issue:45230 is too generic to establish identity, and the specific Gemma-4 ZeRO3 from_pretrained bug is distinct."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:46726",
+            "accept": false,
+            "reason": "Both involve missing vision deps, but one is AutoProcessor for video models and the other is an unrecognized image processor path; not the same concrete bug."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:30990",
+            "accept": false,
+            "reason": "PyTorch-missing ImportError and Sentence Transformers loading hang are unrelated failures."
+          },
+          {
+            "left": "issue:46877",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "Doc-lint stdout noise at import time is unrelated to RTDetrLoss infeasible cost-matrix crashes."
+          },
+          {
+            "left": "issue:45310",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Related Qwen3.5 family, but one is a from_pretrained error and the other is a specific missing config field being dropped; too different to treat as one bug."
+          },
+          {
+            "left": "issue:44373",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "Wrong docstring vs an actual runtime bug in variable-length position_ids handling."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe incompatibility and processor.save_pretrained missing files are different issues."
+          },
+          {
+            "left": "issue:46463",
+            "right": "issue:46877",
+            "accept": false,
+            "reason": "Incorrect T5 validation error message is unrelated to doc-lint stdout diagnostics."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Both are import-related, but they appear to be different missing symbols in different modules, not the same underlying export bug."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Quantized model saving failure is unrelated to Llama4Config missing pad_token_id."
+          },
+          {
+            "left": "issue:44869",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Whisper timestamp decoding crash and Kimi tokenizer codec regression are different tokenizer bugs."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "AutoTokenizer loading issue and mistral_common continue_final_message bug are separate tokenizer paths."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "TF32 settings API guidance is unrelated to InternVL2 meta-initialization incompatibility."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Quantization offloading and pipeline-parallel output corruption are different runtime problems."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Sentence-transformers import NameError and Qwen3.5 kernel registration crash are unrelated."
+          },
+          {
+            "left": "issue:45137",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO3 deque underflow and RTDetrLoss infeasible cost matrix are different crashes."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are missing-import reports, but they concern different symbols and likely different export surfaces."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "AutoModel missing PyTorch and AutoImageProcessor missing import are different dependency vs export failures."
+          },
+          {
+            "left": "issue:44945",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Pipeline parallelism output corruption and Qwen3_5 infinite-loop-on-image are unrelated generation bugs."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Quantized saving failure is unrelated to a missing Qwen2_5_VLForConditionalGeneration export."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:46877",
+            "accept": false,
+            "reason": "AMP/CUDA loss crash and import-time doc-lint stdout are unrelated."
+          },
+          {
+            "left": "issue:45561",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "pytest-xdist race on captured_info.txt and CLI failure with huggingface_hub are different test/CLI issues."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Both mention Qwen-related models, but one is a broader transformers 5.0 VL-Moe bug and the other is a specific from_pretrained error; not the same code-path problem."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Mistral regex patch tokenizer crash and auto_docstring annotations crash are unrelated AttributeErrors."
+          },
+          {
+            "left": "issue:45593",
+            "right": "issue:46877",
+            "accept": false,
+            "reason": "D-FINE auxiliary-loss behavior and doc-lint stdout diagnostics are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 48,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:30333",
+        "issue:30990",
+        "issue:33290",
+        "issue:35141",
+        "issue:39692",
+        "issue:42175",
+        "issue:42673",
+        "issue:42831",
+        "issue:42898",
+        "issue:43232",
+        "issue:43479",
+        "issue:43704",
+        "issue:43723",
+        "issue:43784",
+        "issue:43824",
+        "issue:43976",
+        "issue:44016",
+        "issue:44079",
+        "issue:44186",
+        "issue:44222",
+        "issue:44496",
+        "issue:44568",
+        "issue:44623",
+        "issue:44661",
+        "issue:44743",
+        "issue:44779",
+        "issue:44792",
+        "issue:44869",
+        "issue:44933",
+        "issue:45310",
+        "issue:45405",
+        "issue:45406",
+        "issue:45561",
+        "issue:45600",
+        "issue:45712",
+        "issue:45812",
+        "issue:46077",
+        "issue:46123",
+        "issue:46202",
+        "issue:46231",
+        "issue:46693",
+        "issue:46701",
+        "issue:46829",
+        "issue:46877",
+        "issue:46909",
+        "issue:47059",
+        "issue:47068",
+        "issue:47085"
+      ],
+      "soft_pairs": [
+        "issue:45406|issue:46202",
+        "issue:44661|issue:47059",
+        "issue:44222|issue:45310",
+        "issue:46202|issue:46701",
+        "issue:42673|issue:43704",
+        "issue:43232|issue:44792",
+        "issue:42831|issue:44222",
+        "issue:42175|issue:44933",
+        "issue:44496|issue:45561",
+        "issue:45405|issue:46877",
+        "issue:30333|issue:33290",
+        "issue:44743|issue:47068",
+        "issue:43976|issue:44933",
+        "issue:30990|issue:35141",
+        "issue:42898|issue:44568",
+        "issue:45561|issue:45600",
+        "issue:43479|issue:46829",
+        "issue:42175|issue:45712",
+        "issue:44568|issue:47085",
+        "issue:44623|issue:45310",
+        "issue:42673|issue:46693",
+        "issue:44186|issue:47085",
+        "issue:45600|issue:47059",
+        "issue:46077|issue:46123",
+        "issue:42898|issue:45812",
+        "issue:43723|issue:44568",
+        "issue:43479|issue:44869",
+        "issue:44016|issue:46231",
+        "issue:39692|issue:46909",
+        "issue:43784|issue:43824",
+        "issue:42898|issue:44779",
+        "issue:43479|issue:44079"
+      ],
+      "prepared_review_unit_hash": "2d5494f48bd40d8ddb442d92cb2e4fea2e2c21c8a7fd0049d500a5ee77936064",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37344,
+        "estimated_input_tokens": 9336,
+        "estimated_eval_tokens": 18928
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly unrelated issue cluster; the only clear duplicate pair is the Qwen3ForCausalLM VRAM leak report (#42673/#43704).",
+        "confidence": 0.94,
+        "canonical_issue_reason": "#42673 is the earliest and clearest report of the exact duplicate VRAM-leak bug, matched verbatim by #43704.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Best issue anchor in this set: it has the strongest duplicate signal and a concrete, reproducible bug description.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45406",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "Related to processor/tokenization behavior, but these are different bugs and different code paths."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "Tokenizer mapping/model-addition failure vs CLI completion API break; not the same issue."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Both involve FP8/MoE, but one is save_pretrained and the other is from_pretrained/config loading; too different to merge."
+          },
+          {
+            "left": "issue:46202",
+            "right": "issue:46701",
+            "accept": false,
+            "reason": "Both touch generation/pipeline behavior, but the underlying failures are unrelated."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:43704",
+            "accept": true,
+            "reason": "Exact same title and same VRAM leak in Qwen3ForCausalLM with multiple dataloader threads; clear duplicate."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44792",
+            "accept": false,
+            "reason": "Generation kwargs/sync_gpus bug vs Janus image generation test failure; different concrete problems."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Accuracy regression in FineGrainedFP8 is not the same as FP8 MoE save_pretrained behavior."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Packaging/backend omission vs nonexistent import from image_utils; unrelated."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Model config recognition error vs pytest-xdist file race; no shared bug."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:46877",
+            "accept": false,
+            "reason": "Dependency pin issue vs noisy doc-lint output at import time; different concerns."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "MLflow reporting on failed training vs an OOM in DeepSpeed Adafactor; not duplicates."
+          },
+          {
+            "left": "issue:44743",
+            "right": "issue:47068",
+            "accept": false,
+            "reason": "Recurrent-state reset in modular_qwen3_5.py vs possible host sync in deepspeed loss; different code paths."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Python version compatibility vs import error from image_utils; unrelated."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Loading hang in Sentence Transformers vs embedding reinitialization after resize; different bugs."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both are tokenization regressions, but one is cleanup-space behavior and the other is add_special_tokens/BOS-EOS handling."
+          },
+          {
+            "left": "issue:45561",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "pytest-xdist race condition vs stale auto_mappings references; not the same issue."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:46829",
+            "accept": false,
+            "reason": "Multimodal config defaults vs hidden_states assignment in image classifier output; different failures."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Missing TensorFlow backend in extras vs leftover dummy classes leaking into dir(); unrelated."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:47085",
+            "accept": false,
+            "reason": "Special-token regression vs tokenizer thread-safety regression; same area, different bugs."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "processor.save_pretrained missing files vs Qwen3.5 MoE from_pretrained error; not the same concrete path."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:46693",
+            "accept": false,
+            "reason": "Same model family, but VRAM leak in dataloader threads is different from a long-context FA2 performance regression."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:47085",
+            "accept": false,
+            "reason": "LayoutLMv2 tokenizer crash on NER/padding vs Rust tokenizer thread safety; different symptoms and paths."
+          },
+          {
+            "left": "issue:45600",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "CI config reference cleanup vs CLI completion regression; unrelated."
+          },
+          {
+            "left": "issue:46077",
+            "right": "issue:46123",
+            "accept": false,
+            "reason": "Bidirectional attention mask bug vs partial-batch is_last bug in a pipeline; different code."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are tokenization regressions in v5, but they affect different models and behaviors."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Tokenizer loading from_pretrained issue vs add_special_tokens regression; not duplicates."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44869",
+            "accept": false,
+            "reason": "Multimodal config initialization bug vs Whisper timestamp decode crash; unrelated."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:46231",
+            "accept": false,
+            "reason": "Notebook syntax error vs documentation parameter mismatch; different scope."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:46909",
+            "accept": false,
+            "reason": "SigLIP2 example errors vs DIA documentation example breakage; not the same underlying bug."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Sentence-transformers import NameError vs Qwen2.5-VL import error; unrelated."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both involve v5 tokenizer behavior, but DeepSeek token correctness and clean_up_tokenization_spaces are distinct issues."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Config default initialization bug vs ModelOutput key assignment bug; different code paths."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet metadata. The only accepted duplicate pair, issue:42673/issue:43704, has identical titles and clearly matches the same VRAM-leak report; the rest of the soft-edge rejections are conservative and do not overstate equivalence. No mergeability concerns apply since this is an issue cluster, not a PR pair."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 49,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:22355",
+        "issue:27957",
+        "issue:28282",
+        "issue:30333",
+        "issue:34689",
+        "issue:35141",
+        "issue:41628",
+        "issue:42222",
+        "issue:42673",
+        "issue:42757",
+        "issue:42796",
+        "issue:42831",
+        "issue:42898",
+        "issue:43116",
+        "issue:43295",
+        "issue:43479",
+        "issue:43644",
+        "issue:43723",
+        "issue:43874",
+        "issue:44060",
+        "issue:44075",
+        "issue:44190",
+        "issue:44222",
+        "issue:44295",
+        "issue:44297",
+        "issue:44360",
+        "issue:44373",
+        "issue:44448",
+        "issue:44485",
+        "issue:44623",
+        "issue:44734",
+        "issue:44843",
+        "issue:44933",
+        "issue:45092",
+        "issue:45103",
+        "issue:45125",
+        "issue:45200",
+        "issue:45245",
+        "issue:45310",
+        "issue:45356",
+        "issue:45405",
+        "issue:45656",
+        "issue:45712",
+        "issue:46132",
+        "issue:46231",
+        "issue:46762",
+        "issue:47085"
+      ],
+      "soft_pairs": [
+        "issue:44843|issue:47085",
+        "issue:27957|issue:30333",
+        "issue:42673|issue:47085",
+        "issue:30333|issue:34689",
+        "issue:42898|issue:43295",
+        "issue:43874|issue:45092",
+        "issue:42222|issue:45200",
+        "issue:44295|issue:45656",
+        "issue:44734|issue:45103",
+        "issue:42898|issue:45356",
+        "issue:30333|issue:35141",
+        "issue:22355|issue:28282",
+        "issue:44222|issue:44297",
+        "issue:44075|issue:46231",
+        "issue:43116|issue:44190",
+        "issue:45405|issue:45712",
+        "issue:44297|issue:45656",
+        "issue:44933|issue:45712",
+        "issue:42831|issue:44623",
+        "issue:44448|issue:46132",
+        "issue:44060|issue:45125",
+        "issue:44222|issue:44623",
+        "issue:42796|issue:43723",
+        "issue:43479|issue:43644",
+        "issue:45245|issue:46132",
+        "issue:44373|issue:44485",
+        "issue:44360|issue:46762",
+        "issue:41628|issue:42757",
+        "issue:44297|issue:45310"
+      ],
+      "prepared_review_unit_hash": "2489a127e51753f959348b9acdfc563cbdb43878f8af513019a3938318b98da2",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 35707,
+        "estimated_input_tokens": 8927,
+        "estimated_eval_tokens": 18110
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is heterogeneous: most items are unrelated Transformers bugs that only share broad vocabulary like tokenizer, loading, v5, or model-specific regressions. None of the soft pairs look like the same concrete underlying defect or change.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "If forced to pick one issue as a representative anchor, issue 47085 is the most broadly framed, current low-level runtime bug; however, the set is too mixed to have a true canonical duplicate issue.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 47085 is the best single issue to anchor the cluster because it is open, concrete, and about core runtime behavior, but it is not a good duplicate representative for the other items.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44843",
+            "right": "issue:47085",
+            "accept": false,
+            "reason": "Both mention tokenizer/runtime behavior, but one is offline hub lookup in AutoTokenizer and the other is tokenizers thread-safety; different code paths and bugs."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:30333",
+            "accept": false,
+            "reason": "Unrelated issues: Flash Attention support for XLMRoberta vs MLflow reporting after failed training jobs."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:47085",
+            "accept": false,
+            "reason": "One is a Qwen3ForCausalLM VRAM leak in dataloader threads; the other is tokenizer thread-safety. Similar theme, different defect."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "MLflow failure reporting and Llama 3.2 Vision model loading regression are unrelated."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Both are v5 regressions around tokenization/processor handling, but they concern different behaviors and code paths."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Missing image-patch method in a processor vs remote-code checkpoint/meta-init incompatibility; not the same bug."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "VitPose model breakage and Gemma 4 mm_token_type_ids defaulting are distinct model-family issues."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Position_ids buffer access error and optimizer step being called twice are unrelated."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Serve KV-cache indexing crash vs auto_docstring crash under future annotations are unrelated."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Both concern tokenizer regressions, but one is clean_up_tokenization_spaces behavior and the other is Kimi-K2.5 codec/regex handling."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "MLflow job finalization and embedding reinitialization after resizing are unrelated."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Missing transformers.onnx module vs missing PyTorch dependency for AutoModel are different import failures."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "FP8 save_pretrained behavior and tokenizer_class mismatch in tokenizer_config are different save_pretrained bugs."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:46231",
+            "accept": false,
+            "reason": "Optimizer SGD arguments not used vs a docstring naming mismatch are unrelated."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Example script empty multi-label outputs vs local dataset loading failure are different example-script issues."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Released-version dependency pinning vs leftover dummy classes in repo checks are unrelated."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Tokenizer save metadata mismatch and duplicate optimizer steps are unrelated."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Missing image_utils import and leftover dummy classes in dummy_pt_objects.py are different cleanup bugs."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "FineGrainedFP8 accuracy issue and processor.save_pretrained missing files are unrelated."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Pegasus v4/v5 output difference and AttentionInterface.register behavior change are different functional regressions."
+          },
+          {
+            "left": "issue:44060",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "Wrong tied-weights warning in Qwen3-Next vs missing _tp_plan for Qwen3_5Moe are distinct model-integration problems."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Both involve saving, but one is FP8 MoE save_pretrained and the other is processor save_pretrained missing files; different artifacts."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43723",
+            "accept": false,
+            "reason": "Tokenizer BPE marker handling and AutoTokenizer loading in v5 are related to tokenizers but not the same bug."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:43644",
+            "accept": false,
+            "reason": "Multimodal config defaults and non-persistent buffer initialization are unrelated."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Category cardinality error and AttentionInterface.register behavior change are unrelated."
+          },
+          {
+            "left": "issue:44373",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Docstring correction for position_ids vs GLM-5 RoPE implementation are not the same issue."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:46762",
+            "accept": false,
+            "reason": "DSA indexer missing ReLU and Minimax MSA attention numerical differences are unrelated."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Cannot import AutoImageProcessor from transformers vs cannot import is_offline_mode from huggingface_hub are separate import regressions."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Tokenizer config save mismatch and Qwen3.5 MoE from_pretrained failure are different bugs."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 50,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:22355",
+        "issue:27957",
+        "issue:30990",
+        "issue:38175",
+        "issue:41628",
+        "issue:42831",
+        "issue:42907",
+        "issue:43122",
+        "issue:43232",
+        "issue:43408",
+        "issue:43475",
+        "issue:43504",
+        "issue:43723",
+        "issue:43824",
+        "issue:44038",
+        "issue:44190",
+        "issue:44222",
+        "issue:44261",
+        "issue:44263",
+        "issue:44279",
+        "issue:44295",
+        "issue:44360",
+        "issue:44373",
+        "issue:44484",
+        "issue:44623",
+        "issue:44829",
+        "issue:44977",
+        "issue:45071",
+        "issue:45092",
+        "issue:45310",
+        "issue:45357",
+        "issue:45381",
+        "issue:45834",
+        "issue:46002",
+        "issue:46050",
+        "issue:46082",
+        "issue:46209",
+        "issue:46305",
+        "issue:46328",
+        "issue:46369",
+        "issue:46402",
+        "issue:46424",
+        "issue:46431",
+        "issue:46614",
+        "issue:46655",
+        "issue:46909",
+        "issue:46932",
+        "issue:47059"
+      ],
+      "soft_pairs": [
+        "issue:45071|issue:46082",
+        "issue:43504|issue:47059",
+        "issue:44222|issue:44484",
+        "issue:43723|issue:44977",
+        "issue:43122|issue:46209",
+        "issue:46614|issue:46932",
+        "issue:44263|issue:45092",
+        "issue:46209|issue:46431",
+        "issue:44295|issue:44623",
+        "issue:43232|issue:46424",
+        "issue:43232|issue:45834",
+        "issue:44263|issue:44373",
+        "issue:30990|issue:38175",
+        "issue:44190|issue:46082",
+        "issue:46369|issue:46909",
+        "issue:43475|issue:46305",
+        "issue:44263|issue:44360",
+        "issue:46305|issue:46655",
+        "issue:42831|issue:45310",
+        "issue:42907|issue:46050",
+        "issue:22355|issue:27957",
+        "issue:44279|issue:46328",
+        "issue:45092|issue:45357",
+        "issue:42831|issue:44038",
+        "issue:44829|issue:46002",
+        "issue:41628|issue:43824",
+        "issue:44261|issue:45092",
+        "issue:45381|issue:46305",
+        "issue:43408|issue:46402"
+      ],
+      "prepared_review_unit_hash": "5fe9936af2dfa32c2d01c601b7aab6973ef42e843de34ac5cb21ff393217856f",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36016,
+        "estimated_input_tokens": 9004,
+        "estimated_eval_tokens": 18264
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is mostly a loose semantic cluster rather than true duplicates. The soft pairs share broad themes like v5 regressions, save/load, or model-family names, but they describe different concrete failures and should not be merged as duplicates.",
+        "confidence": 0.31,
+        "canonical_issue_reason": "No single issue cleanly represents the whole cluster. If one must be chosen, issue:45092 is the broadest and most detailed v5 regression report, but the surrounding items are largely unrelated.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45092 is the strongest standalone report in the set, with a clear, reproducible compatibility failure and enough detail to serve as the nearest representative, though this is still a weak fit overall.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45071",
+            "right": "issue:46082",
+            "accept": false,
+            "reason": "Both are config-related v5 regressions, but one is a generic PretrainedConfig type-checking break while the other is a specific Llama head_dim validation error."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "Unrelated failures: BEiT legacy-field loading vs a CLI completion compatibility error with huggingface_hub."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Both mention save_pretrained, but one is an FP8 MoE save bug and the other is a question about shard size defaults."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Tokenizer loading in v5 vs Qwen3.5 flash-attention generation failure are different code paths and symptoms."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Tokenization changes across versions versus FP8 inference failure on B200 are unrelated bugs."
+          },
+          {
+            "left": "issue:46614",
+            "right": "issue:46932",
+            "accept": false,
+            "reason": "Stale phonemizer backend state is unrelated to Gemma4 quantized embedder dtype casting."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "GlmMoeDsaIndexer torch.split behavior and InternVL2 meta-init loading incompatibility are different failures."
+          },
+          {
+            "left": "issue:46209",
+            "right": "issue:46431",
+            "accept": false,
+            "reason": "FP8 inference on B200 and missing chat template in a processor are unrelated."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "A position_ids buffer read issue is not the same as processor.save_pretrained missing files."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Generation kwarg sync behavior and cache-length handling are different bugs in adjacent areas."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Generation kwargs after sync_gpus and Kosmos2.5 OCR index errors do not share the same underlying defect."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "A functional DSA indexer issue is not the same as a wrong docstring for position_ids."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:38175",
+            "accept": false,
+            "reason": "Sentence Transformers loading stall and SigLIP2 zero-probability outputs are distinct model issues."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:46082",
+            "accept": false,
+            "reason": "Local dataset loading in an example script is unrelated to LlamaConfig head_dim validation."
+          },
+          {
+            "left": "issue:46369",
+            "right": "issue:46909",
+            "accept": false,
+            "reason": "SAM3 memory overflow and a broken DIA docs example are not the same bug."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:46305",
+            "accept": false,
+            "reason": "SAM3 vision encoder attribute access and Moonshine streaming mask off-by-one are unrelated."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Same component family, but torch.split return handling and missing ReLU are different concrete defects."
+          },
+          {
+            "left": "issue:46305",
+            "right": "issue:46655",
+            "accept": false,
+            "reason": "Streaming window off-by-one and video processor padding math are separate bugs."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "FineGrainedFP8 accuracy concerns do not match a Qwen3.5 MoE from_pretrained error."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Dequantized Ministrals/Devstrals save failure and sinq save/load breakage are different quantization paths."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:27957",
+            "accept": false,
+            "reason": "Missing transformers.onnx module and XLMRoberta Flash Attention 2 are unrelated."
+          },
+          {
+            "left": "issue:44279",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "A general dependency issue and a performance inefficiency in SequenceFeatureExtractor.pad are not duplicates."
+          },
+          {
+            "left": "issue:45092",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "InternVL2 meta-initialization incompatibility is different from Qwen3.5 save_pretrained key handling."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44038",
+            "accept": false,
+            "reason": "FineGrainedFP8 accuracy issues and Qwen3-VL-Moe loading bugs are different problems."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Flash Attention 3 training degeneration and NaNs after checkout/init are not the same root cause."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are missing imports, but they refer to different exported symbols and likely different packaging regressions."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "MLA rms_norm_eps precision drift and remote-code meta-init incompatibility are unrelated."
+          },
+          {
+            "left": "issue:45381",
+            "right": "issue:46305",
+            "accept": false,
+            "reason": "Qwen2.5-VL vision_position_ids issues and Moonshine streaming mask off-by-one affect different model code paths."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "A class-type warning during SAM3 instantiation is not the same as SAM3 text encoder weights not loading."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Same file area does not imply same defect; split-output handling and missing activation are distinct."
+          },
+          {
+            "left": "issue:46369",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "SAM3-LiteText memory overflow and a CLI completion API break are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 51,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:22355",
+        "issue:27957",
+        "issue:28218",
+        "issue:30990",
+        "issue:33357",
+        "issue:39379",
+        "issue:39401",
+        "issue:41628",
+        "issue:42371",
+        "issue:42757",
+        "issue:42796",
+        "issue:42831",
+        "issue:42886",
+        "issue:42907",
+        "issue:42994",
+        "issue:43066",
+        "issue:43122",
+        "issue:43232",
+        "issue:43329",
+        "issue:43408",
+        "issue:43540",
+        "issue:43576",
+        "issue:43644",
+        "issue:43646",
+        "issue:43784",
+        "issue:43976",
+        "issue:44188",
+        "issue:44263",
+        "issue:44295",
+        "issue:44297",
+        "issue:44327",
+        "issue:44373",
+        "issue:44483",
+        "issue:44568",
+        "issue:44945",
+        "issue:44964",
+        "issue:45278",
+        "issue:45381",
+        "issue:45406",
+        "issue:45593",
+        "issue:45753",
+        "issue:46002",
+        "issue:46172",
+        "issue:46305",
+        "issue:46369",
+        "issue:46909",
+        "issue:47059"
+      ],
+      "soft_pairs": [
+        "issue:44964|issue:47059",
+        "issue:39379|issue:46172",
+        "issue:44483|issue:45406",
+        "issue:43644|issue:46305",
+        "issue:42831|issue:45753",
+        "issue:22355|issue:28218",
+        "issue:44327|issue:47059",
+        "issue:39401|issue:44568",
+        "issue:42757|issue:45278",
+        "issue:44188|issue:44373",
+        "issue:43408|issue:46369",
+        "issue:42886|issue:43066",
+        "issue:42886|issue:42994",
+        "issue:27957|issue:30990",
+        "issue:42796|issue:43122",
+        "issue:45381|issue:46909",
+        "issue:43232|issue:45593",
+        "issue:43540|issue:46909",
+        "issue:43646|issue:46002",
+        "issue:42907|issue:44945",
+        "issue:44295|issue:44297",
+        "issue:30990|issue:33357",
+        "issue:43576|issue:43784",
+        "issue:43976|issue:44945",
+        "issue:42907|issue:44964",
+        "issue:41628|issue:45278",
+        "issue:42371|issue:44263",
+        "issue:43329|issue:47059"
+      ],
+      "prepared_review_unit_hash": "0d86ec258facb7aac854adb1799e716191b6eb56ac96d661d1f2c08529cdddd9",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 28,
+        "serialized_chars": 35005,
+        "estimated_input_tokens": 8752,
+        "estimated_eval_tokens": 17760
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "All soft pairs look like false positives: the titles point to different bugs, regressions, or docs/CLI issues rather than the same underlying change. No pull requests are present.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "Issue 42886 is the most self-contained and actionable report: it has a precise symptom, exact affected version (4.57.3), a clear offline-cache reproduction, and a narrow code-path scope.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 42886 is a strong representative issue because it is concrete, reproducible, and specific to a single failure mode rather than a broad or ambiguous regression.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44964",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "Model loading failure vs CLI completion AttributeError; different subsystems and root causes."
+          },
+          {
+            "left": "issue:39379",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "mllama mask handling bug vs DeepSeek tensor transpose issue; unrelated code paths."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Chat/completions request rejection vs Gemma4Processor missing _tokenizer; different failures."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:46305",
+            "accept": false,
+            "reason": "Non-persistent buffer junk vs Moonshine sliding-window off-by-one; not the same bug."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "FP8 accuracy regression vs Qwen3_5 infinite loop on a specific image; distinct symptoms and causes."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:28218",
+            "accept": false,
+            "reason": "Missing transformers.onnx module vs tokenizer adds extra space after added token; unrelated."
+          },
+          {
+            "left": "issue:44327",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "QA decode_spans top-k bounds error vs CLI completion attribute error; different code paths."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Offset mapping bug vs BOS/EOS not added with add_special_tokens; separate tokenizer issues."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Specific is_offline_mode import break vs broad import failures after upgrade; too different to merge."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "Torch.compile attention-kernel divergence vs a wrong docstring; one is runtime behavior, the other documentation."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Model-type warning mismatch vs memory overflow in another SAM3 model; not the same bug."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:43066",
+            "accept": false,
+            "reason": "Offline tokenizer cache loading failure vs wrong tokenizer decoder type in v5; distinct issues."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:42994",
+            "accept": false,
+            "reason": "Offline cache loading vs quantized model saving failure; different phases and failures."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:30990",
+            "accept": false,
+            "reason": "Flash Attention 2 support question vs loading hang in Sentence Transformers; unrelated."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43122",
+            "accept": false,
+            "reason": "BPE marker handling bug vs version-to-version tokenization drift; similar domain but not the same concrete defect."
+          },
+          {
+            "left": "issue:45381",
+            "right": "issue:46909",
+            "accept": false,
+            "reason": "Qwen2.5-VL video position_ids bug vs broken documentation example; unrelated."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Generation kwargs sync issue vs D-FINE auxiliary-loss behavior; different components."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:46909",
+            "accept": false,
+            "reason": "Video-input ValueError in Qwen3OmniMoe vs documentation example failure; unrelated."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Custom model initialization breakage vs NaNs in classification heads after init; different failure modes."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Quantized/dequantized save regression vs incorrect output under pipeline parallelism; different code paths."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "position_ids buffer read error vs tokenizer_class mismatch on save_pretrained; distinct tokenizer/model-save issues."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Loading hang vs MacOS bus error; unrelated runtime failures."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:43784",
+            "accept": false,
+            "reason": "Broken `transformers env` CLI vs sentence-transformers import NameError; different areas."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Python version compatibility regression vs pipeline-parallel output corruption; not the same bug."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Quantized model saving failure vs multimodal model loading failure; different operations and symptoms."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Missing AutoImageProcessor import vs many import errors after upgrade; broad similarity only, not a single concrete fix."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "TF32 settings guidance issue vs torch.split return-value bug; unrelated."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "Multimodal token counting video-branch bug vs CLI completion attribute error; different subsystems."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 52,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:22355",
+        "issue:29127",
+        "issue:29159",
+        "issue:36296",
+        "issue:39379",
+        "issue:39401",
+        "issue:42175",
+        "issue:42757",
+        "issue:42907",
+        "issue:43116",
+        "issue:43122",
+        "issue:43232",
+        "issue:43278",
+        "issue:43408",
+        "issue:43452",
+        "issue:43475",
+        "issue:43493",
+        "issue:43756",
+        "issue:43784",
+        "issue:43824",
+        "issue:43937",
+        "issue:44038",
+        "issue:44060",
+        "issue:44117",
+        "issue:44222",
+        "issue:44295",
+        "issue:44297",
+        "issue:44351",
+        "issue:44360",
+        "issue:44485",
+        "issue:44488",
+        "issue:44521",
+        "issue:44625",
+        "issue:44704",
+        "issue:44779",
+        "issue:44829",
+        "issue:44938",
+        "issue:44945",
+        "issue:45200",
+        "issue:45216",
+        "issue:45290",
+        "issue:45478",
+        "issue:45600",
+        "issue:45753",
+        "issue:46172",
+        "issue:46326",
+        "issue:46614",
+        "issue:46762"
+      ],
+      "soft_pairs": [
+        "issue:44117|issue:44704",
+        "issue:43756|issue:44060",
+        "issue:42907|issue:43937",
+        "issue:43784|issue:45600",
+        "issue:39401|issue:44488",
+        "issue:39401|issue:44521",
+        "issue:44625|issue:44829",
+        "issue:44704|issue:45290",
+        "issue:46172|issue:46762",
+        "issue:44360|issue:44485",
+        "issue:43122|issue:46326",
+        "issue:43408|issue:43475",
+        "issue:43122|issue:44521",
+        "issue:44038|issue:44297",
+        "issue:42907|issue:43452",
+        "issue:22355|issue:29159",
+        "issue:44038|issue:44295",
+        "issue:44038|issue:44222",
+        "issue:45216|issue:45478",
+        "issue:43824|issue:44351",
+        "issue:43116|issue:45200",
+        "issue:39401|issue:46614",
+        "issue:43122|issue:44779",
+        "issue:42757|issue:44351",
+        "issue:36296|issue:43116",
+        "issue:43278|issue:44945",
+        "issue:22355|issue:29127",
+        "issue:42175|issue:44938",
+        "issue:39379|issue:43232",
+        "issue:43493|issue:45753"
+      ],
+      "prepared_review_unit_hash": "edc2eedf0f638889d0142137168e52e74146e5c133961ed6d79985b5f520e02a",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36435,
+        "estimated_input_tokens": 9109,
+        "estimated_eval_tokens": 18474
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The items are largely unrelated; the soft pairs are mostly same-subsystem but not the same underlying bug/change, so I would reject all duplicate candidates. If forced to pick one issue as the best representative, the open multimodal `apply_chat_template` report is the strongest standalone anchor.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "Issue 44521 is the most actionable open report here: it has the strongest current discussion/inbound activity and a clear, reproducible multimodal bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44521 is the best standalone representative of the set because it is open, well-scoped, and likely to benefit from direct triage.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44117",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "Both are API/loading bugs, but one is tokenizer mapping resolution and the other is `AutoProcessor` kwargs plumbing; different code paths and failure modes."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44060",
+            "accept": false,
+            "reason": "Different models and different bugs: Smollm3 RoPE layer handling vs Qwen3-Next tied-weights warning."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43937",
+            "accept": false,
+            "reason": "One is a save/load regression for dequantized Ministrals/Devstrals; the other is a GLM-5 generation config validation error."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Both are import-time failures, but they point to different missing symbols/config references in different code paths."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Qwen3 tokenizer offset mapping and loading `cjvt/sleng-bert` are separate tokenizer/model issues."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Tokenizer offset mapping for Qwen3 is unrelated to multimodal assistant-mask generation in `apply_chat_template`."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Num-label propagation in Qwen3.5 config is unrelated to flash-attention training degeneration."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "`AutoProcessor` caching/kwargs handling is unrelated to chat-template crashes with tool-call assistant messages."
+          },
+          {
+            "left": "issue:46172",
+            "right": "issue:46762",
+            "accept": false,
+            "reason": "Both are numerical/model-computation complaints, but for different models and different attention implementations."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "DSA indexer activation choice and GLM-5 RoPE implementation are unrelated subsystems."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Tokenizer-version differences and `continue_final_message` handling are different tokenizer behaviors, not the same bug."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "One is a model-type warning/mismatch; the other is a missing attribute on `Sam3VisionEncoderOutput`."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "General tokenization drift across versions is not the same issue as multimodal assistant-mask generation."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe behavior in 5.0 and tokenizer save metadata mismatch are different bugs."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "Dequantized save failures and GGUF loading failures are both serialization-related but not the same concrete bug."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:29159",
+            "accept": false,
+            "reason": "Missing `transformers.onnx` import and slow/fast tokenizer inconsistencies are unrelated."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44295",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe issue and `position_ids` buffer read error affect different components."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Both mention MoE, but one is a 5.0 model bug and the other is an FP8 save-pretrained issue."
+          },
+          {
+            "left": "issue:45216",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both concern Qwen3.5 on newer transformers, but one is a save_pretrained regression and the other is a from_pretrained load error."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "These are different missing imports (`Qwen2_5_VLForConditionalGeneration` vs `HybridCache`) with different causes."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "A multi-label classification example bug is unrelated to Gemma 4 mm token-type defaults."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:46614",
+            "accept": false,
+            "reason": "Qwen3 offset mapping and stale phonemizer language state are separate tokenizer correctness problems."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they involve different models and different failure mechanisms."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Different missing imports from different packages; no shared underlying bug."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43116",
+            "accept": false,
+            "reason": "Tensor-parallel training and classification-example output are unrelated."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Training/eval dtype mismatch for embeddings is not the same as pipeline-parallel output corruption."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "One is a missing module import, the other is an error-message clarity issue in LayoutLMv3."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "Packaging backend omission and Python 3.14 import failure are different compatibility problems."
+          },
+          {
+            "left": "issue:39379",
+            "right": "issue:43232",
+            "accept": false,
+            "reason": "mllama mask handling and generation kwarg updates after `sync_gpus` are different code paths."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "SigLIP2 numerical divergence and Qwen3_5 image infinite loop are distinct multimodal model bugs."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 53,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28348",
+        "issue:38617",
+        "issue:42371",
+        "issue:42673",
+        "issue:42796",
+        "issue:42898",
+        "issue:43066",
+        "issue:43116",
+        "issue:43493",
+        "issue:43611",
+        "issue:43618",
+        "issue:43825",
+        "issue:43827",
+        "issue:43874",
+        "issue:43937",
+        "issue:43976",
+        "issue:43992",
+        "issue:44038",
+        "issue:44075",
+        "issue:44188",
+        "issue:44355",
+        "issue:44485",
+        "issue:44610",
+        "issue:44625",
+        "issue:44704",
+        "issue:44779",
+        "issue:44936",
+        "issue:44995",
+        "issue:45072",
+        "issue:45083",
+        "issue:45278",
+        "issue:45357",
+        "issue:45446",
+        "issue:45478",
+        "issue:46002",
+        "issue:46231",
+        "issue:46328",
+        "issue:46614",
+        "issue:46640",
+        "issue:46697",
+        "issue:46710",
+        "issue:46726",
+        "issue:46829",
+        "issue:46877",
+        "issue:46909",
+        "issue:46981",
+        "issue:47116"
+      ],
+      "soft_pairs": [
+        "issue:46231|issue:46877",
+        "issue:44188|issue:46328",
+        "issue:43937|issue:44485",
+        "issue:42898|issue:46614",
+        "issue:38617|issue:45278",
+        "issue:43976|issue:45446",
+        "issue:43611|issue:46002",
+        "issue:44038|issue:45478",
+        "issue:43066|issue:44779",
+        "issue:42673|issue:45072",
+        "issue:44355|issue:46231",
+        "issue:45357|issue:45478",
+        "issue:44625|issue:46002",
+        "issue:43116|issue:43493",
+        "issue:44610|issue:46909",
+        "issue:45083|issue:45446",
+        "issue:42371|issue:43874",
+        "issue:44075|issue:44936",
+        "issue:46726|issue:46981",
+        "issue:43618|issue:46829",
+        "issue:46328|issue:47116",
+        "issue:43992|issue:44704",
+        "issue:43066|issue:46710",
+        "issue:43116|issue:43825",
+        "issue:28348|issue:46640",
+        "issue:42673|issue:42796",
+        "issue:43827|issue:45446",
+        "issue:44995|issue:46697"
+      ],
+      "prepared_review_unit_hash": "b1cc815f4864b5a5941ef67e6d0bf5a2a14fb7932a1c72e7140b5aa3356bd677",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 28,
+        "serialized_chars": 35241,
+        "estimated_input_tokens": 8811,
+        "estimated_eval_tokens": 17878
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The items are mostly unrelated issues that only share broad Transformers-v5 or model-family wording. I do not see any soft pair that clearly represents the same underlying bug or change, so all soft edges are rejected.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "issue:45478 is the most concrete, actionable bug report in the set: it names the affected version range, model family, and a specific from_pretrained failure path.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45478 is a strong anchor for triage because it is narrowly scoped and reproducible, unlike the broader regression or documentation reports nearby.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46231",
+            "right": "issue:46877",
+            "accept": false,
+            "reason": "Docs parameter-name mismatch vs import-time doc-lint output; same documentation area, but not the same bug."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "Attention-kernel divergence under torch.compile vs padding performance issue; different code paths and symptoms."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "GenerationConfig validation error vs RoPE implementation discussion; both GLM-5 related, but distinct problems."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:46614",
+            "accept": false,
+            "reason": "Tokenization cleanup behavior change vs stale phonemizer backend state; unrelated tokenizer subsystems."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "A specific missing import vs a broad collection of import errors after upgrade; too general to be the same bug."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Python version incompatibility report vs flex_attention version-check bug; different failure causes."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Model loading with base_model_prefix vs NaNs in classification heads; different runtime failures."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both concern Qwen/MoE loading, but they affect different model variants and failure modes; not enough to treat as one bug."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both are tokenizer regressions in v5, but one is decoder-type selection and the other is DeepSeek output correctness; separate issues."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "VRAM leak in Qwen3ForCausalLM threading vs dtype mismatches in inference; unrelated."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:46231",
+            "accept": false,
+            "reason": "Compiled Python execution errors vs a docs typo; no shared underlying defect."
+          },
+          {
+            "left": "issue:45357",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Save-pretrained key regression vs from_pretrained loading error; related model family, but different code paths and bugs."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Config propagation bug vs NaNs in classification heads; different symptoms and likely causes."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Example script returning empty results vs SigLIP2 implementation discrepancy; unrelated."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:46909",
+            "accept": false,
+            "reason": "Processor size/padding mismatch vs broken docs example; one is runtime preprocessing, the other is documentation."
+          },
+          {
+            "left": "issue:45083",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Qwen3 Omni helper behavior vs flex_attention version check; unrelated subsystems."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:43874",
+            "accept": false,
+            "reason": "TF32 API/settings change vs missing multimodal image-patch method; different features."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Optimizer argument handling vs trainer.evaluate after train; different functional areas."
+          },
+          {
+            "left": "issue:46726",
+            "right": "issue:46981",
+            "accept": false,
+            "reason": "Missing-vision-backend messaging vs retry behavior on HTTP 429; both AutoProcessor.from_pretrained, but distinct bugs."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:46829",
+            "accept": false,
+            "reason": "Different output-field assignment bugs in different model classes; too broad to merge as duplicates."
+          },
+          {
+            "left": "issue:46328",
+            "right": "issue:47116",
+            "accept": false,
+            "reason": "Feature-extraction padding inefficiency vs streaming pipeline regression; unrelated."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "Missing weight in UMT5Encoder load vs kwargs not forwarded in AutoProcessor; different from_pretrained problems."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:46710",
+            "accept": false,
+            "reason": "Tokenizer decoder-type regression vs model accuracy regression; not the same underlying bug."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Example script behavior vs pipeline error-message wording; not a duplicate."
+          },
+          {
+            "left": "issue:28348",
+            "right": "issue:46640",
+            "accept": false,
+            "reason": "Feature request for GPT2 Flash Attention 2 support vs an unspecified Flash Attention 2 issue; too vague and not the same change."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:42796",
+            "accept": false,
+            "reason": "VRAM leak in model inference vs tokenizer BPE-marker handling; unrelated."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Pipeline-removal docs cleanup vs flex_attention version check; unrelated."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:46697",
+            "accept": false,
+            "reason": "Stale cache crash on second forward pass vs image over-padding; different components and failure modes."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 54,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:33453",
+        "issue:39379",
+        "issue:39401",
+        "issue:42222",
+        "issue:42617",
+        "issue:43064",
+        "issue:43408",
+        "issue:43441",
+        "issue:43450",
+        "issue:43673",
+        "issue:43688",
+        "issue:43742",
+        "issue:43824",
+        "issue:43937",
+        "issue:43986",
+        "issue:44117",
+        "issue:44263",
+        "issue:44291",
+        "issue:44380",
+        "issue:44410",
+        "issue:44466",
+        "issue:44704",
+        "issue:44779",
+        "issue:44829",
+        "issue:44977",
+        "issue:45071",
+        "issue:45081",
+        "issue:45092",
+        "issue:45216",
+        "issue:45310",
+        "issue:45341",
+        "issue:45405",
+        "issue:45446",
+        "issue:45561",
+        "issue:45563",
+        "issue:45588",
+        "issue:45632",
+        "issue:46002",
+        "issue:46077",
+        "issue:46328",
+        "issue:46512",
+        "issue:46752",
+        "issue:46762",
+        "issue:46897",
+        "issue:46981",
+        "issue:47000",
+        "issue:47059",
+        "issue:47110"
+      ],
+      "soft_pairs": [
+        "issue:45092|issue:46762",
+        "issue:33453|issue:39401",
+        "issue:44263|issue:46762",
+        "issue:43824|issue:46981",
+        "issue:43064|issue:43673",
+        "issue:43673|issue:46328",
+        "issue:42222|issue:43441",
+        "issue:39379|issue:43688",
+        "issue:42222|issue:47000",
+        "issue:42222|issue:46752",
+        "issue:45405|issue:45446",
+        "issue:43937|issue:45563",
+        "issue:44380|issue:46077",
+        "issue:44117|issue:45588",
+        "issue:44291|issue:45071",
+        "issue:44466|issue:46002",
+        "issue:39379|issue:46897",
+        "issue:44410|issue:45310",
+        "issue:39401|issue:44779",
+        "issue:43450|issue:43986",
+        "issue:43937|issue:44977",
+        "issue:44704|issue:47110",
+        "issue:45071|issue:47059",
+        "issue:45341|issue:46328",
+        "issue:44829|issue:45216",
+        "issue:45071|issue:46512",
+        "issue:45341|issue:45561",
+        "issue:43408|issue:43742",
+        "issue:45632|issue:46981",
+        "issue:42617|issue:45341",
+        "issue:44117|issue:45081"
+      ],
+      "prepared_review_unit_hash": "bb7c97eadb7db390157fd56964eb6ed3b34eca45f477ace478f55d5b6dac7b33",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 37009,
+        "estimated_input_tokens": 9253,
+        "estimated_eval_tokens": 18762
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "No soft edge looks like the same underlying bug/change. The pairs mostly share only broad subsystem keywords (tokenizers, generation, vision, training, loading) but differ in concrete failure mode, model, or code path, so all are rejected.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45092",
+            "right": "issue:46762",
+            "accept": false,
+            "reason": "Different failures: v5 meta-init incompatibility vs Minimax attention numeric mismatch."
+          },
+          {
+            "left": "issue:33453",
+            "right": "issue:39401",
+            "accept": false,
+            "reason": "Generic tokenizer-loading regression vs Qwen3 offset_mapping bug."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:46762",
+            "accept": false,
+            "reason": "GLM MoE indexer split semantics are unrelated to attention numerical differences."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:46981",
+            "accept": false,
+            "reason": "Missing Qwen2.5-VL import is unrelated to repeated HTTP 429 retries."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "FSDP/PEFT optimizer-state corruption is not the same as missing GenerationMixin cache."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "Generation cache bug vs SequenceFeatureExtractor.pad performance issue."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "VitPose breakage and Ministral FlashAttention failure are different models/code paths."
+          },
+          {
+            "left": "issue:39379",
+            "right": "issue:43688",
+            "accept": false,
+            "reason": "mllama mask handling bug is unrelated to auxiliary-loss normalization."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "VitPose model breakage and RTDetr loss infeasibility are different bugs."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:46752",
+            "accept": false,
+            "reason": "VitPose loading failure vs empty-list chat template crash."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Released PEFT version bump is unrelated to AuxRequest PyTorch version gating."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Invalid GenerationConfig vs stale paged-generate warning."
+          },
+          {
+            "left": "issue:44380",
+            "right": "issue:46077",
+            "accept": false,
+            "reason": "GPT2 attention scaling and Gemma4 bidirectional masking are different attention bugs."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "TOKENIZER_MAPPING_NAMES None handling is unrelated to flash-attention s_aux=None crash."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Unexpected _is_hf_initialized arg vs PretrainedConfig type-checking regression."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "lm_head serialization inconsistency is not the same as NaNs after checkout/init."
+          },
+          {
+            "left": "issue:39379",
+            "right": "issue:46897",
+            "accept": false,
+            "reason": "mllama row-mask bug vs Florence2 double-shift training-loss bug."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Missing qkv/gate projections in qwen3next is distinct from Qwen3.5 MoE from_pretrained failure."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Qwen3 offset_mapping bug vs Deepseek tokenizer incorrect output."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:43986",
+            "accept": false,
+            "reason": "Batched video shape bug vs missing-torchvision AutoProcessor import crash."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Invalid GenerationConfig is not the same as Qwen3.5 flash-attention generation failure."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:47110",
+            "accept": false,
+            "reason": "cached_file kwarg forwarding bug vs special_tokens list/dict type error."
+          },
+          {
+            "left": "issue:45071",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "PretrainedConfig type-checking regression vs CLI completion API breakage."
+          },
+          {
+            "left": "issue:45341",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "testing_utils bug vs numpy-padding performance regression."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "FlashAttention-3 training degeneration and Qwen3.5 save_pretrained corruption are unrelated."
+          },
+          {
+            "left": "issue:45071",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "Type-checking failure vs stale model.config.dtype after from_config."
+          },
+          {
+            "left": "issue:45341",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "A small testing_utils bug is not the same as the xdist captured_info.txt race."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "sam3_video/sam3_tracker warning differs from MobileLLM KeyError on load."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:46981",
+            "accept": false,
+            "reason": "trust_remote_code cache collision is unrelated to 429 retry behavior."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "3d_parallel.py runtime failure vs testing_utils bug."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Tokenizer-mapping None issue is unrelated to Mistral regex patch AttributeError."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 55,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:27957",
+        "issue:38617",
+        "issue:39401",
+        "issue:41628",
+        "issue:42222",
+        "issue:42371",
+        "issue:43012",
+        "issue:43066",
+        "issue:43097",
+        "issue:43122",
+        "issue:43208",
+        "issue:43352",
+        "issue:43577",
+        "issue:43673",
+        "issue:43874",
+        "issue:43931",
+        "issue:43937",
+        "issue:44075",
+        "issue:44077",
+        "issue:44220",
+        "issue:44261",
+        "issue:44263",
+        "issue:44410",
+        "issue:44483",
+        "issue:44484",
+        "issue:44485",
+        "issue:44493",
+        "issue:44496",
+        "issue:44877",
+        "issue:44933",
+        "issue:45071",
+        "issue:45278",
+        "issue:45310",
+        "issue:45341",
+        "issue:45464",
+        "issue:45478",
+        "issue:45656",
+        "issue:46172",
+        "issue:46178",
+        "issue:46209",
+        "issue:46231",
+        "issue:46298",
+        "issue:46617",
+        "issue:46752",
+        "issue:46762",
+        "issue:46894",
+        "issue:46981",
+        "issue:47116"
+      ],
+      "soft_pairs": [
+        "issue:38617|issue:41628",
+        "issue:43122|issue:46298",
+        "issue:42222|issue:46178",
+        "issue:44877|issue:46231",
+        "issue:39401|issue:46209",
+        "issue:42371|issue:44261",
+        "issue:43937|issue:44496",
+        "issue:44484|issue:45310",
+        "issue:44261|issue:46762",
+        "issue:44933|issue:45278",
+        "issue:42371|issue:46617",
+        "issue:44410|issue:45478",
+        "issue:44075|issue:45656",
+        "issue:38617|issue:44220",
+        "issue:44263|issue:44485",
+        "issue:43066|issue:43122",
+        "issue:43208|issue:46894",
+        "issue:43097|issue:44493",
+        "issue:44077|issue:46981",
+        "issue:46752|issue:47116",
+        "issue:43673|issue:45341",
+        "issue:43931|issue:44410",
+        "issue:41628|issue:45071",
+        "issue:44483|issue:45464",
+        "issue:43874|issue:44261",
+        "issue:44484|issue:45478",
+        "issue:27957|issue:43352",
+        "issue:44263|issue:46172",
+        "issue:43012|issue:43577"
+      ],
+      "prepared_review_unit_hash": "c4da3af09b7aeb34cbfdc1cf9f9fb020caef38f9fb6283b8fe4d7dccf693c111",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 35777,
+        "estimated_input_tokens": 8945,
+        "estimated_eval_tokens": 18146
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a grab bag of unrelated Transformers issues. A few are broadly in the same general area (v5 regressions, tokenizer/import problems), but none of the soft pairs look like the same concrete bug or change from the titles alone.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "No single canonical issue stands out: the items span unrelated import errors, tokenizer regressions, model-loading failures, optimizer behavior, and API/runtime bugs.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the closest umbrella candidate because it is the broadest regression report in the set (many import errors after an upgrade), but it still should not be treated as the same bug as the others.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:38617",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "Both are import failures, but they concern different missing symbols and likely different code paths."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:46298",
+            "accept": false,
+            "reason": "Tokenization regression vs noisy logging; no shared underlying bug."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "VitPose model breakage and a backend-mapping error are unrelated."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:46231",
+            "accept": false,
+            "reason": "Strict config loading and a documentation parameter mismatch are not the same bug."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Tokenizer offset mapping bug vs FP8 inference failure on B200; different subsystems."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44261",
+            "accept": false,
+            "reason": "TF32 settings behavior and MLA normalization epsilon precision are different issues."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "GenerationConfig validation for GLM-5 and unrecognized model loading for Olmo-Hybrid are unrelated."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Save-pretrained shard sizing and Qwen3.5 MoE from_pretrained failure are different problems."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:46762",
+            "accept": false,
+            "reason": "Both mention numerical behavior, but they are different model/code-path bugs."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "One is a specific bad import, the other is a broader upgrade regression; not the same concrete failure."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "TF32 configuration and cache FileNotFoundError are unrelated."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Different Qwen model-loading failures with different symptoms and likely different root causes."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Unused SGD args and double optimizer stepping under deepspeed are distinct optimizer bugs."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "ImportError and audio feature extraction failure are unrelated."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "torch.split return handling and GLM-5 RoPE implementation are different code paths."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:43122",
+            "accept": false,
+            "reason": "Both are tokenizer-related v5 regressions, but the titles point to different concrete symptoms and fixes."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:46894",
+            "accept": false,
+            "reason": "xLSTM training blockers and a CohereASR double-shift loss bug are unrelated."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Removed tie_embeddings API and unexpected position-id keys are different regressions."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:46981",
+            "accept": false,
+            "reason": "PatchTSMixer post_init policy and HF_ENDPOINT retry behavior do not share a code-path bug."
+          },
+          {
+            "left": "issue:46752",
+            "right": "issue:47116",
+            "accept": false,
+            "reason": "Empty-list chat template crash and streaming pipeline regression are different failures."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Missing generation cache and a testing_utils bug are unrelated."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44410",
+            "accept": false,
+            "reason": "Both involve Qwen-family models, but one is a weight-shape mismatch and the other is missing projections; not the same bug."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "ImportError for AutoImageProcessor vs PreTrainedConfig type-checking regression are different issues."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Both hit chat/completions, but the model scope and reported failure differ; not enough to treat as one bug."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:44261",
+            "accept": false,
+            "reason": "Missing image-patch count and MLA epsilon precision issues are unrelated."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "save_pretrained shard-size question vs from_pretrained load error; different operations and bugs."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:43352",
+            "accept": false,
+            "reason": "Flash Attention 2 support issues on different models are not the same concrete bug."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "torch.split behavior in GlmMoeDsaIndexer and a deepseek tensor transpose/cast issue are unrelated."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "PyTorch precision warning during compilation and Blip2 dtype propagation are different problems."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 56,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:38617",
+        "issue:42757",
+        "issue:42886",
+        "issue:42890",
+        "issue:43012",
+        "issue:43064",
+        "issue:43097",
+        "issue:43208",
+        "issue:43441",
+        "issue:43452",
+        "issue:43531",
+        "issue:43673",
+        "issue:43873",
+        "issue:43874",
+        "issue:43931",
+        "issue:43986",
+        "issue:44038",
+        "issue:44062",
+        "issue:44117",
+        "issue:44261",
+        "issue:44263",
+        "issue:44315",
+        "issue:44351",
+        "issue:44360",
+        "issue:44410",
+        "issue:44484",
+        "issue:44485",
+        "issue:44829",
+        "issue:44987",
+        "issue:45003",
+        "issue:45341",
+        "issue:45372",
+        "issue:45375",
+        "issue:45440",
+        "issue:45588",
+        "issue:45753",
+        "issue:46050",
+        "issue:46123",
+        "issue:46172",
+        "issue:46298",
+        "issue:46369",
+        "issue:46655",
+        "issue:46664",
+        "issue:46701",
+        "issue:46752",
+        "issue:46821",
+        "issue:46897",
+        "issue:46981"
+      ],
+      "soft_pairs": [
+        "issue:44360|issue:46172",
+        "issue:43441|issue:44829",
+        "issue:43931|issue:44038",
+        "issue:43097|issue:43452",
+        "issue:43986|issue:45341",
+        "issue:44261|issue:44485",
+        "issue:45003|issue:45341",
+        "issue:42886|issue:46981",
+        "issue:45753|issue:46369",
+        "issue:43531|issue:45341",
+        "issue:44117|issue:44987",
+        "issue:43874|issue:44263",
+        "issue:43097|issue:45588",
+        "issue:43673|issue:46123",
+        "issue:44410|issue:45375",
+        "issue:45440|issue:46172",
+        "issue:38617|issue:42757",
+        "issue:44038|issue:44410",
+        "issue:43986|issue:45372",
+        "issue:43097|issue:46050",
+        "issue:43064|issue:44315",
+        "issue:43208|issue:46897",
+        "issue:46655|issue:46664",
+        "issue:43986|issue:44484",
+        "issue:46298|issue:46752",
+        "issue:38617|issue:44351",
+        "issue:42890|issue:46298",
+        "issue:44062|issue:44117",
+        "issue:43097|issue:43873",
+        "issue:43012|issue:46821",
+        "issue:44117|issue:46701"
+      ],
+      "prepared_review_unit_hash": "204f4c763571ae04ff708b6a8ae1514cbd3441ef59a0b9a1740e1ef97c9923a5",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 36853,
+        "estimated_input_tokens": 9214,
+        "estimated_eval_tokens": 18684
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous: most items are unrelated Transformers regressions across different models, loaders, processors, and utilities. None of the soft pairs look like the same underlying bug, so I would not merge any of them as duplicates.",
+        "confidence": 0.9,
+        "canonical_issue_reason": "Issue 43097 is the closest cluster anchor because it\u2019s a broad Transformers v5 regression report, but the set is too mixed for it to be a true duplicate canonical.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43097 is the least-specific, most central representative of the general v5 breakage theme, though it still does not match the other reports closely enough to justify deduping.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44360",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "Different bugs: one is a DSA indexer activation issue, the other is a DeepSeek-v4 dtype/transpose problem."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Both mention flash attention, but one is a Ministral-3 load failure and the other is a sequence-classification training pathology."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44038",
+            "accept": false,
+            "reason": "Different Qwen3-VL model families and different failure modes; weight-shape mismatch is not the same as the Qwen3-VL-MoE bug."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "Both are v5-era regressions, but one is an API removal notice and the other is a gguf/from_pretrained loading breakage."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Unrelated subsystems: video-model processor loading vs a small testing utility bug."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Different model internals: MLA layer norm epsilon handling vs GLM-5 RoPE implementation."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "One is unsafe sys.modules access in modeling_utils; the other is a testing_utils issue."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:46981",
+            "accept": false,
+            "reason": "Both involve loading/network behavior, but offline cache handling and repeated 429 retries are distinct code paths."
+          },
+          {
+            "left": "issue:45753",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Different models and symptoms: infinite-loop generation on Qwen3_5 vs memory overflow in sam3-litetext."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding-window behavior is unrelated to a testing utility bug."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Both are loading-related, but TOKENIZER_MAPPING_NAMES returning None and a specific model load failure are different defects."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "Same broad GLM area, but one is a missing image-processor method and the other is a torch.split return-value issue."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Both touch flash/v5 regressions, but they affect different mechanisms: API removal versus flash_attention.py handling of s_aux=None."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:46123",
+            "accept": false,
+            "reason": "Different behaviors: missing GenerationMixin cache in chunked_prefill vs a batch bookkeeping bug in MaskGenerationPipeline."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Different model configs and bugs: missing attention projections vs a strict-config field being dropped."
+          },
+          {
+            "left": "issue:45440",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "Same DeepSeek family, but one is remote/native implementation divergence and the other is a specific comb.transpose dtype issue."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are import errors, but they are different missing symbols from different modules."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44410",
+            "accept": false,
+            "reason": "Both involve Qwen3-family models, but the reported failures are different architecture/config mismatches."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "Both are processor-loading crashes, but the missing torchvision dependency and the mistral_common import error are unrelated."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "A removed API in v5 is not the same as quantized model persistence breaking after save/load."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44315",
+            "accept": false,
+            "reason": "Different training paths and symptoms: optimizer-state corruption under FSDP2/PEFT vs Liger Kernel not applying with model_init."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:46897",
+            "accept": false,
+            "reason": "Completely different models and bugs; xLSTM training blockers are unrelated to Florence2 loss shifting."
+          },
+          {
+            "left": "issue:46655",
+            "right": "issue:46664",
+            "accept": false,
+            "reason": "Same video processor area, but padding calculation and num_frames semantics are separate issues."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Processor loading without torchvision has nothing to do with save_pretrained shard-size defaults."
+          },
+          {
+            "left": "issue:46298",
+            "right": "issue:46752",
+            "accept": false,
+            "reason": "Noisy logging and an IndexError in apply_chat_template are unrelated."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but for different names and different compatibility problems."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:46298",
+            "accept": false,
+            "reason": "A flaky integration test and a logging-message bug are not the same defect."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44117",
+            "accept": false,
+            "reason": "AddedToken keyword duplication and tokenizer-mapping None handling are different tokenizer bugs."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Both are v5-era regressions, but one is API removal and the other is quantization/offloading behavior."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:46821",
+            "accept": false,
+            "reason": "A PyTorch precision warning from bfloat16 compilation is unrelated to a window_size AttributeError."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:46701",
+            "accept": false,
+            "reason": "Tokenizer mapping fallback and a DiffusionGemma pipeline failure are different code paths and failures."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 57,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:36683",
+        "issue:42757",
+        "issue:42890",
+        "issue:42907",
+        "issue:43012",
+        "issue:43064",
+        "issue:43066",
+        "issue:43097",
+        "issue:43208",
+        "issue:43352",
+        "issue:43502",
+        "issue:43550",
+        "issue:43582",
+        "issue:43824",
+        "issue:43825",
+        "issue:43827",
+        "issue:43906",
+        "issue:43976",
+        "issue:43986",
+        "issue:43992",
+        "issue:44016",
+        "issue:44077",
+        "issue:44162",
+        "issue:44222",
+        "issue:44368",
+        "issue:44380",
+        "issue:44451",
+        "issue:44484",
+        "issue:44655",
+        "issue:44704",
+        "issue:44734",
+        "issue:44936",
+        "issue:44938",
+        "issue:45103",
+        "issue:45478",
+        "issue:45479",
+        "issue:45584",
+        "issue:45712",
+        "issue:45753",
+        "issue:46132",
+        "issue:46172",
+        "issue:46178",
+        "issue:46298",
+        "issue:46365",
+        "issue:46421",
+        "issue:46710",
+        "issue:47059",
+        "issue:47116"
+      ],
+      "soft_pairs": [
+        "issue:46172|issue:46710",
+        "issue:43502|issue:43992",
+        "issue:44704|issue:45103",
+        "issue:44936|issue:46365",
+        "issue:44451|issue:47116",
+        "issue:44380|issue:46421",
+        "issue:44938|issue:46178",
+        "issue:43986|issue:44016",
+        "issue:43064|issue:47116",
+        "issue:42907|issue:45753",
+        "issue:42907|issue:44655",
+        "issue:43012|issue:43550",
+        "issue:45712|issue:46298",
+        "issue:43012|issue:44368",
+        "issue:43992|issue:44484",
+        "issue:43827|issue:47116",
+        "issue:43208|issue:43550",
+        "issue:43352|issue:43976",
+        "issue:43352|issue:44380",
+        "issue:43825|issue:44655",
+        "issue:43066|issue:43906",
+        "issue:44734|issue:47116",
+        "issue:43097|issue:46298",
+        "issue:45584|issue:46365",
+        "issue:44380|issue:46132",
+        "issue:44162|issue:46172",
+        "issue:44077|issue:46365",
+        "issue:42890|issue:47059",
+        "issue:44222|issue:45478",
+        "issue:42757|issue:43824",
+        "issue:44380|issue:45479",
+        "issue:36683|issue:43582",
+        "issue:43064|issue:44380"
+      ],
+      "prepared_review_unit_hash": "6a23443cb31ba4c592db964535defb69823b4b7812c64b21c299eaa5f19113ed",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 37470,
+        "estimated_input_tokens": 9368,
+        "estimated_eval_tokens": 18992
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "All soft pairs are false positives: they share broad Transformers/v5 or model-family wording, but each reports a different bug, model, or code path. No pull requests are present.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "issue:47116 is the broadest, most recent open report and best matches the cluster\u2019s recurring v5 regression theme.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:47116 is the most general and user-facing issue here, so it is the best representative even though the set is not a true duplicate cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46172",
+            "right": "issue:46710",
+            "accept": false,
+            "reason": "Different problems: a tensor dtype/transpose bug vs an accuracy regression after upgrade."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:43992",
+            "accept": false,
+            "reason": "Unrelated code paths: offline file access vs missing encoder weights on load."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Different subsystems: AutoProcessor kwarg forwarding vs docstring parsing crash."
+          },
+          {
+            "left": "issue:44936",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Different failures: trainer evaluate-after-train vs importing Trainer being blocked by datasets."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:47116",
+            "accept": false,
+            "reason": "Different bug classes: model loading failure vs streaming pipeline regression."
+          },
+          {
+            "left": "issue:44380",
+            "right": "issue:46421",
+            "accept": false,
+            "reason": "Different issues: backend config ignored in attention vs cache lazy-init recompilation."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "Different causes: Python 3.14 import/load failure vs invalid backend mapping."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44016",
+            "accept": false,
+            "reason": "Different areas: AutoProcessor/video model crash vs a notebook syntax error."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:47116",
+            "accept": false,
+            "reason": "Different problems: FSDP2/PEFT optimizer state bug vs streaming pipeline regression."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Different model-specific bugs: xLSTM training issues vs Qwen3_5 image infinite loop."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Training xLSTM bugs are unrelated to Pipeline save_pretrained behavior."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Both mention compile/precision, but they are different model/backend-specific warnings."
+          },
+          {
+            "left": "issue:45712",
+            "right": "issue:46298",
+            "accept": false,
+            "reason": "Left is dummy-class cleanup/repo hygiene; right is a noisy logging message."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Different warnings: bfloat16 compile precision vs tie_word_embeddings config warning."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Different serialization issues: missing state dict weight vs shard-size default question."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:47116",
+            "accept": false,
+            "reason": "Docs referencing pipeline after removals is not the same as a runtime streaming regression."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Different model families and symptoms: xLSTM training blockers vs Bamba SDPA compile failure."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "Different failures: unsupported Flash Attention vs Python version compatibility."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Unsupported attention backend vs ignored attention scaling are distinct issues."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Different pipeline-related bugs: wrong error message vs inability to save Pipeline objects."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "Tokenizer decoder type bug is unrelated to an isolated reproduction of a different issue."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:47116",
+            "accept": false,
+            "reason": "Serve/KV-cache tensor indexing crash is a different code path from streaming pipeline regression."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:46298",
+            "accept": false,
+            "reason": "Removed API warning is unrelated to a noisy logging message."
+          },
+          {
+            "left": "issue:45584",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Whisper generation bug is unrelated to importing Trainer."
+          },
+          {
+            "left": "issue:44380",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "GPT2 attention backend behavior is unrelated to AttentionInterface.register semantics."
+          },
+          {
+            "left": "issue:44162",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "ESM2 breakage is unrelated to the deepseek dtype/transpose bug."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "PatchTSMixer post_init policy and Trainer import failure are unrelated."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "Seed-related integration-test flakiness is unrelated to the huggingface_hub CLI attribute error."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both concern MoE/FP8, but one is save_pretrained and the other is from_pretrained; not the same bug."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Different import failures from different packages and symbols."
+          },
+          {
+            "left": "issue:44380",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Attention-scaling backend behavior is unrelated to degenerate zero loss in sequence classification."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "Gemma3 config attribute error is unrelated to an Apple Silicon allocator TypeError."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Trainer/FSDP optimizer-state corruption is unrelated to GPT2 attention scaling."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 58,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:29870",
+        "issue:36010",
+        "issue:36296",
+        "issue:36683",
+        "issue:38617",
+        "issue:42175",
+        "issue:42757",
+        "issue:42994",
+        "issue:43012",
+        "issue:43064",
+        "issue:43208",
+        "issue:43299",
+        "issue:43352",
+        "issue:43475",
+        "issue:43502",
+        "issue:43525",
+        "issue:43526",
+        "issue:43611",
+        "issue:43618",
+        "issue:43756",
+        "issue:43761",
+        "issue:43867",
+        "issue:43937",
+        "issue:43976",
+        "issue:44038",
+        "issue:44077",
+        "issue:44230",
+        "issue:44355",
+        "issue:44380",
+        "issue:44483",
+        "issue:44485",
+        "issue:44492",
+        "issue:44514",
+        "issue:44625",
+        "issue:44655",
+        "issue:44683",
+        "issue:44938",
+        "issue:45216",
+        "issue:45276",
+        "issue:45290",
+        "issue:46050",
+        "issue:46077",
+        "issue:46082",
+        "issue:46365",
+        "issue:46877",
+        "issue:47059",
+        "issue:47110"
+      ],
+      "soft_pairs": [
+        "issue:42175|issue:46365",
+        "issue:43012|issue:44683",
+        "issue:44483|issue:44514",
+        "issue:38617|issue:43502",
+        "issue:43618|issue:44380",
+        "issue:43756|issue:44485",
+        "issue:43618|issue:43761",
+        "issue:45276|issue:46077",
+        "issue:43208|issue:43526",
+        "issue:36010|issue:42757",
+        "issue:44038|issue:44230",
+        "issue:42994|issue:44655",
+        "issue:43299|issue:43867",
+        "issue:44483|issue:45290",
+        "issue:44625|issue:46365",
+        "issue:43352|issue:43937",
+        "issue:44625|issue:46077",
+        "issue:44355|issue:46877",
+        "issue:43611|issue:44655",
+        "issue:36683|issue:43475",
+        "issue:44655|issue:46050",
+        "issue:43976|issue:44938",
+        "issue:44938|issue:47059",
+        "issue:36683|issue:44077",
+        "issue:43867|issue:45216",
+        "issue:36683|issue:47110",
+        "issue:46077|issue:46082",
+        "issue:43618|issue:45216",
+        "issue:36683|issue:43525",
+        "issue:29870|issue:36296",
+        "issue:43064|issue:44492"
+      ],
+      "prepared_review_unit_hash": "6ec74e778f6a050430a8f008d782729e0565874d2878a805e022cd73275a2cdf",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36105,
+        "estimated_input_tokens": 9027,
+        "estimated_eval_tokens": 18310
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly heterogeneous issue set with only one plausible duplicate edge: two CLIP v5 regression reports about missing optional outputs. The rest are superficial overlaps (imports, save/load, config attrs, backend/model-family mentions) but different concrete bugs.",
+        "confidence": 0.63,
+        "canonical_issue_reason": "Issue 43761 is the clearest, most specific report in the only plausible duplicate pair: it names the exact model path (CLIPVisionModel.forward), the regression framing, and the missing output field behavior.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43761 is the best representative because it is concrete, scoped to one code path, and most likely to map cleanly to a fix/test in a duplicate-triage workflow.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42175",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Both are import/dependency complaints, but they concern different failures and different package relationships (backend extras vs Trainer import shadowing)."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "A warning about bfloat16 precision is not the same bug as a compiled flex_attention failure on torch >= 2.9."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "Both involve chat-related APIs, but one is request acceptance in the server endpoint and the other is a tokenizer/chat-template crash on batched input."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:43502",
+            "accept": false,
+            "reason": "One is a missing import symbol, the other is unwanted network access despite local_files_only=True; different code paths and fixes."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Both mention attention/backends, but CLIP attentions missing and GPT2 attention scaling ignored are distinct model-specific regressions."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Both mention RoPE, but they are different models and different problems (Smollm3 layer count vs GLM-5 implementation discussion)."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:43761",
+            "accept": true,
+            "reason": "Both are CLIP v5 regressions about optional outputs not being populated in the forward/output path, and a single fix could plausibly address both attentions and hidden_states."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:46077",
+            "accept": false,
+            "reason": "Gemma4 embedding resize propagation and Gemma4 attention-mask construction are unrelated bugs."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Completely different models and failures; xLSTM training bugs vs BeitImageProcessorFast label reduction."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are import errors, but they involve different symbols and dependencies, so they are not the same underlying issue."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "Same model family, but one is a general transformers/Qwen3-VL-Moe bug while the other is an fp8 inference support request."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Both relate to saving, but quantized model save failure and pipeline save_pretrained failure are different objects and failure modes."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43867",
+            "accept": false,
+            "reason": "Both are loading failures, but one is a base_model_prefix regression and the other is a state_dict sorting issue."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both are chat-related, but one is request validation and the other is assistant-message tool-call tokenization; different bugs."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Qwen3.5 num_labels propagation and Trainer import resolution are unrelated."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:43937",
+            "accept": false,
+            "reason": "Different models and failure types: FlashAttention 2 support gap vs invalid GenerationConfig."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:46077",
+            "accept": false,
+            "reason": "Different Gemma4/Qwen3.5 configuration problems with no shared concrete code path."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:46877",
+            "accept": false,
+            "reason": "One is about compiled Python file runtime errors; the other is noisy doc-lint output at import time."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Loading models with base_model_prefix and saving Pipeline objects are different serialization problems."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Both mention missing attributes, but they are different model families and different missing fields."
+          },
+          {
+            "left": "issue:44655",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Pipeline save_pretrained failure and quantized-model save/from_pretrained breakage are different save paths and bugs."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "Different Python-version compatibility reports; the failing versions and symptoms are not the same."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "Python 3.14 loading failure and CLI tests breaking with a huggingface_hub change are unrelated."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "Gemma3Config missing vocab_size and patchtsmixer post_init validation are unrelated."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both concern model serialization/loading, but one is a load error tied to state_dict ordering and the other is a Qwen3.5 save_pretrained regression."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:47110",
+            "accept": false,
+            "reason": "Missing Gemma3Config vocab_size and a special_tokens list/dict type error are unrelated."
+          },
+          {
+            "left": "issue:46077",
+            "right": "issue:46082",
+            "accept": false,
+            "reason": "Gemma4 attention-mask behavior and LlamaConfig head_dim validation are different configuration issues."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "CLIP output regression and Qwen3.5 checkpoint corruption are unrelated."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Both are missing config attributes, but they affect different configs and different fields."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "Bart generate() decoder_attention_mask bug and tensor-parallel training bug are different failures."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "Optimizer-state corruption under FSDP2/PEFT is unrelated to a typo in cache strategies."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The analysis is mostly grounded and conservative. The single accepted pair, issue:43618 with issue:43761, is a plausible same-bug match because both concern CLIP optional outputs not being populated. The other pairwise rejections are also cautious. Minor tightening: avoid calling issue:43618 explicitly a v5 regression unless the issue body supports that, since the title alone does not."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 59,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28282",
+        "issue:29870",
+        "issue:33453",
+        "issue:36296",
+        "issue:41669",
+        "issue:42890",
+        "issue:42898",
+        "issue:43010",
+        "issue:43065",
+        "issue:43502",
+        "issue:43575",
+        "issue:43618",
+        "issue:43723",
+        "issue:43867",
+        "issue:43874",
+        "issue:43976",
+        "issue:43992",
+        "issue:44062",
+        "issue:44075",
+        "issue:44077",
+        "issue:44112",
+        "issue:44222",
+        "issue:44246",
+        "issue:44355",
+        "issue:44360",
+        "issue:44393",
+        "issue:44485",
+        "issue:44556",
+        "issue:44734",
+        "issue:44855",
+        "issue:44908",
+        "issue:44964",
+        "issue:45125",
+        "issue:45278",
+        "issue:45310",
+        "issue:45357",
+        "issue:45375",
+        "issue:45440",
+        "issue:45478",
+        "issue:45561",
+        "issue:45850",
+        "issue:46202",
+        "issue:46285",
+        "issue:46365",
+        "issue:46617",
+        "issue:46664",
+        "issue:46851",
+        "issue:47000"
+      ],
+      "soft_pairs": [
+        "issue:43010|issue:45278",
+        "issue:43992|issue:45310",
+        "issue:44075|issue:46365",
+        "issue:44908|issue:46851",
+        "issue:29870|issue:33453",
+        "issue:44393|issue:47000",
+        "issue:43867|issue:46202",
+        "issue:44355|issue:44855",
+        "issue:43010|issue:45850",
+        "issue:36296|issue:43575",
+        "issue:43867|issue:45375",
+        "issue:41669|issue:44246",
+        "issue:42890|issue:45561",
+        "issue:43618|issue:44077",
+        "issue:43010|issue:43976",
+        "issue:43867|issue:45357",
+        "issue:43992|issue:45478",
+        "issue:28282|issue:33453",
+        "issue:43618|issue:46285",
+        "issue:44360|issue:45440",
+        "issue:42890|issue:44964",
+        "issue:42890|issue:44112",
+        "issue:43502|issue:44062",
+        "issue:42898|issue:46664",
+        "issue:44734|issue:46617",
+        "issue:44222|issue:45125",
+        "issue:43723|issue:44556",
+        "issue:43502|issue:44485",
+        "issue:43010|issue:43065",
+        "issue:43874|issue:46617"
+      ],
+      "prepared_review_unit_hash": "01cbbe9eda0a06715c99374ed7d1df2a5fc67970dcf87d5139ecb7c20734f865",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36306,
+        "estimated_input_tokens": 9077,
+        "estimated_eval_tokens": 18410
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This is a heterogeneous cluster of mostly unrelated issues: import-time regressions, model-loading failures, generation bugs, tensor-parallel memory problems, tokenizer behavior changes, and test flakiness. I do not see a single underlying bug/change that ties these together, so the soft edges should all be rejected.",
+        "confidence": 0.96,
+        "canonical_issue_reason": "Issue 45278 is the broadest umbrella among the set because it describes a version-upgrade regression with multiple import failures. That said, the cluster is not actually cohesive, so this is only a loose representative, not a true duplicate anchor.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "45278 is the best overall representative by breadth, but the issues are too diverse for a strong canonical choice.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43010",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Cache update no-grad decoration and broad import failures are different problems with different code paths."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Both are from_pretrained/loading issues, but they involve different models and different missing-key failures."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "SGD args being ignored is unrelated to Trainer import failures caused by datasets."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:46851",
+            "accept": false,
+            "reason": "Scheduler kwargs propagation and Qwen3.5 variable-length position_ids are unrelated."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Bart generate decoder_attention_mask bug and tokenizer loading regression are separate issues."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:47000",
+            "accept": false,
+            "reason": "Qwen3-VL bounding-box output behavior and RTDetrLoss infeasible cost matrix are different vision failures."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "State-dict sorting load failure is unrelated to Gemma pipeline roundtrip tokenization on \u2581."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "Compiled-file import errors and the DebertaV2Model Python 3.13 decorator parsing issue are distinct causes."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "A cache update decorator change does not match a broad \"many models break\" report."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Tensor-parallel training bug and Qwen2-57B-A14B-Instruct OOM are not the same concrete failure."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "State-dict ordering and missing Qwen3.5MoeVisionConfig field are unrelated loading/config bugs."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "Import-star removal/perf regression is not the same as intermittent slow import of transformers."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Missing seeding in a model integration test and xdist race on captured_info.txt are different flaky-test issues."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "CLIPOutput attentions assignment and patchtsmixer post_init policy are unrelated."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "no_grad decoration for caches/layers is unrelated to the Python-version compatibility regression."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "State-dict loading order and incorrect visual encoder keys in save_pretrained are different serialization bugs."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Missing PyTorch dependency error and tokenizer loading regression are not the same issue."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "Missing attentions assignment in CLIPOutput is unrelated to the CLIP text-model structure change."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "A DSA indexer activation mismatch and DeepseekV3MoE divergence are different implementation issues."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Seed-related test flakiness does not match a Phi-4 multimodal model loading failure."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Seed omissions in integration tests are unrelated to the GraniteSpeech device-override CI failure."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44062",
+            "accept": false,
+            "reason": "local_files_only API calls and AddedToken special-argument duplication are unrelated tokenizer/loading bugs."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:46664",
+            "accept": false,
+            "reason": "Tokenizer cleanup behavior changes and SmolVLM video num_frames semantics are unrelated."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "KV-cache continuation tensor indexing and trust_remote_code cache FileNotFoundError are different runtime failures."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "FP8 save_pretrained MoE behavior and missing _tp_plan for tensor parallelism are not the same bug."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Tokenizer loading regression and checkpoint reload incompatibility after upgrade are related only broadly, not the same concrete failure."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "local_files_only network leakage and GLM-5 RoPE implementation are unrelated."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "Missing get_number_of_image_patches in GLM46V is unrelated to cache FileNotFoundError when loading trust_remote_code models."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 60,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:29870",
+        "issue:30333",
+        "issue:30990",
+        "issue:31795",
+        "issue:33453",
+        "issue:34689",
+        "issue:36296",
+        "issue:38617",
+        "issue:41628",
+        "issue:42548",
+        "issue:42757",
+        "issue:43010",
+        "issue:43232",
+        "issue:43299",
+        "issue:43352",
+        "issue:43475",
+        "issue:43502",
+        "issue:43576",
+        "issue:43824",
+        "issue:43827",
+        "issue:43867",
+        "issue:43901",
+        "issue:44222",
+        "issue:44230",
+        "issue:44315",
+        "issue:44351",
+        "issue:44355",
+        "issue:44360",
+        "issue:44387",
+        "issue:44485",
+        "issue:44556",
+        "issue:44749",
+        "issue:44961",
+        "issue:45071",
+        "issue:45125",
+        "issue:45375",
+        "issue:45468",
+        "issue:45478",
+        "issue:45850",
+        "issue:46132",
+        "issue:46143",
+        "issue:46144",
+        "issue:46209",
+        "issue:46617",
+        "issue:46664",
+        "issue:46726",
+        "issue:47059",
+        "issue:47110"
+      ],
+      "soft_pairs": [
+        "issue:45850|issue:47110",
+        "issue:43824|issue:46617",
+        "issue:43576|issue:46132",
+        "issue:44387|issue:44556",
+        "issue:42548|issue:44351",
+        "issue:45125|issue:45478",
+        "issue:30333|issue:33453",
+        "issue:44355|issue:44749",
+        "issue:30990|issue:31795",
+        "issue:45125|issue:45375",
+        "issue:43502|issue:44222",
+        "issue:43502|issue:44961",
+        "issue:43827|issue:44961",
+        "issue:46664|issue:46726",
+        "issue:43010|issue:43232",
+        "issue:30333|issue:36296",
+        "issue:43299|issue:44556",
+        "issue:29870|issue:31795",
+        "issue:33453|issue:34689",
+        "issue:44355|issue:45071",
+        "issue:42757|issue:43576",
+        "issue:43576|issue:45071",
+        "issue:43576|issue:47059",
+        "issue:31795|issue:38617",
+        "issue:43576|issue:43824",
+        "issue:44315|issue:44360",
+        "issue:44961|issue:45468",
+        "issue:44485|issue:44961",
+        "issue:44351|issue:46617",
+        "issue:43352|issue:44961",
+        "issue:36296|issue:43901",
+        "issue:43475|issue:46617",
+        "issue:46143|issue:46144",
+        "issue:41628|issue:42548",
+        "issue:45375|issue:47110",
+        "issue:44222|issue:44230",
+        "issue:43867|issue:47110",
+        "issue:44315|issue:46209"
+      ],
+      "prepared_review_unit_hash": "0d9a30fbe88ab9732dedacf285f60ad32c100af7fbf435c27f0835dfb5bac699",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 38,
+        "serialized_chars": 38509,
+        "estimated_input_tokens": 9628,
+        "estimated_eval_tokens": 19512
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is noisy: it mixes broad v5 regression reports, import/load failures, docs issues, and model-specific bugs. None of the soft edges look like true duplicates of the same underlying fix.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "Issue 43576 is the closest thing to a hub: it is broad, v5-related, and appears in several soft matches, so it is the best representative even though the cluster is not truly duplicate-heavy.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43576 is the most central umbrella-style report among the candidates, but it still only loosely groups nearby v5 breakage complaints rather than one concrete bug.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45850",
+            "right": "issue:47110",
+            "accept": false,
+            "reason": "Both are vague breakage reports, but one is a generic v5 complaint and the other is a specific special_tokens error; not the same bug."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "Different failure modes: missing Qwen2.5-VL export vs FileNotFoundError loading from cache."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "CLI env command breakage is unrelated to AttentionInterface.register behavior changes."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "One is int4 quantization OOM from higher reserved memory; the other is checkpoint reload incompatibility after upgrade."
+          },
+          {
+            "left": "issue:42548",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and likely different export/regression causes."
+          },
+          {
+            "left": "issue:45125",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both involve Qwen3.5-Moe, but one is tensor-parallel support and the other is from_pretrained loading regression."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "MLFlow reporting on failed training jobs is unrelated to tokenizer loading regressions."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "Compiled Python file errors are not the same as the data filtering slowdown regression."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:31795",
+            "accept": false,
+            "reason": "Sentence Transformers loading hang and docs confusion about forward args are unrelated."
+          },
+          {
+            "left": "issue:45125",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both mention Qwen3.5-Moe, but one is tensor-parallel support while the other is missing a config field silently dropped by strict parsing."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Local-files-only network requests and FP8 moe save_pretrained are different code paths."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "No shared underlying bug; local_files_only network leakage is unrelated to the placeholder title."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "Docs still mentioning pipeline() and the unrelated placeholder issue are not duplicates."
+          },
+          {
+            "left": "issue:46664",
+            "right": "issue:46726",
+            "accept": false,
+            "reason": "Video frame count semantics and AutoProcessor backend-missing messaging are separate processor issues."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43232",
+            "accept": false,
+            "reason": "Both concern generation/cache behavior, but one is a no_grad decoration request and the other is a sync_gpus ordering bug."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "MLFlow failure reporting and tensor-parallel training are unrelated."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Qwen3VL MoE loading regression and reload-after-upgrade failure are not the same concrete defect."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:31795",
+            "accept": false,
+            "reason": "A BART generate() UnboundLocalError is unrelated to documentation confusion about forward inputs."
+          },
+          {
+            "left": "issue:33453",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Tokenizer loading regression and Llama 3.2 90B Vision model loading failure are different model/load bugs."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Compiled-file errors and PretrainedConfig type-checking regression do not share the same underlying cause."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:43576",
+            "accept": false,
+            "reason": "huggingface_hub import failure is a dependency/export issue, not the transformers env command bug."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "CLI env command breakage and PretrainedConfig type checking are unrelated regressions."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "CLI env command regression is not the same as the _add_completion attribute error in tests/cli."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Docs confusion about forward parameters is unrelated to the missing layer_type_validation import."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "A broken env CLI command and a missing model export are separate issues."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Liger Kernel not applied with model_init and a DSA indexer ReLU discussion are unrelated."
+          },
+          {
+            "left": "issue:44961",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "The placeholder title and Gemma-4 audio positional encoding bug are clearly unrelated."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "GLM-5 RoPE implementation discussion does not match the placeholder issue."
+          },
+          {
+            "left": "issue:44351",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "Missing HybridCache export and cache FileNotFoundError are different import vs filesystem failures."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "Flash Attention support error for NemotronH is unrelated to the placeholder issue."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Tensor parallel training bug and TextClassificationPipeline docs mismatch are unrelated."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "SAM 3 Video missing attribute and trust_remote_code cache loading failure are different bugs."
+          },
+          {
+            "left": "issue:46143",
+            "right": "issue:46144",
+            "accept": false,
+            "reason": "Missing kwargs passthrough in RoFormer methods and not using the attention interface are related thematically but not the same concrete bug."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:42548",
+            "accept": false,
+            "reason": "AutoImageProcessor and PreTrainedModel import errors are separate missing-export regressions."
+          },
+          {
+            "left": "issue:45375",
+            "right": "issue:47110",
+            "accept": false,
+            "reason": "Missing config field in Qwen3.5MoeVisionConfig is unrelated to the special_tokens list/dict error."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "FP8 save_pretrained moe and FP8 inference support for Qwen3 VL/Qwen3.5 Moe are different phases and likely different bugs."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:47110",
+            "accept": false,
+            "reason": "State_dict sorting load error is unrelated to the special_tokens attribute error."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Applying Liger Kernel in model_init and FP8 inference on B200 are different execution paths and problems."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 61,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:31795",
+        "issue:33453",
+        "issue:41669",
+        "issue:42371",
+        "issue:42890",
+        "issue:43352",
+        "issue:43408",
+        "issue:43525",
+        "issue:43572",
+        "issue:43576",
+        "issue:43673",
+        "issue:43976",
+        "issue:44038",
+        "issue:44077",
+        "issue:44230",
+        "issue:44246",
+        "issue:44273",
+        "issue:44295",
+        "issue:44297",
+        "issue:44315",
+        "issue:44351",
+        "issue:44360",
+        "issue:44393",
+        "issue:44485",
+        "issue:44492",
+        "issue:44556",
+        "issue:44623",
+        "issue:44704",
+        "issue:44829",
+        "issue:44861",
+        "issue:44908",
+        "issue:45003",
+        "issue:45071",
+        "issue:45092",
+        "issue:45125",
+        "issue:45230",
+        "issue:45341",
+        "issue:45362",
+        "issue:45468",
+        "issue:45542",
+        "issue:45753",
+        "issue:45850",
+        "issue:46095",
+        "issue:46218",
+        "issue:46328",
+        "issue:46664",
+        "issue:47059",
+        "issue:47110"
+      ],
+      "soft_pairs": [
+        "issue:43525|issue:47110",
+        "issue:45362|issue:45850",
+        "issue:43673|issue:44492",
+        "issue:44485|issue:46664",
+        "issue:44273|issue:45542",
+        "issue:44623|issue:45230",
+        "issue:44038|issue:45125",
+        "issue:44273|issue:45003",
+        "issue:44360|issue:46664",
+        "issue:41669|issue:43673",
+        "issue:44393|issue:45071",
+        "issue:44393|issue:45753",
+        "issue:43576|issue:44556",
+        "issue:43525|issue:45850",
+        "issue:44556|issue:45092",
+        "issue:43572|issue:45850",
+        "issue:44230|issue:46095",
+        "issue:44492|issue:45850",
+        "issue:45850|issue:47059",
+        "issue:42371|issue:44492",
+        "issue:41669|issue:46218",
+        "issue:43408|issue:44315",
+        "issue:44246|issue:44273",
+        "issue:42890|issue:44908",
+        "issue:43576|issue:44908",
+        "issue:43352|issue:44393",
+        "issue:43976|issue:44393",
+        "issue:44295|issue:45230",
+        "issue:44297|issue:45230",
+        "issue:44908|issue:45071",
+        "issue:44492|issue:46328",
+        "issue:41669|issue:44351",
+        "issue:44704|issue:44908",
+        "issue:41669|issue:45341",
+        "issue:43976|issue:46218",
+        "issue:44077|issue:44861",
+        "issue:31795|issue:33453",
+        "issue:43576|issue:44273",
+        "issue:44360|issue:45468",
+        "issue:44315|issue:44829"
+      ],
+      "prepared_review_unit_hash": "28a93d0526d13cea73e65a226533419ea33b50ed7eccedbba3d3ef32ef65322f",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 40,
+        "serialized_chars": 39119,
+        "estimated_input_tokens": 9780,
+        "estimated_eval_tokens": 19816
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are a heterogeneous set of unrelated issues. None of the soft pairs look like the same underlying bug or change, so all soft edges are rejected and there is no clear canonical representative.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43525",
+            "right": "issue:47110",
+            "accept": false,
+            "reason": "Different failures: missing config attribute vs. special_tokens being a list instead of a mapping."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "One is a specific Qwen3.5 chat crash; the other is a broad complaint about many models breaking in v5."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "Generation cache missing in chunked prefill is unrelated to a typo in cache strategy docs/code."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:46664",
+            "accept": false,
+            "reason": "RoPE implementation discussion and video num_frames semantics are different model/component bugs."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Lazy loading problems are unrelated to the TensorFlow backend error from tensorboard-only installs."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "Missing processor save files is not the same as a vague generic bug report."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe issue vs missing tensor-parallel plan for Qwen3.5-Moe are different concrete problems."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Lazy loading/import behavior is not the same as unsafe sys.modules access."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:46664",
+            "accept": false,
+            "reason": "DSA indexer/ReLU discussion is unrelated to video frame count semantics."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "Import-star performance issue is unrelated to generation cache handling."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Qwen3-VL bbox hallucination is unrelated to PretrainedConfig type-checking regressions."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Both mention Qwen3/VL image behavior, but one is bbox output errors and the other is an infinite loop on a specific image."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Broken transformers env CLI and checkpoint reload incompatibility are different failure modes."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "A specific pad_token_id attribute error is not the same as a broad v5 compatibility complaint."
+          },
+          {
+            "left": "issue:44556",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Both involve v5 upgrade incompatibilities, but they concern different loading paths and model families."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "StableLm pad_token index regression is not the same as a generic 'many models break' report."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:46095",
+            "accept": false,
+            "reason": "FP8 inference support for Qwen models is unrelated to DeepSeek weight conversion tooling."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "A cache-strategy typo is not the same as a broad compatibility complaint."
+          },
+          {
+            "left": "issue:45850",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "General v5 model breakage is unrelated to the CLI completion attribute error from huggingface_hub 1.22.0."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "TF32 API migration guidance is unrelated to cache strategy typos."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Import-star performance regression is not the same as a missing Qwen3_5 module import."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:44315",
+            "accept": false,
+            "reason": "Model type mismatch warning and Liger kernel application under model_init are different issues."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44273",
+            "accept": false,
+            "reason": "Slow import sometimes and lazy loading not working properly overlap conceptually, but they are not the same concrete bug."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Missing set_seed in tests is unrelated to inverse_sqrt scheduler kwargs being ignored."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "CLI env breakage does not match a scheduler configuration bug."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Flash Attention support gating is unrelated to tied-weight key handling crashing on list input."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44393",
+            "accept": false,
+            "reason": "Python version compatibility failure is unrelated to Qwen3-VL bbox hallucination."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "Position_ids buffer access error is not the same as a generic bug report."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "Tokenizer save_pretrained class mismatch is unrelated to a generic bug report."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Scheduler kwarg handling and config type checking are unrelated bugs."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "AutoProcessor cached_file kwargs forwarding is unrelated to scheduler timescale handling."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Import-star performance issue is unrelated to a testing_utils bug."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Python compatibility issues are unrelated to a missing model import."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Optional post_init enforcement is unrelated to tied-weight key list handling."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Documentation confusion about forward args is unrelated to tokenizer loading regression."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44273",
+            "accept": false,
+            "reason": "Broken CLI env command and lazy loading problems are different subsystems and symptoms."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "DSA indexer/ReLU discussion is unrelated to Gemma-4 audio positional encoding."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Liger kernel application under model_init is unrelated to flash_attention_3 degenerate training."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 62,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:28348",
+        "issue:29870",
+        "issue:30333",
+        "issue:31515",
+        "issue:31795",
+        "issue:34689",
+        "issue:36296",
+        "issue:38617",
+        "issue:41669",
+        "issue:42371",
+        "issue:42913",
+        "issue:43066",
+        "issue:43298",
+        "issue:43352",
+        "issue:43502",
+        "issue:43519",
+        "issue:43576",
+        "issue:43618",
+        "issue:43704",
+        "issue:43717",
+        "issue:44230",
+        "issue:44261",
+        "issue:44263",
+        "issue:44273",
+        "issue:44351",
+        "issue:44485",
+        "issue:44492",
+        "issue:44623",
+        "issue:44655",
+        "issue:44861",
+        "issue:44908",
+        "issue:44938",
+        "issue:44998",
+        "issue:45230",
+        "issue:45250",
+        "issue:45290",
+        "issue:45310",
+        "issue:45468",
+        "issue:45850",
+        "issue:46018",
+        "issue:46095",
+        "issue:46144",
+        "issue:46172",
+        "issue:46218",
+        "issue:46655",
+        "issue:46894",
+        "issue:47059",
+        "issue:47110"
+      ],
+      "soft_pairs": [
+        "issue:43519|issue:46655",
+        "issue:44273|issue:44908",
+        "issue:45230|issue:45310",
+        "issue:43519|issue:44485",
+        "issue:45468|issue:46894",
+        "issue:44655|issue:44861",
+        "issue:43576|issue:46218",
+        "issue:43519|issue:44908",
+        "issue:41669|issue:43704",
+        "issue:31795|issue:34689",
+        "issue:44485|issue:45468",
+        "issue:43704|issue:44485",
+        "issue:46018|issue:46172",
+        "issue:43502|issue:43519",
+        "issue:42371|issue:43704",
+        "issue:45850|issue:46218",
+        "issue:43519|issue:44623",
+        "issue:44263|issue:45468",
+        "issue:44261|issue:45468",
+        "issue:43298|issue:46095",
+        "issue:43519|issue:43704",
+        "issue:43352|issue:46218",
+        "issue:44351|issue:46218",
+        "issue:44861|issue:47110",
+        "issue:42913|issue:43066",
+        "issue:44861|issue:47059",
+        "issue:28348|issue:45250",
+        "issue:44908|issue:44998",
+        "issue:44492|issue:44998",
+        "issue:44998|issue:45850",
+        "issue:43618|issue:44998",
+        "issue:44938|issue:44998",
+        "issue:44861|issue:45290",
+        "issue:31515|issue:38617",
+        "issue:44230|issue:44623",
+        "issue:28348|issue:43352",
+        "issue:29870|issue:31515",
+        "issue:31515|issue:36296",
+        "issue:30333|issue:31515",
+        "issue:31515|issue:31795",
+        "issue:44230|issue:45310",
+        "issue:43298|issue:44230",
+        "issue:43717|issue:46144"
+      ],
+      "prepared_review_unit_hash": "41be1038d8425153f8b06b185a78b6dfffdce240552829048c3f41958ac4e442",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 43,
+        "serialized_chars": 39697,
+        "estimated_input_tokens": 9925,
+        "estimated_eval_tokens": 20106
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This is a loose, similarity-driven cluster of mostly unrelated Transformers issues around v5 regressions, model support, loading, and docs. None of the suggested soft pairs look like safe duplicates under a strict triage standard.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "No clear duplicate canonical exists; the items are too heterogeneous. If a representative is needed, issue:45850 is the broadest v5-breakage umbrella, but it is too generic to be a true canonical.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45850 is the best cluster representative because it is the broadest cross-cutting complaint about v5 regressions, but it should not be treated as a precise duplicate target.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43519",
+            "right": "issue:46655",
+            "accept": false,
+            "reason": "Different processors and different bugs: timestamp math vs padding calculation."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Unrelated areas: lazy loading vs scheduler kwargs handling."
+          },
+          {
+            "left": "issue:45230",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Too generic vs specific model-load failure; no shared code-path shown."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Different model components and bug types: processor timestamping vs RoPE implementation."
+          },
+          {
+            "left": "issue:45468",
+            "right": "issue:46894",
+            "accept": false,
+            "reason": "Different models and symptoms; not the same loss/precision bug."
+          },
+          {
+            "left": "issue:44655",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Both touch save paths, but one is pipeline serialization and the other tied-weight key handling."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Broken env command vs import failure for a specific model module."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Processor timestamp bug and LR scheduler bug are unrelated."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:43704",
+            "accept": false,
+            "reason": "Import-star cleanup vs VRAM leak in threaded inference are different changes."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Documentation confusion vs model loading regression."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "Both mention positional encoding, but they are different models and concrete defects."
+          },
+          {
+            "left": "issue:43704",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "VRAM leak in multithreading is unrelated to RoPE implementation."
+          },
+          {
+            "left": "issue:46018",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "Same model family, but shared expert gating and tensor conversion are different bugs."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:43519",
+            "accept": false,
+            "reason": "Offline-download behavior vs processor timestamp calculation."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:43704",
+            "accept": false,
+            "reason": "TF32 API guidance issue is unrelated to VRAM leakage."
+          },
+          {
+            "left": "issue:45850",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Broad v5 complaint vs a specific module import failure."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Processor computation bug vs missing save_pretrained artifacts."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "Different model internals and different failure modes."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "RMS norm epsilon precision issue is unrelated to Gemma-4 positional encoding."
+          },
+          {
+            "left": "issue:43298",
+            "right": "issue:46095",
+            "accept": false,
+            "reason": "Continuous batching audio support and DeepSeek-V4 weight conversion are unrelated."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:43704",
+            "accept": false,
+            "reason": "Processor timestamp bug vs VRAM leak; no shared underlying defect."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Flash Attention support request and import error are different issues."
+          },
+          {
+            "left": "issue:44351",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Different missing symbols/modules; not the same import regression."
+          },
+          {
+            "left": "issue:44861",
+            "right": "issue:47110",
+            "accept": false,
+            "reason": "Tied-weight key handling is unrelated to special_tokens list shape."
+          },
+          {
+            "left": "issue:42913",
+            "right": "issue:43066",
+            "accept": false,
+            "reason": "Both tokenizer-v5 related, but one is general behavior drift and the other is decoder-type mismatch; not enough to treat as the same bug."
+          },
+          {
+            "left": "issue:44861",
+            "right": "issue:47059",
+            "accept": false,
+            "reason": "Save helper AttributeError vs CLI completion failure from huggingface_hub."
+          },
+          {
+            "left": "issue:28348",
+            "right": "issue:45250",
+            "accept": false,
+            "reason": "Specific GPT2 flash-attention support vs broad feature request; not the same concrete change."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Scheduler kwargs bug is unrelated to the placeholder issue."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Cache typo issue is unrelated to the placeholder issue."
+          },
+          {
+            "left": "issue:44998",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "Generic placeholder issue and broad v5 complaint do not describe the same bug."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "CLIPOutput attentions regression is unrelated to the placeholder issue."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Python 3.14 load failure is unrelated to the placeholder issue."
+          },
+          {
+            "left": "issue:44861",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Tied-weights crash is unrelated to chat-template tool-call handling."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Slow from_pretrained loading and missing import symbol are different problems."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "FP8 inference support and processor save_pretrained missing files are unrelated."
+          },
+          {
+            "left": "issue:28348",
+            "right": "issue:43352",
+            "accept": false,
+            "reason": "Both mention Flash Attention 2.0, but they are different models and different support requests."
+          },
+          {
+            "left": "issue:29870",
+            "right": "issue:31515",
+            "accept": false,
+            "reason": "Generation mask crash vs checkpoint loading slowness are unrelated."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "Loading slowness vs tensor-parallel training bug are different code paths."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:31515",
+            "accept": false,
+            "reason": "MLflow reporting on failed jobs is unrelated to loading performance."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:31795",
+            "accept": false,
+            "reason": "Performance issue and documentation issue are not duplicates."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Both involve Qwen3.5/MoE, but one is FP8 inference support and the other is a from_pretrained error."
+          },
+          {
+            "left": "issue:43298",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "Audio continuous batching and Qwen3-VL FP8 support are unrelated."
+          },
+          {
+            "left": "issue:43717",
+            "right": "issue:46144",
+            "accept": false,
+            "reason": "Init-weights behavior in Mamba vs RoFormer attention interface are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 27,
+      "review_unit_index": 63,
+      "review_unit_count": 63,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-514",
+      "nodes": [
+        "issue:37934",
+        "issue:42936",
+        "issue:43240",
+        "issue:43668",
+        "issue:43813",
+        "issue:44031",
+        "issue:46018",
+        "issue:46144"
+      ],
+      "soft_pairs": [
+        "issue:42936|issue:46018",
+        "issue:43668|issue:44031",
+        "issue:42936|issue:43240",
+        "issue:43240|issue:46144",
+        "issue:42936|issue:44031",
+        "issue:42936|issue:46144",
+        "issue:42936|issue:43813",
+        "issue:43668|issue:43813",
+        "issue:43240|issue:43668",
+        "issue:37934|issue:43668",
+        "issue:42936|issue:43668"
+      ],
+      "prepared_review_unit_hash": "16ac7e6b8b3d3365a208685ace08810752de3ea96c587e4c9f3bf6c8a841ca22",
+      "component_budget": {
+        "node_count": 514,
+        "item_count": 514,
+        "soft_pair_count": 1891,
+        "serialized_chars": 980935,
+        "estimated_input_tokens": 245234,
+        "estimated_eval_tokens": 490724
+      },
+      "budget": {
+        "node_count": 8,
+        "item_count": 8,
+        "soft_pair_count": 11,
+        "serialized_chars": 7770,
+        "estimated_input_tokens": 1943,
+        "estimated_eval_tokens": 4142
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These are unrelated issue reports across different models/components (normalization, Mask2Former loss handling, ModernBERT config typing, tokenizer warning, DeepSeek gating, RoFormer attention, etc.). I don\u2019t see a true duplicate cluster here.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No single issue stands out as a canonical duplicate target because the reports describe different bugs in different code paths and subsystems.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "None of these is a clear duplicate hub; if one must be retained as the most substantive standalone report, issue 43240 is the most concrete and actionable, but it is not a duplicate match for the others.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42936",
+            "right": "issue:46018",
+            "accept": false,
+            "reason": "Different models and failures: Mask2Former ignore_value handling vs DeepSeek-V4 shared expert gating."
+          },
+          {
+            "left": "issue:43668",
+            "right": "issue:44031",
+            "accept": false,
+            "reason": "Unrelated bug types: a config type-hint mistake vs a tokenizer regex warning regression."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43240",
+            "accept": false,
+            "reason": "Both concern loss-related code only at a very broad level, but one is a specific Mask2Former ignore_value bug and the other is a missing kwargs passthrough."
+          },
+          {
+            "left": "issue:43240",
+            "right": "issue:46144",
+            "accept": false,
+            "reason": "Different subsystems and symptoms: loss-function kwargs forwarding vs RoFormer attention interface usage."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:44031",
+            "accept": false,
+            "reason": "Mask2Former loss ignore_value issue is unrelated to tokenizer warning behavior."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:46144",
+            "accept": false,
+            "reason": "Different architectures and bug classes; no shared concrete code path."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43813",
+            "accept": false,
+            "reason": "Loss-parameter usage bug vs a typo in a quantization attribute name."
+          },
+          {
+            "left": "issue:43668",
+            "right": "issue:43813",
+            "accept": false,
+            "reason": "Both are small correctness issues, but one is ModernBERT config typing and the other is a spelling/attribute-name error in a different area."
+          },
+          {
+            "left": "issue:43240",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Missing kwargs forwarding in loss computation is not the same as an incorrect type hint on a config field."
+          },
+          {
+            "left": "issue:37934",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Llama4TextL2Norm normalization semantics vs ModernBERTConfig type annotation are different problems."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "These do not appear to share a concrete implementation bug beyond both involving model configuration/behavior."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10519-3",
+      "nodes": [
+        "issue:10519",
+        "issue:11005",
+        "issue:11307"
+      ],
+      "soft_pairs": [
+        "issue:10519|issue:11307",
+        "issue:10519|issue:11005",
+        "issue:11005|issue:11307"
+      ],
+      "prepared_review_unit_hash": "88726b9a5a684cad1c203669db0da2fe728675adb1b583e05f4a40676c4139b9",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 2878,
+        "estimated_input_tokens": 720,
+        "estimated_eval_tokens": 1696
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 2878,
+        "estimated_input_tokens": 720,
+        "estimated_eval_tokens": 1696
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These three issues are about different features in different parts of the library, so I do not see a duplicate cluster here.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "No single canonical issue fits this cluster because the items are unrelated feature requests rather than duplicates of the same bug or change.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 10519 is the most concrete and well-scoped request, but it is still distinct from the other two topics.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:10519",
+            "right": "issue:11307",
+            "accept": false,
+            "reason": "Different topics: token truncation behavior vs. Wav2Vec2 word time offsets."
+          },
+          {
+            "left": "issue:10519",
+            "right": "issue:11005",
+            "accept": false,
+            "reason": "Different subsystems and requests: tokenizer truncation option vs. ReduceLROnPlateau-like scheduler behavior."
+          },
+          {
+            "left": "issue:11005",
+            "right": "issue:11307",
+            "accept": false,
+            "reason": "Unrelated feature requests in separate areas: learning-rate scheduling vs. speech model word alignment offsets."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 3,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-13244-3",
+      "nodes": [
+        "issue:13244",
+        "issue:15980",
+        "issue:16618"
+      ],
+      "soft_pairs": [
+        "issue:13244|issue:16618",
+        "issue:15980|issue:16618"
+      ],
+      "prepared_review_unit_hash": "e8a61c0b2275e65fde6d72b74522855747d848e14caf8a4ad6cebbaa05bb30b5",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 2477,
+        "estimated_input_tokens": 620,
+        "estimated_eval_tokens": 1496
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 2477,
+        "estimated_input_tokens": 620,
+        "estimated_eval_tokens": 1496
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster looks heterogeneous rather than duplicate-heavy: one issue is about Tapas tokenization, one about an unauthenticated private-model download error message, and one about a general model-loading failure for bert-base-uncased. They do not appear to be the same underlying bug.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "Issue #16618 is the broadest and most central-looking model-loading report in the set, so it is the best single representative if one must be chosen.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#16618 is the strongest candidate for the cluster\u2019s main thread because it concerns the core symptom of failing to load a model, whereas the others are about unrelated tokenization or a different auth/error-message scenario.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:13244",
+            "right": "issue:16618",
+            "accept": false,
+            "reason": "Different problem domains: Tapas tokenization vs. a model load failure for bert-base-uncased. No evidence they describe the same bug."
+          },
+          {
+            "left": "issue:15980",
+            "right": "issue:16618",
+            "accept": false,
+            "reason": "#15980 is specifically about the error message when downloading a private model without login, while #16618 is a generic cannot-load-bert-base-uncased report. Similar area, but not the same concrete issue."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 4,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-26350-2",
+      "nodes": [
+        "issue:26350",
+        "issue:33260"
+      ],
+      "soft_pairs": [
+        "issue:26350|issue:33260"
+      ],
+      "prepared_review_unit_hash": "1520105c39a9d92474498f76f0bf6641cce5907180f74540fe4f6da69af30e62",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1590,
+        "estimated_input_tokens": 398,
+        "estimated_eval_tokens": 1052
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1590,
+        "estimated_input_tokens": 398,
+        "estimated_eval_tokens": 1052
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two separate community-contribution feature requests: one for Flash Attention 2 support on more architectures, the other for GGUF support on more architectures. They share a generic wording pattern but target different subsystems and changes.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No single canonical issue: these are distinct feature requests with different underlying code paths and implementation goals.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:26350",
+            "right": "issue:33260",
+            "accept": false,
+            "reason": "Both are architecture-support requests, but for different features (Flash Attention 2 vs GGUF). They are not the same underlying bug or change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 5,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-27088-9",
+      "nodes": [
+        "issue:27088",
+        "issue:28936",
+        "issue:38435",
+        "issue:43834",
+        "issue:44365",
+        "issue:44370",
+        "issue:46222",
+        "issue:46226",
+        "issue:46931"
+      ],
+      "soft_pairs": [
+        "issue:27088|issue:44370",
+        "issue:27088|issue:44365",
+        "issue:27088|issue:43834",
+        "issue:27088|issue:38435",
+        "issue:27088|issue:28936",
+        "issue:28936|issue:43834",
+        "issue:28936|issue:44365",
+        "issue:28936|issue:38435",
+        "issue:28936|issue:44370",
+        "issue:38435|issue:44370",
+        "issue:38435|issue:44365",
+        "issue:38435|issue:43834",
+        "issue:38435|issue:46222",
+        "issue:38435|issue:46226",
+        "issue:43834|issue:44370",
+        "issue:43834|issue:44365",
+        "issue:43834|issue:46222",
+        "issue:43834|issue:46226",
+        "issue:43834|issue:46931",
+        "issue:44365|issue:44370",
+        "issue:44365|issue:46222",
+        "issue:44365|issue:46226",
+        "issue:44365|issue:46931",
+        "issue:44370|issue:46222",
+        "issue:44370|issue:46226",
+        "issue:44370|issue:46931",
+        "issue:46222|issue:46226",
+        "issue:46222|issue:46931",
+        "issue:46226|issue:46931"
+      ],
+      "prepared_review_unit_hash": "6aa28ce7826d74121c68559c71df0576dfd70881da17fff7d8ac51b7c2033ad8",
+      "component_budget": {
+        "node_count": 9,
+        "item_count": 9,
+        "soft_pair_count": 29,
+        "serialized_chars": 15980,
+        "estimated_input_tokens": 3995,
+        "estimated_eval_tokens": 8246
+      },
+      "budget": {
+        "node_count": 9,
+        "item_count": 9,
+        "soft_pair_count": 29,
+        "serialized_chars": 15980,
+        "estimated_input_tokens": 3995,
+        "estimated_eval_tokens": 8246
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster is mostly a set of i18n translation-request issues, not a true duplicate set. The high similarity appears to come from repeated template text rather than the same underlying issue. I\u2019d treat 43834 as the best representative of the recurring template-style issues, but keep the language-specific requests separate.",
+        "confidence": 0.82,
+        "canonical_issue_reason": "issue:43834 is the cleanest representative of the repeated generic i18n translation template and is the earliest of the near-identical placeholder-style issues in this cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43834 best represents the recurring template-like translation-request subgroup, but the cluster as a whole should not be collapsed into one duplicate issue.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:27088",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Different language translation requests; similar boilerplate does not make them the same issue."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Different language translation requests; no evidence they describe the same underlying task."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:43834",
+            "accept": false,
+            "reason": "Turkish-specific request vs a generic/template-like translation issue; not the same concrete issue."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:38435",
+            "accept": false,
+            "reason": "Different languages (Turkish vs Romanian); these are separate translation requests."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:28936",
+            "accept": false,
+            "reason": "Different languages (Turkish vs Spanish); similar wording is just the shared i18n template."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:43834",
+            "accept": false,
+            "reason": "Spanish translation request is distinct from the generic/template-style issue."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Spanish translation request vs a template-like generic issue; not the same underlying issue."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:38435",
+            "accept": false,
+            "reason": "Different languages (Spanish vs Romanian); these should not be merged."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Different language request; similarity is only from shared boilerplate."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Romanian request and generic/template issue are not the same concrete bug/change."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Romanian request vs generic/template issue; no evidence of duplication."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:43834",
+            "accept": false,
+            "reason": "Romanian translation request is distinct from the placeholder-style issue title."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:46222",
+            "accept": false,
+            "reason": "Different language/request content; similarity is not enough to establish a duplicate."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:46226",
+            "accept": false,
+            "reason": "Different language/request content; not enough evidence of the same underlying issue."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "These look like repeated template-form issues, but there is no concrete evidence they are the same actionable issue."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "High textual similarity comes from boilerplate/template text, not a shared underlying bug or change."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:46222",
+            "accept": false,
+            "reason": "Similar generic phrasing, but the actual issue content is unclear and not demonstrably the same."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:46226",
+            "accept": false,
+            "reason": "Template-level similarity only; insufficient evidence these are the same issue."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:46931",
+            "accept": false,
+            "reason": "Different titles/content; the similarity score appears driven by shared template structure."
+          },
+          {
+            "left": "issue:44365",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Near-identical template text is not enough to prove a single underlying issue."
+          },
+          {
+            "left": "issue:44365",
+            "right": "issue:46222",
+            "accept": false,
+            "reason": "Similar boilerplate, but not enough evidence they refer to the same concrete change."
+          },
+          {
+            "left": "issue:44365",
+            "right": "issue:46226",
+            "accept": false,
+            "reason": "Template similarity only; do not merge without clearer duplicate evidence."
+          },
+          {
+            "left": "issue:44365",
+            "right": "issue:46931",
+            "accept": false,
+            "reason": "Different issue content; similarity is likely from the shared template."
+          },
+          {
+            "left": "issue:44370",
+            "right": "issue:46222",
+            "accept": false,
+            "reason": "Looks like the same issue template, but not the same underlying request."
+          },
+          {
+            "left": "issue:44370",
+            "right": "issue:46226",
+            "accept": false,
+            "reason": "Template similarity only; no concrete evidence of duplication."
+          },
+          {
+            "left": "issue:44370",
+            "right": "issue:46931",
+            "accept": false,
+            "reason": "Different issue content; similarity appears superficial."
+          },
+          {
+            "left": "issue:46222",
+            "right": "issue:46226",
+            "accept": false,
+            "reason": "Identical titles suggest possible duplication, but the issue content is too unclear to confirm the same underlying actionable request."
+          },
+          {
+            "left": "issue:46222",
+            "right": "issue:46931",
+            "accept": false,
+            "reason": "Similar language-like text, but not enough evidence these are the same issue."
+          },
+          {
+            "left": "issue:46226",
+            "right": "issue:46931",
+            "accept": false,
+            "reason": "Similar language-like text, but not enough evidence these are the same issue."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 6,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-36718-21",
+      "nodes": [
+        "issue:36718",
+        "issue:36817",
+        "issue:36879",
+        "issue:37015",
+        "issue:41115",
+        "issue:41211",
+        "issue:42503",
+        "issue:42738",
+        "issue:42954",
+        "issue:43671",
+        "issue:43837",
+        "issue:43909",
+        "issue:44121",
+        "issue:44205",
+        "issue:45027",
+        "issue:45177",
+        "issue:45322",
+        "issue:46133",
+        "issue:46645",
+        "issue:46653",
+        "issue:47004"
+      ],
+      "soft_pairs": [
+        "issue:36718|issue:47004",
+        "issue:36718|issue:41115",
+        "issue:36817|issue:41211",
+        "issue:36817|issue:43671",
+        "issue:36817|issue:36879",
+        "issue:36817|issue:42954",
+        "issue:36879|issue:41211",
+        "issue:36879|issue:43909",
+        "issue:36879|issue:44121",
+        "issue:36879|issue:43837",
+        "issue:36879|issue:43671",
+        "issue:37015|issue:43837",
+        "issue:37015|issue:44121",
+        "issue:37015|issue:43909",
+        "issue:36879|issue:37015",
+        "issue:37015|issue:42503",
+        "issue:41115|issue:43837",
+        "issue:41115|issue:44121",
+        "issue:41115|issue:42503",
+        "issue:41115|issue:45027",
+        "issue:41211|issue:43671",
+        "issue:42503|issue:43837",
+        "issue:42503|issue:43671",
+        "issue:42503|issue:42954",
+        "issue:42738|issue:45027",
+        "issue:42503|issue:42738",
+        "issue:42738|issue:42954",
+        "issue:42738|issue:46653",
+        "issue:42954|issue:44205",
+        "issue:42954|issue:43671",
+        "issue:42954|issue:45177",
+        "issue:42954|issue:46653",
+        "issue:43671|issue:43837",
+        "issue:43671|issue:45027",
+        "issue:43671|issue:45177",
+        "issue:43671|issue:44121",
+        "issue:43837|issue:43909",
+        "issue:43837|issue:45322",
+        "issue:43837|issue:44121",
+        "issue:43909|issue:44205",
+        "issue:43909|issue:46645",
+        "issue:44121|issue:45322",
+        "issue:43837|issue:45027",
+        "issue:37015|issue:45027",
+        "issue:43837|issue:45177",
+        "issue:45322|issue:46133",
+        "issue:42503|issue:46653",
+        "issue:45027|issue:46653"
+      ],
+      "prepared_review_unit_hash": "d6a7819405bf5c6224358cbc0652663bb53bf3dbdd16321fcc4afc7bf5d873cb",
+      "component_budget": {
+        "node_count": 21,
+        "item_count": 21,
+        "soft_pair_count": 48,
+        "serialized_chars": 28616,
+        "estimated_input_tokens": 7154,
+        "estimated_eval_tokens": 14564
+      },
+      "budget": {
+        "node_count": 21,
+        "item_count": 21,
+        "soft_pair_count": 48,
+        "serialized_chars": 28616,
+        "estimated_input_tokens": 7154,
+        "estimated_eval_tokens": 14564
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This is a loose thematic cluster of distinct model-request issues. The titles mostly share generic wording like \"Add\"/\"Support\"/\"Proposal,\" but the underlying artifacts span unrelated models, architectures, and modalities, so I found no true duplicates to merge.",
+        "confidence": 0.96,
+        "canonical_issue_reason": "No single canonical issue: the items are separate requests for different models/features, and the similarity is mostly in generic request phrasing rather than the same underlying change.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:42503 is the cleanest representative of the broad \"add a model\" pattern, but it is still not a duplicate anchor for the rest of the set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:36718",
+            "right": "issue:47004",
+            "accept": false,
+            "reason": "Both are keypoint-related, but RoMa matcher vs LoMa detector/matcher are different requested components, not the same underlying change."
+          },
+          {
+            "left": "issue:36718",
+            "right": "issue:41115",
+            "accept": false,
+            "reason": "RoMa keypoint matching and MiniCPM3 architecture support are unrelated requests."
+          },
+          {
+            "left": "issue:36817",
+            "right": "issue:41211",
+            "accept": false,
+            "reason": "EuroBert config support and DEIMv2 are different model additions."
+          },
+          {
+            "left": "issue:36817",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "EuroBert config and Qwen3-TTS support are unrelated features."
+          },
+          {
+            "left": "issue:36817",
+            "right": "issue:36879",
+            "accept": false,
+            "reason": "Different model families and different integration work; not the same change."
+          },
+          {
+            "left": "issue:36817",
+            "right": "issue:42954",
+            "accept": false,
+            "reason": "EuroBert and XiaoMi MiMo are separate model requests."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:41211",
+            "accept": false,
+            "reason": "RF-DETR and DEIMv2 are distinct model additions."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "RF-DETR object detection and LFM2.5 Audio support are different code paths."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "RF-DETR and weight-sparse transformer requests are unrelated."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "RF-DETR and Qwen3-ASR are different model/support requests."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "RF-DETR and Qwen3-TTS are unrelated additions."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "NeoBERT and Qwen3-ASR are different model requests."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "NeoBERT and the OpenAI weight-sparse transformer are unrelated."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "NeoBERT and LFM2.5 Audio target different model types and code paths."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:37015",
+            "accept": false,
+            "reason": "RF-DETR and NeoBERT are separate model additions."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:42503",
+            "accept": false,
+            "reason": "NeoBERT and ModernVBERT models are different requests, not one concrete change."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "MiniCPM3 architecture support and Qwen3-ASR support are unrelated."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "MiniCPM3 and the weight-sparse transformer request are different features."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:42503",
+            "accept": false,
+            "reason": "MiniCPM3 architecture and ModernVBERT models are not the same underlying change."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "MiniCPM3 architecture support and Voxtral TTS support are different model integrations."
+          },
+          {
+            "left": "issue:41211",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "DEIMv2 and Qwen3-TTS are unrelated model additions."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "ModernVBERT models and Qwen3-ASR support are distinct requests."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "ModernVBERT and Qwen3-TTS are different model families/capabilities."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:42954",
+            "accept": false,
+            "reason": "ModernVBERT models and XiaoMi MiMo are unrelated requests."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:46653",
+            "accept": false,
+            "reason": "A model request and a person-named issue are clearly not the same underlying change."
+          },
+          {
+            "left": "issue:42738",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "BERT-like models with RoPE and Voxtral TTS support are unrelated changes."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:42738",
+            "accept": false,
+            "reason": "Both mention BERT-like model families, but they request different model support and are not the same change."
+          },
+          {
+            "left": "issue:42738",
+            "right": "issue:42954",
+            "accept": false,
+            "reason": "BERT-like models with RoPE and XiaoMi MiMo are different requests."
+          },
+          {
+            "left": "issue:42738",
+            "right": "issue:46653",
+            "accept": false,
+            "reason": "BERT-like model support and a person-named issue are unrelated."
+          },
+          {
+            "left": "issue:42954",
+            "right": "issue:44205",
+            "accept": false,
+            "reason": "XiaoMi MiMo and SAM3-LiteText are distinct model requests."
+          },
+          {
+            "left": "issue:42954",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "MiMo and Qwen3-TTS are different model capabilities."
+          },
+          {
+            "left": "issue:42954",
+            "right": "issue:45177",
+            "accept": false,
+            "reason": "MiMo and DeepSeek-OCR2 are unrelated model requests."
+          },
+          {
+            "left": "issue:42954",
+            "right": "issue:46653",
+            "accept": false,
+            "reason": "MiMo and a person-named issue are not the same underlying change."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "Qwen3-TTS and Qwen3-ASR are related by family, but they are different task-specific support requests."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "Both are TTS-related, but they request support for different model families and code paths."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:45177",
+            "accept": false,
+            "reason": "Qwen3-TTS and DeepSeek-OCR2 are unrelated."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "Qwen3-TTS and the sparse transformer request are different changes."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "ASR support and LFM2.5 Audio support overlap only at a broad audio level, not the same concrete change."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45322",
+            "accept": false,
+            "reason": "Qwen3-ASR and EUPE are unrelated model requests."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "Qwen3-ASR support and the sparse transformer request are different code paths."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "NeoBERT and Voxtral TTS are unrelated model additions."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "Qwen3-ASR and Voxtral TTS are different support requests."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45177",
+            "accept": false,
+            "reason": "Qwen3-ASR and DeepSeek-OCR2 are unrelated changes."
+          },
+          {
+            "left": "issue:45322",
+            "right": "issue:46133",
+            "accept": false,
+            "reason": "EUPE and TIPSv2 are distinct vision-language model requests."
+          },
+          {
+            "left": "issue:45027",
+            "right": "issue:46653",
+            "accept": false,
+            "reason": "Voxtral TTS support and a person-named issue are not duplicates."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 7,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-38740-2",
+      "nodes": [
+        "issue:38740",
+        "issue:46443"
+      ],
+      "soft_pairs": [
+        "issue:38740|issue:46443"
+      ],
+      "prepared_review_unit_hash": "aec52a746e5d4fa28f519c8e1f88b70f6fd9640ff91534f0718941808ff016b9",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1506,
+        "estimated_input_tokens": 377,
+        "estimated_eval_tokens": 1010
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1506,
+        "estimated_input_tokens": 377,
+        "estimated_eval_tokens": 1010
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two issues are unrelated: one is a docs request about adding Pruna as an optimization framework, and the other is a Vision RoPE refactor. The soft link should be rejected.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No canonical issue across the pair; the items describe different topics and different kinds of work.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 46443 is the more substantive technical issue, but it is not a duplicate of the docs request.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:38740",
+            "right": "issue:46443",
+            "accept": false,
+            "reason": "Docs request vs. Vision RoPE refactor; no shared bug, code path, or concrete change target."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 8,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-41084-2",
+      "nodes": [
+        "issue:41084",
+        "issue:43037"
+      ],
+      "soft_pairs": [
+        "issue:41084|issue:43037"
+      ],
+      "prepared_review_unit_hash": "bc18442f93c8bfc41c9dc559d43026305e89c95088297e3ece60f009be17714e",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1480,
+        "estimated_input_tokens": 370,
+        "estimated_eval_tokens": 996
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1480,
+        "estimated_input_tokens": 370,
+        "estimated_eval_tokens": 996
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two issues are unrelated: one asks for Block Decoding support, the other asks for DeepSeek v3.2 support. There is no duplicate or same underlying bug signal here.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "If one issue must be kept as the most representative, issue 43037 is clearer and more specific; issue 41084 is more ambiguous and terse. That said, they are not duplicates.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41084",
+            "right": "issue:43037",
+            "accept": false,
+            "reason": "Different requests with no shared code path or explicit linkage: Block Decoding vs DeepSeek v3.2 support."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 9,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-41306-2",
+      "nodes": [
+        "issue:41306",
+        "issue:41826"
+      ],
+      "soft_pairs": [
+        "issue:41306|issue:41826"
+      ],
+      "prepared_review_unit_hash": "f61a0fedafd1e8cec20d956a8e98dbb4702e0dc60f041cdf3b0b2370be143f40",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1584,
+        "estimated_input_tokens": 396,
+        "estimated_eval_tokens": 1048
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1584,
+        "estimated_input_tokens": 396,
+        "estimated_eval_tokens": 1048
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues are about different features in different model components: one asks for a `from_pretrained()` flag for TiledFusedLogitsLoss/FLCE, the other for integrating TiledMLP to reduce memory footprint. They are not duplicates.",
+        "confidence": 0.96,
+        "canonical_issue_reason": "No canonical issue here: the requests target different code paths and different functionality.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "41306 is the more concrete and narrowly scoped request, with a specific API change (`from_pretrained()` flag) and named target component.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41306",
+            "right": "issue:41826",
+            "accept": false,
+            "reason": "Different underlying feature requests: FLCE/TiledFusedLogitsLoss configuration vs TiledMLP integration for memory reduction."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 10,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43598-2",
+      "nodes": [
+        "issue:43598",
+        "issue:43600"
+      ],
+      "soft_pairs": [
+        "issue:43598|issue:43600"
+      ],
+      "prepared_review_unit_hash": "ac421878a2842b710a6318cd2be8f63a079a48d5608e15e9b6477872c70694d2",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1584,
+        "estimated_input_tokens": 396,
+        "estimated_eval_tokens": 1048
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1584,
+        "estimated_input_tokens": 396,
+        "estimated_eval_tokens": 1048
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are related thematically around Trainer customization, but they describe different problems: one is about `remove_unused_column` behavior, the other about FLOPS estimation for custom trainers. I would not treat them as duplicates.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "Issue 43598 is the cleaner representative because it states a specific Trainer customization bug with a concrete knob (`remove_unused_column`) and a focused expected behavior.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43598 is the best representative issue for the cluster since it is more concrete and actionable, while 43600 is a separate FLOPS-accounting concern.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43598",
+            "right": "issue:43600",
+            "accept": false,
+            "reason": "They concern different code paths and outcomes: column-pruning customizability vs FLOPS estimation for custom trainers. Shared subject matter is not enough to call them the same underlying bug."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 11,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43684-2",
+      "nodes": [
+        "issue:43684",
+        "issue:44405"
+      ],
+      "soft_pairs": [
+        "issue:43684|issue:44405"
+      ],
+      "prepared_review_unit_hash": "639c400077ad65303e42332ab318898a9879007022440bde6dfc914de471a175",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1612,
+        "estimated_input_tokens": 403,
+        "estimated_eval_tokens": 1062
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1612,
+        "estimated_input_tokens": 403,
+        "estimated_eval_tokens": 1062
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are related only at a very high level (both concern adding AutoModel support for Qwen variants), but they target different model families and different auto classes, so they are not duplicates.",
+        "confidence": 0.95,
+        "canonical_issue_reason": "Issue #43684 is the broader, more foundational request: registering Qwen3-Omni with AutoModel and AutoModelForConditionalGeneration.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #43684 is the best representative for the cluster because it covers the general model-registration gap rather than a narrower task-specific head.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43684",
+            "right": "issue:44405",
+            "accept": false,
+            "reason": "Different underlying requests: Qwen3-Omni registration for AutoModel/AutoModelForConditionalGeneration vs. Qwen3.5 sequence-classification support for Qwen3_5Config. Same ecosystem, but not the same bug/change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 12,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43835-4",
+      "nodes": [
+        "issue:43835",
+        "issue:43845",
+        "issue:43846",
+        "issue:43859"
+      ],
+      "soft_pairs": [
+        "issue:43835|issue:43845",
+        "issue:43835|issue:43846",
+        "issue:43835|issue:43859",
+        "issue:43845|issue:43846",
+        "issue:43845|issue:43859",
+        "issue:43846|issue:43859"
+      ],
+      "prepared_review_unit_hash": "9d0608a492736c6549cdaf340bcd9ac9afadd50ba6996d5f741da0af4ed22bfb",
+      "component_budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4267,
+        "estimated_input_tokens": 1067,
+        "estimated_eval_tokens": 2390
+      },
+      "budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4267,
+        "estimated_input_tokens": 1067,
+        "estimated_eval_tokens": 2390
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster looks like a set of very sparse, low-signal issues: three near-identical \"huggingface\" reports and one unrelated \"Fett\" issue. There isn\u2019t enough concrete bug detail to safely merge any pair.",
+        "confidence": 0.44,
+        "canonical_issue_reason": "issue:43845 is the best canonical issue because it is one of the repeated \"huggingface\" reports, is the earliest among that repeated set, and has the most representative title for the cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43845 is the most suitable overall issue because it best matches the repeated title pattern and is the earliest of the near-duplicates; however, the cluster content is too thin to be strongly informative.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43835",
+            "right": "issue:43845",
+            "accept": false,
+            "reason": "Different titles and no substantive description or target overlap; soft similarity alone is not enough to treat these as the same issue."
+          },
+          {
+            "left": "issue:43835",
+            "right": "issue:43846",
+            "accept": false,
+            "reason": "Different titles and no concrete evidence of the same underlying bug or request."
+          },
+          {
+            "left": "issue:43835",
+            "right": "issue:43859",
+            "accept": false,
+            "reason": "The only evidence is vague similarity; there is no actionable detail tying this to issue:43835."
+          },
+          {
+            "left": "issue:43845",
+            "right": "issue:43846",
+            "accept": false,
+            "reason": "Both are titled \"huggingface\", but there is no description-level evidence that they report the same specific problem, so this is too weak to merge."
+          },
+          {
+            "left": "issue:43845",
+            "right": "issue:43859",
+            "accept": false,
+            "reason": "Same broad title, but no concrete bug/change details, targets, or references to confirm a true duplicate."
+          },
+          {
+            "left": "issue:43846",
+            "right": "issue:43859",
+            "accept": false,
+            "reason": "Same as above: matching generic titles are not enough without a shared underlying failure mode or change request."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 13,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-44371-6",
+      "nodes": [
+        "issue:44371",
+        "issue:44418",
+        "issue:44572",
+        "issue:44573",
+        "issue:44937",
+        "issue:47036"
+      ],
+      "soft_pairs": [
+        "issue:44371|issue:44572",
+        "issue:44371|issue:44573",
+        "issue:44371|issue:44418",
+        "issue:44371|issue:47036",
+        "issue:44371|issue:44937",
+        "issue:44418|issue:44572",
+        "issue:44418|issue:44573",
+        "issue:44572|issue:44573",
+        "issue:44572|issue:47036",
+        "issue:44572|issue:44937",
+        "issue:44573|issue:47036",
+        "issue:44573|issue:44937"
+      ],
+      "prepared_review_unit_hash": "b82a91cac411854796e8c6d6be4ae72158862a36bc2370506161377ff7d3cb35",
+      "component_budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 12,
+        "serialized_chars": 7352,
+        "estimated_input_tokens": 1838,
+        "estimated_eval_tokens": 3932
+      },
+      "budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 12,
+        "serialized_chars": 7352,
+        "estimated_input_tokens": 1838,
+        "estimated_eval_tokens": 3932
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "All items are low-signal closed issues with generic or spammy titles; there is no evidence they describe the same underlying bug or request, so none of the soft links should be merged.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No issue is a strong canonical representative. If forced to pick a single item as the least-spammy standalone record, issue 44418 is the only one with a somewhat meaningful title, but it still appears unrelated to the others.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44371",
+            "right": "issue:44572",
+            "accept": false,
+            "reason": "Both are just '<spam>' placeholders; no concrete shared bug or request is visible."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "Both are spam-placeholder issues with no description to indicate the same underlying problem."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44418",
+            "accept": false,
+            "reason": "A spam placeholder and a documentation-suggestion title are not the same issue."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:47036",
+            "accept": false,
+            "reason": "A spam placeholder and 'Community benchmark tip' do not appear to describe the same request."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "A spam placeholder and 'Check out \"Google LLC\"' are unrelated on face value."
+          },
+          {
+            "left": "issue:44418",
+            "right": "issue:44572",
+            "accept": false,
+            "reason": "Documentation enhancement request vs. spam placeholder; no shared concrete topic."
+          },
+          {
+            "left": "issue:44418",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "Documentation enhancement request vs. spam placeholder; not the same underlying issue."
+          },
+          {
+            "left": "issue:44572",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "Both are spam placeholders, but there is no evidence they refer to the same specific bug or change."
+          },
+          {
+            "left": "issue:44572",
+            "right": "issue:47036",
+            "accept": false,
+            "reason": "Spam placeholder and benchmark-tip title are unrelated."
+          },
+          {
+            "left": "issue:44572",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "Spam placeholder and the Google LLC title do not indicate the same issue."
+          },
+          {
+            "left": "issue:44573",
+            "right": "issue:47036",
+            "accept": false,
+            "reason": "Spam placeholder and benchmark-tip title are unrelated."
+          },
+          {
+            "left": "issue:44573",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "Spam placeholder and the Google LLC title do not describe the same underlying problem."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 14,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45145-2",
+      "nodes": [
+        "issue:45145",
+        "issue:45151"
+      ],
+      "soft_pairs": [
+        "issue:45145|issue:45151"
+      ],
+      "prepared_review_unit_hash": "b2c679f1be725fc7d7f9f86445864a482dc392972573339c94fc45f150089d1e",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1646,
+        "estimated_input_tokens": 412,
+        "estimated_eval_tokens": 1080
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1646,
+        "estimated_input_tokens": 412,
+        "estimated_eval_tokens": 1080
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are highly similar closed issues with the same title and close body lengths, so they likely duplicate the same energy-reduction topic. #45145 is the best canonical thread because it was created first and already has an inbound reference.",
+        "confidence": 0.89,
+        "canonical_issue_reason": "#45145 is the better canonical issue because it was created earlier and already has an inbound reference, making it the more established duplicate target.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#45145 is the strongest representative issue for the cluster for the same reasons: earliest creation and more thread signal than #45151.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45145",
+            "right": "issue:45151",
+            "accept": true,
+            "reason": "The issues share an identical title, similar body size, and both are closed, which is enough to treat them as highly similar issues that likely duplicate the same topic."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The canonical-choice reasoning is grounded, but the soft-edge summary is a bit too strong for the evidence in the packet. We only have identical titles, similar body lengths, and closed state\u2014not the bodies themselves\u2014so phrases like \"near-identical\" and \"same underlying request/report\" overstate certainty. A more conservative phrasing would be \"highly similar issues that likely duplicate the same topic.\""
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 15,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45295-2",
+      "nodes": [
+        "issue:45295",
+        "issue:45373"
+      ],
+      "soft_pairs": [],
+      "prepared_review_unit_hash": "f5fd95fada2f530ab0999e343c17c16266cde0c7182bb22509972ca44827cd49",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 0,
+        "serialized_chars": 1286,
+        "estimated_input_tokens": 322,
+        "estimated_eval_tokens": 900
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 0,
+        "serialized_chars": 1286,
+        "estimated_input_tokens": 322,
+        "estimated_eval_tokens": 900
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues describe the same missing Gemma 4 sequence-classification support; the later, more specific open issue is the better cluster representative.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "issue #45373 is the clearest canonical tracker: it is open, explicitly names the missing class `Gemma4ForSequenceClassification`, and states the module-level gap relative to Gemma 2/3.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue #45373 best captures the concrete request and likely remaining work, while #45295 is a shorter earlier duplicate with less detail and is already closed.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": []
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 16,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45762-2",
+      "nodes": [
+        "issue:45762",
+        "issue:45763"
+      ],
+      "soft_pairs": [
+        "issue:45762|issue:45763"
+      ],
+      "prepared_review_unit_hash": "71ed701a24b2bac69d0a77b0551dbda4b6d71ae6feab348a0ddeb16bd3949dfe",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1477,
+        "estimated_input_tokens": 370,
+        "estimated_eval_tokens": 996
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1477,
+        "estimated_input_tokens": 370,
+        "estimated_eval_tokens": 996
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two closed issues with only personal-name titles and no bodies, links, or shared technical context. They do not appear to describe the same bug or feature and should not be clustered as duplicates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45762",
+            "right": "issue:45763",
+            "accept": false,
+            "reason": "The titles are different person/date strings with no evidence of a shared underlying bug, request, or code path; soft text similarity alone is not enough."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 17,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45764-2",
+      "nodes": [
+        "issue:45764",
+        "issue:45765"
+      ],
+      "soft_pairs": [
+        "issue:45764|issue:45765"
+      ],
+      "prepared_review_unit_hash": "3d17492d02b82588cce7ee6879158c0ec217054e25e67701ba66ee8012b33c8f",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1481,
+        "estimated_input_tokens": 371,
+        "estimated_eval_tokens": 998
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1481,
+        "estimated_input_tokens": 371,
+        "estimated_eval_tokens": 998
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two closed issues are just personal-name/date titles with no body content or shared issue context. They do not appear to describe the same bug, task, or change, so they should not be clustered as duplicates.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45764",
+            "right": "issue:45765",
+            "accept": false,
+            "reason": "The titles refer to different people and different dates, with no body text or linked context indicating a shared underlying issue."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 18,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45810-2",
+      "nodes": [
+        "issue:45810",
+        "issue:46295"
+      ],
+      "soft_pairs": [
+        "issue:45810|issue:46295"
+      ],
+      "prepared_review_unit_hash": "8b7b358d68ce1f9b36b7b165b1021a07d7b62fe49d929a7c3d8e6c5aa142b3fb",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1542,
+        "estimated_input_tokens": 386,
+        "estimated_eval_tokens": 1028
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1542,
+        "estimated_input_tokens": 386,
+        "estimated_eval_tokens": 1028
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues are not duplicates: one requests a new Qwen3_5 token-classification/value-model class, while the other requests FP8 training support. They concern different features and different code paths.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45810",
+            "right": "issue:46295",
+            "accept": false,
+            "reason": "Different requests with no shared concrete bug or change target: model-class support for token classification vs FP8 training support."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 19,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45869-5",
+      "nodes": [
+        "issue:45869",
+        "issue:45870",
+        "issue:45871",
+        "issue:45872",
+        "issue:45873"
+      ],
+      "soft_pairs": [
+        "issue:45869|issue:45870",
+        "issue:45869|issue:45871",
+        "issue:45870|issue:45871",
+        "issue:45870|issue:45872",
+        "issue:45870|issue:45873",
+        "issue:45871|issue:45872",
+        "issue:45871|issue:45873",
+        "issue:45872|issue:45873"
+      ],
+      "prepared_review_unit_hash": "52ceabfe956e7b03b1d8275b0af81c307030b158ef749ef20c8785221d7200a6",
+      "component_budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 8,
+        "serialized_chars": 5724,
+        "estimated_input_tokens": 1431,
+        "estimated_eval_tokens": 3118
+      },
+      "budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 8,
+        "serialized_chars": 5724,
+        "estimated_input_tokens": 1431,
+        "estimated_eval_tokens": 3118
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The cluster splits into two exact duplicate pairs: 45870/45871 and 45872/45873. Issue 45869 is a separate, similar-looking dataset-promotion issue, but the packet does not show enough to treat it as a duplicate of the DataForge example requests.",
+        "confidence": 0.92,
+        "canonical_issue_reason": "issue:45870 is a clear representative of the exact duplicate pair with issue:45871.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45870 is a good representative for the 'Example: Financial time series prediction with DataForge' duplicate pair, but it should not be treated as the canonical representative of the entire cluster because 45872/45873 form a separate exact duplicate pair.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45869",
+            "right": "issue:45870",
+            "accept": false,
+            "reason": "Only moderate textual similarity and no shared targets/files; this looks like a different request, not the same underlying issue."
+          },
+          {
+            "left": "issue:45869",
+            "right": "issue:45871",
+            "accept": false,
+            "reason": "Only moderate textual similarity and no shared targets/files; this looks like a different request, not the same underlying issue."
+          },
+          {
+            "left": "issue:45870",
+            "right": "issue:45871",
+            "accept": true,
+            "reason": "Exact same title and matching body length; these are the same issue."
+          },
+          {
+            "left": "issue:45870",
+            "right": "issue:45872",
+            "accept": false,
+            "reason": "Both mention DataForge, but the requests differ: a plain example request versus an example plus scraping service."
+          },
+          {
+            "left": "issue:45870",
+            "right": "issue:45873",
+            "accept": false,
+            "reason": "Both mention DataForge, but the requests differ: a plain example request versus an example plus scraping service."
+          },
+          {
+            "left": "issue:45871",
+            "right": "issue:45872",
+            "accept": false,
+            "reason": "The titles indicate different requests, so they are not the same underlying issue."
+          },
+          {
+            "left": "issue:45871",
+            "right": "issue:45873",
+            "accept": false,
+            "reason": "The titles indicate different requests, so they are not the same underlying issue."
+          },
+          {
+            "left": "issue:45872",
+            "right": "issue:45873",
+            "accept": true,
+            "reason": "Exact same title and matching body length; these are the same issue."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The duplicate-pair summary is mostly grounded, but the canonical_issue_reason overstates what the packet supports: from titles/body lengths alone, 45870 and 45871 look like exact duplicates, but there is no evidence that 45870 is the \"cleaner, broader version\". I\u2019d also soften any claims of full unrelatedness for 45869; the packet only shows it is a different title with moderate similarity, not definitively unrelated."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 20,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46439-2",
+      "nodes": [
+        "issue:46439",
+        "issue:46441"
+      ],
+      "soft_pairs": [
+        "issue:46439|issue:46441"
+      ],
+      "prepared_review_unit_hash": "8d9640e74a1f73ebff9121180323ca2a4b7053840baa762c267229793e68a27d",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1641,
+        "estimated_input_tokens": 411,
+        "estimated_eval_tokens": 1078
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1641,
+        "estimated_input_tokens": 411,
+        "estimated_eval_tokens": 1078
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues are related static-cache/linear-attention crashes, but the packet does not prove they are the same underlying bug. Issue 46439 is the broader report; 46441 is a more specific Qwen3.5 symptom. The evidence is enough for clustering, not for a duplicate merge.",
+        "confidence": 0.68,
+        "canonical_issue_reason": "46439 is the better canonical issue because its title describes the broader static-cache/linear-attention failure mode, while 46441 is a narrower model-specific crash report.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "46439 is the best representative for the cluster because it is more general and likely to subsume related reports about static-cache and linear-attention breakage.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46439",
+            "right": "issue:46441",
+            "accept": false,
+            "reason": "Related, but the packet only confirms overlap on `cache_implementation=\"static\"` and `linear_attention`. There is not enough evidence here to conclude both reports describe the same root cause."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The pairing is plausible, but the analyst overstates the evidence. From the packet we only know both issues involve `cache_implementation=\"static\"` and `linear_attention`; we do not have body text, shared files, or explicit cross-links proving the same root cause. The summary and canonical reasons go beyond the packet by asserting a specific mechanism (`early_initialization` corrupting state / `KeyError` as a manifestation of the same bug). A more conservative read is that these are related static-cache/linear-attention crashes, but not yet confirmed to be the same underlying issue."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 21,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46551-2",
+      "nodes": [
+        "issue:46551",
+        "issue:46552"
+      ],
+      "soft_pairs": [
+        "issue:46551|issue:46552"
+      ],
+      "prepared_review_unit_hash": "25d55738087b1a5a8a313dfe9a1dedaa4e3be03a9318401517cfb47be18927d5",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1707,
+        "estimated_input_tokens": 427,
+        "estimated_eval_tokens": 1110
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1707,
+        "estimated_input_tokens": 427,
+        "estimated_eval_tokens": 1110
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues are near-perfect duplicates: identical title/body about Wav2Vec2CTCTokenizer v5 incorrectly treating the word delimiter as a special token. The open issue is the better canonical record.",
+        "confidence": 0.99,
+        "canonical_issue_reason": "Issue #46552 is open and has inbound references, so it is the better living tracker for the same bug reported in #46551.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #46552 is the best issue candidate because it is active/open and already linked from other discussions, making it the more useful canonical duplicate target.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46551",
+            "right": "issue:46552",
+            "accept": true,
+            "reason": "Same title and same described bug; this is a clear duplicate pair."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The conclusion is well grounded: both issues have the same title and matching body length, and the packet shows a perfect soft-similarity signal (jaccard 1.0). Choosing the open issue with inbound references as the canonical tracker is conservative and supported by the evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 22,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46657-2",
+      "nodes": [
+        "issue:46657",
+        "issue:46692"
+      ],
+      "soft_pairs": [
+        "issue:46657|issue:46692"
+      ],
+      "prepared_review_unit_hash": "2bed8ad3d8ab0024142723701f69b934fbde84e5cbbf1adf635966984d9845d1",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1569,
+        "estimated_input_tokens": 393,
+        "estimated_eval_tokens": 1042
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1569,
+        "estimated_input_tokens": 393,
+        "estimated_eval_tokens": 1042
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues describe different training failures: one is a tensor-size mismatch during object detection training, the other is an out-of-memory condition that appears after several epochs in RT-DETR v2. They do not look like the same underlying bug, so the soft link should be rejected.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No clear canonical issue for a duplicate cluster here; the reports point to different failure modes and should remain separate.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "If one had to be kept as the more concrete issue, #46657 is more specific and actionable, but it still should not be treated as a duplicate of #46692.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46657",
+            "right": "issue:46692",
+            "accept": false,
+            "reason": "Different symptom and likely different root cause: tensor shape mismatch versus progressive memory exhaustion/OOM. No evidence they are the same bug or a single fixable code-path issue."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 23,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46822-2",
+      "nodes": [
+        "issue:46822",
+        "issue:46823"
+      ],
+      "soft_pairs": [
+        "issue:46822|issue:46823"
+      ],
+      "prepared_review_unit_hash": "fd8a7c6da3d2e622f81fddf017b1f4c7e6cd08b2c73bc23142b92ca47d03084d",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1671,
+        "estimated_input_tokens": 418,
+        "estimated_eval_tokens": 1092
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1671,
+        "estimated_input_tokens": 418,
+        "estimated_eval_tokens": 1092
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are two different bug reports: one about ZeRO-3 zero.Init failing to partition a composite model submodule and causing OOM, the other about 4-bit loading with device_map='auto' and max_memory incorrectly offloading to CPU and erroring. They do not look like the same underlying defect.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "Issue 46822 is the more concrete representative bug report, with a specific ZeRO-3 zero.Init partitioning failure and a clear OOM symptom during multi-GPU load.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 46822 is the stronger standalone issue to keep as the representative item because it names the exact code path and failure mode more precisely than the 4-bit dispatch report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46822",
+            "right": "issue:46823",
+            "accept": false,
+            "reason": "Different loading paths and failure modes: ZeRO-3 zero.Init partitioning/OOM versus 4-bit device_map/max_memory CPU dispatch ValueError. They share only a broad theme of model loading bugs, not the same underlying issue."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 24,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46906-12",
+      "nodes": [
+        "issue:46906",
+        "issue:46940",
+        "issue:46948",
+        "issue:46969",
+        "issue:46985",
+        "issue:46998",
+        "issue:47037",
+        "issue:47049",
+        "issue:47053",
+        "issue:47091",
+        "issue:47102",
+        "issue:47120"
+      ],
+      "soft_pairs": [
+        "issue:46906|issue:46940",
+        "issue:46906|issue:46948",
+        "issue:46906|issue:46969",
+        "issue:46906|issue:46985",
+        "issue:46906|issue:47037",
+        "issue:46940|issue:46948",
+        "issue:46940|issue:46969",
+        "issue:46940|issue:46985",
+        "issue:46940|issue:47049",
+        "issue:46948|issue:46969",
+        "issue:46948|issue:46985",
+        "issue:46948|issue:46998",
+        "issue:46969|issue:46985",
+        "issue:46969|issue:47037",
+        "issue:46969|issue:47091",
+        "issue:46985|issue:47037",
+        "issue:46985|issue:47049",
+        "issue:46998|issue:47053",
+        "issue:46998|issue:47037",
+        "issue:46998|issue:47049",
+        "issue:46985|issue:46998",
+        "issue:46998|issue:47091",
+        "issue:47037|issue:47049",
+        "issue:47037|issue:47091",
+        "issue:47037|issue:47053",
+        "issue:47049|issue:47053",
+        "issue:47049|issue:47091",
+        "issue:47053|issue:47091",
+        "issue:47053|issue:47102",
+        "issue:47091|issue:47102",
+        "issue:46985|issue:47091",
+        "issue:47037|issue:47102",
+        "issue:46985|issue:47102",
+        "issue:47049|issue:47102",
+        "issue:47102|issue:47120",
+        "issue:47091|issue:47120",
+        "issue:47037|issue:47120",
+        "issue:47053|issue:47120",
+        "issue:46998|issue:47120"
+      ],
+      "prepared_review_unit_hash": "7f0332ccdde14bd65c9b8189d011407407a0a4e2df8678654c024f6de38d8979",
+      "component_budget": {
+        "node_count": 12,
+        "item_count": 12,
+        "soft_pair_count": 39,
+        "serialized_chars": 20777,
+        "estimated_input_tokens": 5195,
+        "estimated_eval_tokens": 10646
+      },
+      "budget": {
+        "node_count": 12,
+        "item_count": 12,
+        "soft_pair_count": 39,
+        "serialized_chars": 20777,
+        "estimated_input_tokens": 5195,
+        "estimated_eval_tokens": 10646
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster is a run of date-stamped daily integration-triage issues, not a single duplicate bug report. The latest open issue is the best representative for the current thread.",
+        "confidence": 0.96,
+        "canonical_issue_reason": "Issue 47120 is the latest open triage thread and likely the current active entry; the earlier issues are closed, date-stamped daily reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "47120 is the only open item and has the most recent inbound references, making it the best cluster representative.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46906",
+            "right": "issue:46940",
+            "accept": false,
+            "reason": "Same triage-series title, but different days; these are separate daily reports rather than one underlying bug."
+          },
+          {
+            "left": "issue:46906",
+            "right": "issue:46948",
+            "accept": false,
+            "reason": "Different daily triage entries with no concrete shared defect evidence."
+          },
+          {
+            "left": "issue:46906",
+            "right": "issue:46969",
+            "accept": false,
+            "reason": "Date-stamped triage reports for different days; not the same issue instance."
+          },
+          {
+            "left": "issue:46906",
+            "right": "issue:46985",
+            "accept": false,
+            "reason": "They look like successive daily triage threads, not duplicates of one bug."
+          },
+          {
+            "left": "issue:46906",
+            "right": "issue:47037",
+            "accept": false,
+            "reason": "Separate day-specific triage issues; same template/title only."
+          },
+          {
+            "left": "issue:46940",
+            "right": "issue:46948",
+            "accept": false,
+            "reason": "Both are daily triage logs, but on different dates and likely different failure snapshots."
+          },
+          {
+            "left": "issue:46940",
+            "right": "issue:46969",
+            "accept": false,
+            "reason": "No evidence they report the same underlying failure; only the same triage series."
+          },
+          {
+            "left": "issue:46940",
+            "right": "issue:46985",
+            "accept": false,
+            "reason": "Distinct date-stamped triage issues, not one bug report duplicated."
+          },
+          {
+            "left": "issue:46940",
+            "right": "issue:47049",
+            "accept": false,
+            "reason": "Different daily integration triage entries; insufficient evidence of a shared concrete defect."
+          },
+          {
+            "left": "issue:46948",
+            "right": "issue:46969",
+            "accept": false,
+            "reason": "Same broad topic, but these are separate daily triage issues from different days."
+          },
+          {
+            "left": "issue:46948",
+            "right": "issue:46985",
+            "accept": false,
+            "reason": "No sign of a single underlying bug; these are day-specific triage threads."
+          },
+          {
+            "left": "issue:46948",
+            "right": "issue:46998",
+            "accept": false,
+            "reason": "Different date markers indicate separate reports, not duplicates."
+          },
+          {
+            "left": "issue:46969",
+            "right": "issue:46985",
+            "accept": false,
+            "reason": "Sequential triage issues, but each appears to be a distinct daily instance."
+          },
+          {
+            "left": "issue:46969",
+            "right": "issue:47037",
+            "accept": false,
+            "reason": "Same series title only; different dates and likely different failing jobs."
+          },
+          {
+            "left": "issue:46969",
+            "right": "issue:47091",
+            "accept": false,
+            "reason": "No concrete same-bug evidence beyond the shared triage label."
+          },
+          {
+            "left": "issue:46985",
+            "right": "issue:46998",
+            "accept": false,
+            "reason": "Separate daily triage reports with no explicit common failure target."
+          },
+          {
+            "left": "issue:46985",
+            "right": "issue:47037",
+            "accept": false,
+            "reason": "Different day reports; too little evidence they describe the exact same defect."
+          },
+          {
+            "left": "issue:46985",
+            "right": "issue:47049",
+            "accept": false,
+            "reason": "Same triage series, but not the same underlying bug/change."
+          },
+          {
+            "left": "issue:46985",
+            "right": "issue:47091",
+            "accept": false,
+            "reason": "Daily triage entries should not be merged absent a shared concrete failure."
+          },
+          {
+            "left": "issue:46985",
+            "right": "issue:47102",
+            "accept": false,
+            "reason": "Distinct date-stamped integration triage issues, not duplicates."
+          },
+          {
+            "left": "issue:46998",
+            "right": "issue:47037",
+            "accept": false,
+            "reason": "These appear to be different daily snapshots of integration triage."
+          },
+          {
+            "left": "issue:46998",
+            "right": "issue:47049",
+            "accept": false,
+            "reason": "No evidence of the same code-path problem; only similar titles."
+          },
+          {
+            "left": "issue:46998",
+            "right": "issue:47053",
+            "accept": false,
+            "reason": "Separate daily triage issues; not enough to call them duplicates."
+          },
+          {
+            "left": "issue:46998",
+            "right": "issue:47091",
+            "accept": false,
+            "reason": "Same triage series, but different day and likely different failures."
+          },
+          {
+            "left": "issue:46998",
+            "right": "issue:47120",
+            "accept": false,
+            "reason": "Current open triage thread vs prior dated report; different issue instances."
+          },
+          {
+            "left": "issue:47037",
+            "right": "issue:47049",
+            "accept": false,
+            "reason": "Different date-stamped integration triage issues, not one bug report."
+          },
+          {
+            "left": "issue:47037",
+            "right": "issue:47053",
+            "accept": false,
+            "reason": "Same subsystem/series, but no concrete evidence of the same underlying failure."
+          },
+          {
+            "left": "issue:47037",
+            "right": "issue:47091",
+            "accept": false,
+            "reason": "These look like successive daily triage logs rather than duplicates."
+          },
+          {
+            "left": "issue:47037",
+            "right": "issue:47102",
+            "accept": false,
+            "reason": "Different dates and likely different integration failure snapshots."
+          },
+          {
+            "left": "issue:47037",
+            "right": "issue:47120",
+            "accept": false,
+            "reason": "Open current triage thread versus prior dated report; not the same artifact."
+          },
+          {
+            "left": "issue:47049",
+            "right": "issue:47053",
+            "accept": false,
+            "reason": "Adjacent daily triage issues, but still separate reports."
+          },
+          {
+            "left": "issue:47049",
+            "right": "issue:47091",
+            "accept": false,
+            "reason": "No shared concrete bug/change evidence; only the same triage label."
+          },
+          {
+            "left": "issue:47049",
+            "right": "issue:47102",
+            "accept": false,
+            "reason": "Different daily integration failure reports; should remain separate."
+          },
+          {
+            "left": "issue:47053",
+            "right": "issue:47091",
+            "accept": false,
+            "reason": "Date-stamped triage entries for different days, not duplicates."
+          },
+          {
+            "left": "issue:47053",
+            "right": "issue:47102",
+            "accept": false,
+            "reason": "Same broad theme, but no indication they describe the same failure."
+          },
+          {
+            "left": "issue:47053",
+            "right": "issue:47120",
+            "accept": false,
+            "reason": "Current day triage thread should not be merged with prior day reports."
+          },
+          {
+            "left": "issue:47091",
+            "right": "issue:47102",
+            "accept": false,
+            "reason": "Different daily triage issues; no concrete same-bug evidence."
+          },
+          {
+            "left": "issue:47091",
+            "right": "issue:47120",
+            "accept": false,
+            "reason": "Separate date-specific triage reports, not one underlying bug."
+          },
+          {
+            "left": "issue:47102",
+            "right": "issue:47120",
+            "accept": false,
+            "reason": "Prior-day triage report vs current open thread; distinct issue instances."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 25,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46983-2",
+      "nodes": [
+        "issue:46983",
+        "issue:46984"
+      ],
+      "soft_pairs": [
+        "issue:46983|issue:46984"
+      ],
+      "prepared_review_unit_hash": "3d68debd24140fd073e4f1666033ae42e79d46137045e8cd9a6bb0808361d886",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1490,
+        "estimated_input_tokens": 373,
+        "estimated_eval_tokens": 1002
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1490,
+        "estimated_input_tokens": 373,
+        "estimated_eval_tokens": 1002
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two issues are highly similar and likely track the same RWKV7 model support request, but the evidence is limited to identical titles and strong textual overlap. I would treat #46984 as the probable canonical issue because it is open and likely the current active discussion, while #46983 is a probable duplicate rather than a proven one.",
+        "confidence": 0.87,
+        "canonical_issue_reason": "Issue #46984 is the probable canonical issue: it is open, has the same title as #46983, and appears to be the currently active report for this request. The evidence supports treating #46983 as likely redundant, though not conclusively duplicate.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #46984 is the best issue to keep as the primary tracker because it is open and likely represents the latest active discussion for the same RWKV7 support request.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46983",
+            "right": "issue:46984",
+            "accept": true,
+            "reason": "Likely the same underlying request: identical titles and very high textual similarity strongly suggest duplication, though the evidence is not definitive beyond text overlap."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The conclusion is plausible, but a bit too definitive for the evidence provided. The packet only shows identical titles plus a high soft-similarity score; it does not prove the two issues are duplicates or the same underlying request. Please hedge the duplicate/canonical language (e.g., \"likely duplicate\" / \"probable canonical issue\") and avoid stating that redundancy is established as fact."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 26,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-47007-2",
+      "nodes": [
+        "issue:47007",
+        "issue:47015"
+      ],
+      "soft_pairs": [
+        "issue:47007|issue:47015"
+      ],
+      "prepared_review_unit_hash": "c47e6b966b001d9bc6ef256be84565e43f43bf60be8af5a6b2dfdfd89c3fa046",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1626,
+        "estimated_input_tokens": 407,
+        "estimated_eval_tokens": 1070
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1626,
+        "estimated_input_tokens": 407,
+        "estimated_eval_tokens": 1070
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two closed issues are thematically similar AI/cybersecurity writeups, but there\u2019s no evidence they describe the same concrete bug, feature request, or fixable change. They read more like separate article-style submissions than duplicates of one underlying issue.",
+        "confidence": 0.63,
+        "canonical_issue_reason": "Issue #47015 is the better representative because its title is broader and more succinct, making it the cleaner cluster anchor if these are grouped as related content.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #47015 is slightly more suitable as the cluster\u2019s best issue because it is the most general and least tied to a specific framing, but the evidence is weak for treating either item as a true duplicate.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:47007",
+            "right": "issue:47015",
+            "accept": false,
+            "reason": "The overlap is only topical: both mention AI-assisted vulnerability discovery and cybersecurity, but the titles suggest different writeups and there is no concrete shared bug, target, or fix path."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 27,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-47138-2",
+      "nodes": [
+        "issue:47138",
+        "issue:47139"
+      ],
+      "soft_pairs": [
+        "issue:47138|issue:47139"
+      ],
+      "prepared_review_unit_hash": "ac613b5fe6be9567e99d03792efcedee6a5af0d4f70ec05120efb06e6f5ac2f1",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1591,
+        "estimated_input_tokens": 398,
+        "estimated_eval_tokens": 1052
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1591,
+        "estimated_input_tokens": 398,
+        "estimated_eval_tokens": 1052
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues appear to be duplicates: they have identical titles and matching body lengths, and issue 47139 has inbound references, making it the better canonical thread.",
+        "confidence": 0.96,
+        "canonical_issue_reason": "issue:47139 is the better canonical issue because it has inbound references, which suggests it is the more established thread for this same report.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:47139 is the best representative issue since it matches issue 47138 exactly in title and has the stronger thread metadata.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:47138",
+            "right": "issue:47139",
+            "accept": true,
+            "reason": "Identical titles and matching report shape strongly indicate the same issue; the inbound references on 47139 make it the more canonical thread."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The match is strongly supported by the identical titles and the inbound-reference difference, but some wording goes beyond the packet. In particular, \"more conversation around it\" is not evidenced here, and \"same underlying bug with no distinguishing details\" is stronger than what we can verify from titles/body lengths alone. A more conservative explanation would note that the issues appear duplicate based on identical titles and similar metadata, with 47139 having inbound references that make it the better canonical thread."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 1,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 2,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-40546-44",
+      "nodes": [
+        "pull_request:43451",
+        "pull_request:44517",
+        "pull_request:46196",
+        "pull_request:46266",
+        "pull_request:46658",
+        "pull_request:46696",
+        "pull_request:46707",
+        "pull_request:46836",
+        "pull_request:46895",
+        "pull_request:46903",
+        "pull_request:46912",
+        "pull_request:46954",
+        "pull_request:46957",
+        "pull_request:46958",
+        "pull_request:46974",
+        "pull_request:46995",
+        "pull_request:47025",
+        "pull_request:47034",
+        "pull_request:47035",
+        "pull_request:47069",
+        "pull_request:47072",
+        "pull_request:47077",
+        "pull_request:47078",
+        "pull_request:47082",
+        "pull_request:47089",
+        "pull_request:47090",
+        "pull_request:47093",
+        "pull_request:47096",
+        "pull_request:47100",
+        "pull_request:47108",
+        "pull_request:47113",
+        "pull_request:47114",
+        "pull_request:47134",
+        "pull_request:47135",
+        "pull_request:47136",
+        "pull_request:47137",
+        "pull_request:47142",
+        "pull_request:47143",
+        "pull_request:47148",
+        "pull_request:47149"
+      ],
+      "soft_pairs": [
+        "pull_request:46696|pull_request:47134",
+        "pull_request:46903|pull_request:47090",
+        "pull_request:46903|pull_request:46958",
+        "pull_request:46954|pull_request:46958",
+        "pull_request:46958|pull_request:46974",
+        "pull_request:47072|pull_request:47089",
+        "pull_request:47077|pull_request:47078",
+        "pull_request:47072|pull_request:47148",
+        "pull_request:46266|pull_request:46957",
+        "pull_request:47114|pull_request:47148",
+        "pull_request:47072|pull_request:47134",
+        "pull_request:47114|pull_request:47135",
+        "pull_request:44517|pull_request:47082",
+        "pull_request:47034|pull_request:47035",
+        "pull_request:47069|pull_request:47089",
+        "pull_request:46196|pull_request:46995",
+        "pull_request:46266|pull_request:47148",
+        "pull_request:47093|pull_request:47114",
+        "pull_request:46707|pull_request:47108",
+        "pull_request:46707|pull_request:47114",
+        "pull_request:47077|pull_request:47149",
+        "pull_request:46707|pull_request:47148",
+        "pull_request:47089|pull_request:47113",
+        "pull_request:46266|pull_request:46707",
+        "pull_request:46995|pull_request:47143",
+        "pull_request:47082|pull_request:47108",
+        "pull_request:47114|pull_request:47149",
+        "pull_request:47108|pull_request:47114",
+        "pull_request:47077|pull_request:47114",
+        "pull_request:46912|pull_request:47025",
+        "pull_request:47093|pull_request:47143",
+        "pull_request:47134|pull_request:47135",
+        "pull_request:46658|pull_request:47114",
+        "pull_request:46196|pull_request:47148",
+        "pull_request:46707|pull_request:47082",
+        "pull_request:46266|pull_request:47114",
+        "pull_request:47069|pull_request:47136",
+        "pull_request:46912|pull_request:47082",
+        "pull_request:47082|pull_request:47148",
+        "pull_request:46912|pull_request:47078",
+        "pull_request:47114|pull_request:47136",
+        "pull_request:47035|pull_request:47089",
+        "pull_request:46895|pull_request:47134",
+        "pull_request:47100|pull_request:47134",
+        "pull_request:46196|pull_request:47025",
+        "pull_request:43451|pull_request:46995",
+        "pull_request:47100|pull_request:47142",
+        "pull_request:46895|pull_request:47137",
+        "pull_request:46836|pull_request:46995",
+        "pull_request:46707|pull_request:47025",
+        "pull_request:47108|pull_request:47148",
+        "pull_request:46266|pull_request:47034",
+        "pull_request:47034|pull_request:47108",
+        "pull_request:47035|pull_request:47069",
+        "pull_request:46836|pull_request:47135",
+        "pull_request:47035|pull_request:47113",
+        "pull_request:47100|pull_request:47113",
+        "pull_request:46995|pull_request:47096",
+        "pull_request:47142|pull_request:47149",
+        "pull_request:47069|pull_request:47113",
+        "pull_request:47025|pull_request:47114",
+        "pull_request:46912|pull_request:47136",
+        "pull_request:46912|pull_request:47077",
+        "pull_request:47069|pull_request:47082",
+        "pull_request:47089|pull_request:47134",
+        "pull_request:47135|pull_request:47137",
+        "pull_request:43451|pull_request:47078",
+        "pull_request:47078|pull_request:47096",
+        "pull_request:43451|pull_request:47136",
+        "pull_request:47096|pull_request:47136",
+        "pull_request:46707|pull_request:46957",
+        "pull_request:47100|pull_request:47149"
+      ],
+      "prepared_review_unit_hash": "447aa0f8b1af9136370bf1c82818957af07c341e6fae39b81e53a9ac9d616cc4",
+      "component_budget": {
+        "node_count": 44,
+        "item_count": 44,
+        "soft_pair_count": 130,
+        "serialized_chars": 154587,
+        "estimated_input_tokens": 38647,
+        "estimated_eval_tokens": 77550
+      },
+      "budget": {
+        "node_count": 40,
+        "item_count": 40,
+        "soft_pair_count": 72,
+        "serialized_chars": 114038,
+        "estimated_input_tokens": 28510,
+        "estimated_eval_tokens": 57276
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous: model additions, documentation-only updates, and a variety of unrelated bugfixes. None of the soft edges look like true duplicates or mergeable into one PR.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46696",
+            "right": "pull_request:47134",
+            "accept": false,
+            "reason": "Same file, but different bugs: 46696 fixes tracing/eager int handling for FA2 decode, while 47134 fixes a separate max_seqlen .item() performance regression."
+          },
+          {
+            "left": "pull_request:46903",
+            "right": "pull_request:47090",
+            "accept": false,
+            "reason": "Both touch CohereASR, but 46903 is a broader multi-model loss-shift change that also affects PPFormulaNet; 47090 is a CohereASR-only fix with different implementation details."
+          },
+          {
+            "left": "pull_request:46903",
+            "right": "pull_request:46958",
+            "accept": false,
+            "reason": "Same subsystem, but not the same standalone change: 46903 bundles CohereASR with PPFormulaNet, while 46958 is a separate CohereASR loss fix with a different implementation."
+          },
+          {
+            "left": "pull_request:46954",
+            "right": "pull_request:46958",
+            "accept": false,
+            "reason": "Unrelated fixes: Pytest-v9 strict FP32-module handling in GlmASR vs CohereASR training-loss double shift."
+          },
+          {
+            "left": "pull_request:46958",
+            "right": "pull_request:46974",
+            "accept": false,
+            "reason": "Different models and code paths: CohereASR loss computation vs MusicGen generate cache-length typo."
+          },
+          {
+            "left": "pull_request:47072",
+            "right": "pull_request:47089",
+            "accept": false,
+            "reason": "Different issues in different utilities: chat-schema enum parsing vs AqlmConfig type-error messages."
+          },
+          {
+            "left": "pull_request:47077",
+            "right": "pull_request:47078",
+            "accept": false,
+            "reason": "Both are Japanese docs edits, but they cover different documentation topics and wording changes, not the same change."
+          },
+          {
+            "left": "pull_request:47072",
+            "right": "pull_request:47148",
+            "accept": false,
+            "reason": "Different areas: schema enum extraction vs auto-mapping registration for string keys."
+          },
+          {
+            "left": "pull_request:46266",
+            "right": "pull_request:46957",
+            "accept": false,
+            "reason": "Different modernizations in unrelated code paths: DINOv2 conversion mapping vs generation decode switching for finegrained FP8."
+          },
+          {
+            "left": "pull_request:47114",
+            "right": "pull_request:47148",
+            "accept": false,
+            "reason": "Both relate to conversion/auto-mapping docs or mechanics, but one is documentation and the other is an auto-factory API fix; not the same bug/change."
+          },
+          {
+            "left": "pull_request:47072",
+            "right": "pull_request:47134",
+            "accept": false,
+            "reason": "Unrelated: JSON schema enum parsing vs flash-attention performance fix."
+          },
+          {
+            "left": "pull_request:47114",
+            "right": "pull_request:47135",
+            "accept": false,
+            "reason": "Different docs/utilities: custom-code conversion mapping docs vs quote rendering in the attention interface docs."
+          },
+          {
+            "left": "pull_request:44517",
+            "right": "pull_request:47082",
+            "accept": false,
+            "reason": "Different model additions in different domains: Qwen3 TTS vs RoMa keypoint matcher."
+          },
+          {
+            "left": "pull_request:47034",
+            "right": "pull_request:47035",
+            "accept": false,
+            "reason": "Related platform, but not the same change: a Neuron finetuning experiment vs torch.compile backend default selection."
+          },
+          {
+            "left": "pull_request:47069",
+            "right": "pull_request:47089",
+            "accept": false,
+            "reason": "Completely different fixes: num_labels returning None vs AqlmConfig error-message text."
+          },
+          {
+            "left": "pull_request:46196",
+            "right": "pull_request:46995",
+            "accept": false,
+            "reason": "Different attention-related capabilities: wav2vec2 SDPA dispatch vs MLU support for flash-linear-attention availability."
+          },
+          {
+            "left": "pull_request:46266",
+            "right": "pull_request:47148",
+            "accept": false,
+            "reason": "Different scope: DINO modernization/conversion work vs a LazyAutoMapping string-key registration fix."
+          },
+          {
+            "left": "pull_request:47093",
+            "right": "pull_request:47114",
+            "accept": false,
+            "reason": "Unrelated: OpenPangu-V2 model addition vs docs about custom-code conversion mapping."
+          },
+          {
+            "left": "pull_request:46707",
+            "right": "pull_request:47108",
+            "accept": false,
+            "reason": "Different classes of change: native FSDP2 support vs a documentation comment tweak in doc-builder workflow pins."
+          },
+          {
+            "left": "pull_request:46707",
+            "right": "pull_request:47114",
+            "accept": false,
+            "reason": "Unrelated: FSDP2 runtime migration vs conversion-mapping documentation."
+          },
+          {
+            "left": "pull_request:46707",
+            "right": "pull_request:47148",
+            "accept": false,
+            "reason": "Unrelated: native FSDP2 migration vs auto-factory string-key handling."
+          },
+          {
+            "left": "pull_request:47089",
+            "right": "pull_request:47113",
+            "accept": false,
+            "reason": "Different issues: AqlmConfig validation messages vs audio feature/processor backend requirements."
+          },
+          {
+            "left": "pull_request:46707",
+            "right": "pull_request:46957",
+            "accept": false,
+            "reason": "Different subsystems: distributed FSDP2 migration vs MoE decode-switching/FP8 generation changes."
+          },
+          {
+            "left": "pull_request:47096",
+            "right": "pull_request:47136",
+            "accept": false,
+            "reason": "Different changes: export support for Kimi vs a Quicktour inference wording clarification."
+          },
+          {
+            "left": "pull_request:46912",
+            "right": "pull_request:47025",
+            "accept": false,
+            "reason": "Different goals: broader RoPE testing vs Diffusion Gemma test adjustments."
+          },
+          {
+            "left": "pull_request:47093",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "OpenPangu-V2 model support vs NeuCodec model support; distinct model additions."
+          },
+          {
+            "left": "pull_request:47134",
+            "right": "pull_request:47135",
+            "accept": false,
+            "reason": "Different fixes in different docs/utilities: flash-attention performance regression vs quote rendering in docs."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47078",
+            "accept": false,
+            "reason": "Molmo2 model addition vs Japanese documentation wording cleanup; unrelated."
+          },
+          {
+            "left": "pull_request:47078",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Different content: Japanese docs wording vs exportability changes for Kimi."
+          },
+          {
+            "left": "pull_request:46707",
+            "right": "pull_request:47136",
+            "accept": false,
+            "reason": "Distributed/FSDP2 migration vs Quicktour docs wording; unrelated."
+          },
+          {
+            "left": "pull_request:47069",
+            "right": "pull_request:47113",
+            "accept": false,
+            "reason": "Num_labels API tweak vs audio backend requirements; unrelated."
+          },
+          {
+            "left": "pull_request:47025",
+            "right": "pull_request:47114",
+            "accept": false,
+            "reason": "Diffusion Gemma test cleanup vs conversion-mapping docs; unrelated."
+          },
+          {
+            "left": "pull_request:46912",
+            "right": "pull_request:47136",
+            "accept": false,
+            "reason": "RoPE testing changes vs Quicktour documentation wording; unrelated."
+          },
+          {
+            "left": "pull_request:46912",
+            "right": "pull_request:47077",
+            "accept": false,
+            "reason": "RoPE test coverage vs Japanese DistilBERT docs translation; unrelated."
+          },
+          {
+            "left": "pull_request:47069",
+            "right": "pull_request:47082",
+            "accept": false,
+            "reason": "Num_labels property fix vs RoMa model addition; unrelated."
+          },
+          {
+            "left": "pull_request:47089",
+            "right": "pull_request:47134",
+            "accept": false,
+            "reason": "AqlmConfig error-text cleanup vs flash-attention performance fix; unrelated."
+          },
+          {
+            "left": "pull_request:47135",
+            "right": "pull_request:47137",
+            "accept": false,
+            "reason": "Different fixes: documentation quote rendering vs tokenizer legacy special-token handling."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47078",
+            "accept": false,
+            "reason": "Molmo2 model addition vs Japanese documentation wording cleanup; unrelated."
+          },
+          {
+            "left": "pull_request:47078",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Different content: Japanese docs wording vs exportability changes for Kimi."
+          },
+          {
+            "left": "pull_request:46707",
+            "right": "pull_request:46957",
+            "accept": false,
+            "reason": "Distributed/FSDP2 migration vs generation expert-kernel switching; unrelated."
+          },
+          {
+            "left": "pull_request:47142",
+            "right": "pull_request:47149",
+            "accept": false,
+            "reason": "Pipelines generator handling cleanup vs MoE decode switching docs; unrelated."
+          },
+          {
+            "left": "pull_request:47069",
+            "right": "pull_request:47113",
+            "accept": false,
+            "reason": "Num_labels API tweak vs audio backend requirements; unrelated."
+          },
+          {
+            "left": "pull_request:47025",
+            "right": "pull_request:47114",
+            "accept": false,
+            "reason": "Diffusion Gemma test cleanup vs conversion-mapping docs; unrelated."
+          },
+          {
+            "left": "pull_request:46912",
+            "right": "pull_request:47136",
+            "accept": false,
+            "reason": "RoPE testing changes vs Quicktour documentation wording; unrelated."
+          },
+          {
+            "left": "pull_request:46912",
+            "right": "pull_request:47077",
+            "accept": false,
+            "reason": "RoPE test coverage vs Japanese DistilBERT docs translation; unrelated."
+          },
+          {
+            "left": "pull_request:47069",
+            "right": "pull_request:47082",
+            "accept": false,
+            "reason": "Num_labels property fix vs RoMa model addition; unrelated."
+          },
+          {
+            "left": "pull_request:47089",
+            "right": "pull_request:47134",
+            "accept": false,
+            "reason": "AqlmConfig error-text cleanup vs flash-attention performance fix; unrelated."
+          },
+          {
+            "left": "pull_request:47135",
+            "right": "pull_request:47137",
+            "accept": false,
+            "reason": "Different fixes: documentation quote rendering vs tokenizer legacy special-token handling."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47078",
+            "accept": false,
+            "reason": "Molmo2 model addition vs Japanese documentation wording cleanup; unrelated."
+          },
+          {
+            "left": "pull_request:47078",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Different content: Japanese docs wording vs exportability changes for Kimi."
+          },
+          {
+            "left": "pull_request:46707",
+            "right": "pull_request:46957",
+            "accept": false,
+            "reason": "Distributed/FSDP2 migration vs generation expert-kernel switching; unrelated."
+          },
+          {
+            "left": "pull_request:47142",
+            "right": "pull_request:47149",
+            "accept": false,
+            "reason": "Pipelines generator handling cleanup vs MoE decode switching docs; unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 1,
+      "component_count": 34,
+      "review_unit_index": 2,
+      "review_unit_count": 2,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-40546-44",
+      "nodes": [
+        "pull_request:40546",
+        "pull_request:43451",
+        "pull_request:44517",
+        "pull_request:46196",
+        "pull_request:46658",
+        "pull_request:46707",
+        "pull_request:46836",
+        "pull_request:46880",
+        "pull_request:46895",
+        "pull_request:46912",
+        "pull_request:46957",
+        "pull_request:46958",
+        "pull_request:47003",
+        "pull_request:47025",
+        "pull_request:47034",
+        "pull_request:47035",
+        "pull_request:47077",
+        "pull_request:47078",
+        "pull_request:47090",
+        "pull_request:47092",
+        "pull_request:47096",
+        "pull_request:47100",
+        "pull_request:47113",
+        "pull_request:47134",
+        "pull_request:47135",
+        "pull_request:47137",
+        "pull_request:47142",
+        "pull_request:47143",
+        "pull_request:47148",
+        "pull_request:47149"
+      ],
+      "soft_pairs": [
+        "pull_request:47003|pull_request:47078",
+        "pull_request:43451|pull_request:47135",
+        "pull_request:43451|pull_request:44517",
+        "pull_request:47100|pull_request:47143",
+        "pull_request:46912|pull_request:47113",
+        "pull_request:46912|pull_request:47142",
+        "pull_request:44517|pull_request:47143",
+        "pull_request:47113|pull_request:47137",
+        "pull_request:44517|pull_request:46836",
+        "pull_request:44517|pull_request:47096",
+        "pull_request:43451|pull_request:46957",
+        "pull_request:46957|pull_request:47096",
+        "pull_request:46957|pull_request:47003",
+        "pull_request:47134|pull_request:47137",
+        "pull_request:47148|pull_request:47149",
+        "pull_request:46658|pull_request:46895",
+        "pull_request:43451|pull_request:47142",
+        "pull_request:47096|pull_request:47142",
+        "pull_request:46880|pull_request:47092",
+        "pull_request:46895|pull_request:47090",
+        "pull_request:46958|pull_request:47090",
+        "pull_request:46895|pull_request:46958",
+        "pull_request:47025|pull_request:47092",
+        "pull_request:46707|pull_request:47034",
+        "pull_request:46658|pull_request:46836",
+        "pull_request:43451|pull_request:46836",
+        "pull_request:46658|pull_request:47143",
+        "pull_request:46836|pull_request:47143",
+        "pull_request:46196|pull_request:47143",
+        "pull_request:46836|pull_request:47092",
+        "pull_request:43451|pull_request:47077",
+        "pull_request:46836|pull_request:47096",
+        "pull_request:43451|pull_request:46196",
+        "pull_request:46196|pull_request:47096",
+        "pull_request:40546|pull_request:43451",
+        "pull_request:43451|pull_request:46658",
+        "pull_request:46658|pull_request:47096",
+        "pull_request:43451|pull_request:47143",
+        "pull_request:43451|pull_request:47092",
+        "pull_request:47092|pull_request:47143",
+        "pull_request:43451|pull_request:47096",
+        "pull_request:43451|pull_request:47003",
+        "pull_request:43451|pull_request:47137",
+        "pull_request:47003|pull_request:47096",
+        "pull_request:47096|pull_request:47137",
+        "pull_request:47096|pull_request:47143",
+        "pull_request:43451|pull_request:47034",
+        "pull_request:47034|pull_request:47096",
+        "pull_request:47092|pull_request:47096",
+        "pull_request:43451|pull_request:47035",
+        "pull_request:47035|pull_request:47096",
+        "pull_request:43451|pull_request:46912",
+        "pull_request:47003|pull_request:47137",
+        "pull_request:47003|pull_request:47034",
+        "pull_request:47003|pull_request:47143",
+        "pull_request:47137|pull_request:47143",
+        "pull_request:46836|pull_request:47003",
+        "pull_request:46836|pull_request:47137"
+      ],
+      "prepared_review_unit_hash": "cdc3b4ba341ef626a5fa4f4dfcca6bbf516c8660900b12bc9605af44b385fa74",
+      "component_budget": {
+        "node_count": 44,
+        "item_count": 44,
+        "soft_pair_count": 130,
+        "serialized_chars": 154587,
+        "estimated_input_tokens": 38647,
+        "estimated_eval_tokens": 77550
+      },
+      "budget": {
+        "node_count": 30,
+        "item_count": 30,
+        "soft_pair_count": 58,
+        "serialized_chars": 93214,
+        "estimated_input_tokens": 23304,
+        "estimated_eval_tokens": 46864
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The set is mostly a mix of unrelated model additions, docs changes, backend plumbing, and test tweaks. The only clear duplicate cluster is the CohereASR training-loss double-shift fixes (46895, 46958, 47090), all targeting the same issue and code path.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 47090 is the clearest and most complete fix for the CohereASR double-shift bug, with an explicit explanation and correct loss-call wiring; 46895 and 46958 are earlier variants of the same change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 47090 is the best representative because it directly fixes the underlying CohereASR training-loss bug without the earlier ambiguity, and it cleanly subsumes the previous variants.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:47003",
+            "right": "pull_request:47078",
+            "accept": false,
+            "reason": "OpenVINO exporter work and Japanese DistilBERT docs are unrelated changes."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47135",
+            "accept": false,
+            "reason": "Different domains: Molmo2 model support vs a docs quote-rendering fix."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:44517",
+            "accept": false,
+            "reason": "Separate model additions for different architectures; only broad auto-mapping/docs overlap."
+          },
+          {
+            "left": "pull_request:47100",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "Kernel API cleanup is unrelated to NeuCodec model support."
+          },
+          {
+            "left": "pull_request:46912",
+            "right": "pull_request:47113",
+            "accept": false,
+            "reason": "RoPE testing updates and audio requirements fixes are unrelated."
+          },
+          {
+            "left": "pull_request:46912",
+            "right": "pull_request:47142",
+            "accept": false,
+            "reason": "RoPE test coverage and pipeline generator cleanup are different code paths."
+          },
+          {
+            "left": "pull_request:44517",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "Qwen3-TTS and NeuCodec are distinct model integrations."
+          },
+          {
+            "left": "pull_request:47113",
+            "right": "pull_request:47137",
+            "accept": false,
+            "reason": "Audio requirements changes and tokenizer special-token fix are unrelated."
+          },
+          {
+            "left": "pull_request:44517",
+            "right": "pull_request:46836",
+            "accept": false,
+            "reason": "Different model additions with no same bug/change."
+          },
+          {
+            "left": "pull_request:44517",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Qwen3-TTS model support and Kimi exportability are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:46957",
+            "accept": false,
+            "reason": "Molmo2 support and FP8 decode optimization are unrelated."
+          },
+          {
+            "left": "pull_request:46957",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Different subsystems: FP8/MoE generation vs exportable Kimi."
+          },
+          {
+            "left": "pull_request:46957",
+            "right": "pull_request:47003",
+            "accept": false,
+            "reason": "FP8/MoE decode optimization is unrelated to OpenVINO exporter work."
+          },
+          {
+            "left": "pull_request:47134",
+            "right": "pull_request:47137",
+            "accept": false,
+            "reason": "Flash-attention performance fix and tokenizer method regression are unrelated."
+          },
+          {
+            "left": "pull_request:47148",
+            "right": "pull_request:47149",
+            "accept": false,
+            "reason": "Auto-mapping registration typing fix is unrelated to MoE decode switching docs."
+          },
+          {
+            "left": "pull_request:46658",
+            "right": "pull_request:46895",
+            "accept": false,
+            "reason": "Step-3.7-Flash model integration and CohereASR loss fix are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47142",
+            "accept": false,
+            "reason": "Model addition work and pipeline generator cleanup are unrelated."
+          },
+          {
+            "left": "pull_request:47096",
+            "right": "pull_request:47142",
+            "accept": false,
+            "reason": "Exportability changes for Kimi and pipeline generator cleanup are unrelated."
+          },
+          {
+            "left": "pull_request:46880",
+            "right": "pull_request:47092",
+            "accept": false,
+            "reason": "Both add XPU expectations, but for different models and test cases."
+          },
+          {
+            "left": "pull_request:46895",
+            "right": "pull_request:47090",
+            "accept": true,
+            "reason": "Both address the same CohereASR training-loss double-shift bug in the same files and target the same issue."
+          },
+          {
+            "left": "pull_request:46958",
+            "right": "pull_request:47090",
+            "accept": true,
+            "reason": "Same CohereASR double-shift fix, just a later refinement of the same code path."
+          },
+          {
+            "left": "pull_request:46895",
+            "right": "pull_request:46958",
+            "accept": true,
+            "reason": "Both are variants of the same CohereASR loss-shift correction."
+          },
+          {
+            "left": "pull_request:47025",
+            "right": "pull_request:47092",
+            "accept": false,
+            "reason": "Diffusion Gemma test cleanup and MiniCPM3 XPU expectations are unrelated."
+          },
+          {
+            "left": "pull_request:46707",
+            "right": "pull_request:47034",
+            "accept": false,
+            "reason": "Native FSDP2 migration and Neuron SFT experiment are different backend initiatives."
+          },
+          {
+            "left": "pull_request:46658",
+            "right": "pull_request:46836",
+            "accept": false,
+            "reason": "Step-3.7-Flash and Unlimited OCR are different model additions."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:46836",
+            "accept": false,
+            "reason": "Molmo2 and Unlimited OCR are separate model integrations."
+          },
+          {
+            "left": "pull_request:46658",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "Step-3.7-Flash and NeuCodec are unrelated model support PRs."
+          },
+          {
+            "left": "pull_request:46836",
+            "right": "pull_request:47092",
+            "accept": false,
+            "reason": "Unlimited OCR model support is unrelated to MiniCPM3 test expectations."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47077",
+            "accept": false,
+            "reason": "Molmo2 model support is unrelated to Japanese DistilBERT docs."
+          },
+          {
+            "left": "pull_request:46836",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Unlimited OCR support and Kimi exportability are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:46196",
+            "accept": false,
+            "reason": "Molmo2 support and wav2vec2 SDPA dispatch are unrelated."
+          },
+          {
+            "left": "pull_request:46196",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Attention SDPA dispatch and exportable Kimi are different code paths."
+          },
+          {
+            "left": "pull_request:40546",
+            "right": "pull_request:43451",
+            "accept": false,
+            "reason": "VibeVoice and Molmo2 are separate model support additions."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:46658",
+            "accept": false,
+            "reason": "Different new-model integration work, not the same change."
+          },
+          {
+            "left": "pull_request:46658",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Step-3.7-Flash and Kimi exportability are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "Molmo2 and NeuCodec are different model additions."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47092",
+            "accept": false,
+            "reason": "Molmo2 model support and MiniCPM3 XPU expectations are unrelated."
+          },
+          {
+            "left": "pull_request:47092",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "MiniCPM3 test expectations and NeuCodec model support are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Molmo2 support and Kimi exportability are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47003",
+            "accept": false,
+            "reason": "Molmo2 model support and OpenVINO exporter are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47137",
+            "accept": false,
+            "reason": "Model addition work and tokenizer property/test fix are unrelated."
+          },
+          {
+            "left": "pull_request:47003",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "OpenVINO exporter and Kimi exportability touch different export stacks."
+          },
+          {
+            "left": "pull_request:47096",
+            "right": "pull_request:47137",
+            "accept": false,
+            "reason": "Kimi exportability and tokenizer special-token fix are unrelated."
+          },
+          {
+            "left": "pull_request:47096",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "Kimi exportability and NeuCodec support are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47034",
+            "accept": false,
+            "reason": "Molmo2 support and Neuron SFT experiment are unrelated."
+          },
+          {
+            "left": "pull_request:47034",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Neuron training experiment and Kimi exportability are unrelated."
+          },
+          {
+            "left": "pull_request:47092",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Different models and test/export concerns; not the same underlying change."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:47035",
+            "accept": false,
+            "reason": "Molmo2 support and Neuron torch-compile default are unrelated."
+          },
+          {
+            "left": "pull_request:47035",
+            "right": "pull_request:47096",
+            "accept": false,
+            "reason": "Neuron compile default and Kimi exportability are unrelated."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:46912",
+            "accept": false,
+            "reason": "Molmo2 model support and RoPE test work are unrelated."
+          },
+          {
+            "left": "pull_request:47003",
+            "right": "pull_request:47137",
+            "accept": false,
+            "reason": "OpenVINO exporter and tokenizer property fix are unrelated."
+          },
+          {
+            "left": "pull_request:47003",
+            "right": "pull_request:47034",
+            "accept": false,
+            "reason": "OpenVINO exporter and Neuron SFT experiment are unrelated."
+          },
+          {
+            "left": "pull_request:47003",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "OpenVINO exporter and NeuCodec model support are unrelated."
+          },
+          {
+            "left": "pull_request:47137",
+            "right": "pull_request:47143",
+            "accept": false,
+            "reason": "Tokenizer special-token fix and NeuCodec support are unrelated."
+          },
+          {
+            "left": "pull_request:46836",
+            "right": "pull_request:47003",
+            "accept": false,
+            "reason": "Unlimited OCR model support and OpenVINO exporter are unrelated."
+          },
+          {
+            "left": "pull_request:46836",
+            "right": "pull_request:47137",
+            "accept": false,
+            "reason": "Unlimited OCR support and tokenizer special-token fix are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 2,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-40755-2",
+      "nodes": [
+        "pull_request:40755",
+        "pull_request:41224"
+      ],
+      "soft_pairs": [
+        "pull_request:40755|pull_request:41224"
+      ],
+      "prepared_review_unit_hash": "dcdf09dbc2c1b83666e5cfd824d732813dd0fa1245667936e4b27caa76c90040",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4675,
+        "estimated_input_tokens": 1169,
+        "estimated_eval_tokens": 2594
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4675,
+        "estimated_input_tokens": 1169,
+        "estimated_eval_tokens": 2594
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two PRs are not duplicates: one adds TimesFM covariate forecasting support, while the other adds DINOv3 image classification support. They touch different models, code paths, and use cases.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 41224 is the cleaner representative of this cluster if one must be chosen: it is a self-contained feature addition for DINOv3 image classification with matching code, docs, and tests.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 41224 is the stronger standalone PR to surface because its change is narrowly scoped and clearly validated by the affected model and tests, whereas 40755 is a separate TimesFM feature request.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:40755",
+            "right": "pull_request:41224",
+            "accept": false,
+            "reason": "Different subsystems and different concrete changes: TimesFM covariate forecasting vs DINOv3 image classification. They do not fix the same bug or implement the same feature."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 3,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-41356-2",
+      "nodes": [
+        "pull_request:41356",
+        "pull_request:44339"
+      ],
+      "soft_pairs": [
+        "pull_request:41356|pull_request:44339"
+      ],
+      "prepared_review_unit_hash": "aead1308c167b0bae6ff11bf9358340baa5e76f816618383418b505a10a9cd01",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6050,
+        "estimated_input_tokens": 1513,
+        "estimated_eval_tokens": 3282
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6050,
+        "estimated_input_tokens": 1513,
+        "estimated_eval_tokens": 3282
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two DEIMv2 pull requests target the same model family, but they are not the same change: PR 41356 is an early/basic model + image processor add, while PR 44339 is a much broader follow-up that adds loss code, auto-mappings, conversion tooling, modular files, and more integration work.",
+        "confidence": 0.91,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 44339 is the more complete and representative DEIMv2 implementation; PR 41356 looks like an earlier partial version of the same model bring-up.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 44339 is the strongest single artifact for the cluster because it subsumes the basic model work and includes the broader integration needed for a full Transformers addition.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:41356",
+            "right": "pull_request:44339",
+            "accept": false,
+            "reason": "Same model area, but not the same concrete change: 44339 expands the implementation substantially and adds new code paths beyond the basic PR 41356. They look like related iterations, not mergeable duplicates."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 4,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-41855-12",
+      "nodes": [
+        "pull_request:41855",
+        "pull_request:42668",
+        "pull_request:43251",
+        "pull_request:43254",
+        "pull_request:43424",
+        "pull_request:43823",
+        "pull_request:43995",
+        "pull_request:44827",
+        "pull_request:45170",
+        "pull_request:45213",
+        "pull_request:45655",
+        "pull_request:46048"
+      ],
+      "soft_pairs": [
+        "pull_request:41855|pull_request:43424",
+        "pull_request:41855|pull_request:43823",
+        "pull_request:42668|pull_request:43823",
+        "pull_request:43251|pull_request:43823",
+        "pull_request:43251|pull_request:44827",
+        "pull_request:43424|pull_request:44827",
+        "pull_request:43424|pull_request:43823",
+        "pull_request:43823|pull_request:45655",
+        "pull_request:43823|pull_request:43995",
+        "pull_request:43995|pull_request:45213",
+        "pull_request:43995|pull_request:44827",
+        "pull_request:44827|pull_request:45170",
+        "pull_request:44827|pull_request:46048",
+        "pull_request:44827|pull_request:45655",
+        "pull_request:44827|pull_request:45213",
+        "pull_request:43823|pull_request:45213",
+        "pull_request:43251|pull_request:43254"
+      ],
+      "prepared_review_unit_hash": "a08b617106a1567988a5969491313c2b251a6cb3143fc75210273334fe3e1681",
+      "component_budget": {
+        "node_count": 12,
+        "item_count": 12,
+        "soft_pair_count": 17,
+        "serialized_chars": 31596,
+        "estimated_input_tokens": 7899,
+        "estimated_eval_tokens": 16054
+      },
+      "budget": {
+        "node_count": 12,
+        "item_count": 12,
+        "soft_pair_count": 17,
+        "serialized_chars": 31596,
+        "estimated_input_tokens": 7899,
+        "estimated_eval_tokens": 16054
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous PRs with superficial similarity; the only clear duplicate pair is the two `fixed_cross_entropy` changes for issue 43240.",
+        "confidence": 0.89,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 43251 is the clearest canonical representative of the only true duplicate pair: it targets the same issue (43240) and updates the same loss helper with the broader kwargs forwarding change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 43251 is the best overall PR candidate because it is the most explicit, issue-linked version of the one genuine duplicate change in this set; the rest are unrelated feature/fix PRs.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:41855",
+            "right": "pull_request:43424",
+            "accept": false,
+            "reason": "Different features: Mistral tokenizer methods vs executorch dynamic-shape export tests."
+          },
+          {
+            "left": "pull_request:41855",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Unrelated models and code paths; tokenizer utilities vs adding MobileLLM."
+          },
+          {
+            "left": "pull_request:42668",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Both are processor/model-loading related only at a very broad level, but they fix different model families and issues."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Loss helper kwargs forwarding is unrelated to adding MobileLLM."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "Different code paths: cross-entropy helper vs Mistral4 test/model fixes."
+          },
+          {
+            "left": "pull_request:43424",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "Executorch export tests and Mistral4 fixes are not the same underlying change."
+          },
+          {
+            "left": "pull_request:43424",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Exportability tests vs new model implementation are unrelated."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:45655",
+            "accept": false,
+            "reason": "MobileLLM addition and config-resolution fix solve different problems."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:43995",
+            "accept": false,
+            "reason": "Different models and objectives; MobileLLM creation vs Falcon output-collection refactor."
+          },
+          {
+            "left": "pull_request:43995",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "Falcon refactor and model-creation skill are unrelated."
+          },
+          {
+            "left": "pull_request:43995",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "Falcon output interface refactor is not the same as Mistral4 test fixes."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45170",
+            "accept": false,
+            "reason": "Mistral4 test fixes and `layrnorm` renaming are different bug fixes."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:46048",
+            "accept": false,
+            "reason": "Mistral4 test fixes and TDT loss kernel addition are unrelated."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45655",
+            "accept": false,
+            "reason": "Mistral4 test fixes and config-resolution order fix address different code paths."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "A model test/fix PR is not the same as a model-creation skill PR."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "Adding MobileLLM is unrelated to the model-creation skill draft."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:43254",
+            "accept": true,
+            "reason": "Same helper function, same issue target 43240, and both add support for passing cross-entropy kwargs; they are competing implementations of the same fix."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The only accepted pair, 43251/43254, is clearly mergeable as the same `fixed_cross_entropy` issue on the same file and issue target 43240. The rest of the false verdicts are appropriately cautious and do not overclaim duplicate status."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 5,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-42975-2",
+      "nodes": [
+        "pull_request:42975",
+        "pull_request:42976"
+      ],
+      "soft_pairs": [
+        "pull_request:42975|pull_request:42976"
+      ],
+      "prepared_review_unit_hash": "76c059122abc6cf823918f4037d7b61addcb7135a01d8fb7455f24526fb2ce65",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5934,
+        "estimated_input_tokens": 1484,
+        "estimated_eval_tokens": 3224
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5934,
+        "estimated_input_tokens": 1484,
+        "estimated_eval_tokens": 3224
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open PRs both upgrade GitHub Actions in workflow files, but they target different action families and different concrete maintenance goals: Node 24 compatibility versus docker/build-push-action bumps. They overlap at a high level, not as the same underlying change.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 42975 is the better representative PR: it has the broader, more explicit goal (Node 24 compatibility) and updates several core workflow actions across many files.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 42975 is the best single PR for the cluster because it spans more of the workflow-action upgrade work and most clearly captures the main maintenance theme.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:42975",
+            "right": "pull_request:42976",
+            "accept": false,
+            "reason": "Both are GitHub Actions upgrade PRs, but they fix different concrete changes: 42975 updates actions/checkout, actions/cache, actions/upload-artifact for Node 24 compatibility, while 42976 bumps docker/build-push-action to v6 in docker build workflows. They are related maintenance, not the same underlying change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 6,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-42978-4",
+      "nodes": [
+        "pull_request:42978",
+        "pull_request:43448",
+        "pull_request:45490",
+        "pull_request:46562"
+      ],
+      "soft_pairs": [
+        "pull_request:42978|pull_request:43448",
+        "pull_request:42978|pull_request:45490",
+        "pull_request:43448|pull_request:45490",
+        "pull_request:45490|pull_request:46562"
+      ],
+      "prepared_review_unit_hash": "a230a01ccf2e14f3a5d23cdad9c42f9238c4cb04e32d4f02656bbc597714ce3d",
+      "component_budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 4,
+        "serialized_chars": 13047,
+        "estimated_input_tokens": 3262,
+        "estimated_eval_tokens": 6780
+      },
+      "budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 4,
+        "serialized_chars": 13047,
+        "estimated_input_tokens": 3262,
+        "estimated_eval_tokens": 6780
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "All four PRs are separate model-addition changes, not duplicates. They share the usual Hugging Face model-registration/docs/test pattern, but each targets a different architecture and code path.",
+        "confidence": 0.99,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No true canonical duplicate exists here: ViT NEPA, Molmo, CTSM, and Step3p7 are distinct model integrations rather than the same fix/change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "45490 is the strongest standalone representative because it has the highest review activity and broad model-registration changes, but it is still not a duplicate of the others.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:43448",
+            "accept": false,
+            "reason": "Different model integrations (ViT NEPA vs Molmo); shared docs/auto-mapping edits are generic add-model boilerplate, not the same bug or change."
+          },
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "ViT NEPA and CTSM are unrelated architectures with different model code and conversion paths; overlap is only in registration/docs infrastructure."
+          },
+          {
+            "left": "pull_request:43448",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "Molmo and CTSM are distinct model additions touching different implementations; same subsystem is not enough to merge as duplicates."
+          },
+          {
+            "left": "pull_request:45490",
+            "right": "pull_request:46562",
+            "accept": false,
+            "reason": "CTSM vs Step3p7 are different time-series/vision-model additions with separate concrete code paths, so they are not the same PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 7,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43291-7",
+      "nodes": [
+        "pull_request:43291",
+        "pull_request:43665",
+        "pull_request:44815",
+        "pull_request:46380",
+        "pull_request:46539",
+        "pull_request:46886",
+        "pull_request:46922"
+      ],
+      "soft_pairs": [
+        "pull_request:43291|pull_request:43665",
+        "pull_request:43291|pull_request:44815",
+        "pull_request:43291|pull_request:46380",
+        "pull_request:43291|pull_request:46922",
+        "pull_request:43665|pull_request:46539",
+        "pull_request:43665|pull_request:44815",
+        "pull_request:43665|pull_request:46380",
+        "pull_request:43665|pull_request:46922",
+        "pull_request:44815|pull_request:46539",
+        "pull_request:44815|pull_request:46380",
+        "pull_request:44815|pull_request:46922",
+        "pull_request:46380|pull_request:46539",
+        "pull_request:46380|pull_request:46922",
+        "pull_request:46539|pull_request:46922",
+        "pull_request:46886|pull_request:46922"
+      ],
+      "prepared_review_unit_hash": "b6f88835bf3bac2cd8990f7fa8c687125067563baffaa6ff4098db3559087e09",
+      "component_budget": {
+        "node_count": 7,
+        "item_count": 7,
+        "soft_pair_count": 15,
+        "serialized_chars": 19599,
+        "estimated_input_tokens": 4900,
+        "estimated_eval_tokens": 10056
+      },
+      "budget": {
+        "node_count": 7,
+        "item_count": 7,
+        "soft_pair_count": 15,
+        "serialized_chars": 19599,
+        "estimated_input_tokens": 4900,
+        "estimated_eval_tokens": 10056
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The cluster is mostly heterogeneous: several PRs are unrelated test, CI, generation, FP8, or model-specific changes. The only likely duplicate pair is the YOLOS device-mismatch fixes (46886 and 46922), which target the same issue and file; 46922 is the more complete version.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "46922 is the best canonical PR: it targets the concrete YOLOS device-mismatch bug, shares the explicit issue target 46869, and contains the fix plus the supporting split-module change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "46922 is the strongest standalone PR because it directly addresses the same YOLOS runtime bug and is more complete than 46886.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:43665",
+            "accept": false,
+            "reason": "Different models and different bugs; both are vague/test-focused PRs, but they do not fix the same code-path problem."
+          },
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:44815",
+            "accept": false,
+            "reason": "Whisper tokenization/tests versus FP8 dequant/loading logic are unrelated changes."
+          },
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:46380",
+            "accept": false,
+            "reason": "Whisper decode/test changes are unrelated to generation-config watermark validation."
+          },
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:46922",
+            "accept": false,
+            "reason": "Whisper tokenization/test work and YOLOS device-mismatch work are different bugs in different subsystems."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:46539",
+            "accept": false,
+            "reason": "Vision model test edits versus Phimoe CI/memory handling are unrelated."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:44815",
+            "accept": false,
+            "reason": "CLIP/ViT test changes do not match the FP8 dequant/loading fix in core model loading."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:46380",
+            "accept": false,
+            "reason": "These touch unrelated areas: vision model tests versus generation/watermark validation and Whisper code."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:46922",
+            "accept": false,
+            "reason": "Generic vision test changes are not the same underlying bug as the YOLOS device-mismatch fix."
+          },
+          {
+            "left": "pull_request:44815",
+            "right": "pull_request:46539",
+            "accept": false,
+            "reason": "FP8 dequant/loading changes and Phimoe CI memory tweaks are different issues and code paths."
+          },
+          {
+            "left": "pull_request:44815",
+            "right": "pull_request:46380",
+            "accept": false,
+            "reason": "No overlap in bug or change: FP8 quantization versus generation config validation."
+          },
+          {
+            "left": "pull_request:44815",
+            "right": "pull_request:46922",
+            "accept": false,
+            "reason": "Different subsystems and different runtime failures; not the same concrete bug."
+          },
+          {
+            "left": "pull_request:46380",
+            "right": "pull_request:46539",
+            "accept": false,
+            "reason": "Generation-config watermark validation and Phimoe CI behavior are unrelated."
+          },
+          {
+            "left": "pull_request:46380",
+            "right": "pull_request:46922",
+            "accept": false,
+            "reason": "Generation/watermark validation is unrelated to YOLOS device placement."
+          },
+          {
+            "left": "pull_request:46539",
+            "right": "pull_request:46922",
+            "accept": false,
+            "reason": "CI memory-limit patches for Phimoe are unrelated to the YOLOS runtime device-mismatch fix."
+          },
+          {
+            "left": "pull_request:46886",
+            "right": "pull_request:46922",
+            "accept": true,
+            "reason": "Both PRs target the same YOLOS device_map='auto' mismatch in the same file and same issue target; 46922 subsumes 46886 with the additional hidden-state device placement fix."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative about the soft edges. The only accepted pair, 46886 vs 46922, is well supported by the shared YOLOS file and explicit issue target 46869, and the two diffs are mergeable as one bugfix. The other pairwise rejections are appropriately cautious given the distinct subsystems and code paths."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 8,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43488-3",
+      "nodes": [
+        "pull_request:43488",
+        "pull_request:43532",
+        "pull_request:46747"
+      ],
+      "soft_pairs": [
+        "pull_request:43488|pull_request:46747",
+        "pull_request:43488|pull_request:43532",
+        "pull_request:43532|pull_request:46747"
+      ],
+      "prepared_review_unit_hash": "4e70d9ceaae39774dc6bed193b1d3cccfb163179521399e4b5e7b5eb4bc316be",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 6349,
+        "estimated_input_tokens": 1588,
+        "estimated_eval_tokens": 3432
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 6349,
+        "estimated_input_tokens": 1588,
+        "estimated_eval_tokens": 3432
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are three unrelated throwaway PRs: one tweaks packaging/versioning and a vit test, one changes several workflow/testing files to show diffs, and one alters a security-gate workflow trigger. They do not look like the same underlying bug or change.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No canonical PR: the three PRs target different code paths and workflows, so there is no true duplicate representative.",
+        "best_issue_reason": null,
+        "best_pr_reason": "43488 is the narrowest self-test/packaging PR, but it does not subsume the other two unrelated workflow changes; there is no strong cluster-level best PR.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43488",
+            "right": "pull_request:46747",
+            "accept": false,
+            "reason": "Different files and intent: setup/version + vit test vs workflow trigger change in pr-ci-security-gate.yml."
+          },
+          {
+            "left": "pull_request:43488",
+            "right": "pull_request:43532",
+            "accept": false,
+            "reason": "43488 is packaging/versioning plus a test tweak; 43532 rewrites several CI/notification workflows and testing utilities. Not the same change."
+          },
+          {
+            "left": "pull_request:43532",
+            "right": "pull_request:46747",
+            "accept": false,
+            "reason": "Both touch workflows, but 43532 is about comment/diff reporting plumbing while 46747 changes the security gate trigger. Different underlying code-paths."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 9,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43612-2",
+      "nodes": [
+        "pull_request:43612",
+        "pull_request:43613"
+      ],
+      "soft_pairs": [
+        "pull_request:43612|pull_request:43613"
+      ],
+      "prepared_review_unit_hash": "46f90e94225e7dd9e2983a90ca151b4d055d69fcab30b792de2503f4d54f5c79",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6082,
+        "estimated_input_tokens": 1521,
+        "estimated_eval_tokens": 3298
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6082,
+        "estimated_input_tokens": 1521,
+        "estimated_eval_tokens": 3298
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two distinct feature PRs: one adds a Promptable Concept Segmentation pipeline, the other adds a Promptable Visual Segmentation pipeline. They share documentation and pipeline scaffolding, but they target different tasks, models, and tests, so they are not duplicates.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": "pull_request:43613 is the broader and more complete integration, touching more model backends and tracker support, so it is the better representative PR if one must be chosen.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43612",
+            "right": "pull_request:43613",
+            "accept": false,
+            "reason": "Different underlying features: concept segmentation vs visual segmentation. Overlap is only in shared pipeline/docs scaffolding, not the same concrete change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 10,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43651-2",
+      "nodes": [
+        "pull_request:43651",
+        "pull_request:43663"
+      ],
+      "soft_pairs": [
+        "pull_request:43651|pull_request:43663"
+      ],
+      "prepared_review_unit_hash": "c1d436fa62a2d4dc8b92107a67e120fab9915b712c6a6e6b3f7c36067f650f9a",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4292,
+        "estimated_input_tokens": 1073,
+        "estimated_eval_tokens": 2402
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4292,
+        "estimated_input_tokens": 1073,
+        "estimated_eval_tokens": 2402
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open `trainer.py` PRs, but they address different Trainer extension points: one changes gradient-accumulation loss scaling, the other refactors signature-column filtering for custom trainers. Same file, different code paths, not duplicates.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #43651 is the better representative of the cluster only in the sense that it is a concrete behavior change in training-step loss handling, but it is not a duplicate of #43663; the two PRs solve unrelated Trainer customization problems.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43651",
+            "right": "pull_request:43663",
+            "accept": false,
+            "reason": "Different changes: #43651 adds `_loss_is_scaled_for_ga` and alters loss normalization during training, while #43663 adds `_get_signature_columns` to change column filtering. They touch the same file but not the same underlying bug or mergeable change set."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 11,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43656-2",
+      "nodes": [
+        "pull_request:43656",
+        "pull_request:43842"
+      ],
+      "soft_pairs": [
+        "pull_request:43656|pull_request:43842"
+      ],
+      "prepared_review_unit_hash": "a6a36337ad5b8dc3f5f1586b6878b623ccbe70f2d36ed9b3b3da85b48cbc6f99",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 3664,
+        "estimated_input_tokens": 916,
+        "estimated_eval_tokens": 2088
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 3664,
+        "estimated_input_tokens": 916,
+        "estimated_eval_tokens": 2088
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs touch `src/transformers/cli/serve.py` and are about `TypeAdapter`/annotation-related NameErrors, but they address different concrete failure modes with different fixes. They are related, not duplicates.",
+        "confidence": 0.84,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "43842 is the clearest anchor: it explicitly targets the `TypeAdapter` NameError when `pydantic` is absent and uses a narrowly scoped runtime-safety fix.",
+        "best_issue_reason": null,
+        "best_pr_reason": "43842 is the best representative because its title and patch directly describe the bug and the fix is self-contained.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43656",
+            "right": "pull_request:43842",
+            "accept": false,
+            "reason": "Same file and same broad `TypeAdapter` NameError theme, but different underlying problems: 43656 adds postponed annotations/type-hint cleanup, while 43842 specifically guards the `TypeAdapter` annotation for the `pydantic`-missing path. They could coexist, so this is not a duplicate."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 12,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43779-2",
+      "nodes": [
+        "pull_request:43779",
+        "pull_request:43816"
+      ],
+      "soft_pairs": [
+        "pull_request:43779|pull_request:43816"
+      ],
+      "prepared_review_unit_hash": "e7c674adcc031aa8883109a5dfc678b4407fb20b10764302092e384d031ba83a",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4362,
+        "estimated_input_tokens": 1091,
+        "estimated_eval_tokens": 2438
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4362,
+        "estimated_input_tokens": 1091,
+        "estimated_eval_tokens": 2438
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both pull requests appear to implement the same SwanLab integration change: adding run id/resume support in `integration_utils.py` for issue #43698, with one variant also expanding the docs. They are strong duplicate PR candidates.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #43816 is the cleaner representative: it carries the same SwanLab id/resume support plus the accompanying documentation updates, making it the more complete version of the shared change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #43816 is the best global representative because it appears to include the full feature change and docs for SwanLab run id/resume support, while #43779 is the narrower implementation-focused variant.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43779",
+            "right": "pull_request:43816",
+            "accept": true,
+            "reason": "Same underlying SwanLab callback enhancement for id/resume support in the same file and same issue target; the doc differences do not change the core code-path change, so these are duplicate PRs."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: both PRs target issue #43698, touch the same file, and the previews indicate the same SwanLab id/resume support with #43816 adding docs. The duplicate/mergeable verdict is conservative enough given the shared target and overlapping code change."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 13,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 2,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43996-28",
+      "nodes": [
+        "pull_request:43996",
+        "pull_request:43997",
+        "pull_request:44007",
+        "pull_request:44010",
+        "pull_request:44013",
+        "pull_request:44017",
+        "pull_request:44018",
+        "pull_request:44019",
+        "pull_request:44024",
+        "pull_request:44044",
+        "pull_request:44056",
+        "pull_request:44059",
+        "pull_request:44066",
+        "pull_request:44068",
+        "pull_request:44071",
+        "pull_request:44072",
+        "pull_request:44073",
+        "pull_request:44074",
+        "pull_request:44076",
+        "pull_request:44085",
+        "pull_request:44086",
+        "pull_request:44098",
+        "pull_request:44101",
+        "pull_request:44116",
+        "pull_request:44129",
+        "pull_request:44154",
+        "pull_request:44161",
+        "pull_request:44722"
+      ],
+      "soft_pairs": [
+        "pull_request:43996|pull_request:44085",
+        "pull_request:43996|pull_request:44044",
+        "pull_request:44066|pull_request:44085",
+        "pull_request:44007|pull_request:44072",
+        "pull_request:44072|pull_request:44722",
+        "pull_request:44066|pull_request:44072",
+        "pull_request:44013|pull_request:44044",
+        "pull_request:44066|pull_request:44086",
+        "pull_request:44066|pull_request:44071",
+        "pull_request:44018|pull_request:44068",
+        "pull_request:44066|pull_request:44068",
+        "pull_request:44018|pull_request:44066",
+        "pull_request:44019|pull_request:44722",
+        "pull_request:44086|pull_request:44722",
+        "pull_request:44019|pull_request:44085",
+        "pull_request:44018|pull_request:44019",
+        "pull_request:44019|pull_request:44071",
+        "pull_request:44018|pull_request:44116",
+        "pull_request:44068|pull_request:44116",
+        "pull_request:44068|pull_request:44722",
+        "pull_request:44071|pull_request:44722",
+        "pull_request:44018|pull_request:44085",
+        "pull_request:44018|pull_request:44086",
+        "pull_request:44068|pull_request:44086",
+        "pull_request:44071|pull_request:44085",
+        "pull_request:44098|pull_request:44101",
+        "pull_request:44085|pull_request:44086",
+        "pull_request:44059|pull_request:44086",
+        "pull_request:44018|pull_request:44071",
+        "pull_request:44059|pull_request:44722",
+        "pull_request:44017|pull_request:44161",
+        "pull_request:44017|pull_request:44073",
+        "pull_request:44017|pull_request:44722",
+        "pull_request:44068|pull_request:44071",
+        "pull_request:43997|pull_request:44073",
+        "pull_request:44017|pull_request:44071",
+        "pull_request:44071|pull_request:44086",
+        "pull_request:44073|pull_request:44161",
+        "pull_request:44073|pull_request:44101",
+        "pull_request:44059|pull_request:44071",
+        "pull_request:44056|pull_request:44101",
+        "pull_request:44018|pull_request:44059",
+        "pull_request:44059|pull_request:44066",
+        "pull_request:44101|pull_request:44161",
+        "pull_request:43997|pull_request:44017",
+        "pull_request:43997|pull_request:44056",
+        "pull_request:44071|pull_request:44076",
+        "pull_request:44074|pull_request:44101",
+        "pull_request:44056|pull_request:44161",
+        "pull_request:43997|pull_request:44018",
+        "pull_request:43997|pull_request:44161",
+        "pull_request:44074|pull_request:44161",
+        "pull_request:44010|pull_request:44071",
+        "pull_request:44018|pull_request:44129",
+        "pull_request:44010|pull_request:44073",
+        "pull_request:43997|pull_request:44074",
+        "pull_request:44010|pull_request:44018",
+        "pull_request:44071|pull_request:44129",
+        "pull_request:43997|pull_request:44129",
+        "pull_request:43996|pull_request:44161",
+        "pull_request:43996|pull_request:44017",
+        "pull_request:44010|pull_request:44044",
+        "pull_request:44044|pull_request:44074",
+        "pull_request:44066|pull_request:44722",
+        "pull_request:44085|pull_request:44722",
+        "pull_request:44085|pull_request:44129",
+        "pull_request:43996|pull_request:44007",
+        "pull_request:43996|pull_request:44129",
+        "pull_request:44044|pull_request:44129",
+        "pull_request:44044|pull_request:44154",
+        "pull_request:44024|pull_request:44076",
+        "pull_request:44073|pull_request:44074"
+      ],
+      "prepared_review_unit_hash": "d13eb9f09d557950649627ed98414cdd2c86b282f0414d98d95b018716e08bd7",
+      "component_budget": {
+        "node_count": 28,
+        "item_count": 28,
+        "soft_pair_count": 76,
+        "serialized_chars": 84116,
+        "estimated_input_tokens": 21029,
+        "estimated_eval_tokens": 42314
+      },
+      "budget": {
+        "node_count": 28,
+        "item_count": 28,
+        "soft_pair_count": 72,
+        "serialized_chars": 76938,
+        "estimated_input_tokens": 19235,
+        "estimated_eval_tokens": 38726
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a theme cluster around standardized output tracing, not true duplicates across models. The only strong duplicate-like overlaps are within the same model/code path: GPT-Neo (44018/44068) and GPT-J/CodeGen (44066/44085/44722).",
+        "confidence": 0.87,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 44722 is the most consolidated representative of the GPT-J/CodeGen output-tracing work and overlaps the earlier GPT-J-specific variants.",
+        "best_issue_reason": null,
+        "best_pr_reason": "44722 is the best cluster anchor: latest, broadest among the near-duplicate GPT-J-side changes, and it directly targets the shared output-tracing refactor path.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "Same umbrella output-tracing theme, but different model families; not the same concrete change."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "Different models and code paths; shared issue target alone is not enough."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44085",
+            "accept": true,
+            "reason": "Both touch the GPT-J output-tracing code path in the same file and look like overlapping implementations of the same refactor."
+          },
+          {
+            "left": "pull_request:44007",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "RegNet/ResNet vs EfficientNet-style refactors are different model changes, not one duplicate bug."
+          },
+          {
+            "left": "pull_request:44072",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "Different model families; both are output-tracing refactors but not the same concrete change."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "Different model files and different refactors."
+          },
+          {
+            "left": "pull_request:44013",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "MobileNetV2 and DeBERTa are unrelated implementations despite similar tracing cleanup."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "Different models; same general standardized-output theme only."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-J/CodeGen vs MPT are different code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44068",
+            "accept": true,
+            "reason": "Same GPT-Neo model file and same output-tracing refactor; these look like overlapping versions of one change."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "GPT-J/CodeGen versus GPT-Neo; same theme, different code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44066",
+            "accept": false,
+            "reason": "Different model families; not mergeable as one concrete fix."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "ResNet vs GPT-J/CodeGen; too broad to be duplicates."
+          },
+          {
+            "left": "pull_request:44086",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "MGP-STR and GPT-J/CodeGen are unrelated changes."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "ResNet and GPT-J are different model-specific refactors."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44019",
+            "accept": false,
+            "reason": "GPT-Neo and ResNet are different code paths."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "Different models; shared output-tracing theme is too broad."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44116",
+            "accept": false,
+            "reason": "GPT-Neo and Flaubert are different implementations."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44116",
+            "accept": false,
+            "reason": "Different models and different concrete fixes."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "GPT-Neo is not the same code path as GPT-J/CodeGen."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "MPT and GPT-J/CodeGen are unrelated model changes."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J are distinct model-specific changes."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "Different models and files."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo vs MGP-STR; not the same underlying bug."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "MPT and GPT-J are different code paths."
+          },
+          {
+            "left": "pull_request:44098",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "ViLT and Flaubert/XLM are separate model refactors."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-J and MGP-STR are unrelated implementations."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT2 and MGP-STR are different model changes."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "SegFormer and LongT5 are different model-specific refactors."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "SegFormer and VisualBert are unrelated model changes."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "Different models; only the generic tracing theme overlaps."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-Neo and MPT are different code paths."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "RegNet and VisualBert are unrelated model refactors."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "SegFormer and MPT are different concrete changes."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "Different models; not mergeable as one PR."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "VisualBert and LongT5 are different model families."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "VisualBert and Flaubert/XLM are separate code paths."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT2 and MPT are different model implementations."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "MPNet and Flaubert/XLM are unrelated changes."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "RegNet and GPT-Neo are different models."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44056",
+            "accept": false,
+            "reason": "RegNet and MPNet are different code paths."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "MPT and ImageGPT are different model-specific refactors."
+          },
+          {
+            "left": "pull_request:44074",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "TextNet and Flaubert/XLM are unrelated changes."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "MPNet and LongT5 are different model families."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "RegNet and SpeechT5 are different code paths."
+          },
+          {
+            "left": "pull_request:44074",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "TextNet and LongT5 are unrelated model changes."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "SqueezeBert and MPT are different models."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "SqueezeBert and VisualBert are different code paths."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44017",
+            "accept": false,
+            "reason": "FNet/CVT and SegFormer are different model-specific refactors."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "SqueezeBert and DeBERTa are unrelated changes."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "DeBERTa and TextNet are different models."
+          },
+          {
+            "left": "pull_request:44024",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "FocalNet and ImageGPT are different concrete implementations."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "VisualBert and TextNet are different model-specific refactors."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "GPT-Neo and SpeechT5 are unrelated changes."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "SqueezeBert and GPT-Neo are different model code paths."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "DeBERTa and SpeechT5 are different models."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "MPT and SpeechT5 are unrelated model refactors."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "Different models and different concrete changes."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44154",
+            "accept": false,
+            "reason": "DeBERTa and Vits are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "Different model families; shared issue target is not enough."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44722",
+            "accept": true,
+            "reason": "Both overlap on the GPT-J output-tracing refactor in the same file; 44722 looks like the broader consolidated version."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: it correctly distinguishes the broad cross-model output-tracing theme from the smaller same-code-path clusters. The accepted soft edges are conservative enough because they stay within the same model/file families (GPT-Neo and GPT-J/CodeGen), and the cross-model pairs are rejected appropriately. The canonical PR choice is also supported by the packet as the latest, broadest GPT-J/CodeGen variant."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 13,
+      "component_count": 34,
+      "review_unit_index": 2,
+      "review_unit_count": 2,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43996-28",
+      "nodes": [
+        "pull_request:43997",
+        "pull_request:44010",
+        "pull_request:44056",
+        "pull_request:44073",
+        "pull_request:44074",
+        "pull_request:44129"
+      ],
+      "soft_pairs": [
+        "pull_request:44056|pull_request:44073",
+        "pull_request:44056|pull_request:44074",
+        "pull_request:43997|pull_request:44010",
+        "pull_request:44010|pull_request:44129"
+      ],
+      "prepared_review_unit_hash": "5025ab2700dd76a18b68676fccf735462e4a308a7df8418d770e84dcddd598a0",
+      "component_budget": {
+        "node_count": 28,
+        "item_count": 28,
+        "soft_pair_count": 76,
+        "serialized_chars": 84116,
+        "estimated_input_tokens": 21029,
+        "estimated_eval_tokens": 42314
+      },
+      "budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 4,
+        "serialized_chars": 12756,
+        "estimated_input_tokens": 3189,
+        "estimated_eval_tokens": 6634
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These PRs share a broad theme\u2014migrating different model implementations to standardized output tracing/capture\u2014but they touch different architectures and code paths, so they are related maintenance changes rather than duplicates.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 44010 is the clearest representative of the cluster: it directly frames the shared output-collection refactor and is a straightforward model-specific example of the pattern.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 44010 is the best overall representative because its title and diff most directly describe the common maintenance goal without extra model-specific complications or an explicit dependency on another issue.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "Both are output-tracing refactors, but they affect different models (MPNet vs VisualBert) and different internal layers; same theme, not the same concrete change."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "Different model code paths and different refactor details; they modernize output capture in separate implementations, so they are not duplicates."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44010",
+            "accept": false,
+            "reason": "RegNet and SqueezeBert are unrelated model implementations; the PRs share a refactor pattern but not one underlying bug or one mergeable fix."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "SqueezeBert and SpeechT5 are different subsystems with distinct output-shaping changes; they are not the same concrete fix and should not be merged as duplicates."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 14,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43998-13",
+      "nodes": [
+        "pull_request:43998",
+        "pull_request:43999",
+        "pull_request:44000",
+        "pull_request:44001",
+        "pull_request:44002",
+        "pull_request:44003",
+        "pull_request:44004",
+        "pull_request:44025",
+        "pull_request:44026",
+        "pull_request:44027",
+        "pull_request:44028",
+        "pull_request:44029",
+        "pull_request:44030"
+      ],
+      "soft_pairs": [
+        "pull_request:43998|pull_request:44002",
+        "pull_request:43998|pull_request:44000",
+        "pull_request:43998|pull_request:44001",
+        "pull_request:43998|pull_request:44003",
+        "pull_request:43998|pull_request:44029",
+        "pull_request:43999|pull_request:44001",
+        "pull_request:43999|pull_request:44002",
+        "pull_request:43999|pull_request:44025",
+        "pull_request:43999|pull_request:44026",
+        "pull_request:43999|pull_request:44027",
+        "pull_request:44000|pull_request:44001",
+        "pull_request:44000|pull_request:44002",
+        "pull_request:44000|pull_request:44003",
+        "pull_request:44000|pull_request:44025",
+        "pull_request:44001|pull_request:44002",
+        "pull_request:44001|pull_request:44003",
+        "pull_request:44001|pull_request:44025",
+        "pull_request:44002|pull_request:44029",
+        "pull_request:44002|pull_request:44003",
+        "pull_request:44003|pull_request:44028",
+        "pull_request:44003|pull_request:44004",
+        "pull_request:43998|pull_request:44004",
+        "pull_request:44001|pull_request:44004",
+        "pull_request:44002|pull_request:44004",
+        "pull_request:44000|pull_request:44004",
+        "pull_request:44025|pull_request:44026",
+        "pull_request:44025|pull_request:44027",
+        "pull_request:44025|pull_request:44028",
+        "pull_request:44025|pull_request:44029",
+        "pull_request:44025|pull_request:44030",
+        "pull_request:44026|pull_request:44029",
+        "pull_request:44026|pull_request:44027",
+        "pull_request:44026|pull_request:44028",
+        "pull_request:44026|pull_request:44030",
+        "pull_request:44027|pull_request:44028",
+        "pull_request:44027|pull_request:44029",
+        "pull_request:44027|pull_request:44030",
+        "pull_request:44028|pull_request:44029",
+        "pull_request:44028|pull_request:44030",
+        "pull_request:44029|pull_request:44030"
+      ],
+      "prepared_review_unit_hash": "72c8633fe4ca5175fbce9537f6784749ab6fa8827b16fa4816549d4338c00c5f",
+      "component_budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 40,
+        "serialized_chars": 40178,
+        "estimated_input_tokens": 10045,
+        "estimated_eval_tokens": 20346
+      },
+      "budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 40,
+        "serialized_chars": 40178,
+        "estimated_input_tokens": 10045,
+        "estimated_eval_tokens": 20346
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are thematically similar PRs for refactoring output tracing, but they target different model implementations and code paths. They look like a family of related changes, not duplicate fixes or mergeable single changes.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No true duplicate canonical PR here; 44003 is the broadest representative because it spans two related model files (mamba and falcon_mamba) and shows the core output-tracing refactor pattern.",
+        "best_issue_reason": null,
+        "best_pr_reason": "44003 is the best single representative of the cluster\u2019s theme, but it should not be treated as a duplicate of the others since each PR is model-specific.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Different model code paths (timm_backbone vs upernet); same refactor theme only."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44000",
+            "accept": false,
+            "reason": "Different architectures and files; not the same concrete bug/change."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Distinct model implementations; only a shared output-tracing refactor pattern."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "Different model families (timm_backbone vs mamba/falcon_mamba); not mergeable as one fix."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Codegen-specific forward signature change is unrelated to timm_backbone."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "RWKV refactor touches a different model path and cannot be treated as the same change."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "MobileNetV1 and UnivNet are unrelated code paths; same refactor idea only."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Different model subsystems; not the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "MobileNetV1 and Depth Anything are separate implementations, so this is not a duplicate."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44026",
+            "accept": false,
+            "reason": "Vision-encoder-decoder is a different code path from MobileNetV1."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "Speech-encoder-decoder refactor is separate from MobileNetV1."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Vision-text dual encoder and UnivNet are unrelated model paths."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Different model families; no shared concrete fix."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "Same high-level refactor theme, but different model code paths."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "CodeGen forward/output handling is unrelated to vision-text dual encoder."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "Different models and files; not the same change."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "UnivNet and UperNet are unrelated implementations."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "Different model families; cannot plausibly be merged into one PR."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "UnivNet and Depth Anything are separate code paths."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "UperNet and mamba/falcon_mamba are unrelated concrete changes."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Different architecture and file; only same broad refactor theme."
+          },
+          {
+            "left": "pull_request:44003",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "mamba/falcon_mamba and superpoint are different model paths, not duplicates."
+          },
+          {
+            "left": "pull_request:44003",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "CodeGen output tracing is a separate change from mamba/falcon_mamba."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44026",
+            "accept": false,
+            "reason": "Depth Anything and vision-encoder-decoder are distinct code paths."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "Depth Anything and speech-encoder-decoder are separate implementations."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Different model families; not the same underlying fix."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Depth Anything and RWKV are unrelated model code paths."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Depth Anything and DPR are different concrete changes."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "Vision-encoder-decoder and speech-encoder-decoder are separate implementations despite similar wrappers."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Different model code paths; not a duplicate."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Vision-encoder-decoder and RWKV are unrelated concrete fixes."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Different models and files; same refactor pattern only."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Speech-encoder-decoder and superpoint do not share the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Different model paths; not mergeable as one PR."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Speech-encoder-decoder and DPR are distinct changes."
+          },
+          {
+            "left": "pull_request:44028",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "SuperPoint and RWKV are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:44028",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Different code paths and model types; not duplicates."
+          },
+          {
+            "left": "pull_request:44029",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "RWKV and DPR are separate model-specific refactors."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 15,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-44662-2",
+      "nodes": [
+        "pull_request:44662",
+        "pull_request:45401"
+      ],
+      "soft_pairs": [
+        "pull_request:44662|pull_request:45401"
+      ],
+      "prepared_review_unit_hash": "11541ef619306b165e9f825e12d8075a3ed46c188a64924f730ee720c328207e",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6551,
+        "estimated_input_tokens": 1638,
+        "estimated_eval_tokens": 3532
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6551,
+        "estimated_input_tokens": 1638,
+        "estimated_eval_tokens": 3532
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are two unrelated model-addition PRs: one introduces PenguinVL, the other adds Voxtral TTS support. They share some auto/docs scaffolding, but they target different model families, code paths, and release goals, so this is not a duplicate pair.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No canonical PR: 44662 is a PenguinVL implementation, while 45401 is Voxtral TTS support with an explicit issue target; they address different underlying changes.",
+        "best_issue_reason": null,
+        "best_pr_reason": "No single best PR for the cluster because the two PRs are not duplicates and should not be merged into one representative change.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44662",
+            "right": "pull_request:45401",
+            "accept": false,
+            "reason": "Different model additions and concrete code paths (PenguinVL vs Voxtral TTS); shared auto/docs files are just framework plumbing, not the same bug or change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 16,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-44979-2",
+      "nodes": [
+        "pull_request:44979",
+        "pull_request:45363"
+      ],
+      "soft_pairs": [
+        "pull_request:44979|pull_request:45363"
+      ],
+      "prepared_review_unit_hash": "e8522f71cc5acd9970d518dd633a561d583abb00b5a87df5869bdf114cd83875",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4407,
+        "estimated_input_tokens": 1102,
+        "estimated_eval_tokens": 2460
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4407,
+        "estimated_input_tokens": 1102,
+        "estimated_eval_tokens": 2460
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are related fusion efforts, but they are not the same change: PR #44979 introduces a new generic Module Fusion API, while PR #45363 adds n-to-1 kernel fusion support through KernelConfig plus an example and integration changes. They overlap conceptually, but they target different layers and code paths.",
+        "confidence": 0.89,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #44979 is the better canonical representative because it introduces the foundational Module Fusion API that the other work can build on, whereas PR #45363 is a more specific KernelConfig-based application.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #44979 is the strongest cluster representative: it is the core abstraction change, with tests, and is broader than the downstream kernel-fusion integration in PR #45363.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44979",
+            "right": "pull_request:45363",
+            "accept": false,
+            "reason": "Reject: both are about fusion, but #44979 adds a generic module fusion API and #45363 implements a specific n-to-1 kernel-fusion path via KernelConfig; they are not the same concrete change and would not be merged as one PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 17,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45086-2",
+      "nodes": [
+        "pull_request:45086",
+        "pull_request:45317"
+      ],
+      "soft_pairs": [
+        "pull_request:45086|pull_request:45317"
+      ],
+      "prepared_review_unit_hash": "483af6812c1a1435d48821518b83cdfad61b7cd146807c70b1c878f198f62440",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4414,
+        "estimated_input_tokens": 1104,
+        "estimated_eval_tokens": 2464
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4414,
+        "estimated_input_tokens": 1104,
+        "estimated_eval_tokens": 2464
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both pull requests fix the same AttributeError in `_patch_mistral_regex` by switching from `tokenizer.backend_tokenizer.pre_tokenizer` to `tokenizer.pre_tokenizer`; 45317 is the more complete version because it also adds tests.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "45317 is the better canonical PR because it addresses the same concrete bug as 45086 but includes accompanying test coverage and looks like the more complete patch.",
+        "best_issue_reason": null,
+        "best_pr_reason": "45317 is the strongest representative for the cluster: same fix, same target issue, and it adds tests, making it the most merge-ready version.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45086",
+            "right": "pull_request:45317",
+            "accept": true,
+            "reason": "Same underlying `_patch_mistral_regex` AttributeError fix on the same code path and shared issue target; the second PR is a superset with tests, so these are plausibly one merged change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 18,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45115-2",
+      "nodes": [
+        "pull_request:45115",
+        "pull_request:45569"
+      ],
+      "soft_pairs": [
+        "pull_request:45115|pull_request:45569"
+      ],
+      "prepared_review_unit_hash": "27d236b7c7298d915e46d6a19dbb33d4c640bade82694d0faf42f6994fcc20d5",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6066,
+        "estimated_input_tokens": 1517,
+        "estimated_eval_tokens": 3290
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6066,
+        "estimated_input_tokens": 1517,
+        "estimated_eval_tokens": 3290
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs touch the Nemotron H codepath, but they are not the same change: #45115 is a smaller refactor/conversion-mapping update toward GraniteMoeHybrid compatibility, while #45569 introduces the dedicated NemotronH dense/sparse model split and associated docs/tests. The overlap is mostly shared model files, not a single concrete bug fix.",
+        "confidence": 0.92,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No issue-linked canonical PR exists here; if one PR must represent the cluster, #45569 is the broader end-state implementation because it adds the dedicated model variants and user-facing docs/tests.",
+        "best_issue_reason": null,
+        "best_pr_reason": "#45569 is the best overall representative because it covers the full NemotronH family restructuring and documentation, whereas #45115 is a narrower conversion/compatibility refactor.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45115",
+            "right": "pull_request:45569",
+            "accept": false,
+            "reason": "Not the same underlying change: #45115 is about conversion key remapping and inheritance/refactor details, while #45569 adds separate dense/sparse model implementations plus docs/tests. Shared files alone are not enough to treat these as duplicates."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 19,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45332-2",
+      "nodes": [
+        "pull_request:45332",
+        "pull_request:45333"
+      ],
+      "soft_pairs": [
+        "pull_request:45332|pull_request:45333"
+      ],
+      "prepared_review_unit_hash": "8b0722d862c2d330641bb7e183a1506f114f415aed58d03d88a7bc5597944289",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6081,
+        "estimated_input_tokens": 1521,
+        "estimated_eval_tokens": 3298
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6081,
+        "estimated_input_tokens": 1521,
+        "estimated_eval_tokens": 3298
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are related PRs in the same heterogeneous-modeling effort, but they are not duplicates: #45333 appears to be the config-only slice, while #45332 includes that plus broader modeling/cache/integration changes. I would not merge them under a soft duplicate edge.",
+        "confidence": 0.91,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "#45332 is the broader, more complete implementation and subsumes the config work shown in #45333, so it is the better canonical representative for this cluster.",
+        "best_issue_reason": null,
+        "best_pr_reason": "#45332 is the strongest representative because it covers the larger underlying change set (per-layer config plus heterogeneous modeling/integration), whereas #45333 is a narrower subset focused on configuration only.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45332",
+            "right": "pull_request:45333",
+            "accept": false,
+            "reason": "Related topic, but not the same concrete change: #45333 is config-only, while #45332 adds modeling/cache/integration support as well. They could coexist as separate PRs rather than being merged as duplicates."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 20,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45415-2",
+      "nodes": [
+        "pull_request:45415",
+        "pull_request:45425"
+      ],
+      "soft_pairs": [
+        "pull_request:45415|pull_request:45425"
+      ],
+      "prepared_review_unit_hash": "2b4daaa6f748601149058f08cbbced2fbd7c41c058b1056aa0c5b7f67d82fcb9",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5514,
+        "estimated_input_tokens": 1379,
+        "estimated_eval_tokens": 3014
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5514,
+        "estimated_input_tokens": 1379,
+        "estimated_eval_tokens": 3014
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are related typing-focused PRs, but they address different scopes: #45415 is a broad type-checking sweep across many transformer modules, while #45425 adds specific typing protocols/aliases and updates a smaller set of typing-sensitive paths. They share a few files, but not the same concrete change or bug fix.",
+        "confidence": 0.88,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "#45425 is the more focused typing-infrastructure PR; #45415 is a much broader repository-wide type-checking cleanup. They overlap in theme, but not in a single underlying change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "#45425 is the better representative of the cluster because it introduces concrete typing support abstractions, whereas #45415 is a sweeping type-checking pass with many unrelated file edits.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45415",
+            "right": "pull_request:45425",
+            "accept": false,
+            "reason": "Both are typing-related and share a few touched files, but #45415 is a broad type-checking sweep while #45425 adds specific typing aliases/protocols and updates targeted code paths. They are not the same underlying fix and would not plausibly be merged as one PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 21,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45548-2",
+      "nodes": [
+        "pull_request:45548",
+        "pull_request:45662"
+      ],
+      "soft_pairs": [
+        "pull_request:45548|pull_request:45662"
+      ],
+      "prepared_review_unit_hash": "0ad5f8078275230420a1efbe410f1ae63e0af1361ad1b7028acf49a9b2d3f695",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4508,
+        "estimated_input_tokens": 1127,
+        "estimated_eval_tokens": 2510
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4508,
+        "estimated_input_tokens": 1127,
+        "estimated_eval_tokens": 2510
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two PRs are related to expert-parallel distributed loading, but they fix different failure modes in different code paths: one targets DeepSpeed ZeRO-3/accelerate launch serialization and config handling, the other targets FSDP2/rank-0 broadcast overwriting experts during sharded parameter loading.",
+        "confidence": 0.92,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45662 is the better canonical representative because it addresses a more concrete, multi-file runtime bug in expert-parallel loading and includes the more direct fix path around parameter sharding/wrapping.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45662 is the strongest overall artifact in this cluster since its fix is more directly tied to the observed data-loss symptom and spans the actual loading path rather than a narrower config-serialization adjustment.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45548",
+            "right": "pull_request:45662",
+            "accept": false,
+            "reason": "Not the same underlying bug: #45548 focuses on DeepSpeed ZeRO-3/accelerate launch config serialization and EP detection, while #45662 fixes experts being overwritten during FSDP2/post-shard parameter assignment. Shared EP context and one common file are not enough to merge them as duplicates."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 22,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45639-2",
+      "nodes": [
+        "pull_request:45639",
+        "pull_request:45645"
+      ],
+      "soft_pairs": [
+        "pull_request:45639|pull_request:45645"
+      ],
+      "prepared_review_unit_hash": "446aa5b53dc5a045399cf21597918e5d3652d35ce02dab86ab04b6b6a546bb0d",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4673,
+        "estimated_input_tokens": 1169,
+        "estimated_eval_tokens": 2594
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4673,
+        "estimated_input_tokens": 1169,
+        "estimated_eval_tokens": 2594
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs address the same xdist-related captured_info problem around patched testing debug logs: one makes the writer use worker-specific output files, and the other adds workflow-side collection of multiple captured_info files plus the same underlying writer fix. They are close enough to be considered the same change family.",
+        "confidence": 0.92,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45645 is the broader, more complete fix: it includes the xdist-safe output handling in testing_utils and also updates CI log collection so the new per-worker files are preserved and displayed.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45645 is the best representative because it covers the full end-to-end artifact-handling change, not just the low-level write-side fix.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45639",
+            "right": "pull_request:45645",
+            "accept": true,
+            "reason": "Same underlying xdist collision bug for captured_info artifacts; #45645 clearly subsumes #45639 by adding the workflow-side aggregation on top of the shared testing_utils fix."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 23,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45737-2",
+      "nodes": [
+        "pull_request:45737",
+        "pull_request:45738"
+      ],
+      "soft_pairs": [
+        "pull_request:45737|pull_request:45738"
+      ],
+      "prepared_review_unit_hash": "cc2ade73a49ec6a8883a7a24c433988645588bcd049e8aad2bce56b1b24daa1c",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4131,
+        "estimated_input_tokens": 1033,
+        "estimated_eval_tokens": 2322
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4131,
+        "estimated_input_tokens": 1033,
+        "estimated_eval_tokens": 2322
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs address the same MusicgenMelody cache initialization bug in `forward()`: replacing `EncoderDecoderCache(DynamicCache, DynamicCache)` with a single `DynamicCache`. PR #45737 is the more complete duplicate because it bundles the fix with a regression test, while #45738 is the fix-only variant.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45737 is the best canonical PR because it contains the same code fix as #45738 plus a regression test, making it the most complete representation of the bug fix.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45737 is the strongest overall candidate: it fixes the concrete cache bug and adds a targeted regression test, whereas #45738 only changes the implementation.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45737",
+            "right": "pull_request:45738",
+            "accept": true,
+            "reason": "Same underlying code-path bug in `musicgen_melody` cache setup; the diffs match on the core fix and the test-only addition in #45737 could plausibly live in the same PR as #45738."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The assessment is grounded in the packet: both PRs make the same `musicgen_melody` cache initialization change, and #45737 only adds a regression test on top of that fix. The soft-edge accept is conservative enough because the diffs are mergeable into one PR for the same bugfix."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 24,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46126-2",
+      "nodes": [
+        "pull_request:46126",
+        "pull_request:46269"
+      ],
+      "soft_pairs": [
+        "pull_request:46126|pull_request:46269"
+      ],
+      "prepared_review_unit_hash": "deb492424e89ab074ececcd3cb6213f43b39d57547c8d05839d49071f6b0c73a",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5526,
+        "estimated_input_tokens": 1382,
+        "estimated_eval_tokens": 3020
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5526,
+        "estimated_input_tokens": 1382,
+        "estimated_eval_tokens": 3020
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are two different distributed-feature PRs with some overlap in shared files, but they target different changes: one is a focused EP/Trainer integration, the other is a broad distributed branch base with FSDP/CI and many unrelated edits.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #46126 is the more specific, self-contained change with a clear feature scope (EP + Trainer integration).",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #46126 is the best representative of the cluster because it has the narrowest, clearest concrete code-path change; #46269 is too broad and mixed to serve as a canonical duplicate target.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46126",
+            "right": "pull_request:46269",
+            "accept": false,
+            "reason": "Shared distributed filenames alone are not enough: #46126 is about EP routing/trainer integration, while #46269 is a broad distributed branch base covering FSDP, CI, and multiple subsystems. They do not look like the same underlying bug or mergeable change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 25,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46138-10",
+      "nodes": [
+        "pull_request:46138",
+        "pull_request:46432",
+        "pull_request:46501",
+        "pull_request:46502",
+        "pull_request:46503",
+        "pull_request:46504",
+        "pull_request:46554",
+        "pull_request:46558",
+        "pull_request:46563",
+        "pull_request:46716"
+      ],
+      "soft_pairs": [
+        "pull_request:46138|pull_request:46554",
+        "pull_request:46138|pull_request:46558",
+        "pull_request:46138|pull_request:46563",
+        "pull_request:46138|pull_request:46716",
+        "pull_request:46138|pull_request:46504",
+        "pull_request:46432|pull_request:46503",
+        "pull_request:46432|pull_request:46501",
+        "pull_request:46432|pull_request:46502",
+        "pull_request:46432|pull_request:46554",
+        "pull_request:46432|pull_request:46558",
+        "pull_request:46501|pull_request:46503",
+        "pull_request:46501|pull_request:46502",
+        "pull_request:46501|pull_request:46504",
+        "pull_request:46501|pull_request:46554",
+        "pull_request:46502|pull_request:46503",
+        "pull_request:46502|pull_request:46504",
+        "pull_request:46502|pull_request:46554",
+        "pull_request:46503|pull_request:46504",
+        "pull_request:46503|pull_request:46554",
+        "pull_request:46504|pull_request:46554",
+        "pull_request:46504|pull_request:46558",
+        "pull_request:46554|pull_request:46558",
+        "pull_request:46554|pull_request:46563",
+        "pull_request:46554|pull_request:46716",
+        "pull_request:46558|pull_request:46563",
+        "pull_request:46558|pull_request:46716",
+        "pull_request:46563|pull_request:46716",
+        "pull_request:46504|pull_request:46563",
+        "pull_request:46432|pull_request:46563",
+        "pull_request:46504|pull_request:46716",
+        "pull_request:46432|pull_request:46716"
+      ],
+      "prepared_review_unit_hash": "350892a7ceab14370c304c1f5683b2170f335e83fb49e47a01b1907ee046e8e2",
+      "component_budget": {
+        "node_count": 10,
+        "item_count": 10,
+        "soft_pair_count": 31,
+        "serialized_chars": 26687,
+        "estimated_input_tokens": 6672,
+        "estimated_eval_tokens": 13600
+      },
+      "budget": {
+        "node_count": 10,
+        "item_count": 10,
+        "soft_pair_count": 31,
+        "serialized_chars": 26687,
+        "estimated_input_tokens": 6672,
+        "estimated_eval_tokens": 13600
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are generic workflow-update PRs, not true duplicates. The closest cluster is the pr-ci-security-gate.yml edits, but each PR changes a different security-gate detail rather than the same concrete fix.",
+        "confidence": 0.9,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #46554 is the most substantive representative of the security-gate mini-cluster, since it changes the actual gate logic; the other pr-ci-security-gate PRs are separate hardening tweaks.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #46554 is the best cluster representative overall because it contains the main functional workflow change and is more central than the smaller permission/comment/secrets adjustments.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Different workflows and different behavior; one updates self-comment input handling, the other changes SSH runner security setup."
+          },
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46558",
+            "accept": false,
+            "reason": "Unrelated workflow updates in different files; no shared concrete bug or change."
+          },
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Different workflow files and distinct changes; not the same underlying fix."
+          },
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "Different workflow files and separate workflow-hardening changes; not duplicates."
+          },
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46504",
+            "accept": false,
+            "reason": "Different CI workflows and different command-safety tweak; not the same change."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46503",
+            "accept": false,
+            "reason": "Different workflow files and different trigger logic; no duplicate bug/change."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46501",
+            "accept": false,
+            "reason": "Different workflow files and unrelated permission placement changes."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46502",
+            "accept": false,
+            "reason": "Different workflows; one changes SSH runner secrets handling, the other reviewer assignment checkout behavior."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "SSH-runner hardening is unrelated to PR CI security-gate logic."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46558",
+            "accept": false,
+            "reason": "Different workflows and different security concerns; not the same fix."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
+          },
+          {
+            "left": "pull_request:46501",
+            "right": "pull_request:46503",
+            "accept": false,
+            "reason": "Different workflow files; one is anti-slop permissions, the other benchmark dispatch trigger."
+          },
+          {
+            "left": "pull_request:46501",
+            "right": "pull_request:46502",
+            "accept": false,
+            "reason": "Different workflow files and different behavior; not duplicates."
+          },
+          {
+            "left": "pull_request:46501",
+            "right": "pull_request:46504",
+            "accept": false,
+            "reason": "Different workflow files and unrelated CI tweaks."
+          },
+          {
+            "left": "pull_request:46501",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Different files and different workflow concerns."
+          },
+          {
+            "left": "pull_request:46502",
+            "right": "pull_request:46503",
+            "accept": false,
+            "reason": "Different workflow files and separate CI updates."
+          },
+          {
+            "left": "pull_request:46502",
+            "right": "pull_request:46504",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
+          },
+          {
+            "left": "pull_request:46502",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Different workflows and different concrete changes."
+          },
+          {
+            "left": "pull_request:46503",
+            "right": "pull_request:46504",
+            "accept": false,
+            "reason": "Different workflow files and distinct CI behavior changes."
+          },
+          {
+            "left": "pull_request:46503",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Different workflows and no shared concrete bug."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Different workflow files and unrelated fixes."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46558",
+            "accept": false,
+            "reason": "Different workflows; one is build-image tagging, the other PR CI gate permissions."
+          },
+          {
+            "left": "pull_request:46554",
+            "right": "pull_request:46558",
+            "accept": false,
+            "reason": "Same file, but different concrete changes: one adds security checks, the other adjusts permissions/approval flow."
+          },
+          {
+            "left": "pull_request:46554",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Same workflow file, but one changes executable security-gate logic while the other only adds explanatory comments."
+          },
+          {
+            "left": "pull_request:46554",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "Same workflow file, but different fixes: one changes gate logic, the other removes inherited secrets."
+          },
+          {
+            "left": "pull_request:46558",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Same workflow file, but different changes: permissions/guard condition versus documentation comment."
+          },
+          {
+            "left": "pull_request:46558",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "Same workflow file, but different concrete changes to the reusable workflow call."
+          },
+          {
+            "left": "pull_request:46563",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "One is a comment-only clarification; the other changes secrets inheritance. Not the same fix."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Different workflow files and different changes; not duplicates."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46716",
+            "accept": false,
+            "reason": "Different workflow files and unrelated changes."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 26,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46174-2",
+      "nodes": [
+        "pull_request:46174",
+        "pull_request:46780"
+      ],
+      "soft_pairs": [
+        "pull_request:46174|pull_request:46780"
+      ],
+      "prepared_review_unit_hash": "a48ce4ca162f555682f79411d7229ac1ccc08d7822dfab654bb881639fccddbc",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6030,
+        "estimated_input_tokens": 1508,
+        "estimated_eval_tokens": 3272
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6030,
+        "estimated_input_tokens": 1508,
+        "estimated_eval_tokens": 3272
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two PRs target the same reported MPS crash in the sinusoidal position embedding helper, but they implement different fixes and are not safe to treat as duplicates.",
+        "confidence": 0.89,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 46174 is the broader, more complete fix for the float64-on-MPS crash: it updates the helper across all affected model files and keeps the internal math on CPU float64 before casting back.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 46174 is the best representative of the cluster because it is the original, more comprehensive fix and has more discussion/review context than the narrower follow-up.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46174",
+            "right": "pull_request:46780",
+            "accept": false,
+            "reason": "They share the same issue target and helper, but the patches are different solutions: 46174 avoids float64 on MPS by doing the arithmetic on CPU first, while 46780 switches the helper to use dtype for internal arithmetic. Those are alternative implementations, not one mergeable duplicate PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 27,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46281-2",
+      "nodes": [
+        "pull_request:46281",
+        "pull_request:46890"
+      ],
+      "soft_pairs": [
+        "pull_request:46281|pull_request:46890"
+      ],
+      "prepared_review_unit_hash": "0433c4cd12a70a1fe8ac748396c8f68e3f4158bab735a8726ebef6cf8caaf277",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4280,
+        "estimated_input_tokens": 1070,
+        "estimated_eval_tokens": 2396
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4280,
+        "estimated_input_tokens": 1070,
+        "estimated_eval_tokens": 2396
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open PRs that both harden path handling against traversal, but they target different code paths and different inputs: sharded checkpoint index filenames vs the `transformers_weights` config field. They are related security fixes, not duplicates.",
+        "confidence": 0.89,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 46281 is the better cluster representative because it addresses the more specific checkpoint-shard file resolution path and directly validates index-derived filenames against directory escape.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 46281 is the strongest standalone representative of the cluster theme: a concrete path-traversal fix in shard loading with a clear, narrowly-scoped code change.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46281",
+            "right": "pull_request:46890",
+            "accept": false,
+            "reason": "Both are path-traversal hardening PRs, but they fix different mechanisms in different loaders (`weight_map` shard filenames vs `transformers_weights` config). They do not look like the same underlying bug or a PR that could be merged into one change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 28,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46574-2",
+      "nodes": [
+        "pull_request:46574",
+        "pull_request:46580"
+      ],
+      "soft_pairs": [
+        "pull_request:46574|pull_request:46580"
+      ],
+      "prepared_review_unit_hash": "d79b388ab23b97bec0ba17354356195ce5c2b7a757f740f722db1472a25670ef",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4410,
+        "estimated_input_tokens": 1103,
+        "estimated_eval_tokens": 2462
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4410,
+        "estimated_input_tokens": 1103,
+        "estimated_eval_tokens": 2462
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The cluster contains two unrelated pull requests: one fixes CodeLlama decode behavior removing real leading spaces, while the other fixes DeepSeek-Coder/DeepSeek-LLM checkpoint routing to the correct tokenizer class.",
+        "confidence": 0.99,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46574",
+            "right": "pull_request:46580",
+            "accept": false,
+            "reason": "Different bugs and different code paths: 46574 changes CodeLlama tokenizer decoding in `tokenization_code_llama.py`, while 46580 changes auto-tokenizer checkpoint mapping in `tokenization_auto.py` for DeepSeek models. They are not the same underlying fix and would not be merged into one PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 29,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46773-2",
+      "nodes": [
+        "pull_request:46773",
+        "pull_request:47087"
+      ],
+      "soft_pairs": [
+        "pull_request:46773|pull_request:47087"
+      ],
+      "prepared_review_unit_hash": "661f1e21191d284ac8230a5f38c2606ea782fba09184f37204de05c3a9a52154",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5432,
+        "estimated_input_tokens": 1358,
+        "estimated_eval_tokens": 2972
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5432,
+        "estimated_input_tokens": 1358,
+        "estimated_eval_tokens": 2972
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These PRs are both about padding-mask handling, but they fix different bugs in different code paths: PR 46773 corrects a batch-size-1 masking guard in linear-attention model code, while PR 47087 changes recurrent masking utility behavior for cache/chunked continued forwards. They are related conceptually but not duplicates.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 47087 is the better representative of the cluster because it fixes the shared masking utility at the framework level and includes an explicit issue reference, while PR 46773 is a model-specific fast-path condition fix.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 47087 is the strongest standalone PR in this set because it addresses a broader reusable masking path, has clearer problem framing, and is more central than the model-specific fix in PR 46773.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46773",
+            "right": "pull_request:47087",
+            "accept": false,
+            "reason": "Different underlying bugs: 46773 fixes a condition that skipped padding masking for batch size 1 in linear-attention model code, whereas 47087 fixes recurrent-mask creation during continued/cache forwards and single-token decode. They do not appear mergeable into one PR for the same concrete code-path problem."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 30,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46781-3",
+      "nodes": [
+        "pull_request:46781",
+        "pull_request:46993",
+        "pull_request:46994"
+      ],
+      "soft_pairs": [
+        "pull_request:46781|pull_request:46993",
+        "pull_request:46993|pull_request:46994"
+      ],
+      "prepared_review_unit_hash": "1182ceb2c520f0b0f4134d33ba790ff0fb174ea061fe22ebb34dba23e6b2954d",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 6603,
+        "estimated_input_tokens": 1651,
+        "estimated_eval_tokens": 3558
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 6603,
+        "estimated_input_tokens": 1651,
+        "estimated_eval_tokens": 3558
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The cluster contains three unrelated pull requests: one for QuantizedCache validation, one for prompt lookup decoding length capping, and one for empty-list validation in generation logits processing. They share only broad generation/theme similarity and should not be treated as duplicates.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46781",
+            "right": "pull_request:46993",
+            "accept": false,
+            "reason": "Different code paths and bugs: cache layer-type validation vs prompt lookup decoding max_length handling. No plausible single merged fix."
+          },
+          {
+            "left": "pull_request:46993",
+            "right": "pull_request:46994",
+            "accept": false,
+            "reason": "Both are generation-related, but they fix unrelated validation/logic bugs in different modules (candidate generation vs logits processor validation)."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 31,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-47026-2",
+      "nodes": [
+        "pull_request:47026",
+        "pull_request:47046"
+      ],
+      "soft_pairs": [
+        "pull_request:47026|pull_request:47046"
+      ],
+      "prepared_review_unit_hash": "f48a065bc422c4ec5e6aa2960062dc82f69bade5fa35e169b23486ad215e600d",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4441,
+        "estimated_input_tokens": 1111,
+        "estimated_eval_tokens": 2478
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4441,
+        "estimated_input_tokens": 1111,
+        "estimated_eval_tokens": 2478
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open xLSTM PRs target the same train-mode forward path in `modeling_xlstm.py` and the same tracking issue. PR 47046 appears to be the broader, more complete fix (it explicitly covers both `return_last_states=False` and `use_cache`), so it is the best representative for the cluster.",
+        "confidence": 0.71,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 47046 is the more complete and likely superseding fix: it targets the same xLSTM train-mode crash path as 47026 but also covers `use_cache` and includes tests, making it the better canonical PR.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 47046 is the strongest single PR to keep because it addresses the shared xLSTM train-mode failure mode more comprehensively than 47026.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:47026",
+            "right": "pull_request:47046",
+            "accept": true,
+            "reason": "Both PRs fix the same xLSTM train-mode code path in `modeling_xlstm.py` and share the same issue target; 47046 looks like a superset of the 47026 fix rather than a separate change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 32,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-47032-5",
+      "nodes": [
+        "pull_request:47032",
+        "pull_request:47051",
+        "pull_request:47060",
+        "pull_request:47074",
+        "pull_request:47076"
+      ],
+      "soft_pairs": [
+        "pull_request:47032|pull_request:47051",
+        "pull_request:47051|pull_request:47074",
+        "pull_request:47051|pull_request:47060",
+        "pull_request:47060|pull_request:47074",
+        "pull_request:47060|pull_request:47076",
+        "pull_request:47074|pull_request:47076"
+      ],
+      "prepared_review_unit_hash": "1851f642cbcf838d0dab29dfd1a2d8e2ee8d5dff1a6171bb6d33c9b2dec681b5",
+      "component_budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 6,
+        "serialized_chars": 12763,
+        "estimated_input_tokens": 3191,
+        "estimated_eval_tokens": 6638
+      },
+      "budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 6,
+        "serialized_chars": 12763,
+        "estimated_input_tokens": 3191,
+        "estimated_eval_tokens": 6638
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These PRs are thematically related around multimodal pipeline/processor handling, but they fix distinct bugs and changes rather than the same underlying issue. No soft edge is a true duplicate.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "47051 is the broadest umbrella change in the shared multimodal pipeline codepath and most representative of the cluster, but it still addresses a different fix than the other PRs.",
+        "best_issue_reason": null,
+        "best_pr_reason": "47051 is the best single representative PR because it spans the shared pipeline plumbing touched by nearby PRs, even though the cluster does not contain a true duplicate set.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:47032",
+            "right": "pull_request:47051",
+            "accept": false,
+            "reason": "Same area (image_text_to_text) but different bugs: stop_sequence behavior vs direct generation kwarg forwarding."
+          },
+          {
+            "left": "pull_request:47051",
+            "right": "pull_request:47074",
+            "accept": false,
+            "reason": "Both touch AnyToAnyPipeline, but one forwards kwargs while the other normalizes input keys; separate code-path fixes."
+          },
+          {
+            "left": "pull_request:47051",
+            "right": "pull_request:47060",
+            "accept": false,
+            "reason": "Different subsystems and bugs: multimodal pipeline kwarg plumbing vs Qwen Omni batched post-processing."
+          },
+          {
+            "left": "pull_request:47060",
+            "right": "pull_request:47074",
+            "accept": false,
+            "reason": "Unrelated fixes: Qwen batched decoding behavior versus AnyToAnyPipeline input key normalization."
+          },
+          {
+            "left": "pull_request:47060",
+            "right": "pull_request:47076",
+            "accept": false,
+            "reason": "Different model families and failure modes: Qwen post-processing vs Gemma4 audio placeholder validation."
+          },
+          {
+            "left": "pull_request:47074",
+            "right": "pull_request:47076",
+            "accept": false,
+            "reason": "Both are multimodal, but they fix different problems in different components: pipeline input mapping vs processor placeholder consistency."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 33,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-47105-2",
+      "nodes": [
+        "pull_request:47105",
+        "pull_request:47129"
+      ],
+      "soft_pairs": [
+        "pull_request:47105|pull_request:47129"
+      ],
+      "prepared_review_unit_hash": "5e5adcdbccf9a97bd2f73884338dbc2150d095a5f5fcc19aaa4d3114bf3f9515",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4172,
+        "estimated_input_tokens": 1043,
+        "estimated_eval_tokens": 2342
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4172,
+        "estimated_input_tokens": 1043,
+        "estimated_eval_tokens": 2342
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two PRs are unrelated: one re-enables and adjusts a VITS model forward test, while the other updates an RT-DETR v2 weight-conversion script for changed model keys.",
+        "confidence": 0.99,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No clear canonical PR: the cluster contains two unrelated pull requests touching different models and different code paths.",
+        "best_issue_reason": null,
+        "best_pr_reason": "Neither PR is a good cluster representative for deduping because they address separate, non-overlapping changes; if forced, 47129 is slightly more self-contained, but it is not a duplicate of 47105.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:47105",
+            "right": "pull_request:47129",
+            "accept": false,
+            "reason": "Different artifacts and code paths: VITS test changes versus RT-DETR v2 converter fixes. They do not appear to be the same underlying bug or change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 34,
+      "component_count": 34,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-47122-3",
+      "nodes": [
+        "pull_request:47122",
+        "pull_request:47123",
+        "pull_request:47124"
+      ],
+      "soft_pairs": [
+        "pull_request:47122|pull_request:47123",
+        "pull_request:47122|pull_request:47124",
+        "pull_request:47123|pull_request:47124"
+      ],
+      "prepared_review_unit_hash": "24d92246f23caefccd79e3d7932d01b4bcbff295a37656d4f67388169c1a40e3",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 6618,
+        "estimated_input_tokens": 1655,
+        "estimated_eval_tokens": 3566
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 6618,
+        "estimated_input_tokens": 1655,
+        "estimated_eval_tokens": 3566
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are three unrelated pull requests fixing separate model-specific integration-test failures: EdgeTam runtime kwargs handling, MiniMax-M3-VL load/offload test setup, and Falcon-Mamba expected output updates. They are not duplicates of one another.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 47122 is the best canonical PR because it contains the clearest concrete code-path fix: removing unsupported `original_sizes` before calling the EdgeTam backbone. The other PRs are test-only adjustments for different models and failure modes.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 47122 is the strongest representative of the cluster because it fixes the underlying model implementation bug rather than only changing test harness expectations or loader setup.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:47122",
+            "right": "pull_request:47123",
+            "accept": false,
+            "reason": "Different models and different failures: EdgeTam fixes unsupported `original_sizes` passed into the backbone, while MiniMax-M3-VL only adds an offload folder in a test loader. Not the same underlying bug or change."
+          },
+          {
+            "left": "pull_request:47122",
+            "right": "pull_request:47124",
+            "accept": false,
+            "reason": "EdgeTam runtime kwargs handling vs. Falcon-Mamba test expectation updates. They touch different code paths and failure types, so they are not duplicates."
+          },
+          {
+            "left": "pull_request:47123",
+            "right": "pull_request:47124",
+            "accept": false,
+            "reason": "Both are test-related, but for different models and different issues: one addresses model loading/offloading, the other adjusts generation expectations. Not the same underlying fix."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    }
+  ]
+}
diff --git a/snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/manifest.json b/snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..fb26e6a0eb724c603287a01c3f2fde0683d0d062
--- /dev/null
+++ b/snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/manifest.json
@@ -0,0 +1,14 @@
+{
+  "analysis_id": "hybrid-model-20260708t000615z",
+  "artifacts": {
+    "hybrid": "snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.json",
+    "hybrid_reviews": "snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.llm-reviews.json"
+  },
+  "channel": "canonical",
+  "model": null,
+  "published_at": "2026-07-08T00:12:28Z",
+  "repo": "huggingface/transformers",
+  "schema_version": 1,
+  "snapshot_id": "20260708T000615Z",
+  "variant": "hybrid"
+}
diff --git a/snapshots/20260708T000615Z/manifest.json b/snapshots/20260708T000615Z/manifest.json
index 0621694c67d156cedf8872d5065c8c25c0839b41..19462234ed0895afa3b9e8bd40117915be87a9ad 100644
--- a/snapshots/20260708T000615Z/manifest.json
+++ b/snapshots/20260708T000615Z/manifest.json
@@ -32,6 +32,24 @@
     "timeline_events": 338
   },
   "extracted_at": "2026-07-08T00:06:15Z",
+  "published_analysis": {
+    "canonical_analysis_id": "hybrid-model-20260708t000615z",
+    "runs": {
+      "hybrid-model-20260708t000615z": {
+        "analysis_id": "hybrid-model-20260708t000615z",
+        "artifacts": {
+          "hybrid": "snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.json",
+          "hybrid_reviews": "snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/analysis-report-hybrid.llm-reviews.json"
+        },
+        "channel": "canonical",
+        "manifest_path": "snapshots/20260708T000615Z/analysis-runs/hybrid-model-20260708t000615z/manifest.json",
+        "model": null,
+        "published_at": "2026-07-08T00:12:28Z",
+        "variant": "hybrid"
+      }
+    },
+    "schema_version": 1
+  },
   "repo": "huggingface/transformers",
   "snapshot_id": "20260708T000615Z",
   "watermark": {