diff --git a/analysis/current/analysis-report-hybrid.json b/analysis/current/analysis-report-hybrid.json
index 9015ec6c8e4226ae22ab5479378590bad2ed084c..096120647e0bf5bb1ccfb3bfdfe8201f55d44098 100644
--- a/analysis/current/analysis-report-hybrid.json
+++ b/analysis/current/analysis-report-hybrid.json
@@ -1,8 +1,8 @@
 {
   "schema_version": "1.0",
   "repo": "huggingface/transformers",
-  "snapshot_id": "20260507T000029Z",
-  "generated_at": "2026-05-07T00:10:46Z",
+  "snapshot_id": "20260507T060024Z",
+  "generated_at": "2026-05-07T06:05:24Z",
   "evidence_quality": "full",
   "llm_enrichment": true,
   "meta_bugs": [
@@ -1148,6 +1148,204 @@
         }
       ]
     },
+    {
+      "cluster_id": "cluster-43656-4",
+      "summary": "Cluster of 1 issues and 3 PRs centered on issue #43824.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 43824,
+      "canonical_pr_number": 43656,
+      "issue_numbers": [
+        43824
+      ],
+      "pr_numbers": [
+        43656,
+        43836,
+        43842
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target",
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 43656,
+          "right_pr_number": 43836,
+          "code_similarity": 0.36,
+          "size_similarity": 0.176,
+          "file_overlap": 1.0,
+          "area_overlap": 0.051,
+          "patch_similarity": 0.048,
+          "shared_filenames": [
+            "src/transformers/cli/serve.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/cli/serve.py",
+              "left_ranges": [
+                [
+                  11,
+                  18
+                ],
+                [
+                  30,
+                  36
+                ],
+                [
+                  315,
+                  323
+                ],
+                [
+                  665,
+                  671
+                ],
+                [
+                  931,
+                  937
+                ],
+                [
+                  1843,
+                  1849
+                ],
+                [
+                  1868,
+                  1874
+                ]
+              ],
+              "right_ranges": [
+                [
+                  11,
+                  18
+                ],
+                [
+                  359,
+                  429
+                ],
+                [
+                  584,
+                  590
+                ],
+                [
+                  1892,
+                  1910
+                ],
+                [
+                  1917,
+                  1923
+                ]
+              ]
+            }
+          ]
+        },
+        {
+          "left_pr_number": 43656,
+          "right_pr_number": 43842,
+          "code_similarity": 0.405,
+          "size_similarity": 0.5,
+          "file_overlap": 1.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.036,
+          "shared_filenames": [
+            "src/transformers/cli/serve.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/cli/serve.py",
+              "left_ranges": [
+                [
+                  11,
+                  18
+                ],
+                [
+                  30,
+                  36
+                ],
+                [
+                  315,
+                  323
+                ],
+                [
+                  665,
+                  671
+                ],
+                [
+                  931,
+                  937
+                ],
+                [
+                  1843,
+                  1849
+                ],
+                [
+                  1868,
+                  1874
+                ]
+              ],
+              "right_ranges": [
+                [
+                  54,
+                  61
+                ],
+                [
+                  587,
+                  593
+                ]
+              ]
+            }
+          ]
+        },
+        {
+          "left_pr_number": 43836,
+          "right_pr_number": 43842,
+          "code_similarity": 0.332,
+          "size_similarity": 0.088,
+          "file_overlap": 1.0,
+          "area_overlap": 0.033,
+          "patch_similarity": 0.017,
+          "shared_filenames": [
+            "src/transformers/cli/serve.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/cli/serve.py",
+              "left_ranges": [
+                [
+                  11,
+                  18
+                ],
+                [
+                  359,
+                  429
+                ],
+                [
+                  584,
+                  590
+                ],
+                [
+                  1892,
+                  1910
+                ],
+                [
+                  1917,
+                  1923
+                ]
+              ],
+              "right_ranges": [
+                [
+                  54,
+                  61
+                ],
+                [
+                  587,
+                  593
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
     {
       "cluster_id": "cluster-43240-3",
       "summary": "Cluster of 1 issues and 2 PRs centered on issue #43240.",
@@ -1422,76 +1620,6 @@
         }
       ]
     },
-    {
-      "cluster_id": "cluster-43824-3",
-      "summary": "Cluster of 1 issues and 2 PRs centered on issue #43824.",
-      "status": "open",
-      "confidence": 0.75,
-      "canonical_issue_number": 43824,
-      "canonical_pr_number": 43836,
-      "issue_numbers": [
-        43824
-      ],
-      "pr_numbers": [
-        43836,
-        43842
-      ],
-      "evidence_types": [
-        "closing_reference",
-        "shared_issue_target"
-      ],
-      "pr_comparisons": [
-        {
-          "left_pr_number": 43836,
-          "right_pr_number": 43842,
-          "code_similarity": 0.332,
-          "size_similarity": 0.088,
-          "file_overlap": 1.0,
-          "area_overlap": 0.033,
-          "patch_similarity": 0.017,
-          "shared_filenames": [
-            "src/transformers/cli/serve.py"
-          ],
-          "shared_file_areas": [
-            {
-              "filename": "src/transformers/cli/serve.py",
-              "left_ranges": [
-                [
-                  11,
-                  18
-                ],
-                [
-                  359,
-                  429
-                ],
-                [
-                  584,
-                  590
-                ],
-                [
-                  1892,
-                  1910
-                ],
-                [
-                  1917,
-                  1923
-                ]
-              ],
-              "right_ranges": [
-                [
-                  54,
-                  61
-                ],
-                [
-                  587,
-                  593
-                ]
-              ]
-            }
-          ]
-        }
-      ]
-    },
     {
       "cluster_id": "cluster-45081-3",
       "summary": "Cluster of 1 issues and 2 PRs centered on issue #45081.",
@@ -1989,6 +2117,16 @@
       "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-43656-4",
+      "canonical_pr_number": 43656,
+      "duplicate_pr_numbers": [
+        43836,
+        43842
+      ],
+      "target_issue_number": 43824,
+      "reason": "PRs in cluster-43656-4 are treated as duplicates because they converge on issue #43824 with closing_reference, shared_issue_target, soft_similarity evidence."
+    },
     {
       "cluster_id": "cluster-43698-3",
       "canonical_pr_number": 43779,
@@ -1998,15 +2136,6 @@
       "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-43824-3",
-      "canonical_pr_number": 43836,
-      "duplicate_pr_numbers": [
-        43842
-      ],
-      "target_issue_number": 43824,
-      "reason": "PRs in cluster-43824-3 are treated as duplicates because they converge on issue #43824 with closing_reference, shared_issue_target evidence."
-    },
     {
       "cluster_id": "cluster-43979-11",
       "canonical_pr_number": 44007,
diff --git a/analysis/current/analysis-report-hybrid.llm-reviews.json b/analysis/current/analysis-report-hybrid.llm-reviews.json
index 371054f3e49441dba2ed95e83b971b36500d5ba1..be396d9b183ba35eb01a841bfb88fcf6ff419e5b 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": "20260507T000029Z",
-  "generated_at": "2026-05-07T00:10:46Z",
+  "snapshot_id": "20260507T060024Z",
+  "generated_at": "2026-05-07T06:05:24Z",
   "model": "gpt-5.4-mini?service_tier=flex",
   "reviews": [
     {
@@ -13,8 +13,8 @@
       "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": [
@@ -84,8 +84,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -605,8 +605,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -948,8 +948,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -1312,8 +1312,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -1641,8 +1641,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -1970,8 +1970,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -2320,8 +2320,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -2654,8 +2654,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -2980,8 +2980,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -3313,8 +3313,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -3625,8 +3625,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -3958,8 +3958,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -4265,8 +4265,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -4590,8 +4590,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -4888,8 +4888,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -5195,8 +5195,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -5523,8 +5523,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -5824,8 +5824,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -6170,8 +6170,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -6489,8 +6489,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -6779,8 +6779,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -7121,8 +7121,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -7464,8 +7464,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -7780,8 +7780,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -8078,8 +8078,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -8383,8 +8383,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -8680,8 +8680,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -8999,8 +8999,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -9311,8 +9311,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -9624,8 +9624,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -9928,8 +9928,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -10232,8 +10232,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -10543,8 +10543,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -10869,8 +10869,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -11197,8 +11197,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -11544,8 +11544,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -11848,8 +11848,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -12160,8 +12160,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -12464,8 +12464,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -12832,8 +12832,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -13172,8 +13172,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -13515,8 +13515,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -13855,8 +13855,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -14237,8 +14237,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -14638,8 +14638,8 @@
       "review_unit_count": 45,
       "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-15354-388",
       "nodes": [
@@ -14824,8 +14824,8 @@
       "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": [
@@ -14887,8 +14887,8 @@
       "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-27088-4",
       "nodes": [
@@ -14987,8 +14987,8 @@
       "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-36817-14",
       "nodes": [
@@ -15251,8 +15251,8 @@
       "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-37934-6",
       "nodes": [
@@ -15367,8 +15367,8 @@
       "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": [
@@ -15430,8 +15430,8 @@
       "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": [
@@ -15493,8 +15493,8 @@
       "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": [
@@ -15556,8 +15556,8 @@
       "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": [
@@ -15659,8 +15659,8 @@
       "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-44371-5",
       "nodes": [
@@ -15781,8 +15781,8 @@
       "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": [
@@ -15901,8 +15901,8 @@
       "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-45672-2",
       "nodes": [
@@ -15964,8 +15964,8 @@
       "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": [
@@ -16027,8 +16027,8 @@
       "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": [
@@ -16130,12 +16130,12 @@
         "pull_request:43823|pull_request:45795",
         "pull_request:43251|pull_request:43254"
       ],
-      "prepared_review_unit_hash": "c7cc38b8e80ec24d8c74e644a8a1d7d994429d69ffea37e8ef88a3faba640932",
+      "prepared_review_unit_hash": "bfa0cdeac41d327395bac30a546e9bab851b3bf07c915fdb65ed2f1efd942968",
       "component_budget": {
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 19,
-        "serialized_chars": 35200,
+        "serialized_chars": 35198,
         "estimated_input_tokens": 8800,
         "estimated_eval_tokens": 17856
       },
@@ -16143,7 +16143,7 @@
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 19,
-        "serialized_chars": 35200,
+        "serialized_chars": 35198,
         "estimated_input_tokens": 8800,
         "estimated_eval_tokens": 17856
       },
@@ -16152,138 +16152,132 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "This cluster is mostly unrelated PRs. The only clear duplicate pair is the two loss_utils changes (43251 and 43254), which both add supported kwargs to fixed_cross_entropy / pass them through to nn.functional.cross_entropy. All other soft-edge candidates touch different models, docs, tests, or refactors and should not be merged together.",
-        "confidence": 0.98,
+        "summary": "The cluster is mostly heterogeneous and not a true duplicate set. The only clear duplicate pair is the two loss-utils PRs that both extend `fixed_cross_entropy` to pass supported `cross_entropy` kwargs; everything else is an unrelated model addition, docs change, refactor, or test-only PR.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 43251 is the best canonical representative for the duplicate loss-utils fix: it directly addresses issue 43240 and makes the core cross-entropy kwargs plumbing change cleanly. PR 43254 is the same underlying fix but is less clearly framed.",
+        "canonical_pr_reason": "If a canonical PR must be chosen, `pull_request:43251` is the best representative of the only real duplicate pair because it has the explicit issue target (`43240`) and a fuller implementation of the same `fixed_cross_entropy` kwargs fix.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 43251 is the best single representative because it is the most explicit and concrete version of the shared fixed_cross_entropy fix; PR 43254 is effectively the same change.",
+        "best_pr_reason": "`pull_request:43251` is the strongest PR in the cluster to anchor on: it directly addresses the tracked loss bug, targets the same code path as `43254`, and is the only clear duplicate-worthy change here.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:39895",
             "right": "pull_request:43424",
             "accept": false,
-            "reason": "VideoPrism model addition vs an Executorch dynamic-shape export test; different code paths and different goals."
+            "reason": "Unrelated changes: VideoPrism model integration vs. an ExecuTorch dynamic-shapes test. Same repository area, but not the same bug or change."
           },
           {
             "left": "pull_request:41855",
             "right": "pull_request:43424",
             "accept": false,
-            "reason": "Mistral tokenizer method additions vs Executorch export tests; unrelated functionality."
+            "reason": "Tokenizer method additions are unrelated to ExecuTorch export tests; no shared underlying defect."
           },
           {
             "left": "pull_request:41855",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Tokenizer API patch vs adding a new MobileLLM model; not the same change."
+            "reason": "Mistral tokenizer missing methods and adding MobileLLM are separate model/tokenizer workstreams."
           },
           {
             "left": "pull_request:42668",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Processor auto-loading robustness changes vs a new model implementation; different underlying work."
+            "reason": "Processor auto-loading robustness changes are unrelated to adding a new MobileLLM model."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Loss utility kwargs plumbing vs adding MobileLLM; unrelated."
+            "reason": "Loss-function kwargs fix is unrelated to introducing MobileLLM model files and tests."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "Loss utility change vs Mistral4 test fixes; not the same bug or change."
+            "reason": "`fixed_cross_entropy` kwargs support and Mistral4 test fixes are different code paths and different bugs."
           },
           {
             "left": "pull_request:43424",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "Executorch dynamic-shape test vs Mistral4 test fixes; unrelated test areas and code paths."
+            "reason": "ExecuTorch dynamic-shape export test and Mistral4 fixes do not address the same underlying change."
           },
           {
             "left": "pull_request:43424",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Executorch export testing vs new MobileLLM model code; unrelated."
+            "reason": "Dynamic-shape export testing is unrelated to MobileLLM model addition."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:45655",
             "accept": false,
-            "reason": "New MobileLLM model addition vs config-resolution order fix; different subsystems and different defects."
+            "reason": "New model support and a `cls.config` resolution-order fix are unrelated."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:43995",
             "accept": false,
-            "reason": "New model addition vs Falcon output-collection refactor; unrelated."
+            "reason": "MobileLLM model addition and Falcon output-interface refactor are distinct changes."
           },
           {
-            "left": "pull_request:43823",
+            "left": "pull_request:43995",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "New MobileLLM model vs docs/skill scaffolding for model creation; different change types."
-          },
-          {
-            "left": "pull_request:43823",
-            "right": "pull_request:45795",
-            "accept": false,
-            "reason": "New MobileLLM model vs docs for audio/video processors; unrelated docs work."
+            "reason": "Falcon refactor and the model-creation skill draft are not the same concrete bug/change."
           },
           {
             "left": "pull_request:43995",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "Falcon refactor vs Mistral4 tests; not the same concrete code-path problem."
+            "reason": "Falcon standardized-output refactor is unrelated to Mistral4 test fixes."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45170",
             "accept": false,
-            "reason": "Mistral4 test fixes vs layrnorm/layernorm rename across CLIP-like models; unrelated."
+            "reason": "Mistral4 test fixes and `layrnorm` renaming across CLIP-like models are separate changes."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45795",
             "accept": false,
-            "reason": "Mistral4 tests vs audio/video processor docs; unrelated."
+            "reason": "Mistral4 test fixes are unrelated to docs for audio/video processors."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45655",
             "accept": false,
-            "reason": "Mistral4 tests vs config-resolution fix; different issues."
+            "reason": "Mistral4 test work and `cls.config` resolution-order fix are unrelated code paths."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "Mistral4 tests vs model-creation skill docs; unrelated."
+            "reason": "A Mistral4 test fix and the model-creation skill draft do not represent the same underlying change."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "MobileLLM model code vs model-creation skill docs; not the same underlying bug/change."
+            "reason": "Adding MobileLLM and adding the model-creation skill are different concerns; one is product code, the other is tooling/docs."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:45795",
             "accept": false,
-            "reason": "MobileLLM model addition vs docs for audio/video processors; unrelated."
+            "reason": "MobileLLM model implementation and processor docs are unrelated."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:43254",
             "accept": true,
-            "reason": "Both PRs implement the same fixed_cross_entropy update: threading supported kwargs such as weight and label_smoothing into nn.functional.cross_entropy for issue 43240."
+            "reason": "Both PRs make the same `fixed_cross_entropy` change in `loss_utils.py` to pass supported kwargs (`weight`/`label_smoothing`) to `nn.functional.cross_entropy`; they fix the same concrete bug and could be merged into one."
           }
         ]
       },
       "evaluator_result": {
         "accept": true,
-        "feedback": "The summary is grounded in the packet: 43251 and 43254 are the same loss_utils bugfix targeting issue 43240, and the other pairwise rejections are conservative given the different files/subsystems. No overstatement detected, and the accepted pair looks mergeable as a single PR for the same fix."
+        "feedback": "The summary is grounded in the packet and stays conservative: it identifies the only clearly mergeable duplicate pair (43251/43254) based on the shared loss_utils diff and shared issue target, while treating the rest as heterogeneous. The soft-edge rejections are also cautious and consistent with the provided diffs. No overstatement detected."
       },
       "evaluator_used": true,
       "retried": false,
@@ -16310,12 +16304,12 @@
       "soft_pairs": [
         "pull_request:40546|pull_request:42781"
       ],
-      "prepared_review_unit_hash": "0a26441d7da20ff1a22b5e246afaf4d1d08de15a29df6102e060a31c57b48613",
+      "prepared_review_unit_hash": "14418f4f708943ab930c8f0f2cc658605cfd7402cfa03812e2cfaffbb3cb0c3c",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 7226,
+        "serialized_chars": 7225,
         "estimated_input_tokens": 1807,
         "estimated_eval_tokens": 3870
       },
@@ -16323,7 +16317,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 7226,
+        "serialized_chars": 7225,
         "estimated_input_tokens": 1807,
         "estimated_eval_tokens": 3870
       },
@@ -16332,18 +16326,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both items are open VibeVoice PRs, but they target different scopes: PR 40546 is the base VibeVoice implementation, while PR 42781 adds VibeVoice Realtime plus the acoustic tokenizer and related generation changes. They overlap on shared model plumbing, but they are not the same change.",
-        "confidence": 0.91,
+        "summary": "Related VibeVoice PRs, but not duplicates: 40546 is the initial VibeVoice implementation/docs integration, while 42781 expands into VibeVoice Realtime with the acoustic tokenizer and additional generation/model changes.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 40546 is the broader foundational VibeVoice implementation, covering the main model, configs, auto-mapping, generation, processing, docs, and pipeline integration.",
+        "canonical_pr_reason": "PR 40546 is the foundational VibeVoice integration and the clearest representative of the base feature work; it is narrower and more complete as the original model addition.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 40546 is the best representative of the cluster because it is the core VibeVoice implementation rather than a narrower follow-up variant.",
+        "best_pr_reason": "PR 40546 is the best single representative of the cluster because it covers the core VibeVoice import and registration work, while 42781 is a later follow-on for realtime/tokenizer functionality.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:40546",
             "right": "pull_request:42781",
             "accept": false,
-            "reason": "Shared filenames reflect a common VibeVoice codebase, but 42781 adds Realtime-specific functionality and an acoustic tokenizer; this is a related extension, not the same concrete bug/change."
+            "reason": "They share the VibeVoice namespace and some files, but the scope differs: 40546 adds the base model/docs/auto-mapping integration, whereas 42781 adds realtime support plus an acoustic tokenizer and related generation/path updates. These are adjacent feature PRs, not the same concrete change."
           }
         ]
       },
@@ -16373,7 +16367,7 @@
       "soft_pairs": [
         "pull_request:40755|pull_request:41224"
       ],
-      "prepared_review_unit_hash": "6047127cbbbfa17a86cb2ec105c51394bb9ce83797b9eb5ea531eb191c4ab24f",
+      "prepared_review_unit_hash": "9ac404caba0955a15011b005c361a639100d06174e09f08a930691f24a2e7d1b",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -16395,18 +16389,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These PRs are unrelated feature additions: one adds TimesFM covariate forecasting, the other adds DINOv3 image-classification support. They do not fix the same underlying issue or change the same code path.",
-        "confidence": 0.99,
+        "summary": "These two PRs are unrelated feature additions in different model families: TimesFM covariate forecasting vs DINOv3 image classification support. The soft similarity is only superficial.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No canonical PR: the PRs target different models, different APIs, and different files, so they are not duplicates of the same change.",
+        "canonical_pr_reason": "PR #40755 is the better cluster representative because it is the broader, more substantial functional change with core implementation across model code, utils, tests, and docs.",
         "best_issue_reason": null,
-        "best_pr_reason": "Neither PR is a true cluster representative because the pair is unrelated; if forced, 41224 is the more self-contained single-feature change, but it is not comparable to 40755.",
+        "best_pr_reason": "PR #40755 is the strongest standalone PR in the set: it introduces a new forecasting capability with implementation work rather than only model/classification enablement and example updates.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:40755",
             "right": "pull_request:41224",
             "accept": false,
-            "reason": "Different features and code paths: TimesFM covariate forecasting vs DINOv3 image classification. They are not plausibly mergeable into one PR as the same change."
+            "reason": "Different subsystems and problems: TimesFM time-series covariate forecasting vs DINOv3 image classification support. No shared filenames, targets, or code path, so they are not the same underlying change."
           }
         ]
       },
@@ -16501,12 +16495,12 @@
         "pull_request:41797|pull_request:45186",
         "pull_request:45186|pull_request:45630"
       ],
-      "prepared_review_unit_hash": "4979cbca2160bcb11557cc7f8d74335382ae5f0bc0525f8ed0f95ed758662e1b",
+      "prepared_review_unit_hash": "1abe801664b2f08cf6713e8f215b32b6e7a9eacac07616c57599545bd27281f2",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 2,
-        "serialized_chars": 10428,
+        "serialized_chars": 10426,
         "estimated_input_tokens": 2607,
         "estimated_eval_tokens": 5470
       },
@@ -16514,7 +16508,7 @@
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 2,
-        "serialized_chars": 10428,
+        "serialized_chars": 10426,
         "estimated_input_tokens": 2607,
         "estimated_eval_tokens": 5470
       },
@@ -16523,24 +16517,24 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "All three pull requests are separate new-model additions (DeepSeekOCR, Isaac, Kimi2-6). They share the same Transformers integration pattern and some autogenerated registry/docs files, but they target different model families and code paths, so this cluster does not look like a duplicate set.",
-        "confidence": 0.96,
+        "summary": "Three unrelated PRs adding support/docs for different models (DeepSeek OCR, Isaac, and Kimi2-6). They share template-like integration files, but each targets a distinct model family and change set, so they are not duplicates.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No canonical PR: the items are distinct feature additions for different models, not the same underlying change.",
+        "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "No single best PR for deduplication; each PR is a separate model integration with different implementation and test coverage.",
+        "best_pr_reason": null,
         "soft_edge_verdicts": [
           {
             "left": "pull_request:41797",
             "right": "pull_request:45186",
             "accept": false,
-            "reason": "Different model additions (DeepSeekOCR vs Isaac) with different configs, processors, and model files; shared registry/docs edits are generic boilerplate, not the same bug/change."
+            "reason": "Different model families and codepaths: DeepSeek OCR vs Isaac. Shared framework integration files are generic onboarding scaffolding, not the same fix/change."
           },
           {
             "left": "pull_request:45186",
             "right": "pull_request:45630",
             "accept": false,
-            "reason": "Different model additions (Isaac vs Kimi2-6) and different concrete code paths; they are not plausibly mergeable into one PR."
+            "reason": "Different model additions (Isaac vs Kimi2-6) with distinct filenames and model implementations. Similar auto-registration/docs updates are too broad to treat as one PR."
           }
         ]
       },
@@ -16570,7 +16564,7 @@
       "soft_pairs": [
         "pull_request:42975|pull_request:42976"
       ],
-      "prepared_review_unit_hash": "26b4c278104e2aa9b36ad8d25c3daee184fa6d9e71ec766a13e7d8a5f1d49d3e",
+      "prepared_review_unit_hash": "61cc767ab4c9064a76539b77498a9d44cb60ab97e9d72a2b357471f2df5633c2",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -16592,18 +16586,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These two PRs are both workflow dependency bumps, but they target different GitHub Actions and different code paths: one is a Node 24 compatibility sweep across several actions, the other is a docker/build-push-action upgrade. They are related in theme, not duplicates.",
-        "confidence": 0.92,
+        "summary": "Both pull requests are GitHub Actions maintenance upgrades, but they target different actions and different underlying changes: one is mostly checkout/cache/upload-artifact for Node 24 compatibility, while the other bumps docker/build-push-action. They overlap in workflow files but are not the same fix and should not be merged as duplicates.",
+        "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 42975 is the better representative of the cluster because it is the broader, more clearly scoped GitHub Actions compatibility update across multiple workflow actions.",
+        "canonical_pr_reason": "No canonical duplicate PR. If one must be chosen as the broader representative, 42975 is more umbrella-like because it updates several core workflow actions for Node 24 compatibility, whereas 42976 is a narrower docker/build-push-action version bump.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 42975 is the strongest standalone PR here: it has a concrete compatibility goal (Node 24) and touches a coherent set of workflow action upgrades.",
+        "best_pr_reason": "42975 is the better representative of the cluster because it spans more workflow actions and is centered on a specific compatibility goal. 42976 is related maintenance, but it is a separate change focused on docker/build-push-action.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:42975",
             "right": "pull_request:42976",
             "accept": false,
-            "reason": "Different underlying changes: 42975 updates checkout/cache/upload-artifact for Node 24 compatibility, while 42976 only bumps docker/build-push-action in docker workflows. They share workflow files but not the same code-path fix and would not naturally be merged as one PR."
+            "reason": "Both are workflow action version bumps, but they change different third-party actions and address different maintenance concerns. They do not appear to fix the same concrete code-path problem or be safely mergeable into one PR as duplicates."
           }
         ]
       },
@@ -16640,70 +16634,70 @@
         "pull_request:43448|pull_request:45490",
         "pull_request:43451|pull_request:45490"
       ],
-      "prepared_review_unit_hash": "4f6e41ebd6bf1372a149f451b6ceb979f821ac9a8eef98acfdef2b9d15036bd4",
+      "prepared_review_unit_hash": "8085010eb1f29e9977e50251b11ec2d580297cf79d80faf93ba2923ac66877b5",
       "component_budget": {
         "node_count": 4,
         "item_count": 4,
         "soft_pair_count": 6,
-        "serialized_chars": 14074,
-        "estimated_input_tokens": 3519,
-        "estimated_eval_tokens": 7294
+        "serialized_chars": 14079,
+        "estimated_input_tokens": 3520,
+        "estimated_eval_tokens": 7296
       },
       "budget": {
         "node_count": 4,
         "item_count": 4,
         "soft_pair_count": 6,
-        "serialized_chars": 14074,
-        "estimated_input_tokens": 3519,
-        "estimated_eval_tokens": 7294
+        "serialized_chars": 14079,
+        "estimated_input_tokens": 3520,
+        "estimated_eval_tokens": 7296
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "All four pull requests are separate model-addition PRs for different architectures (ViT NEPA, Molmo, Molmo2, CTSM). The overlap is only in shared registry/docs files, so this is not a duplicate cluster.",
+        "summary": "All four pull requests are separate model additions (ViT NEPA, Molmo, Molmo2, CTSM) and do not fix the same bug or implement the same change. The overlap is limited to common docs/auto-registration plumbing, which is not enough to treat them as duplicates.",
         "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No canonical PR: each item introduces a different model and code path, so they should not be merged as duplicates.",
+        "canonical_pr_reason": "No true duplicate PR exists in this cluster; if a representative must be chosen, 43451 is the most substantial and most discussed PR, but it addresses a distinct model add from the others.",
         "best_issue_reason": null,
-        "best_pr_reason": "No single PR is a best duplicate representative because the PRs are unrelated model additions rather than the same change.",
+        "best_pr_reason": "43451 is the strongest representative by scope and review activity, but it should not be considered a duplicate of the others because it adds Molmo2 rather than fixing the same concrete code path as any other PR.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:42978",
             "right": "pull_request:43448",
             "accept": false,
-            "reason": "ViT NEPA and Molmo are different models with separate implementations; shared auto/docs files are incidental registry updates."
+            "reason": "Different model additions: ViT NEPA vs Molmo. Shared auto/docs files are generic registration changes, not the same underlying feature or bug fix."
           },
           {
             "left": "pull_request:42978",
             "right": "pull_request:43451",
             "accept": false,
-            "reason": "ViT NEPA and Molmo2 are different model additions, not the same bug or change."
+            "reason": "Different model additions with different architectures and files. Both touch auto-mappings, but they implement separate model integrations, not one merged change."
           },
           {
             "left": "pull_request:42978",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "ViT NEPA and CTSM are unrelated model implementations; overlapping docs/auto-registration does not make them duplicates."
+            "reason": "ViT NEPA and CTSM are unrelated model integrations. The common toctree/auto-model edits are incidental and do not indicate the same change."
           },
           {
             "left": "pull_request:43448",
             "right": "pull_request:43451",
             "accept": false,
-            "reason": "Molmo and Molmo2 are distinct PRs for different architectures/code paths; they are not the same concrete change."
+            "reason": "Molmo and Molmo2 are related by naming, but they are distinct model families with different code paths, tests, and docs; not the same concrete change."
           },
           {
             "left": "pull_request:43448",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "Molmo and CTSM add different models and cannot plausibly be merged into one PR as a duplicate."
+            "reason": "Molmo and CTSM are unrelated additions. Shared repo plumbing does not make them duplicates."
           },
           {
             "left": "pull_request:43451",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "Molmo2 and CTSM are separate model additions with different files and behavior, not a shared fix or feature."
+            "reason": "Molmo2 and CTSM are separate model implementations. They overlap only in generic registration/docs updates, so they should not be merged as duplicates."
           }
         ]
       },
@@ -16736,7 +16730,7 @@
         "pull_request:43291|pull_request:44815",
         "pull_request:43665|pull_request:44815"
       ],
-      "prepared_review_unit_hash": "febb0e419cf3b89d250b9aa3903f0b75034f7495058a090abd0ff2171f6d8c2a",
+      "prepared_review_unit_hash": "91af6d51b249076d97dcbdb493b8216942ac481dfd657765de91f4ff07a0e228",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
@@ -16758,30 +16752,30 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are three unrelated pull requests despite superficial similarity. The Whisper PR fixes decode behavior/tests, the CLIP/ViT PR appears to be test-only noise, and the FP8/dequant PR changes a separate loading/quantization path.",
-        "confidence": 0.97,
+        "summary": "The three PRs are not duplicates: one is a Whisper tokenizer fix, one is a test-only change across CLIP/ViT, and one is a broader FP8/dequant loading fix. They touch different code paths and would not plausibly be merged into one PR.",
+        "confidence": 0.92,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No clear canonical duplicate PR exists in this set; if one must be chosen as the best representative, #44815 is the most substantive and self-contained change with a clear fix scope.",
+        "canonical_pr_reason": "PR #43291 is the clearest self-contained bug fix: it changes Whisper decoding logic in the tokenizer and updates related tests, with a focused concrete code path.",
         "best_issue_reason": null,
-        "best_pr_reason": "#44815 is the strongest representative of the cluster overall because it has the clearest technical scope and the most complete fix surface, while #43291 is narrow and #43665 is test-noise rather than a real bug fix.",
+        "best_pr_reason": "PR #43291 is the best representative of a real change because it addresses an actual runtime behavior bug rather than only modifying tests or spanning multiple unrelated subsystems.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43291",
             "right": "pull_request:43665",
             "accept": false,
-            "reason": "Different code paths and intents: Whisper tokenization/decode fix vs. CLIP/ViT test edits. The shared similarity is only superficial and they are not the same underlying change."
+            "reason": "Different scope and code paths: Whisper tokenization/decoding versus CLIP/ViT test edits. The second PR appears test-only and does not fix the same underlying bug."
           },
           {
             "left": "pull_request:43291",
             "right": "pull_request:44815",
             "accept": false,
-            "reason": "Whisper decode logic is unrelated to FP8 model loading/dequantization. They touch different subsystems and fix different problems."
+            "reason": "Unrelated fixes: Whisper decoding versus FP8/dequant/model-loading changes for Mistral/quantization. No shared concrete bug or mergeable common patch."
           },
           {
             "left": "pull_request:43665",
             "right": "pull_request:44815",
             "accept": false,
-            "reason": "Test-only CLIP/ViT changes are not the same as the FP8/dequant loading fix, and there is no concrete shared bug or mergeable combined PR here."
+            "reason": "Test-only edits for CLIP/ViT are not the same underlying change as FP8 dequant/model-loading work. They are too far apart to treat as one duplicate PR."
           }
         ]
       },
@@ -16811,7 +16805,7 @@
       "soft_pairs": [
         "pull_request:43488|pull_request:43532"
       ],
-      "prepared_review_unit_hash": "130cfd24620138d4e51ea76cf0d55d77fb63f662506f6c44c22484b69b5c17ad",
+      "prepared_review_unit_hash": "4bc701c8231180896bc58682d82274a2d08ec63dc543fdef4ea1fbc1b3bde073",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -16833,18 +16827,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Two open PRs are grouped only by a weak shared test filename; they address different concerns and are not duplicates. One is a small repo-bot/version/dependency tweak, the other rewires CI/workflow behavior and test utilities.",
-        "confidence": 0.94,
+        "summary": "These two pull requests are not duplicates: one is a small repo-bot/check related change around setup/dependency/test metadata, while the other is a broader CI/workflow and testing-utilities update. They only overlap on a Vit test file and do not appear to fix the same concrete problem.",
+        "confidence": 0.78,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "Use pull_request:43532 as the cluster anchor if needed: it is the more substantive PR, while pull_request:43488 is a tiny bot/check tweak and not the same change.",
+        "canonical_pr_reason": "PR #43488 is the more coherent representative of the cluster because it is a small, self-contained change centered on repository setup/versioning and a test hook, whereas #43532 spans multiple workflows and utilities with a much broader scope.",
         "best_issue_reason": null,
-        "best_pr_reason": "pull_request:43532 is the better representative because it contains the broader functional/workflow changes; pull_request:43488 is narrowly about repo-bot validation and dependency/version metadata.",
+        "best_pr_reason": "PR #43488 is the best single PR candidate in this set: it has a narrow, specific purpose and a consistent change surface. PR #43532 is broader and touches unrelated CI plumbing, so it is less suitable as the cluster representative.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43488",
             "right": "pull_request:43532",
             "accept": false,
-            "reason": "They do not fix the same concrete code-path problem. 43488 changes setup/version/dependency and a test, while 43532 changes GitHub workflow logic and related test utilities; shared test filename alone is not enough."
+            "reason": "Reject: they do not target the same underlying bug or change. The shared Vit test file is incidental; the rest of the diffs are about different areas (setup/versioning vs CI/workflow and notification plumbing)."
           }
         ]
       },
@@ -16874,7 +16868,7 @@
       "soft_pairs": [
         "pull_request:43612|pull_request:43613"
       ],
-      "prepared_review_unit_hash": "8df3dd62bb72ed77d79a16aabe3ccc2a0b2fd2fd6f4236cb84256acae859d653",
+      "prepared_review_unit_hash": "344e611ebb5d894530ad2ebb7db0bc1b3ed6bd86e2e773f03810d838f68a59bc",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -16896,18 +16890,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related but distinct PRs: one adds a Promptable Concept Segmentation pipeline, the other adds a Promptable Visual Segmentation pipeline. They share documentation and pipeline scaffolding, but target different tasks/model integrations and are not duplicates.",
-        "confidence": 0.96,
+        "summary": "These are two closely related but distinct pull requests: one adds a Promptable Concept Segmentation pipeline, the other adds a Promptable Visual Segmentation pipeline. They overlap in framework/docs/test scaffolding, but they implement different pipeline classes and different underlying model paths, so they are not duplicates.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No canonical duplicate PR: the two changes implement different segmentation tasks with different supported model paths and docs.",
+        "canonical_pr_reason": "No clear canonical PR: the two changes target different features and different concrete code paths, even though they share a lot of boilerplate and integration files.",
         "best_issue_reason": null,
-        "best_pr_reason": "Neither PR is a duplicate candidate; if one must be picked as the broader reference, PR #43613 is slightly more expansive, but it still solves a different task than #43612.",
+        "best_pr_reason": "PR #43613 is the broader of the two and touches more model integrations, so it is the better representative of the pair if one must be chosen.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43612",
             "right": "pull_request:43613",
             "accept": false,
-            "reason": "Different underlying features/tasks (concept vs visual segmentation). Shared pipeline/doc plumbing is not enough to treat them as the same fix or merge into one PR."
+            "reason": "Reject: they add different pipelines ('Promptable Concept Segmentation' vs 'Promptable Visual Segmentation') and different model integrations. Shared docs/test infrastructure is not enough to treat them as the same underlying change or safely merge them into one PR."
           }
         ]
       },
@@ -16937,7 +16931,7 @@
       "soft_pairs": [
         "pull_request:43651|pull_request:43663"
       ],
-      "prepared_review_unit_hash": "2d29e43997743b2ff3cdd83e140ab245de5782a3b356fa26619b27da7aba4114",
+      "prepared_review_unit_hash": "4d01db88fa26ff17763b27eeb298d5b71df3f4aa30324176401793332962ce6a",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -16959,18 +16953,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are two separate trainer extensibility PRs touching the same file but different code paths: one changes gradient-accumulation loss scaling, the other factors out signature-column filtering for custom trainers/PEFT models. They are related in area, not the same change.",
+        "summary": "Both items touch `src/transformers/trainer.py`, but they address different trainer extensions: one changes gradient-accumulation loss scaling, the other changes signature-column filtering for custom trainers. They do not look like duplicates or the same fix.",
         "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No duplicate PR here. If one must be named as representative, #43651 is slightly more central because it changes the core training_step loss-scaling behavior and introduces the new hook used there.",
+        "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "#43651 is the more representative PR for the cluster because it has the more direct behavioral change in training_step and clearer standalone impact; #43663 is a separate refactor/extensibility hook for column filtering.",
+        "best_pr_reason": "No single PR is a clear canonical representative here; the two PRs are about separate trainer customization points and should be triaged independently.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43651",
             "right": "pull_request:43663",
             "accept": false,
-            "reason": "Different enhancements in the same trainer module: #43651 adjusts gradient-accumulation loss scaling, while #43663 extracts signature-column filtering. They do not fix the same concrete bug and would not plausibly merge into one PR as duplicates."
+            "reason": "Different code paths and behaviors: `_loss_is_scaled_for_ga` changes loss normalization in `training_step`, while `_get_signature_columns` changes dataset column filtering. Shared filename alone is not enough to treat them as the same change."
           }
         ]
       },
@@ -17002,55 +16996,55 @@
         "pull_request:43656|pull_request:43842",
         "pull_request:43836|pull_request:43842"
       ],
-      "prepared_review_unit_hash": "704d5ada3347f0c31729560d99980faef400af6c8461a8d285f819474e0f63ee",
+      "prepared_review_unit_hash": "146523d779b736cfb6c0e35f0b38e4497b13e11b85d2516274ffe68789cadcf9",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 2,
-        "serialized_chars": 5992,
-        "estimated_input_tokens": 1498,
-        "estimated_eval_tokens": 3252
+        "serialized_chars": 5993,
+        "estimated_input_tokens": 1499,
+        "estimated_eval_tokens": 3254
       },
       "budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 2,
-        "serialized_chars": 5992,
-        "estimated_input_tokens": 1498,
-        "estimated_eval_tokens": 3252
+        "serialized_chars": 5993,
+        "estimated_input_tokens": 1499,
+        "estimated_eval_tokens": 3254
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "PR 43842 is the best canonical representative for the cluster because it directly fixes the `TypeAdapter` NameError in the CLI serve path with a minimal, focused change. PR 43836 looks like the same underlying bug fix in the same file and code path, but with more surrounding typing cleanup. PR 43656 is only loosely related via broader annotation/runtime-typing adjustments and is not strong enough to treat as the same concrete `TypeAdapter` fix.",
-        "confidence": 0.83,
+        "summary": "All three PRs are trying to fix the same import-time TypeAdapter annotation problem in `src/transformers/cli/serve.py`; two are near-direct variants of the same fix, and one broader annotation cleanup appears to cover the same underlying bug path.",
+        "confidence": 0.84,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 43842 is the most focused fix: it explicitly makes `TypeAdapter` safe in the `serve.py` validation path when pydantic is absent, with the smallest direct change set.",
+        "canonical_pr_reason": "PR #43842 is the most focused and explicit fix for the TypeAdapter NameError, with a narrow change set and a clear description of the runtime failure it addresses.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 43842 best represents the cluster because it targets the concrete `TypeAdapter` NameError directly and avoids the broader annotation churn present in the other PRs.",
+        "best_pr_reason": "PR #43842 is the best representative because it directly targets the TypeAdapter NameError, keeps the scope tight, and is less of a general typing cleanup than #43656 or the broader refactor in #43836.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43656",
             "right": "pull_request:43842",
-            "accept": false,
-            "reason": "43656 shows general annotation cleanup and `from __future__ import annotations`, but the diff preview does not show a `TypeAdapter`-specific runtime fix. That is not enough evidence that it is the same concrete bug as 43842."
+            "accept": true,
+            "reason": "Both target the same TypeAdapter NameError in `serve.py` and change annotation/runtime typing behavior in the same code path; they look like alternative fixes for the same bug."
           },
           {
             "left": "pull_request:43836",
             "right": "pull_request:43842",
             "accept": true,
-            "reason": "Both PRs address the same `serve.py` `TypeAdapter` import/runtime issue on the same code path, and 43836 appears to be a broader implementation of the same fix rather than a different bug."
+            "reason": "Both explicitly fix the TypeAdapter annotation failure in the CLI server module, and the diffs are compatible as variants of the same underlying import-time bug."
           }
         ]
       },
       "evaluator_result": {
-        "accept": false,
-        "feedback": "The writeup is mostly grounded for PR 43842 and PR 43836, but it overreaches on PR 43656. The packet only shows 43656 adding `from __future__ import annotations` and changing some `PreTrainedModel`/`ProcessorMixin` annotations; it does not show a `TypeAdapter`-specific change or enough evidence to conclude it is the same `TypeAdapter NameError` fix. So the 43656\u219443842 soft-edge acceptance is too strong, and the summary should not claim all three PRs are overlapping fixes for the same exact TypeAdapter issue. A more conservative cluster would keep 43842 and 43836 together, while treating 43656 as only possibly related via broader annotation/runtime typing cleanup."
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: all three PRs touch `src/transformers/cli/serve.py`, their titles and issue targets point to the same TypeAdapter/annotation NameError family, and the soft-edge pairs look like mergeable alternative fixes rather than unrelated changes. The verdicts are conservative enough."
       },
       "evaluator_used": true,
-      "retried": true,
+      "retried": false,
       "accepted_nontrivial_soft_edge": true,
       "error_kind": null,
       "error_message": null
@@ -17191,8 +17185,8 @@
         "pull_request:44071|pull_request:44085",
         "pull_request:44059|pull_request:44086",
         "pull_request:44085|pull_request:44086",
-        "pull_request:44098|pull_request:44101",
         "pull_request:44059|pull_request:44722",
+        "pull_request:44098|pull_request:44101",
         "pull_request:44018|pull_request:44071",
         "pull_request:44017|pull_request:44722",
         "pull_request:44017|pull_request:44161",
@@ -17212,8 +17206,8 @@
         "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:44056|pull_request:44161",
         "pull_request:43997|pull_request:44161",
         "pull_request:44074|pull_request:44161",
         "pull_request:43997|pull_request:44071",
@@ -17237,403 +17231,385 @@
         "pull_request:44044|pull_request:44154",
         "pull_request:44024|pull_request:44076"
       ],
-      "prepared_review_unit_hash": "5dde28b339b30d513ed869d1a04d81f0919c0e9700237ba27b1abc0adbeaf2ec",
+      "prepared_review_unit_hash": "23b2c7af0a19e20f1883cf03d70eeb623466c397e466faffed77c1431a3160d5",
       "component_budget": {
         "node_count": 28,
         "item_count": 28,
         "soft_pair_count": 77,
-        "serialized_chars": 84504,
-        "estimated_input_tokens": 21126,
-        "estimated_eval_tokens": 42508
+        "serialized_chars": 84498,
+        "estimated_input_tokens": 21125,
+        "estimated_eval_tokens": 42506
       },
       "budget": {
         "node_count": 28,
         "item_count": 28,
         "soft_pair_count": 72,
-        "serialized_chars": 76938,
-        "estimated_input_tokens": 19235,
-        "estimated_eval_tokens": 38726
+        "serialized_chars": 76931,
+        "estimated_input_tokens": 19233,
+        "estimated_eval_tokens": 38722
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "This cluster is mostly a wave of model-specific output-tracing refactors tied to the same tracking issue (#43979). Most pairs are not duplicates because they target different model implementations; the only clear duplicate-like overlaps are within GPT-J/CodeGen and GPT-Neo.",
-        "confidence": 0.89,
-        "canonical_issue_reason": "Issue #43979 is the shared tracker explicitly linked by many PRs and matches the common goal: standardized output-tracing/refactoring across model families.",
-        "canonical_pr_reason": "PR #44722 is the strongest representative of the true duplicate line: it targets the shared issue, covers the GPT-J/CodeGen code paths, and looks like a later consolidated revision of the same refactor.",
-        "best_issue_reason": "#43979 is the best global issue anchor because it is the explicit umbrella ticket for this refactoring effort across models.",
-        "best_pr_reason": "#44722 is the best global PR anchor because it is the broadest concrete implementation among the near-duplicate GPT-J/CodeGen refactors and most directly matches the shared initiative.",
+        "summary": "This cluster splits into two real duplicate/superset chains: GPT-J/CodeGen (44066 \u2194 44722, with 44085 as a narrower GPT-J slice) and GPT-Neo (44018 \u2194 44068). Everything else is the same output-tracing refactor theme, but across different model code paths, so it should stay separate.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "43979 is the shared tracking issue for the standardized output-tracing refactor work across the cluster.",
+        "canonical_pr_reason": "44722 is the best canonical PR for the GPT-J/CodeGen chain because it is the later, broader version of that same concrete refactor on the same code path; it should not be taken as covering the separate GPT-Neo chain.",
+        "best_issue_reason": "43979 is the central issue anchor for the refactor family in this cluster and the cleanest issue-level representative.",
+        "best_pr_reason": "44722 is the best single PR anchor for the cluster because it is the most complete later version of the largest concrete duplicate chain, but the GPT-Neo chain still needs its own representative PR.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43996",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "Different model families; same umbrella issue is not enough to treat them as the same change."
+            "reason": "Different model families; same output-tracing theme only."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "CVT/FNet refactor vs DeBERTa refactor: same theme, different concrete code paths."
+            "reason": "FNet/CVT vs DeBERTaV2; not the same code path."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44085",
             "accept": true,
-            "reason": "Both are GPT-J output-tracing refactors in the same file and could be merged into one PR."
+            "reason": "Same GPT-J output-tracing code path; 44085 is a narrower GPT-J slice of the broader 44066 refactor."
           },
           {
             "left": "pull_request:44007",
             "right": "pull_request:44072",
             "accept": false,
-            "reason": "RegNet and EfficientNet are separate model implementations, not one duplicate fix."
+            "reason": "ResNet/RT-DETR ResNet vs EfficientNet; different concrete implementations."
           },
           {
             "left": "pull_request:44072",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "EfficientNet and GPT-J/CodeGen are unrelated implementations despite both being output-tracing refactors."
+            "reason": "EfficientNet vs GPT-J/CodeGen; unrelated model code."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44072",
             "accept": false,
-            "reason": "GPT-J/CodeGen vs EfficientNet: different code paths and different model families."
+            "reason": "GPT-J/CodeGen vs EfficientNet; only a shared refactor pattern."
           },
           {
             "left": "pull_request:44013",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "MobileNetV2 and DeBERTa are unrelated concrete changes."
+            "reason": "MobileNetV2 and DeBERTaV2 are unrelated code paths."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-J/CodeGen refactor vs MGP-STR refactor; only the broad pattern matches."
+            "reason": "GPT-J/CodeGen vs MGP-STR; different model internals."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-J/CodeGen and MPT are different model implementations."
+            "reason": "GPT-J/CodeGen vs MPT; same theme, different implementation."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44068",
             "accept": true,
-            "reason": "Same GPT-Neo output-tracing refactor on the same model/file family."
+            "reason": "Same GPT-Neo output-tracing refactor on the same file/code path; 44068 looks like the fuller follow-up version."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44068",
             "accept": false,
-            "reason": "GPT-J/CodeGen and GPT-Neo are separate model-specific changes."
+            "reason": "GPT-J/CodeGen vs GPT-Neo; different models."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44066",
             "accept": false,
-            "reason": "Different models and different concrete code paths."
+            "reason": "GPT-Neo vs GPT-J/CodeGen; different model code."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "ResNet and GPT-J/CodeGen are unrelated model families."
+            "reason": "ResNet vs GPT-J/CodeGen; different code paths."
           },
           {
             "left": "pull_request:44086",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "MGP-STR and GPT-J/CodeGen are different implementations; same refactor pattern is too broad."
+            "reason": "MGP-STR vs GPT-J/CodeGen; unrelated models."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "ResNet refactor is not the same underlying change as GPT-J output tracing."
+            "reason": "ResNet vs GPT-J; only broad output-tracing similarity."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44019",
             "accept": false,
-            "reason": "GPT-Neo vs ResNet; same umbrella effort, not the same bug/change."
+            "reason": "GPT-Neo vs ResNet; different model families."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "ResNet and MPT are separate model-specific changes."
+            "reason": "ResNet vs MPT; unrelated code paths."
           },
           {
             "left": "pull_request:44068",
             "right": "pull_request:44116",
             "accept": false,
-            "reason": "GPT-Neo and Flaubert are unrelated code paths."
+            "reason": "GPT-Neo vs Flaubert; different models."
           },
           {
             "left": "pull_request:44068",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "GPT-Neo refactor is distinct from GPT-J/CodeGen refactor."
+            "reason": "GPT-Neo vs GPT-J/CodeGen; not the same refactor."
           },
           {
             "left": "pull_request:44071",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "MPT and GPT-J/CodeGen are different concrete implementations."
+            "reason": "MPT vs GPT-J/CodeGen; different model internals."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44116",
             "accept": false,
-            "reason": "GPT-Neo and Flaubert are separate model changes."
+            "reason": "GPT-Neo vs Flaubert; same pattern only."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-Neo vs MGP-STR; same broad initiative only."
+            "reason": "GPT-Neo vs MGP-STR; unrelated code paths."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "GPT-Neo and GPT-J are different model families."
+            "reason": "GPT-Neo vs GPT-J; different model code."
           },
           {
             "left": "pull_request:44068",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-Neo and MGP-STR are unrelated concrete changes."
+            "reason": "GPT-Neo vs MGP-STR; not a duplicate."
           },
           {
             "left": "pull_request:44071",
-            "right": "pull_request:44086",
-            "accept": false,
-            "reason": "MPT and MGP-STR are separate model implementations."
-          },
-          {
-            "left": "pull_request:43997",
-            "right": "pull_request:44073",
-            "accept": false,
-            "reason": "RegNet and VisualBert are different model families."
-          },
-          {
-            "left": "pull_request:44073",
-            "right": "pull_request:44101",
+            "right": "pull_request:44085",
             "accept": false,
-            "reason": "VisualBert and Flaubert are separate concrete changes."
+            "reason": "MPT vs GPT-J; different concrete changes."
           },
           {
-            "left": "pull_request:44056",
-            "right": "pull_request:44101",
+            "left": "pull_request:44059",
+            "right": "pull_request:44086",
             "accept": false,
-            "reason": "MPNet refactor vs Flaubert refactor; not the same underlying fix."
+            "reason": "GPT-2 vs MGP-STR; unrelated models."
           },
           {
-            "left": "pull_request:44056",
-            "right": "pull_request:44161",
+            "left": "pull_request:44085",
+            "right": "pull_request:44086",
             "accept": false,
-            "reason": "MPNet and LongT5 are unrelated implementations."
+            "reason": "GPT-J vs MGP-STR; only shared refactor style."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44017",
+            "left": "pull_request:44059",
+            "right": "pull_request:44722",
             "accept": false,
-            "reason": "RegNet and Segformer are separate model-specific refactors."
+            "reason": "GPT-2 vs GPT-J/CodeGen; different code paths."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44056",
+            "left": "pull_request:44098",
+            "right": "pull_request:44101",
             "accept": false,
-            "reason": "RegNet vs MPNet; different code paths."
+            "reason": "ViLT vs XLM; different model families."
           },
           {
-            "left": "pull_request:44071",
-            "right": "pull_request:44076",
+            "left": "pull_request:44018",
+            "right": "pull_request:44071",
             "accept": false,
-            "reason": "MPT and ImageGPT are unrelated models."
+            "reason": "GPT-Neo vs MPT; not the same bug/change."
           },
           {
-            "left": "pull_request:44074",
-            "right": "pull_request:44101",
+            "left": "pull_request:44017",
+            "right": "pull_request:44722",
             "accept": false,
-            "reason": "TextNet and Flaubert are different implementations."
+            "reason": "Segformer vs GPT-J/CodeGen; unrelated implementations."
           },
           {
-            "left": "pull_request:44056",
+            "left": "pull_request:44017",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "MPNet and LongT5 are separate model changes."
+            "reason": "Segformer vs LongT5; different code paths."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44018",
+            "left": "pull_request:44017",
+            "right": "pull_request:44073",
             "accept": false,
-            "reason": "RegNet and GPT-Neo are unrelated concrete changes."
+            "reason": "Segformer vs VisualBert; same pattern only."
           },
           {
-            "left": "pull_request:44010",
+            "left": "pull_request:44068",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "SqueezeBert and MPT are separate model families."
+            "reason": "GPT-Neo vs MPT; different models."
           },
           {
-            "left": "pull_request:44010",
-            "right": "pull_request:44073",
-            "accept": false,
-            "reason": "SqueezeBert and VisualBert are different code paths."
-          },
-          {
-            "left": "pull_request:44018",
-            "right": "pull_request:44129",
+            "left": "pull_request:44071",
+            "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-Neo and SpeechT5 are unrelated model implementations."
+            "reason": "MPT vs MGP-STR; unrelated refactors."
           },
           {
             "left": "pull_request:43997",
-            "right": "pull_request:44074",
+            "right": "pull_request:44073",
             "accept": false,
-            "reason": "RegNet and TextNet are separate refactors."
+            "reason": "RegNet vs VisualBert; different code paths."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44129",
+            "left": "pull_request:44056",
+            "right": "pull_request:44101",
             "accept": false,
-            "reason": "RegNet and SpeechT5 are unrelated."
+            "reason": "MPNet vs XLM; not the same underlying change."
           },
           {
-            "left": "pull_request:44059",
-            "right": "pull_request:44086",
+            "left": "pull_request:44073",
+            "right": "pull_request:44161",
             "accept": false,
-            "reason": "GPT-2 and MGP-STR are different model families."
+            "reason": "VisualBert vs LongT5; different implementations."
           },
           {
-            "left": "pull_request:44085",
-            "right": "pull_request:44086",
+            "left": "pull_request:44017",
+            "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-J and MGP-STR do not share the same concrete code path."
+            "reason": "Segformer vs MPT; no shared concrete change."
           },
           {
-            "left": "pull_request:44098",
+            "left": "pull_request:44073",
             "right": "pull_request:44101",
             "accept": false,
-            "reason": "ViLT and Flaubert/XLM are separate model-specific changes."
+            "reason": "VisualBert vs XLM; same broad refactor theme only."
           },
           {
             "left": "pull_request:44059",
-            "right": "pull_request:44722",
-            "accept": false,
-            "reason": "GPT-2 output tracing and GPT-J/CodeGen output tracing are distinct changes."
-          },
-          {
-            "left": "pull_request:44018",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-Neo and MPT are unrelated model implementations."
+            "reason": "GPT-2 vs MPT; different model internals."
           },
           {
-            "left": "pull_request:44017",
-            "right": "pull_request:44722",
+            "left": "pull_request:44059",
+            "right": "pull_request:44068",
             "accept": false,
-            "reason": "Segformer and GPT-J/CodeGen are unrelated concrete changes."
+            "reason": "GPT-2 vs GPT-Neo; not duplicates."
           },
           {
-            "left": "pull_request:44017",
-            "right": "pull_request:44161",
+            "left": "pull_request:44059",
+            "right": "pull_request:44085",
             "accept": false,
-            "reason": "Segformer and LongT5 are separate model refactors."
+            "reason": "GPT-2 vs GPT-J; different concrete code."
           },
           {
-            "left": "pull_request:44017",
-            "right": "pull_request:44073",
+            "left": "pull_request:44056",
+            "right": "pull_request:44017",
             "accept": false,
-            "reason": "Segformer and VisualBert are different model families."
+            "reason": "MPNet vs Segformer; unrelated model paths."
           },
           {
-            "left": "pull_request:44068",
+            "left": "pull_request:44010",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-Neo and MPT are unrelated concrete changes."
+            "reason": "SqueezeBERT vs MPT; different code paths."
           },
           {
-            "left": "pull_request:44071",
-            "right": "pull_request:44076",
+            "left": "pull_request:44010",
+            "right": "pull_request:44073",
             "accept": false,
-            "reason": "MPT and ImageGPT are different code paths."
+            "reason": "SqueezeBERT vs VisualBert; not the same change."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44017",
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
             "accept": false,
-            "reason": "RegNet and Segformer are not the same underlying bug/change."
+            "reason": "DeBERTaV2 vs SpeechT5; different model internals."
           },
           {
-            "left": "pull_request:44010",
-            "right": "pull_request:44071",
+            "left": "pull_request:44044",
+            "right": "pull_request:44154",
             "accept": false,
-            "reason": "SqueezeBert and MPT are separate changes."
+            "reason": "DeBERTaV2 vs VITS; unrelated model code."
           },
           {
-            "left": "pull_request:44010",
-            "right": "pull_request:44073",
+            "left": "pull_request:44024",
+            "right": "pull_request:44076",
             "accept": false,
-            "reason": "SqueezeBert and VisualBert are unrelated."
+            "reason": "FocalNet vs ImageGPT; same refactor pattern only."
           },
           {
-            "left": "pull_request:44018",
-            "right": "pull_request:44129",
-            "accept": false,
-            "reason": "GPT-Neo and SpeechT5 are different model-specific refactors."
+            "left": "pull_request:44066",
+            "right": "pull_request:44722",
+            "accept": true,
+            "reason": "Same GPT-J/CodeGen output-tracing refactor; 44722 is the broader later version."
           },
           {
-            "left": "pull_request:44074",
-            "right": "pull_request:44101",
+            "left": "pull_request:44085",
+            "right": "pull_request:44722",
+            "accept": true,
+            "reason": "Both target GPT-J output tracing on the same code path; they are mergeable as one PR."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44129",
             "accept": false,
-            "reason": "TextNet and Flaubert are separate concrete changes."
+            "reason": "GPT-J vs SpeechT5; different models."
           },
           {
-            "left": "pull_request:44056",
-            "right": "pull_request:44161",
+            "left": "pull_request:43996",
+            "right": "pull_request:44007",
             "accept": false,
-            "reason": "MPNet and LongT5 are unrelated implementations."
+            "reason": "FNet/CVT vs ResNet/RT-DETR ResNet; unrelated code paths."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44018",
+            "left": "pull_request:43996",
+            "right": "pull_request:44129",
             "accept": false,
-            "reason": "RegNet and GPT-Neo are different model families."
+            "reason": "FNet/CVT vs SpeechT5; same refactor theme only."
           },
           {
-            "left": "pull_request:44010",
-            "right": "pull_request:44044",
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
             "accept": false,
-            "reason": "SqueezeBert and DeBERTa are unrelated concrete changes."
+            "reason": "DeBERTaV2 vs SpeechT5; different concrete change."
           },
           {
             "left": "pull_request:44044",
-            "right": "pull_request:44074",
+            "right": "pull_request:44154",
             "accept": false,
-            "reason": "DeBERTa and TextNet are separate model-specific refactors."
+            "reason": "DeBERTaV2 vs VITS; unrelated model code."
           },
           {
             "left": "pull_request:44024",
             "right": "pull_request:44076",
             "accept": false,
-            "reason": "FocalNet and ImageGPT are different code paths; same theme only."
+            "reason": "FocalNet vs ImageGPT; not the same underlying bug/change."
           }
         ]
       },
       "evaluator_result": {
-        "accept": true,
-        "feedback": "The summary is grounded in the packet: #43979 is clearly the umbrella issue, and #44722 is a reasonable canonical PR anchor because it covers the GPT-J/CodeGen path and shares the issue target. The soft-edge verdicts are conservative, and the two accepted pairs (#44066/#44085 and #44018/#44068) are the only ones that clearly look mergeable as the same bugfix within the same model/file family."
+        "accept": false,
+        "feedback": "The cluster summary is mostly grounded, but it overstates uniqueness: the GPT-Neo pair (44018/44068) is also a high-confidence duplicate/superset chain, so 44722 is not the canonical PR for the 'only' strong duplicate chain. The GPT-J/CodeGen chain is a valid anchor, but the writeup should explicitly acknowledge the separate GPT-Neo duplicate chain and avoid implying the rest is only loose theme similarity."
       },
       "evaluator_used": true,
-      "retried": false,
+      "retried": true,
       "accepted_nontrivial_soft_edge": true,
       "error_kind": null,
       "error_message": null
@@ -17665,20 +17641,20 @@
         "pull_request:43997|pull_request:44010",
         "pull_request:44010|pull_request:44129"
       ],
-      "prepared_review_unit_hash": "db0e8c72f854fb4b9fda3436cf05724bb299dc92802d114f0c9a0c00619f26ed",
+      "prepared_review_unit_hash": "f895268fb71b9c3ea4408b6600e063db97e221bd328e3b3f643cd4f527ac5b13",
       "component_budget": {
         "node_count": 28,
         "item_count": 28,
         "soft_pair_count": 77,
-        "serialized_chars": 84504,
-        "estimated_input_tokens": 21126,
-        "estimated_eval_tokens": 42508
+        "serialized_chars": 84498,
+        "estimated_input_tokens": 21125,
+        "estimated_eval_tokens": 42506
       },
       "budget": {
         "node_count": 6,
         "item_count": 6,
         "soft_pair_count": 5,
-        "serialized_chars": 13066,
+        "serialized_chars": 13067,
         "estimated_input_tokens": 3267,
         "estimated_eval_tokens": 6790
       },
@@ -17687,42 +17663,42 @@
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "These PRs are all model-specific migrations to the same output-capture infrastructure, but they touch different code paths and should not be treated as duplicates. The cluster is a thematic similarity group, not one underlying bug/change.",
-        "confidence": 0.89,
+        "summary": "All items are model-specific PRs for the same broad refactor theme\u2014migrating to standardized output tracing/capture\u2014but they touch different architectures and code paths, so they should not be deduped into one change.",
+        "confidence": 0.95,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No issue artifact exists here; if a PR must anchor the cluster, 44129 is the strongest representative because it has an explicit issue target and a concrete standardized output-capture refactor.",
+        "canonical_pr_reason": "PR 44073 is a clean, representative example of the cluster: it clearly migrates one model to standardized output tracing with a focused diff and no extra cross-cutting changes.",
         "best_issue_reason": null,
-        "best_pr_reason": "44129 is the best representative PR: it is explicitly tied to an issue and shows a substantive, model-specific output-tracing refactor rather than just a title-level similarity.",
+        "best_pr_reason": "PR 44129 is the strongest standalone PR in the group because it is explicitly tied to an issue target and has a concrete, well-scoped output-tracing refactor for SpeechT5.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44073",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "Both refactor output tracing, but in different models (VisualBert vs TextNet). Same pattern, not the same underlying code-path change, and they would not naturally merge into one PR."
+            "reason": "Same refactor pattern, but different models and files (VisualBert vs TextNet). The implementations are not the same concrete code-path and would not plausibly merge into one PR."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44073",
             "accept": false,
-            "reason": "Both are standardized output-tracing refactors, but they affect different model implementations (MPNet vs VisualBert). Similar theme only; not a duplicate change."
+            "reason": "Both are standardized output-tracing migrations, but they target different architectures (MPNet vs VisualBert) with model-specific hooks and metadata. Similar theme is not enough for a duplicate."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "MPNet and TextNet are separate model code paths. The shared capture_outputs refactor is broad infrastructure work, not one concrete bug/change."
+            "reason": "These are separate model migrations with different forward/model wiring. They share a refactor style, not the same underlying bug or change."
           },
           {
             "left": "pull_request:43997",
             "right": "pull_request:44010",
             "accept": false,
-            "reason": "RegNet and SqueezeBert each migrate their own forward logic to the new tracing helpers, but they are distinct model-specific edits and not the same fix."
+            "reason": "RegNet and SqueezeBert are unrelated model implementations. The diffs reflect parallel modernization work, not one shared fix."
           },
           {
             "left": "pull_request:44010",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "Both concern output tracing, but SqueezeBert and SpeechT5 are unrelated model paths and the SpeechT5 PR also has an explicit issue target, indicating a separate change."
+            "reason": "SqueezeBert and SpeechT5 are different model code paths. Even though both migrate output tracing, they are independent PRs and should remain separate."
           }
         ]
       },
@@ -17802,274 +17778,274 @@
         "pull_request:44028|pull_request:44030",
         "pull_request:44029|pull_request:44030"
       ],
-      "prepared_review_unit_hash": "ca0407193162629b353a7db16557785254aa406f8da6ea00d5a75541440433f7",
+      "prepared_review_unit_hash": "167f2d1d2a5d5ce6c5d380f21e549561f2e2d73ab2c0036c0fec215e95711c68",
       "component_budget": {
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 40,
-        "serialized_chars": 40168,
-        "estimated_input_tokens": 10042,
-        "estimated_eval_tokens": 20340
+        "serialized_chars": 40176,
+        "estimated_input_tokens": 10044,
+        "estimated_eval_tokens": 20344
       },
       "budget": {
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 40,
-        "serialized_chars": 40168,
-        "estimated_input_tokens": 10042,
-        "estimated_eval_tokens": 20340
+        "serialized_chars": 40176,
+        "estimated_input_tokens": 10044,
+        "estimated_eval_tokens": 20344
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "This cluster is a set of similarly named but model-specific output-tracing refactors. They share a common theme, but they do not look like duplicates or one mergeable PR across the same concrete code path.",
-        "confidence": 0.86,
+        "summary": "These PRs are all variations of the same output-tracing/refactor pattern, but they are model-specific changes to different code paths, not duplicates of one underlying fix.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 44025 is the broadest representative in the set, touching two closely related depth models and reflecting the shared output-tracing refactor pattern.",
+        "canonical_pr_reason": "44025 is the broadest representative of the set: it applies the refactor across two related depth-anything model files and most clearly matches the cluster theme.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 44025 is the clearest exemplar of this cluster\u2019s theme because it is broader than the single-model refactors and still concrete enough to represent the group.",
+        "best_pr_reason": "44025 is the best anchor for the cluster because it is a multi-file refactor and cleanly exemplifies the common output-tracing migration pattern.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43998",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Same refactor wording, but timm_backbone and upernet are different model implementations and code paths."
+            "reason": "Same refactor theme, but different model families and forward paths; not the same concrete change."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44000",
             "accept": false,
-            "reason": "Both are output-tracing refactors, but they target unrelated model architectures and forward paths."
+            "reason": "Both are output-tracing refactors, but they target different model implementations and are not mergeable as one fix."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Different models and different forward logic; only the refactor theme is shared."
+            "reason": "Different model code paths; shared helper migration alone is too broad to treat as a duplicate."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Separate model families with distinct output handling; not the same underlying change."
+            "reason": "These touch different architectures and distinct output-tracing implementations, so they are not the same underlying bug/change."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different architectures and code paths; similarity is only at the refactor pattern level."
+            "reason": "Same broad subsystem, but separate model-specific refactors with different forward code paths."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "MobileNetV1 and UnivNet are unrelated models; these are not the same bug or fix."
+            "reason": "Similar tracing cleanup, but unrelated model paths and no shared concrete defect."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Shared tracing refactor theme, but different model-specific implementations."
+            "reason": "Both are refactors, but they affect different models and should not be merged as duplicates."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Depth models only; MobileNetV1 is a different architecture and forward path."
+            "reason": "Different model families and changes; only the refactor style matches."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44026",
             "accept": false,
-            "reason": "Different model families and code paths; no concrete duplicate change."
+            "reason": "No evidence of a shared bug or identical code path; just a similar output-tracing migration."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Different architectures and output semantics; not mergeable as one PR."
+            "reason": "Same broad cleanup pattern, but distinct wrapper/model implementations."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Vision-text dual encoder and UnivNet are unrelated; shared title language is not enough."
+            "reason": "Different models and forward logic; the similarity is only at the refactor pattern level."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different models and forward signatures; only the refactor motif overlaps."
+            "reason": "Not the same code path; they are separate model-specific output-tracing updates."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different code paths in different model families; not the same underlying change."
+            "reason": "These are separate model implementations with different refactor scopes, so they are not duplicates."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Different tasks and architectures; this is not one shared bug fix."
+            "reason": "Shared helper migration is not enough; these change different model code paths."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "UnivNet and UperNet are unrelated models, so these are separate refactors."
+            "reason": "Different models with no shared concrete fix, only the same refactor theme."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different model stacks and output-capturing changes; not a duplicate fix."
+            "reason": "Distinct architectures and output handling; not one underlying bug/change."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Different models and separate implementation details; only the high-level theme matches."
+            "reason": "Separate model-specific refactors, not the same code path."
           },
           {
             "left": "pull_request:44002",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "UperNet and RWKV have unrelated forward paths and output handling."
+            "reason": "Similar output-tracing cleanup, but unrelated models and implementation details."
           },
           {
             "left": "pull_request:44002",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Distinct model families and concrete changes; not the same PR in disguise."
+            "reason": "These are different model families and should not be collapsed into one PR."
           },
           {
             "left": "pull_request:44003",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Falcon/Mamba output tracing and SuperPoint tracing are different model-specific changes."
+            "reason": "Same general refactor family, but different architectures and forward paths."
           },
           {
             "left": "pull_request:44003",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Both are refactors, but CodeGen and Mamba fix different code paths and output shapes."
+            "reason": "Different model subsystems and changes; not a single concrete bug or fix."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Timm backbone and CodeGen are unrelated architectures; not one mergeable fix."
+            "reason": "Only the broad output-tracing theme overlaps; the code paths are unrelated."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "UnivNet and CodeGen have different forward contracts and output handling."
+            "reason": "These are separate model implementations, so the refactor similarity is too weak for duplication."
           },
           {
             "left": "pull_request:44002",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "UperNet and CodeGen are different models with unrelated output-tracing changes."
+            "reason": "Different target models and code paths; no same underlying change."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Vision-text dual encoder and CodeGen do not share the same concrete bug."
+            "reason": "Shared output-tracing language, but distinct model code and no common concrete bug."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44026",
             "accept": false,
-            "reason": "Depth Anything and Vision-Encoder-Decoder are separate model families with different forward paths."
+            "reason": "Both are wrapper-style refactors, but they are different model wrappers and not the same change."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Depth Anything and Speech-Encoder-Decoder are unrelated implementations despite the same refactor label."
+            "reason": "Different encoder-decoder wrapper implementations; similar refactor, not duplicate."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Depth Anything and SuperPoint are different vision models; not the same underlying change."
+            "reason": "Separate model code paths with only superficial refactor similarity."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different model families and output-capture logic; no single shared fix."
+            "reason": "Same tracing migration pattern, but distinct models and fixes."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Depth Anything and DPR are unrelated code paths, so these are not duplicates."
+            "reason": "Not the same underlying issue; just a similar output-tracing refactor."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Vision-Encoder-Decoder and RWKV are different architectures with separate output tracing behavior."
+            "reason": "Different model classes and code paths; too broad to merge as duplicates."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Encoder-decoder variants, but different modalities and implementations; not one concrete bug."
+            "reason": "Both are wrapper refactors, but they are separate encoder-decoder models and not one concrete fix."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different model families and output contracts; only the theme is shared."
+            "reason": "Distinct model implementations; shared theme only."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Vision-Encoder-Decoder and DPR are unrelated changes in different code paths."
+            "reason": "Different models and changes; no shared underlying bug or change."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Speech-Encoder-Decoder and SuperPoint are unrelated models; not mergeable as one fix."
+            "reason": "Separate model-specific refactors, not a single code-path problem."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different architectures and output semantics; no duplicate underlying issue."
+            "reason": "No evidence of the same concrete change beyond the refactor pattern."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Speech-Encoder-Decoder and DPR are separate model-specific refactors."
+            "reason": "Different model code paths; not mergeable as one duplicate fix."
           },
           {
             "left": "pull_request:44028",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "SuperPoint and RWKV are unrelated code paths; shared title language is too broad."
+            "reason": "Same broad output-tracing cleanup, but unrelated model implementations."
           },
           {
             "left": "pull_request:44028",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "SuperPoint and DPR are different models with different output-tracing details."
+            "reason": "Distinct models and code paths; not the same underlying bug/change."
           },
           {
             "left": "pull_request:44029",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "RWKV and DPR are separate model implementations; not the same concrete change."
+            "reason": "Different models; only the general refactor theme overlaps."
           }
         ]
       },
@@ -18102,52 +18078,52 @@
         "pull_request:44662|pull_request:45075",
         "pull_request:45075|pull_request:45401"
       ],
-      "prepared_review_unit_hash": "ab76a03da229f5d4ef0ba3fb14d87b3e4a5953db6d4c79c333a003c749143087",
+      "prepared_review_unit_hash": "9be26e50c650fa64cad6acfa55495980c3366f546bb4e2f1ebb1908bc24f67ee",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 10912,
-        "estimated_input_tokens": 2728,
-        "estimated_eval_tokens": 5712
+        "serialized_chars": 10914,
+        "estimated_input_tokens": 2729,
+        "estimated_eval_tokens": 5714
       },
       "budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 10912,
-        "estimated_input_tokens": 2728,
-        "estimated_eval_tokens": 5712
+        "serialized_chars": 10914,
+        "estimated_input_tokens": 2729,
+        "estimated_eval_tokens": 5714
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are three separate model-addition PRs for different architectures and code paths, so they are not duplicates of one another.",
+        "summary": "These are three separate model-integration pull requests for different models (PenguinVL, DeepSeek-OCR-2, and Voxtral TTS). They share Transformers boilerplate and Auto/Docs wiring, but they do not target the same code-path problem or the same change.",
         "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": null,
+        "canonical_pr_reason": "PR 45075 is the best representative because it is a mature, non-draft model integration with substantial implementation coverage and the highest review activity among the three.",
         "best_issue_reason": null,
-        "best_pr_reason": null,
+        "best_pr_reason": "PR 45075 is the strongest overall representative of the cluster: it is a full-featured integration, not a draft, and has the most review activity.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44662",
             "right": "pull_request:45401",
             "accept": false,
-            "reason": "Different models and implementations: PenguinVL vs Voxtral TTS. Shared auto-registration/docs are generic integration scaffolding, not the same underlying change."
+            "reason": "Different model integrations for unrelated architectures/tasks (PenguinVL vs Voxtral TTS); shared docs/auto-registration files are just boilerplate, not evidence of the same fix."
           },
           {
             "left": "pull_request:44662",
             "right": "pull_request:45075",
             "accept": false,
-            "reason": "Different model families and file sets: PenguinVL vs DeepSeek-OCR-2. Both add model support, but they do not fix the same bug or change the same code path."
+            "reason": "PenguinVL and DeepSeek-OCR-2 are distinct model additions with different model code, configs, and converters; the overlap is only in common Transformers scaffolding."
           },
           {
             "left": "pull_request:45075",
             "right": "pull_request:45401",
             "accept": false,
-            "reason": "Different tasks and subsystems: OCR model support vs TTS support. The overlap is only in common Transformers registration files, which is not enough to merge them as duplicates."
+            "reason": "DeepSeek-OCR-2 and Voxtral TTS are unrelated model integrations; they touch similar registry/docs areas but do not fix or add the same underlying functionality."
           }
         ]
       },
@@ -18177,12 +18153,12 @@
       "soft_pairs": [
         "pull_request:44979|pull_request:45363"
       ],
-      "prepared_review_unit_hash": "317bdceca16ff6f5fca07a85279edfdc6ff99bc277859ea439dab9c411ca5bb0",
+      "prepared_review_unit_hash": "deb61c381b92e4ceba5a3960f9949310d19ff5b52dbac4719b19dced7343dc45",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4407,
+        "serialized_chars": 4408,
         "estimated_input_tokens": 1102,
         "estimated_eval_tokens": 2460
       },
@@ -18190,7 +18166,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4407,
+        "serialized_chars": 4408,
         "estimated_input_tokens": 1102,
         "estimated_eval_tokens": 2460
       },
@@ -18199,18 +18175,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs are about model/module fusion, but they implement different layers of the feature: #44979 adds a generic `Module Fusion API`, while #45363 adds `KernelConfig`-driven n-to-1 kernel fusion and related integration points. They are related but not the same underlying change, so this should not be treated as a duplicate pair.",
-        "confidence": 0.89,
+        "summary": "These are related feature-area PRs about fusion, but they implement different abstractions and code paths: one adds a generic Module Fusion API, the other adds n-to-1 kernel fusion via KernelConfig and model-loading integration. They are not duplicates.",
+        "confidence": 0.92,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "#45363 is the more concrete, user-facing fusion implementation: it wires `KernelConfig` into modeling/integrations and includes an example, whereas #44979 is an API foundation for module fusion.",
+        "canonical_pr_reason": "PR 45363 is the broader and more concrete implementation of kernel-backed fusion, touching the loader/integration path and KernelConfig plumbing rather than just introducing a standalone API scaffold.",
         "best_issue_reason": null,
-        "best_pr_reason": "#45363 is the best representative of the cluster because it captures the actual kernel-fusion feature and code-path integration; #44979 is a broader precursor/API scaffold rather than the same change.",
+        "best_pr_reason": "PR 45363 is the best representative of the cluster because it implements the end-to-end user-facing fusion behavior, while PR 44979 is a separate API foundation.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44979",
             "right": "pull_request:45363",
             "accept": false,
-            "reason": "Related fusion work, but different concrete changes: #44979 defines a generic module-fusion API, while #45363 implements `KernelConfig`-based kernel fusion integration. They do not appear mergeable as one PR fixing the same code-path problem."
+            "reason": "They target different layers: 44979 adds a new generic Module Fusion API and tests, while 45363 wires n-to-1 kernel fusion through KernelConfig, integrations, and model loading. Related theme, but not the same underlying change or code-path problem."
           }
         ]
       },
@@ -18240,7 +18216,7 @@
       "soft_pairs": [
         "pull_request:45086|pull_request:45317"
       ],
-      "prepared_review_unit_hash": "8345cccf00045c70f9581f3ff52c7530368a10f7490388b5eb722d4155636aa7",
+      "prepared_review_unit_hash": "dc3a2b77fef08234750a5a94b21b4fff98b4ab68146220ab473401b64d8d9a0a",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -18262,18 +18238,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs fix the same Mistral regex AttributeError in the same code path and target the same issue; 45317 is the more complete version with tests, so it is the better canonical PR.",
+        "summary": "Both pull requests patch the same AttributeError in `_patch_mistral_regex` by switching from `tokenizer.backend_tokenizer.pre_tokenizer` to `tokenizer.pre_tokenizer` in the same file; the later PR adds tests and is the more complete representative.",
         "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "Pick PR 45317 as canonical: it makes the same fix as 45086 but also includes test coverage, making it the more complete representative of this bugfix.",
+        "canonical_pr_reason": "PR #45317 is the better canonical PR because it fixes the same concrete bug as #45086 but also includes test coverage and a clearer title describing the exact failing path (`fix_mistral_regex=True`).",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45317 is the best representative because it addresses the same concrete AttributeError in _patch_mistral_regex and includes validation/tests, whereas 45086 is the smaller preliminary patch.",
+        "best_pr_reason": "PR #45317 is the best overall PR for the cluster: same code change, same issue target, and it adds regression tests, making it the most complete and merge-ready representation.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45086",
             "right": "pull_request:45317",
             "accept": true,
-            "reason": "Accept: same underlying AttributeError fix in src/transformers/tokenization_utils_tokenizers.py for _patch_mistral_regex, same issue target, and 45317 is an expanded version with tests rather than a different change."
+            "reason": "Accept: both PRs address the same AttributeError in `_patch_mistral_regex` with the same line-level fix in `tokenization_utils_tokenizers.py`; #45317 is essentially the same patch plus tests, so they are duplicate-change candidates."
           }
         ]
       },
@@ -18303,7 +18279,7 @@
       "soft_pairs": [
         "pull_request:45115|pull_request:45569"
       ],
-      "prepared_review_unit_hash": "42d693cbf17f9c3824b94095ad3fd85f9babcf6b14070454a6659aec6378dd63",
+      "prepared_review_unit_hash": "41070347c6f8aadb9f4bec1ca58fd65b9742ff0652482c83fe08540ff48c1c8c",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -18325,18 +18301,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These two PRs are related to Nemotron-H model support, but they do not look like duplicates: one is a narrow refactor/conversion-mapping update for inheriting GraniteMoeHybrid-compatible structure, while the other splits Nemotron H into dense/sparse variants and adds docs, auto-mapping, and tests. They touch some of the same core files, but address different concrete changes.",
-        "confidence": 0.89,
+        "summary": "These two PRs are related to Nemotron H, but they are not the same underlying change: one is a narrow conversion/refactor PR, while the other is a broader architecture split introducing dedicated dense/sparse model variants and docs/tests.",
+        "confidence": 0.9,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #45569 is the broader, more complete implementation of Nemotron-H model support: it adds the new dense/sparse model packages, updates auto mappings and docs, and includes tests. PR #45115 is a narrower internal refactor around inheritance/conversion mapping.",
+        "canonical_pr_reason": "PR #45569 is the better representative for this cluster because it is the broader, more complete Nemotron H change set (new dense/sparse variants, auto-mapping, docs, and tests), whereas #45115 is a narrower conversion-mapping/refactor PR.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45569 is the better representative of the cluster because it covers the main end-to-end productization work, whereas #45115 is a smaller structural refactor focused on checkpoint conversion and model inheritance.",
+        "best_pr_reason": "PR #45569 is the strongest standalone PR here due to its broader scope and clearer user-facing impact; #45115 looks like a smaller preparatory/internal refactor rather than the main architecture change.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45115",
             "right": "pull_request:45569",
             "accept": false,
-            "reason": "Shared Nemotron-H files and subsystem, but the changes are not the same underlying fix: #45115 is a conversion/inheritance refactor, while #45569 introduces dense/sparse model variants plus docs and tests. They would not plausibly be merged as one PR."
+            "reason": "Both touch Nemotron H files, but they target different changes: #45115 adds checkpoint conversion/key renaming for GraniteMoeHybrid compatibility, while #45569 restructures Nemotron H into dedicated dense/sparse model variants with docs and tests. They are not mergeable as one duplicate PR."
           }
         ]
       },
@@ -18366,12 +18342,12 @@
       "soft_pairs": [
         "pull_request:45332|pull_request:45333"
       ],
-      "prepared_review_unit_hash": "f006308c692f9a3220babab1d3479b8adcd7bd21fa1c38006e9e761885fa200d",
+      "prepared_review_unit_hash": "c55cac4f5bef2da99d6cb51bcf00b279d623ffce9bd542515a6d05d318d9793d",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5398,
+        "serialized_chars": 5397,
         "estimated_input_tokens": 1350,
         "estimated_eval_tokens": 2956
       },
@@ -18379,7 +18355,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5398,
+        "serialized_chars": 5397,
         "estimated_input_tokens": 1350,
         "estimated_eval_tokens": 2956
       },
@@ -18388,18 +18364,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both pull requests are about heterogeneous-layer support in Transformers, but they are not duplicates: one is a config-plumbing PR and the other is a broader modeling/cache implementation PR. They look sequential/complementary rather than the same mergeable change.",
-        "confidence": 0.87,
+        "summary": "Two related but distinct PRs: one adds heterogeneous per-layer config support, the other extends that foundation into broader heterogeneous model/runtime support. They overlap in shared config utilities but are not the same concrete change.",
+        "confidence": 0.84,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #45332 is the better canonical representative because it is the broader umbrella change: it adds heterogeneous model support across cache/modeling paths, whereas #45333 is narrower config-layer plumbing that appears to enable or support the larger implementation.",
+        "canonical_pr_reason": "PR #45332 is the broader, more representative change: it includes heterogeneous config handling plus model/cache/masking/runtime integration and tests, making it the best canonical PR for the cluster.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45332 is the strongest standalone representative of the cluster since it covers the actual runtime code paths and tests for heterogeneous models, not just config serialization/definition.",
+        "best_pr_reason": "PR #45332 is the best representative because it covers the larger end-to-end heterogeneous model support rather than only the config layer plumbing.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45332",
             "right": "pull_request:45333",
             "accept": false,
-            "reason": "Related but distinct scope: #45333 adds per-layer config support in configuration_utils, while #45332 implements heterogeneous modeling/cache behavior across several runtime files. They share a subsystem, but not the same concrete bug/change and would not plausibly be merged as one duplicate PR."
+            "reason": "They share heterogeneous config utilities, but #45333 is only per-layer configuration while #45332 adds broader model/runtime support. This is related foundation work, not the same fix/change that should be merged into one PR."
           }
         ]
       },
@@ -18429,12 +18405,12 @@
       "soft_pairs": [
         "pull_request:45415|pull_request:45425"
       ],
-      "prepared_review_unit_hash": "8f20fc05b870ae8b7eaecb79db4f75881f6e2bbdc20527f1d9d9d771d452d5f0",
+      "prepared_review_unit_hash": "7721d2b2c5d12f2f10ffbdae79837ba5b2183949a7ccb61f9b3dec78324f16e4",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5514,
+        "serialized_chars": 5513,
         "estimated_input_tokens": 1379,
         "estimated_eval_tokens": 3014
       },
@@ -18442,7 +18418,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5514,
+        "serialized_chars": 5513,
         "estimated_input_tokens": 1379,
         "estimated_eval_tokens": 3014
       },
@@ -18451,18 +18427,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "The two PRs are both typing-related, but they are not the same change: #45415 is a broad type-checking sweep across many modules, while #45425 is a focused typing-compatibility refactor centered on `_typing.py` and a few related integrations. I would not merge them as duplicates.",
-        "confidence": 0.91,
+        "summary": "Both PRs are typing-related, but they are not the same change: #45415 is a broad type-checking sweep across many `src/transformers/*.py` files, while #45425 adds specific typing protocol/alias support for a smaller set of code paths. Shared filenames are incidental, not evidence of the same fix.",
+        "confidence": 0.89,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No clear canonical PR for duplication: #45415 is a broad repository-wide type-checking pass, whereas #45425 introduces specific protocol/alias definitions to unblock typing in a smaller set of code paths.",
+        "canonical_pr_reason": "#45425 is the more focused and concrete typing change; it introduces specific protocol/alias definitions to satisfy a narrower set of typing needs, whereas #45415 is a wide cleanup across many files.",
         "best_issue_reason": null,
-        "best_pr_reason": "#45425 is the better representative PR if one must be chosen: it is more concrete and self-contained, while #45415 is a much broader typing sweep with many unrelated file edits.",
+        "best_pr_reason": "#45425 is the best representative for the cluster because it is a self-contained typing improvement with a clear purpose and smaller scope.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45415",
             "right": "pull_request:45425",
             "accept": false,
-            "reason": "Different underlying changes. #45415 is a broad enforcement/configuration pass for type checking across many files; #45425 adds specific typing protocol/alias support and related code adjustments. Same subsystem, but not the same concrete fix."
+            "reason": "They are related only at a high level (typing work) and touch some overlapping files, but they address different problems: a broad type-checking pass versus targeted typing protocol/alias additions. They would not plausibly be merged into one PR as the same concrete change."
           }
         ]
       },
@@ -18492,7 +18468,7 @@
       "soft_pairs": [
         "pull_request:45487|pull_request:45615"
       ],
-      "prepared_review_unit_hash": "84c87dbf6f7c79a39efcdcbf3534a8acfaadf8410c96b747c3d921b7d07d52d0",
+      "prepared_review_unit_hash": "4d452c0889d073b90531568d25de57b9ddccd272cb56d24a2d268f234eadbf09",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -18514,18 +18490,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are unrelated pull requests: one fixes a device placement bug in model parallel code for AltCLIP/ChineseCLIP, while the other only updates Qianfan OCR XPU test expectations. They do not describe the same underlying change and should not be merged as duplicates.",
-        "confidence": 0.97,
+        "summary": "These are unrelated PRs: one fixes a model-parallel device placement bug in AltCLIP/ChineseCLIP code, while the other only adds XPU test expectations for Qianfan OCR. They should not be clustered as duplicates.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45487 is the stronger canonical PR because it makes a concrete runtime fix in model code across multiple affected models, with matching test coverage.",
+        "canonical_pr_reason": "PR 45487 is the substantive code fix, addressing a concrete device mismatch in model-parallel forward logic across AltCLIP/ChineseCLIP and related models.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45487 is the best representative of the cluster since it addresses a substantive bug in the implementation rather than just adjusting expected test outputs.",
+        "best_pr_reason": "PR 45487 is the best representative because it changes runtime model code to fix a real bug; PR 45615 is only updating expected test outputs for another model.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45487",
             "right": "pull_request:45615",
             "accept": false,
-            "reason": "Different problems: 45487 fixes token_type_ids device handling in model parallel execution, while 45615 only adds XPU-specific expected outputs for Qianfan OCR tests. They are not the same code-path bug or change."
+            "reason": "Different models and different change types: 45487 fixes a device-placement bug in model code, while 45615 only updates XPU test expectations for Qianfan OCR. They do not appear to address the same underlying bug or change."
           }
         ]
       },
@@ -18555,12 +18531,12 @@
       "soft_pairs": [
         "pull_request:45548|pull_request:45662"
       ],
-      "prepared_review_unit_hash": "b8d5a3fa39092c019b4edb83a1b7dcaa1a5c4f4885579a7a801d218124fc60a6",
+      "prepared_review_unit_hash": "49fbc7772e9f05e2978776cd59a1c60531632be67281b39adf9aba42154a01ed",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4507,
+        "serialized_chars": 4508,
         "estimated_input_tokens": 1127,
         "estimated_eval_tokens": 2510
       },
@@ -18568,7 +18544,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4507,
+        "serialized_chars": 4508,
         "estimated_input_tokens": 1127,
         "estimated_eval_tokens": 2510
       },
@@ -18577,18 +18553,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs are about expert parallelism, but they fix different bugs: #45548 targets config serialization and an `has_ep` helper for accelerate/ZeRO-3 loading, while #45662 fixes expert parameters being overwritten during distributed/FSDP2 loading. Shared subsystem only, not the same underlying change.",
-        "confidence": 0.92,
+        "summary": "The two pull requests are related to expert parallelism, but they target different failure modes and code paths: one focuses on EP/ZeRO-3 loading and config serialization, while the other fixes experts being overwritten by rank-0 broadcast during distributed loading. They should not be treated as duplicates.",
+        "confidence": 0.91,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "#45662 is the broader, more central EP-loading fix and touches the actual parameter-loading path; #45548 is narrower and mainly addresses config cleanup plus an EP flag helper.",
+        "canonical_pr_reason": "PR #45662 is the better canonical representative because it targets a concrete distributed-loading bug with a clear failure mode (experts being overwritten) and a focused fix in the loading path; PR #45548 is narrower and addresses EP/ZeRO-3 launch/serialization behavior instead.",
         "best_issue_reason": null,
-        "best_pr_reason": "#45662 best represents the cluster because it addresses the concrete distributed loading failure mode with multiple code-path changes; #45548 is a smaller adjacent fix, not the core bug here.",
+        "best_pr_reason": "PR #45662 is the strongest standalone candidate: it fixes a specific correctness issue in distributed model loading and has the most direct, concrete bug description. PR #45548 is useful but more ancillary, centered on launch compatibility and config handling.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45548",
             "right": "pull_request:45662",
             "accept": false,
-            "reason": "Both are EP-related, but #45548 is about config serialization/EP detection for accelerate+ZeRO-3, while #45662 fixes expert parameters being overwritten during distributed loading. Different bug, different patch shape, not mergeable into one PR."
+            "reason": "Both involve expert parallelism, but they fix different bugs: #45548 adds EP config/compatibility handling for ZeRO-3 loading, while #45662 prevents experts from being overwritten during rank-0 broadcast in the loading path. Not the same underlying change."
           }
         ]
       },
@@ -18618,12 +18594,12 @@
       "soft_pairs": [
         "pull_request:45639|pull_request:45645"
       ],
-      "prepared_review_unit_hash": "f1f9d6b6f493a88c895daa1df649c57d2e47a240743ef7c8fa89dec35d73d019",
+      "prepared_review_unit_hash": "719c76c451bec8aae35576e448d0eea85ef44d6884da0d83fcf90965be09f395",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4590,
+        "serialized_chars": 4589,
         "estimated_input_tokens": 1148,
         "estimated_eval_tokens": 2552
       },
@@ -18631,7 +18607,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4590,
+        "serialized_chars": 4589,
         "estimated_input_tokens": 1148,
         "estimated_eval_tokens": 2552
       },
@@ -18640,18 +18616,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs target the same xdist-related captured_info logging problem, with PR #45645 appearing to subsume the narrower fix in #45639 and add CI log-handling updates.",
-        "confidence": 0.88,
+        "summary": "Both PRs address the same xdist-related captured_info logging problem in testing_utils; PR 45645 is the broader, more complete version because it also updates CI handling for multiple captured_info files.",
+        "confidence": 0.92,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "#45645 is the better canonical PR because it covers the same core xdist-safe captured_info fix as #45639 and additionally updates CI/report collection paths to preserve and aggregate the logs.",
+        "canonical_pr_reason": "PR 45645 covers the same core xdist-safe captured_info fix as 45639 and additionally adjusts the CI workflow to consume the new per-worker artifact naming, making it the more complete implementation.",
         "best_issue_reason": null,
-        "best_pr_reason": "#45645 is the strongest representative of the cluster since it contains the underlying fix plus the surrounding workflow changes needed to keep captured_info artifacts visible in CI.",
+        "best_pr_reason": "PR 45645 is the best representative because it includes the core code-path fix plus the downstream CI/log collection changes needed to use it reliably.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45639",
             "right": "pull_request:45645",
             "accept": true,
-            "reason": "Same underlying xdist collision bug for captured_info artifacts; #45645 looks like a superset of #45639 rather than a different change."
+            "reason": "Same underlying captured_info xdist collision fix; 45645 is a superset that includes the testing_utils change and the related CI workflow update."
           }
         ]
       },
diff --git a/analysis/current/manifest.json b/analysis/current/manifest.json
index 632e60cedb0380b648f37307681433dad26550f7..8ad67098800db978d3c50871519af72d4813bc47 100644
--- a/analysis/current/manifest.json
+++ b/analysis/current/manifest.json
@@ -1,8 +1,8 @@
 {
-  "analysis_id": "hybrid-model-20260507t000029z",
+  "analysis_id": "hybrid-model-20260507t060024z",
   "archived_artifacts": {
-    "hybrid": "snapshots/20260507T000029Z/analysis-runs/hybrid-model-20260507t000029z/analysis-report-hybrid.json",
-    "hybrid_reviews": "snapshots/20260507T000029Z/analysis-runs/hybrid-model-20260507t000029z/analysis-report-hybrid.llm-reviews.json"
+    "hybrid": "snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.json",
+    "hybrid_reviews": "snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.llm-reviews.json"
   },
   "artifacts": {
     "hybrid": "analysis/current/analysis-report-hybrid.json",
@@ -10,9 +10,9 @@
   },
   "channel": "canonical",
   "model": null,
-  "published_at": "2026-05-07T00:10:49Z",
+  "published_at": "2026-05-07T06:05:26Z",
   "repo": "huggingface/transformers",
   "schema_version": 1,
-  "snapshot_id": "20260507T000029Z",
+  "snapshot_id": "20260507T060024Z",
   "variant": "hybrid"
 }
diff --git a/snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.json b/snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.json
new file mode 100644
index 0000000000000000000000000000000000000000..096120647e0bf5bb1ccfb3bfdfe8201f55d44098
--- /dev/null
+++ b/snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.json
@@ -0,0 +1,2205 @@
+{
+  "schema_version": "1.0",
+  "repo": "huggingface/transformers",
+  "snapshot_id": "20260507T060024Z",
+  "generated_at": "2026-05-07T06:05:24Z",
+  "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-43656-4",
+      "summary": "Cluster of 1 issues and 3 PRs centered on issue #43824.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 43824,
+      "canonical_pr_number": 43656,
+      "issue_numbers": [
+        43824
+      ],
+      "pr_numbers": [
+        43656,
+        43836,
+        43842
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target",
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 43656,
+          "right_pr_number": 43836,
+          "code_similarity": 0.36,
+          "size_similarity": 0.176,
+          "file_overlap": 1.0,
+          "area_overlap": 0.051,
+          "patch_similarity": 0.048,
+          "shared_filenames": [
+            "src/transformers/cli/serve.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/cli/serve.py",
+              "left_ranges": [
+                [
+                  11,
+                  18
+                ],
+                [
+                  30,
+                  36
+                ],
+                [
+                  315,
+                  323
+                ],
+                [
+                  665,
+                  671
+                ],
+                [
+                  931,
+                  937
+                ],
+                [
+                  1843,
+                  1849
+                ],
+                [
+                  1868,
+                  1874
+                ]
+              ],
+              "right_ranges": [
+                [
+                  11,
+                  18
+                ],
+                [
+                  359,
+                  429
+                ],
+                [
+                  584,
+                  590
+                ],
+                [
+                  1892,
+                  1910
+                ],
+                [
+                  1917,
+                  1923
+                ]
+              ]
+            }
+          ]
+        },
+        {
+          "left_pr_number": 43656,
+          "right_pr_number": 43842,
+          "code_similarity": 0.405,
+          "size_similarity": 0.5,
+          "file_overlap": 1.0,
+          "area_overlap": 0.0,
+          "patch_similarity": 0.036,
+          "shared_filenames": [
+            "src/transformers/cli/serve.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/cli/serve.py",
+              "left_ranges": [
+                [
+                  11,
+                  18
+                ],
+                [
+                  30,
+                  36
+                ],
+                [
+                  315,
+                  323
+                ],
+                [
+                  665,
+                  671
+                ],
+                [
+                  931,
+                  937
+                ],
+                [
+                  1843,
+                  1849
+                ],
+                [
+                  1868,
+                  1874
+                ]
+              ],
+              "right_ranges": [
+                [
+                  54,
+                  61
+                ],
+                [
+                  587,
+                  593
+                ]
+              ]
+            }
+          ]
+        },
+        {
+          "left_pr_number": 43836,
+          "right_pr_number": 43842,
+          "code_similarity": 0.332,
+          "size_similarity": 0.088,
+          "file_overlap": 1.0,
+          "area_overlap": 0.033,
+          "patch_similarity": 0.017,
+          "shared_filenames": [
+            "src/transformers/cli/serve.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/cli/serve.py",
+              "left_ranges": [
+                [
+                  11,
+                  18
+                ],
+                [
+                  359,
+                  429
+                ],
+                [
+                  584,
+                  590
+                ],
+                [
+                  1892,
+                  1910
+                ],
+                [
+                  1917,
+                  1923
+                ]
+              ],
+              "right_ranges": [
+                [
+                  54,
+                  61
+                ],
+                [
+                  587,
+                  593
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "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-41115-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #41115.",
+      "status": "open",
+      "confidence": 0.55,
+      "canonical_issue_number": 41115,
+      "canonical_pr_number": 45613,
+      "issue_numbers": [
+        41115
+      ],
+      "pr_numbers": [
+        41116,
+        45613
+      ],
+      "evidence_types": [
+        "closing_reference"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 41116,
+          "right_pr_number": 45613,
+          "code_similarity": 0.805,
+          "size_similarity": 0.885,
+          "file_overlap": 0.75,
+          "area_overlap": 0.854,
+          "patch_similarity": 0.693,
+          "shared_filenames": [
+            "docs/source/en/_toctree.yml",
+            "docs/source/en/model_doc/minicpm3.md",
+            "src/transformers/models/auto/modeling_auto.py",
+            "src/transformers/models/minicpm3/__init__.py",
+            "src/transformers/models/minicpm3/configuration_minicpm3.py",
+            "src/transformers/models/minicpm3/modeling_minicpm3.py",
+            "src/transformers/models/minicpm3/modular_minicpm3.py",
+            "tests/models/minicpm3/__init__.py",
+            "tests/models/minicpm3/test_modeling_minicpm3.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "docs/source/en/_toctree.yml",
+              "left_ranges": [
+                [
+                  713,
+                  720
+                ]
+              ],
+              "right_ranges": [
+                [
+                  713,
+                  720
+                ]
+              ]
+            },
+            {
+              "filename": "docs/source/en/model_doc/minicpm3.md",
+              "left_ranges": [
+                [
+                  1,
+                  67
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  45
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/auto/modeling_auto.py",
+              "left_ranges": [
+                [
+                  285,
+                  291
+                ],
+                [
+                  700,
+                  706
+                ],
+                [
+                  1298,
+                  1304
+                ]
+              ],
+              "right_ranges": [
+                [
+                  699,
+                  705
+                ],
+                [
+                  1301,
+                  1307
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/minicpm3/__init__.py",
+              "left_ranges": [
+                [
+                  1,
+                  29
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  29
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/minicpm3/configuration_minicpm3.py",
+              "left_ranges": [
+                [
+                  1,
+                  141
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  126
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/minicpm3/modeling_minicpm3.py",
+              "left_ranges": [
+                [
+                  1,
+                  544
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  522
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/minicpm3/modular_minicpm3.py",
+              "left_ranges": [
+                [
+                  1,
+                  444
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  342
+                ]
+              ]
+            },
+            {
+              "filename": "tests/models/minicpm3/test_modeling_minicpm3.py",
+              "left_ranges": [
+                [
+                  1,
+                  109
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  136
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "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-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-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": 0.532,
+          "size_similarity": 0.619,
+          "file_overlap": 0.5,
+          "area_overlap": 0.543,
+          "patch_similarity": 0.451,
+          "shared_filenames": [
+            "src/transformers/testing_utils.py",
+            "tests/utils/test_testing_utils.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/testing_utils.py",
+              "left_ranges": [
+                [
+                  3529,
+                  3536
+                ],
+                [
+                  3753,
+                  3787
+                ]
+              ],
+              "right_ranges": [
+                [
+                  3525,
+                  3558
+                ],
+                [
+                  3782,
+                  3788
+                ]
+              ]
+            },
+            {
+              "filename": "tests/utils/test_testing_utils.py",
+              "left_ranges": [
+                [
+                  1,
+                  86
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  114
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "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-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": [],
+  "duplicate_prs": [
+    {
+      "cluster_id": "cluster-41115-3",
+      "canonical_pr_number": 45613,
+      "duplicate_pr_numbers": [
+        41116
+      ],
+      "target_issue_number": 41115,
+      "reason": "PRs in cluster-41115-3 are treated as duplicates because they converge on issue #41115 with closing_reference evidence."
+    },
+    {
+      "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-43656-4",
+      "canonical_pr_number": 43656,
+      "duplicate_pr_numbers": [
+        43836,
+        43842
+      ],
+      "target_issue_number": 43824,
+      "reason": "PRs in cluster-43656-4 are treated as duplicates because they converge on issue #43824 with closing_reference, shared_issue_target, soft_similarity 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-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."
+    }
+  ],
+  "best_issue": {
+    "cluster_id": "cluster-45647-3",
+    "issue_number": 45647,
+    "reason": "Issue #45647 is the strongest global issue candidate because it is open, belongs to a cluster with 3 artifacts, and carries 12 discussion comments plus 8 inbound references.",
+    "score": 162.0
+  },
+  "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/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.llm-reviews.json b/snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.llm-reviews.json
new file mode 100644
index 0000000000000000000000000000000000000000..be396d9b183ba35eb01a841bfb88fcf6ff419e5b
--- /dev/null
+++ b/snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.llm-reviews.json
@@ -0,0 +1,18771 @@
+{
+  "schema_version": "1.0",
+  "repo": "huggingface/transformers",
+  "snapshot_id": "20260507T060024Z",
+  "generated_at": "2026-05-07T06:05:24Z",
+  "model": "gpt-5.4-mini?service_tier=flex",
+  "reviews": [
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 16,
+      "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": "f4801749e750d1d74298c36ae44478d7c713bf1e268cc0098f85f399d13baf7c",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 2479,
+        "estimated_input_tokens": 620,
+        "estimated_eval_tokens": 1496
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 2479,
+        "estimated_input_tokens": 620,
+        "estimated_eval_tokens": 1496
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These issues do not look like duplicates. One is about Tapas tokenization, while the other two are model-loading/authentication reports, and even those two appear to be different failure modes.",
+        "confidence": 0.87,
+        "canonical_issue_reason": "Issue 16618 is the broadest and most representative of the cluster\u2019s model-loading complaints, so it is the best single issue to anchor the cluster even though it is not a true duplicate of the others.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "16618 is the most central and generally phrased report in the set, making it the best candidate for a cluster representative; however, the cluster is heterogeneous.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:13244",
+            "right": "issue:16618",
+            "accept": false,
+            "reason": "Tapas tokenization differences are unrelated to a 'can't load bert-base-uncased' model-loading failure; they do not share the same underlying bug."
+          },
+          {
+            "left": "issue:15980",
+            "right": "issue:16618",
+            "accept": false,
+            "reason": "Both mention loading/downloading models, but 15980 is specifically about a private-model login error and 16618 is a general inability to load a public model; the concrete failure modes differ."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 1,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:24643",
+        "issue:30277",
+        "issue:34634",
+        "issue:34928",
+        "issue:35545",
+        "issue:36010",
+        "issue:36331",
+        "issue:39290",
+        "issue:41093",
+        "issue:41720",
+        "issue:41762",
+        "issue:42915",
+        "issue:43039",
+        "issue:43404",
+        "issue:43425",
+        "issue:43531",
+        "issue:43541",
+        "issue:43596",
+        "issue:43638",
+        "issue:43716",
+        "issue:43792",
+        "issue:43827",
+        "issue:43828",
+        "issue:43854",
+        "issue:43866",
+        "issue:43901",
+        "issue:43975",
+        "issue:44292",
+        "issue:44322",
+        "issue:44457",
+        "issue:44509",
+        "issue:44512",
+        "issue:44560",
+        "issue:44661",
+        "issue:44805",
+        "issue:44841",
+        "issue:44863",
+        "issue:44918",
+        "issue:45070",
+        "issue:45081",
+        "issue:45084",
+        "issue:45161",
+        "issue:45237",
+        "issue:45362",
+        "issue:45464",
+        "issue:45507",
+        "issue:45685",
+        "issue:45800"
+      ],
+      "soft_pairs": [
+        "issue:44322|issue:45464",
+        "issue:45237|issue:45685",
+        "issue:43716|issue:45237",
+        "issue:43828|issue:45237",
+        "issue:45237|issue:45507",
+        "issue:44292|issue:45237",
+        "issue:30277|issue:35545",
+        "issue:35545|issue:42915",
+        "issue:35545|issue:41720",
+        "issue:44509|issue:44512",
+        "issue:35545|issue:36010",
+        "issue:43638|issue:44805",
+        "issue:43596|issue:44805",
+        "issue:44805|issue:45161",
+        "issue:44805|issue:44918",
+        "issue:44661|issue:44805",
+        "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:34928|issue:44805",
+        "issue:24643|issue:30277",
+        "issue:43638|issue:45161",
+        "issue:34928|issue:36331",
+        "issue:43596|issue:45161",
+        "issue:43716|issue:45507",
+        "issue:39290|issue:45081",
+        "issue:43828|issue:45161",
+        "issue:43425|issue:44292",
+        "issue:44918|issue:45161",
+        "issue:34928|issue:41720",
+        "issue:43716|issue:45685",
+        "issue:43541|issue:43716",
+        "issue:43828|issue:44292",
+        "issue:45507|issue:45685",
+        "issue:43716|issue:44292",
+        "issue:44292|issue:44918",
+        "issue:43039|issue:44918",
+        "issue:41762|issue:43638",
+        "issue:43827|issue:44512",
+        "issue:43716|issue:44560",
+        "issue:43828|issue:45685",
+        "issue:44560|issue:45685",
+        "issue:43039|issue:43828",
+        "issue:30277|issue:36331",
+        "issue:34928|issue:36010",
+        "issue:43039|issue:44805",
+        "issue:41762|issue:43596",
+        "issue:43901|issue:44512",
+        "issue:39290|issue:41720",
+        "issue:43039|issue:44560",
+        "issue:43866|issue:44863",
+        "issue:43854|issue:43866",
+        "issue:39290|issue:43531",
+        "issue:41093|issue:43638",
+        "issue:43425|issue:45070",
+        "issue:43596|issue:45800",
+        "issue:45362|issue:45464",
+        "issue:43792|issue:45070",
+        "issue:43404|issue:43866",
+        "issue:45464|issue:45685",
+        "issue:43541|issue:45685",
+        "issue:43541|issue:43828",
+        "issue:36010|issue:39290",
+        "issue:44841|issue:45084",
+        "issue:41093|issue:41762"
+      ],
+      "prepared_review_unit_hash": "3bbbc868a01b6aa7a53bffdc3be94b29debcb4058767fe457369dd7244bf2697",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 68,
+        "serialized_chars": 47979,
+        "estimated_input_tokens": 11995,
+        "estimated_eval_tokens": 24246
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a heterogeneous set of model-specific runtime bugs, shape/dtype errors, and a couple of v5 documentation regressions. There is one clear duplicate-like pair around DeepSpeed ZeRO-3/BERT init failures; most other soft links are only superficially similar.",
+        "confidence": 0.81,
+        "canonical_issue_reason": "issue:44805 is the most central representative of the cluster's recurring tensor-shape/runtime-failure reports and has the broadest overlap with other size/mask mismatch issues.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44805 is the best overall representative because it sits closest to multiple related failure reports and summarizes the repeated shape-mismatch class better than the more niche model-specific issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44322",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Same model family, but different failures and code paths: config attribute lookup vs streaming chat/completions inference."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Both are backend-specific failures, but the models, hardware targets, and root causes are unrelated."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Image-preprocessor dtype mismatch for Mistral-3 is unrelated to GPT-OSS AMD GPU execution failure."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Autocast dtype mismatch and AMD GPU failure are different bugs with no shared code path."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Different models and different failure modes; not the same underlying bug."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "NVFP4 runtime error and AMD GPU compatibility issue are separate problems."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:35545",
+            "accept": false,
+            "reason": "DeepSpeed collective mismatch during training is unrelated to ModernBERT ONNX export."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "ONNX export failure and Qwen3Moe FP8 failure do not share the same concrete code path."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Export-to-ONNX and auto device mapping cuda assert are different failures."
+          },
+          {
+            "left": "issue:44509",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are v5 docs regressions, but they affect different stale references and would be separate fixes."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "ImportError for GenerationMixin is unrelated to ModernBERT export issues."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Both are shape-related, but one is DeepSpeed ZeRO-3 zero-size index failure and the other is a mask/index size mismatch; not clearly the same bug."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Same broad DeepSpeed training area, but the reported failure signatures are different enough to avoid merging."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Mask/index size mismatch is unrelated to GPT-OSS tensor-parallel behavior."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Tokenizer mapping failure in add-new-model-like is a different code path from the tensor mask mismatch."
+          },
+          {
+            "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": "Both involve DeepSpeed, but collective mismatch and zero-sized index errors are distinct issues."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:43638",
+            "accept": true,
+            "reason": "Both describe the same DeepSpeed ZeRO-3 / BERT index-0-out-of-bounds failure during initialization or training; only the model phrasing differs."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Liger Kernel cross-entropy being called is unrelated to GPT-OSS tensor-parallel failure."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:44457",
+            "accept": false,
+            "reason": "Incorrect detokenization and LoRA merge/save/reload mismatch are different bugs."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44457",
+            "accept": false,
+            "reason": "NVFP4 runtime failure and LoRA weight merge inconsistency do not share a root cause."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Both are tensor-shape related, but activation-checkpointing/FSDP recompute mismatch is a different scenario from a mask/index length error."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:30277",
+            "accept": false,
+            "reason": "Both mention DeepSpeed, but the runtime error and collective mismatch are separate training failures."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO-3 BERT init error is unrelated to GPT-OSS tensor-parallel behavior."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Activation checkpointing size mismatch and custom compute_loss signature breakage are unrelated."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Mistral image-preprocessor dtype mismatch and GraniteMoEHybrid invalid method are different bugs."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Gemma3 sliding-window attribute error is unrelated to the Mistral tokenizer regex crash."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Different model families and different failures; both are dtype/runtime issues only at a very broad level."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Torch version incompatibility is not the same as the Qwen-3 NVFP4 runtime problem."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Qwen3.5 embedding unpacking with TRL is unrelated to GPT-OSS tensor parallelism."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Gemma3 ZeRO-3 loading and Bert ZeRO-3 loading are similar in theme but not the same concrete bug."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are docs regressions, but one is about removed pipeline tasks and the other about the removed run command."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Mistral-3 image-preprocessor dtype mismatch and Qwen3-vl video StopIteration are unrelated."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Autocast dtype mismatch and MPS histogram_int limitation are different backend-specific issues."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Video processing StopIteration and MPS histogram failure do not share a code path."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Liger Kernel cross-entropy invocation is unrelated to Phi-tiny-MoE autocast dtype mismatch."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "DeepSpeed collective mismatch and compute_loss kwarg incompatibility are separate training problems."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "FSDP/activation-checkpointing size mismatch is unrelated to the GenerationMixin import error."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Liger Kernel behavior and mask/index length mismatch are unrelated failures."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:43596",
+            "accept": false,
+            "reason": "Both are ZeRO-3-related, but Gemma3 loading and BERT init errors are not the same bug."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "TextClassificationPipeline docs and transformers run-command docs are separate stale references."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Gemma3 sliding-window attribute error and Qwen3 auto device-map assert are different issues."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Liger Kernel cross-entropy call and Qwen3-vl video StopIteration are unrelated."
+          },
+          {
+            "left": "issue:43866",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Corrupted Ovis checkpoint and NemotronH checkpoint loading failure are different root causes."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "Unit-test model load failure for GLM-4.7-Flash and Ovis checkpoint corruption are not the same bug."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Gemma3 sliding-window attribute error and Qwen3-MoE sliding_window issue are different model/config bugs."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Mask-shape mismatch and ZeRO-3 zero-sized index failure are related only superficially."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Torch compatibility and pydantic PretrainedConfig field breakage are unrelated."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "ZeRO-3 BERT init failure and torch 2.4.1 / transformers 5.8.0 incompatibility are different problems."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Qwen3.5 chat crash and streaming inference API failure are different issues."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Whisper load failure is unrelated to the pydantic model regression."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "Mistral3 lm_head tying bug and Ovis checkpoint corruption are unrelated."
+          },
+          {
+            "left": "issue:45464",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Qwen3.5 streaming inference and MPS MoE histogram failure are unrelated."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "torch dynamo grouped_mm tracing failure and MPS histogram limitation are different backend issues."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Grouped_mm tracing error and autocast dtype mismatch are not the same bug."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:39290",
+            "accept": false,
+            "reason": "GenerationMixin import error and Gemma3 config attribute error are unrelated."
+          },
+          {
+            "left": "issue:44841",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Voxtral processor failure and template-node compilation error are different code paths."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:41762",
+            "accept": false,
+            "reason": "Both are shape-style errors, but one is mask/index mismatch and the other is ZeRO-3 zero-size load failure."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is well grounded in the packet: it correctly characterizes the cluster as mostly heterogeneous, identifies the one clearly duplicate-like DeepSpeed ZeRO-3/BERT pair conservatively, and treats the remaining soft links as superficial rather than overclaiming shared root causes. The canonical issue choice is also supported by the strongest overlap signals in the packet."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 2,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:39186",
+        "issue:39290",
+        "issue:39692",
+        "issue:40990",
+        "issue:41093",
+        "issue:42915",
+        "issue:43054",
+        "issue:43296",
+        "issue:43366",
+        "issue:43404",
+        "issue:43531",
+        "issue:43541",
+        "issue:43572",
+        "issue:43596",
+        "issue:43644",
+        "issue:43645",
+        "issue:43716",
+        "issue:43742",
+        "issue:43828",
+        "issue:43873",
+        "issue:43950",
+        "issue:43975",
+        "issue:43994",
+        "issue:44291",
+        "issue:44292",
+        "issue:44360",
+        "issue:44387",
+        "issue:44492",
+        "issue:44512",
+        "issue:44530",
+        "issue:44534",
+        "issue:44589",
+        "issue:44863",
+        "issue:44918",
+        "issue:44928",
+        "issue:44960",
+        "issue:45005",
+        "issue:45020",
+        "issue:45070",
+        "issue:45161",
+        "issue:45399",
+        "issue:45464",
+        "issue:45507",
+        "issue:45542",
+        "issue:45632",
+        "issue:45685",
+        "issue:45698",
+        "issue:45800"
+      ],
+      "soft_pairs": [
+        "issue:43296|issue:43366",
+        "issue:43572|issue:45070",
+        "issue:44387|issue:45005",
+        "issue:43541|issue:45161",
+        "issue:43828|issue:45464",
+        "issue:44291|issue:45070",
+        "issue:45020|issue:45800",
+        "issue:43366|issue:45464",
+        "issue:43531|issue:45070",
+        "issue:39290|issue:43742",
+        "issue:43541|issue:44589",
+        "issue:41093|issue:43596",
+        "issue:43054|issue:43994",
+        "issue:45542|issue:45800",
+        "issue:43716|issue:44918",
+        "issue:44918|issue:45685",
+        "issue:45632|issue:45698",
+        "issue:43404|issue:45005",
+        "issue:43645|issue:45399",
+        "issue:40990|issue:44960",
+        "issue:44530|issue:45005",
+        "issue:42915|issue:45399",
+        "issue:44360|issue:44512",
+        "issue:44863|issue:45005",
+        "issue:43873|issue:45005",
+        "issue:39186|issue:44918",
+        "issue:44492|issue:44512",
+        "issue:44918|issue:45507",
+        "issue:43975|issue:44292",
+        "issue:43950|issue:44534",
+        "issue:43644|issue:43950",
+        "issue:44292|issue:45507",
+        "issue:39692|issue:43054",
+        "issue:43828|issue:44928"
+      ],
+      "prepared_review_unit_hash": "77b7b92670b26046e8b215f7552f5ade8ddd6d747e015883fdf4a39646370835",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 37554,
+        "estimated_input_tokens": 9389,
+        "estimated_eval_tokens": 19034
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most items are distinct model/configuration regressions rather than duplicates. The only clear duplicate set is the Transformers v5 non-persistent-buffer corruption reports; the SigLIP2 pair is a plausible but less certain duplicate on the same underlying inference pathology.",
+        "confidence": 0.73,
+        "canonical_issue_reason": "issue:43950 is the most specific and actionable report for the v5 non-persistent-buffer regression, with a clear `from_pretrained()` reproduction and explicit diagnosis of `register_buffer(persistent=False)` corruption.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43950 best represents the cluster because it precisely describes the shared regression and is broad enough to subsume the other buffer-corruption reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43296",
+            "right": "issue:43366",
+            "accept": false,
+            "reason": "Different failures: PaddleOCR-VL load/vLLM compatibility versus GGUF gpt-oss architecture support."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Unrelated config regressions: missing `pad_token_idx` in StableLmConfig vs pydantic `PretrainedConfig` field breakage."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Both mention model behavior changes in v5, but one is int4 OOM and the other is tied-weights handling for translation models; not the same bug."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Different code paths: torch-dynamo/grouped_mm tracing error versus GPT-OSS tensor-parallel support issue."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Autocast dtype mismatch in model forward pass is unrelated to a chat/completions streaming inference failure."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "`init_empty_weights` unexpected argument error is a separate loading-path incompatibility from the pydantic model-field regression."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Remote-code breakage and torch/transformers version incompatibility are different failure modes with no shared concrete bug."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "GGUF gpt-oss architecture support is unrelated to Qwen3.5 streaming API failures."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding-window behavior is a model-specific config issue, not the pydantic `PretrainedConfig` regression."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Gemma3TextConfig missing `sliding_window_pattern` and MobileLLM key-error loading are distinct model-specific issues."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Different runtime errors: grouped_mm tracing on Mixtral versus missing Float8 storage object."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:43596",
+            "accept": false,
+            "reason": "Mask-length mismatch and zero3 init index error are unrelated tensor-shape/initialization bugs."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:43994",
+            "accept": true,
+            "reason": "Both describe SigLIP2 producing degraded or nonsensical outputs, likely the same underlying model/processor inference issue."
+          },
+          {
+            "left": "issue:45542",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "TensorBoard-without-TensorFlow backend error is unrelated to torch/transformers version incompatibility."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Both involve multimodal/Qwen-family usage, but the reported failures are different: image-preprocessor dtype mismatch versus SFT embedding unpacking."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Qwen3.5 SFT embedding unpacking and an MPS histogram kernel limitation are unrelated."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Cache-path collision for local models is a different bug from loading the wrong custom module after save/load."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Tied-weights bug in Mistral3 multimodal generation is not the same concrete issue as translation-model tied weights in v5."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Jupyter custom-model initialization breakage is unrelated to flash-attn2 fallback gating."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "High perplexity on gpt-oss and the GLM5 issue are different model-specific behavior reports."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "PagedAttentionCache linear_attention crash and tied-weights regression are unrelated."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Qwen3Moe FineGrainedFP8Config failure is not the same as flash-attn2 fallback blocking."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "DSA indexer ReLU discussion and docs-about-removed-command are unrelated."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "NemotronH checkpoint loading failure is not the same bug as translation-model tied weights."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Quantization/offloading behavior is unrelated to tied-weights handling."
+          },
+          {
+            "left": "issue:39186",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "FSDP weight dimensionality error is unrelated to Qwen3.5 input-embedding unpacking."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Cache strategy typo and removed CLI docs are unrelated documentation issues."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "SFT input embedding unpacking and GraniteMoEHybrid invalid-method calls are different bugs."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Incorrect detokenization and Qwen-3-8B-NVFP4 runtime failure are different model-specific issues."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:44534",
+            "accept": true,
+            "reason": "Both report the same v5 regression where non-persistent buffers are corrupted/fill with junk during loading."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43950",
+            "accept": true,
+            "reason": "Same underlying Transformers v5 non-persistent-buffer corruption bug; 43950 is the more precise reproduction."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Qwen-3-8B-NVFP4 runtime error and GraniteMoEHybrid invalid method are unrelated."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43054",
+            "accept": false,
+            "reason": "SigLIP2 docs/example errors are about example mismatches and quantization, not the same issue as degraded text embeddings."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:44928",
+            "accept": false,
+            "reason": "Autocast dtype mismatch and Qwen3.5 RLHF NaN/position_ids issue are distinct failures."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded overall. The non-persistent-buffer regression cluster around 43644/43950/44534 is well supported, and the summary keeps the SigLIP2 pair appropriately hedged as less certain. No obvious overstatement in the accepted soft-edge verdicts."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 3,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:36331",
+        "issue:38175",
+        "issue:41553",
+        "issue:41628",
+        "issue:41762",
+        "issue:42491",
+        "issue:42915",
+        "issue:42947",
+        "issue:43054",
+        "issue:43257",
+        "issue:43404",
+        "issue:43493",
+        "issue:43643",
+        "issue:43742",
+        "issue:43782",
+        "issue:43792",
+        "issue:43824",
+        "issue:43827",
+        "issue:43828",
+        "issue:43856",
+        "issue:43866",
+        "issue:43872",
+        "issue:43881",
+        "issue:43883",
+        "issue:43940",
+        "issue:43950",
+        "issue:43975",
+        "issue:44220",
+        "issue:44291",
+        "issue:44368",
+        "issue:44387",
+        "issue:44403",
+        "issue:44451",
+        "issue:44488",
+        "issue:44509",
+        "issue:44589",
+        "issue:44661",
+        "issue:44960",
+        "issue:44991",
+        "issue:45020",
+        "issue:45081",
+        "issue:45237",
+        "issue:45313",
+        "issue:45356",
+        "issue:45399",
+        "issue:45520",
+        "issue:45701",
+        "issue:45800"
+      ],
+      "soft_pairs": [
+        "issue:43866|issue:44403",
+        "issue:36331|issue:41762",
+        "issue:41628|issue:45800",
+        "issue:41553|issue:43883",
+        "issue:43872|issue:44589",
+        "issue:38175|issue:43054",
+        "issue:41553|issue:43881",
+        "issue:43828|issue:45800",
+        "issue:43054|issue:45701",
+        "issue:42947|issue:44387",
+        "issue:44403|issue:44451",
+        "issue:43881|issue:43883",
+        "issue:43643|issue:45020",
+        "issue:43054|issue:43493",
+        "issue:43975|issue:45520",
+        "issue:43975|issue:45237",
+        "issue:44403|issue:44488",
+        "issue:42947|issue:43856",
+        "issue:42491|issue:44368",
+        "issue:41762|issue:42915",
+        "issue:43975|issue:44451",
+        "issue:43824|issue:43872",
+        "issue:44661|issue:45020",
+        "issue:44991|issue:45399",
+        "issue:45020|issue:45356",
+        "issue:43782|issue:44960",
+        "issue:45081|issue:45399",
+        "issue:43827|issue:44509",
+        "issue:43792|issue:44220",
+        "issue:43404|issue:43950",
+        "issue:43742|issue:45399",
+        "issue:42915|issue:45020",
+        "issue:43257|issue:43866",
+        "issue:41553|issue:44991",
+        "issue:43742|issue:45020",
+        "issue:43940|issue:45313",
+        "issue:43872|issue:44291"
+      ],
+      "prepared_review_unit_hash": "56d11504772233f935cdd96dd00ee34693c33ea4f897c8ef028757d14d30173f",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 37,
+        "serialized_chars": 38718,
+        "estimated_input_tokens": 9680,
+        "estimated_eval_tokens": 19616
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous transformer loading/compatibility regressions plus one docs cleanup thread. Only the two v5 pipeline-removal documentation issues look like the same underlying bug/change.",
+        "confidence": 0.84,
+        "canonical_issue_reason": "Issue 45020 is the broadest representative of the cluster\u2019s dominant theme: recent-version loading regressions around remote_code/model loading. It\u2019s the best umbrella issue, though most other items are still distinct.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45020 best captures the most recurring user-facing class here (transformers-version loading breakages), and it is active/open, making it the most useful representative issue.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43866",
+            "right": "issue:44403",
+            "accept": false,
+            "reason": "Different defects: a corrupted checkpoint versus generic loading noise. Same general area, not the same bug."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41762",
+            "accept": false,
+            "reason": "Training API kwarg mismatch vs DeepSpeed ZeRO-3 loading failure; unrelated code paths and symptoms."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Import/export issue for AutoImageProcessor vs broad torch/transformers version incompatibility; too general to merge."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "Voxtral tokenizer error message vs Molmo tied-weights attribute error; different model-specific failures."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "bitsandbytes init-arg incompatibility vs missing Float8 storage type; both dependency errors, but not the same failure."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43054",
+            "accept": false,
+            "reason": "Both involve SigLIP2, but one reports zero probabilities while the other reports worse embeddings; not the same concrete bug."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "Different model-loading problems for different models and error messages."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Autocast dtype mismatch in Phi-MoE versus a general torch/transformers incompatibility report; not the same code-path issue."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Embedding quality regression vs tokenization changes across versions; related theme, but different defect."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Gradient checkpointing/LoRA behavior is distinct from int4 memory reservation/OOM."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Generic loading noise vs a specific model load failure; same broad area, different bugs."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "glm-4v-9b load failure and Molmo missing attribute are different failures in different models."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "AutoConfig missing fields with trust_remote_code=True is narrower and different from the broader remote_code breakage report."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Both concern SigLIP2, but one is a behavioral quality discrepancy and the other is an implementation mismatch with JAX."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Tokenizer detokenization bug vs a Python 3.13 flash_attn import KeyError; unrelated subsystems."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Tokenizer regression is unrelated to AMD GPU runtime support."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "One is about loading noise, the other is a specific model that still fails to load; not the same defect."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "LoRA gradient-checkpointing issue vs MoE training memory inefficiency; both performance-related but distinct."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "A LoRA/qwen3_moe compatibility break is not the same as a tie_word_embeddings warning."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Different model families and different ZeRO-3/load failures; too broad to consider duplicates."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Tokenizer detokenization bug vs a model-loading failure for ScandiBERT; different user-visible problems."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43872",
+            "accept": false,
+            "reason": "Import error for Qwen2.5-VL versus bitsandbytes init incompatibility; unrelated."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Registry/add-new-model-like failure is a different issue from remote_code model loading breakage."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Tokenizer loading regression for a specific model vs flash-attn fallback selection logic; no shared concrete bug."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "General remote_code breakage and a Kimi-K2.5 tokenizer codec regression are different root causes."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "Qwen3VL weight_only loading error and GLM5 are different model-specific issues."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Mistral tokenizer regex crash versus flash-attn fallback blocking; different code paths."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44509",
+            "accept": true,
+            "reason": "Same docs problem: both report stale pipeline task references removed in v5."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Whisper model runtime failure and a specific fbank helper issue may be related, but the evidence is not strong enough to treat them as the same bug."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43950",
+            "accept": false,
+            "reason": "lm_head tie-weight bug in a model class versus from_pretrained corrupting non-persistent buffers; different regressions."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "MobileLLM key error and flash-attn fallback logic are unrelated."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Qwen3 MoE ZeRO-3 weight loading failure is separate from the broader remote_code regression report."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "Qwen3 MoE weights not converted under accelerate+deepspeed and Ovis2 checkpoint corruption are different issues."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Voxtral tokenizer error message and est-roberta tokenizer loading failure are different model-specific issues."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "MobileLLM key error is a specific model-loading bug, not the same as the broader remote_code regression."
+          },
+          {
+            "left": "issue:43940",
+            "right": "issue:45313",
+            "accept": false,
+            "reason": "Both are ZeRO-3 load failures, but they affect different Qwen3 variants and likely different failure modes."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Similar `_is_hf_initialized`/init-empty-weights theme, but one is bitsandbytes-specific while the other is a generic init_empty_weights TypeError; not identical."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative: it correctly treats the cluster as mostly heterogeneous loading/compatibility regressions, with the docs-pipeline pair as the only clear same-bug match. The representative issue choice is defensible, and the soft-edge rejections do not overclaim duplicates."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 4,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:24643",
+        "issue:30064",
+        "issue:36010",
+        "issue:36331",
+        "issue:38175",
+        "issue:39692",
+        "issue:41093",
+        "issue:41553",
+        "issue:42617",
+        "issue:42915",
+        "issue:43257",
+        "issue:43317",
+        "issue:43366",
+        "issue:43531",
+        "issue:43550",
+        "issue:43596",
+        "issue:43638",
+        "issue:43645",
+        "issue:43646",
+        "issue:43824",
+        "issue:43828",
+        "issue:43854",
+        "issue:43872",
+        "issue:43950",
+        "issue:43957",
+        "issue:43994",
+        "issue:44155",
+        "issue:44479",
+        "issue:44488",
+        "issue:44530",
+        "issue:44560",
+        "issue:44589",
+        "issue:44661",
+        "issue:44756",
+        "issue:44805",
+        "issue:44863",
+        "issue:44877",
+        "issue:44910",
+        "issue:44918",
+        "issue:44928",
+        "issue:45030",
+        "issue:45305",
+        "issue:45313",
+        "issue:45325",
+        "issue:45357",
+        "issue:45538",
+        "issue:45584",
+        "issue:45588"
+      ],
+      "soft_pairs": [
+        "issue:43854|issue:44863",
+        "issue:42915|issue:43872",
+        "issue:44928|issue:45588",
+        "issue:41553|issue:44488",
+        "issue:43257|issue:45305",
+        "issue:36331|issue:41093",
+        "issue:43872|issue:45538",
+        "issue:41093|issue:44805",
+        "issue:44530|issue:44863",
+        "issue:44479|issue:44560",
+        "issue:24643|issue:30064",
+        "issue:43366|issue:43828",
+        "issue:43366|issue:43531",
+        "issue:43550|issue:44910",
+        "issue:43645|issue:43646",
+        "issue:43366|issue:44589",
+        "issue:43950|issue:45325",
+        "issue:44560|issue:44918",
+        "issue:42617|issue:43366",
+        "issue:42915|issue:43957",
+        "issue:43824|issue:43957",
+        "issue:36010|issue:42915",
+        "issue:38175|issue:43994",
+        "issue:43646|issue:43950",
+        "issue:44155|issue:45584",
+        "issue:44560|issue:44805",
+        "issue:44877|issue:45030",
+        "issue:43638|issue:44661",
+        "issue:43596|issue:44661",
+        "issue:43317|issue:44756",
+        "issue:38175|issue:39692",
+        "issue:45313|issue:45357"
+      ],
+      "prepared_review_unit_hash": "24831c7fdc518da9887e326c7ff3c6a056d87447316957ecabaddcdee9fca6da",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37464,
+        "estimated_input_tokens": 9366,
+        "estimated_eval_tokens": 18988
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Conservative cluster: the true duplicate/near-duplicate anchors are the mask-shape IndexError pair, the custom-model initialization regression pair, and the SigLIP2 bad-output pair. The DeepSpeed ZeRO-3 BertModel index-error reports look superficially related but should remain separate, since those soft edges are rejected and there isn\u2019t enough evidence to collapse them into one cluster item.",
+        "confidence": 0.79,
+        "canonical_issue_reason": "issue:41093 is the best canonical issue anchor because it has the strongest accepted duplicate-style match (issue:44805) for the same mask-shape IndexError. It is more concrete than the other similar-looking index-error reports, which do not have sufficient evidence to merge.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:41093 is the best overall issue anchor for the cluster because it represents a concrete reproducible bug with a strong accepted near-duplicate and a clearly identifiable failure signature.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43854",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Different model-loading failures: GLM-4.7-Flash unit test loading vs NemotronH checkpoint loading. Same broad area, not the same bug."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43872",
+            "accept": false,
+            "reason": "Both are quantization/DeepSpeed-adjacent, but the concrete failures differ: Qwen3MoE FineGrainedFP8 vs bitsandbytes Int8Params incompatibility."
+          },
+          {
+            "left": "issue:44928",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Both mention attention/3D-related issues, but one is RLHF NaN from SDPA fallback and the other is an AttributeError in flash_attention on sink-less models."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both are loading/error-message issues, but they involve different models and different failure modes."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45305",
+            "accept": false,
+            "reason": "Both touch DeepSpeed, but one is missing MOE weight conversion and the other is gradient averaging with model_accepts_loss_kwargs=True."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41093",
+            "accept": false,
+            "reason": "Training API mismatch vs tensor mask-shape mismatch; unrelated failures and fixes."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "bitsandbytes constructor incompatibility vs CLIPTokenizer max length constant; unrelated."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:44805",
+            "accept": true,
+            "reason": "Same mask-shape IndexError class with only the concrete dimensions differing; clearly the same underlying bug."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "PagedAttentionCache linear_attention crash vs NemotronH checkpoint loading; different subsystems and symptoms."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Both are video-related Qwen issues, but one is a broad v5.3.0 regression and the other is a StopIteration in Qwen3-vl-embedding; not the same defect."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "DeepSpeed training weight-dimension error vs image processor void-segmentation-map failure; unrelated."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "GGUF architecture support request vs autocast dtype mismatch on Phi-tiny-MoE; not the same bug."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "GGUF gpt-oss support vs a sliding_window issue in Qwen3-MoE; different changes."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44910",
+            "accept": false,
+            "reason": "Different model families and different attention failures: torch.compile/SDPA vs Flash Attention illegal memory access from 3D position_ids."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:43646",
+            "accept": true,
+            "reason": "Both describe the same Transformers 5.0.0 custom-model initialization regression; the notebook report is just an environment-specific instance."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "GGUF support request vs missing Float8 storage type; unrelated."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Non-persistent buffer corruption from from_pretrained vs Qwen2.5-VL rope-index scaling bug; different code paths."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Video StopIteration in Qwen3-vl-embedding vs TRL SFT unpacking failure on Qwen3.5 input embeddings; different bugs."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:43366",
+            "accept": false,
+            "reason": "3d_parallel.py runtime issue vs GGUF gpt-oss support; no shared underlying bug."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "FineGrainedFP8 Qwen3MoE failure vs meta-device loading regression; separate loading paths."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Missing Qwen2_5_VL export vs meta-device loading regression; not the same defect."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "GenerationMixin import error vs Qwen3MoE FP8 loading failure; unrelated."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43994",
+            "accept": true,
+            "reason": "Both reports concern the same SigLIP2 base model producing wrong outputs; one says zero probabilities and the other nonsensical results."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:43950",
+            "accept": false,
+            "reason": "Custom model init regression vs silent non-persistent buffer corruption; different failures."
+          },
+          {
+            "left": "issue:44155",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "AudioFlamingo batched-inference token leak vs Whisper empty-transcription generation failure; unrelated."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Video StopIteration vs mask/tensor shape mismatch; different code paths."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:45030",
+            "accept": false,
+            "reason": "Both are config-loading regressions, but they affect different model configs and do not clearly share the same concrete bug."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO-3 Bert index error vs tokenizer-mapping add-new-model-like failure; not the same underlying issue."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "Both mention loading edge cases, but the Bert zero-sized index error is distinct from the TOKENIZER_MAPPING_NAMES failure."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "device_map/offload loading failure vs mmap OOM workaround on Strix Halo; related to memory, not the same bug."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:39692",
+            "accept": false,
+            "reason": "SigLIP2 bad outputs vs a docs example with model/processor mismatch and quantization failures; same model family, different issue."
+          },
+          {
+            "left": "issue:45313",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Qwen3.5 ZeRO-3 language_model weight loading vs incorrect visual encoder keys in save_pretrained; different directions of failure."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The cluster is mostly conservative, but the summary overstates one relationship: it groups the DeepSpeed ZeRO-3 BertModel index-error reports as if they were a true duplicate/near-duplicate cluster item, while the soft-edge verdicts actually reject those links. Please keep that area separate unless there is stronger evidence. The accepted pairs for the mask-shape IndexError and the custom-model initialization regression are grounded."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 5,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:36010",
+        "issue:39404",
+        "issue:39692",
+        "issue:41720",
+        "issue:42491",
+        "issue:42915",
+        "issue:42947",
+        "issue:43257",
+        "issue:43335",
+        "issue:43404",
+        "issue:43425",
+        "issue:43454",
+        "issue:43644",
+        "issue:43645",
+        "issue:43708",
+        "issue:43720",
+        "issue:43854",
+        "issue:43856",
+        "issue:43873",
+        "issue:43931",
+        "issue:43957",
+        "issue:43994",
+        "issue:44075",
+        "issue:44164",
+        "issue:44291",
+        "issue:44387",
+        "issue:44402",
+        "issue:44451",
+        "issue:44488",
+        "issue:44534",
+        "issue:44661",
+        "issue:44683",
+        "issue:44756",
+        "issue:44910",
+        "issue:44960",
+        "issue:44977",
+        "issue:44991",
+        "issue:45276",
+        "issue:45292",
+        "issue:45335",
+        "issue:45362",
+        "issue:45406",
+        "issue:45488",
+        "issue:45588",
+        "issue:45663",
+        "issue:45704",
+        "issue:45759",
+        "issue:45812"
+      ],
+      "soft_pairs": [
+        "issue:43644|issue:44534",
+        "issue:44960|issue:44977",
+        "issue:45488|issue:45812",
+        "issue:44451|issue:44488",
+        "issue:42491|issue:43720",
+        "issue:43708|issue:44075",
+        "issue:43854|issue:43931",
+        "issue:41720|issue:42915",
+        "issue:42491|issue:45276",
+        "issue:43856|issue:43873",
+        "issue:43257|issue:43854",
+        "issue:39692|issue:43994",
+        "issue:42915|issue:44756",
+        "issue:42491|issue:43931",
+        "issue:43856|issue:44387",
+        "issue:44661|issue:44683",
+        "issue:43644|issue:43645",
+        "issue:44910|issue:45588",
+        "issue:45362|issue:45406",
+        "issue:42491|issue:45292",
+        "issue:44488|issue:44991",
+        "issue:43404|issue:43454",
+        "issue:36010|issue:39404",
+        "issue:45276|issue:45335",
+        "issue:44164|issue:44291",
+        "issue:44402|issue:44488",
+        "issue:39692|issue:44960",
+        "issue:43425|issue:43957",
+        "issue:43335|issue:45663",
+        "issue:42947|issue:45704",
+        "issue:45663|issue:45759",
+        "issue:43873|issue:44387"
+      ],
+      "prepared_review_unit_hash": "def06d239694e469cd2b38d95754c1209b8122706badea2146064315bfff78b6",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37339,
+        "estimated_input_tokens": 9335,
+        "estimated_eval_tokens": 18926
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous. I only see three likely true duplicate pairs: the non-persistent buffer regression, the lm_head weight-tying regression, and the resize_token_embeddings regression. The rest look like different model-specific bugs or unrelated failures.",
+        "confidence": 0.68,
+        "canonical_issue_reason": "Issue 43644 is the cleanest anchor here: it describes the core Transformers v5 non-persistent-buffer regression succinctly and has an almost exact duplicate in 44534.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43644 is the best overall issue candidate because it is the clearest, broad regression report among the items and maps to the strongest duplicate signal in the set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43644",
+            "right": "issue:44534",
+            "accept": true,
+            "reason": "Same v5 regression about non-persistent buffers being filled with junk; wording differs but the bug is the same."
+          },
+          {
+            "left": "issue:44960",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Both involve Qwen/flash-attention style failures, but they describe different models and different failure modes."
+          },
+          {
+            "left": "issue:45488",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different model families and the titles point to different root causes."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both are model-loading complaints, but they name different checkpoints with no clear shared failure path."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "One is a qwen3_moe LoRA compatibility issue across hf versions; the other is a BitNet packed-weight loading bug."
+          },
+          {
+            "left": "issue:43708",
+            "right": "issue:44075",
+            "accept": false,
+            "reason": "Trainer checkpoint math and SGD argument handling are unrelated problems."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different models and different loading errors: GLM-4.7-Flash unit-test failure vs Qwen3-VL shape mismatch."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "A800 cudaErrorAssert during device mapping is not the same bug as qwen3_moe weight conversion with accelerate/deepspeed."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "qwen3_moe hf5 compatibility and Gemma4 token-embedding resizing are unrelated."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Training memory usage and quantized offloading behavior are different issues."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Qwen3 MoE weight conversion and GLM-4.7-Flash loading are separate bugs."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "One is a docs example/model-processor mismatch; the other is an actual bad inference result for SigLIP2."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Qwen3 MoE FP8 failure and mmap-related OOM on Strix Halo are unrelated."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different model families and different loading failures."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "General MoE training memory usage is not the same as int4 quantization causing reserved-memory OOM."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Tokenizer-mapping registration and compiled flex_attention on torch>=2.9 are unrelated."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Both are Transformers v5 regressions, but one is about buffers during init and the other about defining custom models in Jupyter."
+          },
+          {
+            "left": "issue:44910",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Both mention flash attention, but the failure modes are different: packed position_ids vs s_aux=None handling."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Qwen3.5 chat crashes and Gemma4Processor missing _tokenizer are unrelated."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:45292",
+            "accept": false,
+            "reason": "qwen3_moe compatibility is unrelated to resize_token_embeddings behavior."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Two tokenizer-loading reports for different models; no evidence they share the same root cause."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43454",
+            "accept": true,
+            "reason": "Both report missing lm_head weight tying causing bad generation in multimodal model classes; same underlying regression."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:39404",
+            "accept": false,
+            "reason": "ImportError for GenerationMixin and Whisper return_language pipeline regression are unrelated."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45335",
+            "accept": true,
+            "reason": "Same resize_token_embeddings regression: the resized embeddings are not propagated to the model's actual embedding/output layers."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "save/from_pretrained extra_state handling and init_empty_weights unexpected argument handling are different lifecycle bugs."
+          },
+          {
+            "left": "issue:44402",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Different model-loading failures with no clear shared defect."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "SigLIP2 documentation/model-processor problems are unrelated to GLM5/Qwen3.5 generation issues."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Torch 2.10 incompatibility and meta-device model loading failures are separate compatibility problems."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer creation and Gemma4 FSDP2 shared_kv_states issues are unrelated."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "PEFT gradient-checkpointing behavior and an Apex RMSNorm memory leak are different bugs."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Gemma4 FSDP2 attention state rebuilding and Qwen 3.5/3.6 multimodal config unwrapping are unrelated."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Quantization offloading behavior and int4 reserved-memory OOM are not the same concrete bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The three accepted pairs are plausibly the same regressions based on their titles, and the rejected pairs are generally separated by model family or failure mode without overstating evidence. No mergeability issues here since these are issue-only pairs."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 6,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:39404",
+        "issue:41720",
+        "issue:42947",
+        "issue:43257",
+        "issue:43299",
+        "issue:43381",
+        "issue:43425",
+        "issue:43611",
+        "issue:43645",
+        "issue:43646",
+        "issue:43708",
+        "issue:43716",
+        "issue:43828",
+        "issue:43854",
+        "issue:43901",
+        "issue:43906",
+        "issue:43931",
+        "issue:43957",
+        "issue:44155",
+        "issue:44164",
+        "issue:44265",
+        "issue:44291",
+        "issue:44292",
+        "issue:44387",
+        "issue:44402",
+        "issue:44451",
+        "issue:44479",
+        "issue:44509",
+        "issue:44530",
+        "issue:44589",
+        "issue:44756",
+        "issue:44805",
+        "issue:44863",
+        "issue:44877",
+        "issue:44898",
+        "issue:44912",
+        "issue:44928",
+        "issue:44960",
+        "issue:45084",
+        "issue:45216",
+        "issue:45276",
+        "issue:45292",
+        "issue:45335",
+        "issue:45588",
+        "issue:45663",
+        "issue:45684",
+        "issue:45698",
+        "issue:45701"
+      ],
+      "soft_pairs": [
+        "issue:43854|issue:45216",
+        "issue:44387|issue:44756",
+        "issue:44291|issue:44756",
+        "issue:44292|issue:44912",
+        "issue:42947|issue:43381",
+        "issue:45684|issue:45698",
+        "issue:45276|issue:45292",
+        "issue:43381|issue:44928",
+        "issue:44164|issue:44756",
+        "issue:41720|issue:44155",
+        "issue:45588|issue:45663",
+        "issue:43646|issue:43708",
+        "issue:44479|issue:44877",
+        "issue:43854|issue:44960",
+        "issue:43957|issue:44756",
+        "issue:45292|issue:45335",
+        "issue:44402|issue:44451",
+        "issue:43645|issue:44164",
+        "issue:43828|issue:43957",
+        "issue:43425|issue:43828",
+        "issue:43257|issue:44863",
+        "issue:44589|issue:44756",
+        "issue:44479|issue:45663",
+        "issue:43906|issue:44530",
+        "issue:39404|issue:45701",
+        "issue:43611|issue:43645",
+        "issue:43425|issue:43716",
+        "issue:43931|issue:44863",
+        "issue:44530|issue:45084",
+        "issue:43901|issue:44509",
+        "issue:39404|issue:44912",
+        "issue:44863|issue:45663",
+        "issue:44265|issue:44898",
+        "issue:44479|issue:44805",
+        "issue:43299|issue:43931"
+      ],
+      "prepared_review_unit_hash": "610bd284a89b206f1e313a6eaa6c6d2bdee35994e244e06d9e2ae13355bfeec0",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 35,
+        "serialized_chars": 38238,
+        "estimated_input_tokens": 9560,
+        "estimated_eval_tokens": 19376
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a heterogeneous cluster of Transformers regression reports. The only clear duplicate family is the `resize_token_embeddings` regression across Gemma/T5Gemma variants; the rest are distinct model-loading, quantization, checkpointing, or docs issues.",
+        "confidence": 0.74,
+        "canonical_issue_reason": "issue:45292 is the best canonical among the related `resize_token_embeddings` reports because it states the core, model-agnostic failure most cleanly; the other close items look like model-specific variants of the same embedding-resize bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45292 is the best representative issue for the only true mini-cluster here: it is broad enough to cover the Gemma4 and T5Gemma variants while remaining specific to the shared code path.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43854",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both are loading regressions, but they involve different models and failure modes (`GLM-4.7-Flash` test loading vs Qwen3.5 checkpoint save correctness)."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Both mention memory/OOM, but one is int4 CUDA reserved-memory growth and the other is mmap on Strix Halo; different underlying problems."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "`init_empty_weights` argument handling and Strix Halo mmap/OOM are unrelated code paths."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Different issues: float8 storage lookup vs MXFP4 quantization fallback; no shared bug mechanism."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "Both are about gradient checkpointing, but one is about PEFT/LoRA ineffectiveness and the other is about eval-mode restrictions."
+          },
+          {
+            "left": "issue:45684",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Both touch save/load behavior for custom modules, but one is about read-only permissions during save and the other is about loading the wrong module after save."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45292",
+            "accept": true,
+            "reason": "Same underlying `resize_token_embeddings` bug: Gemma4-specific wording in 45276, but both report the same failure to propagate resized embeddings to output-related modules."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:44928",
+            "accept": false,
+            "reason": "Both mention training behavior, but one is a gradient-checkpointing eval-mode issue and the other is an RLHF NaN/SDPA fallback problem."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "`extra_state` save/load handling is unrelated to mmap-induced OOM on Strix Halo."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44155",
+            "accept": false,
+            "reason": "Different models and failure mechanisms: Qwen3 auto device mapping CUDA assert vs AudioFlamingo3 batched inference token/embedding leakage."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "Flash-attention sink-less crash and Gemma-4 FSDP2 shared-kv-state rebuild are different bugs despite both being attention-related."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:43708",
+            "accept": false,
+            "reason": "Custom model initialization in notebooks is unrelated to `resume_from_checkpoint` max_steps calculation."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Both are regression-style loading issues, but one is a video-input regression for Qwen VL variants and the other is strict config blocking `granite_speech`."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "Different GLM generations and different symptoms; not the same loading bug."
+          },
+          {
+            "left": "issue:43957",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "`torch.device('meta')` load failures and Strix Halo mmap/OOM are not the same underlying defect."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45335",
+            "accept": true,
+            "reason": "Both are the same resize-token-embedding regression family: a shared failure to update downstream embedding modules, with different model-specific surfaces."
+          },
+          {
+            "left": "issue:44402",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Both are model/tokenizer-related load problems, but the affected models and root causes differ."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:44164",
+            "accept": false,
+            "reason": "Jupyter/custom model initialization breakage is unrelated to `extra_state` serialization handling."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Both mention dtype issues, but one is autocast on a MoE model and the other is meta-device loading; different code paths."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Torch version incompatibility is too broad and distinct from the specific autocast dtype mismatch."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Both are model loading problems, but one is Qwen3 MoE weight conversion and the other is NemotronH checkpoint compatibility."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Float8 storage lookup and mmap/OOM are unrelated."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "A Qwen VL video-input regression and Gemma-4 FSDP2 attention-state bug are unrelated."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "An isolated reproduction of an older issue and a new PagedAttentionCache crash are not the same bug."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Whisper pipeline `return_language` regression and tokenization changes across versions are unrelated."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Both concern custom model loading in Transformers 5, but the concrete failures differ: `base_model_prefix` handling vs notebook-defined initialization."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43716",
+            "accept": false,
+            "reason": "Torch 2.10 compatibility and Mistral-3 image-preprocessor dtype mismatch are different issues."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Qwen3-VL weight-shape mismatch and NemotronH checkpoint loading are separate model-compatibility problems."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "A cache crash on Qwen3.5 and a template-node compilation error are not the same underlying defect."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "Docs mentioning removed pipeline tasks is unrelated to a separate `return_all_scores` docs/behavior mismatch."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Whisper pipeline language-return behavior and MXFP4 quantization fallback are unrelated."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "NemotronH checkpoint loading and Gemma-4 FSDP2 shared-kv-state handling are different bugs."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "`torch.export.export` with `torch_compilable_check` and Perceiver non-default-resolution handling are unrelated."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Qwen VL video regression and mask-shape IndexError do not share a concrete bug path."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Both involve Qwen3-VL loading, but one is a MoE conversion issue and the other is a checkpoint shape mismatch; not the same bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The only accepted soft edges are the Gemma/T5Gemma resize_token_embeddings pairings, which are well supported by the issue titles and look like the same bug family. The remaining soft edges are rejected with appropriately cautious reasoning, and the overall cluster description does not overstate evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 7,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:32090",
+        "issue:34567",
+        "issue:36032",
+        "issue:39404",
+        "issue:40444",
+        "issue:40990",
+        "issue:42915",
+        "issue:42947",
+        "issue:43317",
+        "issue:43421",
+        "issue:43425",
+        "issue:43452",
+        "issue:43482",
+        "issue:43541",
+        "issue:43582",
+        "issue:43596",
+        "issue:43611",
+        "issue:43632",
+        "issue:43643",
+        "issue:43644",
+        "issue:43646",
+        "issue:43716",
+        "issue:43856",
+        "issue:43883",
+        "issue:43994",
+        "issue:44079",
+        "issue:44164",
+        "issue:44206",
+        "issue:44220",
+        "issue:44451",
+        "issue:44479",
+        "issue:44488",
+        "issue:44530",
+        "issue:44589",
+        "issue:44671",
+        "issue:44683",
+        "issue:44756",
+        "issue:44811",
+        "issue:44849",
+        "issue:44918",
+        "issue:44987",
+        "issue:44991",
+        "issue:45081",
+        "issue:45084",
+        "issue:45325",
+        "issue:45584",
+        "issue:45685",
+        "issue:45759"
+      ],
+      "soft_pairs": [
+        "issue:43425|issue:44918",
+        "issue:43856|issue:44756",
+        "issue:44589|issue:45084",
+        "issue:40990|issue:42947",
+        "issue:43425|issue:45685",
+        "issue:39404|issue:44671",
+        "issue:43994|issue:44079",
+        "issue:44164|issue:45759",
+        "issue:43596|issue:44683",
+        "issue:36032|issue:43452",
+        "issue:44671|issue:44991",
+        "issue:44991|issue:45081",
+        "issue:32090|issue:39404",
+        "issue:44164|issue:44479",
+        "issue:44488|issue:45081",
+        "issue:43582|issue:43632",
+        "issue:40990|issue:43421",
+        "issue:43883|issue:44488",
+        "issue:43611|issue:43644",
+        "issue:44206|issue:44220",
+        "issue:43716|issue:45084",
+        "issue:44488|issue:44987",
+        "issue:34567|issue:40444",
+        "issue:43643|issue:44530",
+        "issue:44811|issue:45584",
+        "issue:43883|issue:44991",
+        "issue:43644|issue:43646",
+        "issue:44451|issue:44987",
+        "issue:36032|issue:42915",
+        "issue:43541|issue:45084",
+        "issue:43452|issue:43482",
+        "issue:42915|issue:44530",
+        "issue:43317|issue:43482",
+        "issue:44849|issue:45325"
+      ],
+      "prepared_review_unit_hash": "363ee19632ccbfa55c5bee286247786ae740e65a6b43c0acdf4799684a96776f",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 37733,
+        "estimated_input_tokens": 9434,
+        "estimated_eval_tokens": 19124
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is heterogeneous: most pairs share only broad subsystem/model-family similarity, not the same concrete bug or change. I reject all soft edges; there are no PRs to compare.",
+        "confidence": 0.92,
+        "canonical_issue_reason": "No safe canonical issue; the items are mostly unrelated regressions with only broad topical overlap.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43452 is the closest broad representative of a loader/GGUF subset, but it does not unify the cluster as a whole.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43425",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Torch version incompatibility vs a Qwen3.5 TRL embedding-unpack failure; different code paths."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Both mention memory pressure, but one is Qwen3 MoE training memory usage and the other is Strix Halo mmap OOM; different causes."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Unrelated failures: missing Float8 storage vs template-node compilation error."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:42947",
+            "accept": false,
+            "reason": "Perplexity/eval regression is not the same bug as gradient-checkpointing being ineffective."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Torch incompatibility and an MPS histogram kernel gap are separate platform issues."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Whisper pipeline behavior vs CamemBERT MLM predictions are unrelated model bugs."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "One is a SigLIP2 inference regression; the other is a ModelOutput key assignment bug."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Save/load extra_state handling is unrelated to composite Qwen config unwrapping."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "DeepSpeed zero3/BertModel indexing and torch>=2.9 flex_attention compilation are different failures."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "Tokenizer special-token method conflict is not the same as GGUF from_pretrained loading breakage."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Incorrect CamemBERT predictions vs a tokenizer-loading regression for EMBEDDIA/est-roberta."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both involve tokenizer loading, but one is generic version failure and the other is a specific Mistral regex/backend_tokenizer crash."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:39404",
+            "accept": false,
+            "reason": "Trainer GPU broadcast TypeError and Whisper return_language regression are unrelated."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "extra_state save/load handling is unrelated to multimodal video-input regression."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both are model-loading issues, but for different models and different failure mechanisms."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:43632",
+            "accept": false,
+            "reason": "Apple Silicon allocator warmup TypeError is unrelated to the _is_hf_initialized flag regression."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:43421",
+            "accept": false,
+            "reason": "Perplexity anomaly and tokenizer post-processor updates are different bugs."
+          },
+          {
+            "left": "issue:43883",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Missing all_tied_weights_keys in Molmo is unrelated to loading cjvt/sleng-bert."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43644",
+            "accept": false,
+            "reason": "Both are v5 regressions, but one is base_model_prefix loading and the other is non-persistent buffer initialization; not the same bug."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Related audio area, but the concrete failures differ: unsupported center arg vs _torch_extract_fbank_features issue."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Mistral image dtype mismatch and template-node compilation are unrelated."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Both are model-loading failures in recent versions, but they affect different models and likely different loader paths."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "TrainerState token counting and Qwen2.5-VL multi-image IterableDataset failure are unrelated."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "AutoConfig trust_remote_code field loss is unrelated to PagedAttentionCache linear_attention crashes."
+          },
+          {
+            "left": "issue:44811",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Both are Whisper-related, but one ignores skip_special_tokens and the other fails on empty transcription after align_special_tokens."
+          },
+          {
+            "left": "issue:43883",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Molmo tied-weights metadata and tokenizer loading for est-roberta are unrelated."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43646",
+            "accept": false,
+            "reason": "Both are v5 regressions, but they target different initialization paths and different symptoms."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Both are model-loading problems, but for different models and no shared concrete code path is evident."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "T5 tokenizer special-token conflict is unrelated to Qwen3Moe FineGrainedFP8Config failure."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Torch Dynamo grouped_mm tracing failure is not the same as a non-template-node compilation error."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Both concern GGUF loading, but the evidence only shows broad similarity; no shared concrete failure mode is established."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "device_map/offload dequantization and GGUF model loading are different loader failures."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Qwen3.5 hidden-state output bug and Qwen2.5-VL rope indexing bug are different code paths."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 8,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:36032",
+        "issue:40990",
+        "issue:41553",
+        "issue:42175",
+        "issue:42915",
+        "issue:43262",
+        "issue:43296",
+        "issue:43317",
+        "issue:43322",
+        "issue:43377",
+        "issue:43404",
+        "issue:43425",
+        "issue:43531",
+        "issue:43540",
+        "issue:43582",
+        "issue:43611",
+        "issue:43632",
+        "issue:43645",
+        "issue:43646",
+        "issue:43742",
+        "issue:43819",
+        "issue:43856",
+        "issue:43994",
+        "issue:44062",
+        "issue:44220",
+        "issue:44265",
+        "issue:44291",
+        "issue:44403",
+        "issue:44451",
+        "issue:44466",
+        "issue:44479",
+        "issue:44534",
+        "issue:44589",
+        "issue:44610",
+        "issue:44683",
+        "issue:44756",
+        "issue:44811",
+        "issue:44877",
+        "issue:44987",
+        "issue:44991",
+        "issue:45081",
+        "issue:45084",
+        "issue:45127",
+        "issue:45216",
+        "issue:45325",
+        "issue:45357",
+        "issue:45507",
+        "issue:45542"
+      ],
+      "soft_pairs": [
+        "issue:43404|issue:44466",
+        "issue:36032|issue:43645",
+        "issue:45084|issue:45507",
+        "issue:43262|issue:43377",
+        "issue:43646|issue:44534",
+        "issue:43994|issue:44610",
+        "issue:44987|issue:44991",
+        "issue:43322|issue:43540",
+        "issue:44403|issue:45081",
+        "issue:43377|issue:43819",
+        "issue:43645|issue:44534",
+        "issue:45216|issue:45357",
+        "issue:44451|issue:45081",
+        "issue:42915|issue:45084",
+        "issue:44265|issue:44610",
+        "issue:44220|issue:44683",
+        "issue:44987|issue:45081",
+        "issue:41553|issue:43582",
+        "issue:43404|issue:45127",
+        "issue:43632|issue:44756",
+        "issue:43296|issue:43531",
+        "issue:44062|issue:44987",
+        "issue:42915|issue:44291",
+        "issue:36032|issue:43742",
+        "issue:43531|issue:44877",
+        "issue:43611|issue:44534",
+        "issue:42915|issue:44589",
+        "issue:44479|issue:45325",
+        "issue:40990|issue:44811",
+        "issue:43317|issue:43856",
+        "issue:43425|issue:44403",
+        "issue:42175|issue:45542"
+      ],
+      "prepared_review_unit_hash": "d57f53eaf50ea9fffc55d88a39fcfd4115cf744993b858d946c520e0ee95d9e0",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37208,
+        "estimated_input_tokens": 9302,
+        "estimated_eval_tokens": 18860
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The soft pairs are mostly superficially similar but describe different bugs, regressions, or models. I would not deduplicate any of them based on the available evidence.",
+        "confidence": 0.91,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#44403 is the broadest representative issue in the set, but the cluster is too heterogeneous for a true canonical duplicate target.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43404",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Related to tied weights / lm_head serialization, but the reported failures differ and are not clearly the same code-path bug."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Tokenizer attribute conflict vs custom model init in notebooks; unrelated failure modes."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Different error surfaces and different model code paths; no evidence of the same underlying bug."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43377",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate defaulting and MIMI padding-mask batching are distinct issues."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Custom model initialization and junk-filled non-persistent buffers are separate v5 regressions."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Different models and different symptoms; one is bad outputs, the other is processor size mismatch."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are load failures, but for different models with no shared concrete bug."
+          },
+          {
+            "left": "issue:43322",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Segfault while loading Llava Next is unrelated to Qwen3OmniMoe video ValueError."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "General loading noise vs Mistral regex patch crash; not the same underlying defect."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Padding-mask batching mismatch and DAC STE mismatch are unrelated model bugs."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Notebook custom-model init regression is not the same as non-persistent buffer corruption."
+          },
+          {
+            "left": "issue:45216",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Both concern Qwen3.5 save_pretrained regressions, but the specific failures differ enough that they should stay separate."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Different models and different tokenizer/loading failures."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config failure is unrelated to template-node compilation errors."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "torch.export compilation failure vs processor image-size regression; not the same bug."
+          },
+          {
+            "left": "issue:44220",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Audio feature extraction issue and compiled flex_attention failure have no concrete overlap."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Different affected models and different loading paths."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "Tokenizer loading error and Apple Silicon warmup TypeError are unrelated."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45127",
+            "accept": false,
+            "reason": "Both involve tied weights, but one is missing weight tying and the other is LoRA merge collapse; different underlying problems."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Initialization flag regression and Strix Halo mmap OOM are unrelated."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "PaddleOCR-VL load failure and Qwen3-MoE sliding_window bug do not share a concrete code path."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Tokenizer AddedToken argument collision is distinct from loading a specific model failure."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "FP8 config failure vs init_empty_weights/_is_hf_initialized argument error; separate regressions."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "T5 tokenizer load conflict and MobileLLM key error are unrelated."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding window logic and strict config loading are different bugs."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "base_model_prefix loading breakage and junk non-persistent buffers are separate v5 issues."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config and Float8 storage lookup error are different failures."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Video input regression and still-image rope-index scaling bug are not the same issue."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Perplexity regression and Whisper batch_decode skip_special_tokens bug are unrelated."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "device_map/offload dequantized loading failure is not the same as Qwen3 MoE memory usage during training."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44403",
+            "accept": false,
+            "reason": "Torch version incompatibility and loading-noise report do not describe the same concrete bug."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Missing TensorFlow in torch extra and tensorboard-only backend error are related at a high level, but 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": 2,
+      "component_count": 16,
+      "review_unit_index": 9,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:33666",
+        "issue:36032",
+        "issue:40444",
+        "issue:41720",
+        "issue:42175",
+        "issue:42915",
+        "issue:43296",
+        "issue:43377",
+        "issue:43421",
+        "issue:43454",
+        "issue:43482",
+        "issue:43525",
+        "issue:43531",
+        "issue:43611",
+        "issue:43632",
+        "issue:43644",
+        "issue:43646",
+        "issue:43792",
+        "issue:43818",
+        "issue:43828",
+        "issue:43881",
+        "issue:43957",
+        "issue:43994",
+        "issue:44079",
+        "issue:44265",
+        "issue:44291",
+        "issue:44403",
+        "issue:44462",
+        "issue:44466",
+        "issue:44488",
+        "issue:44534",
+        "issue:44568",
+        "issue:44589",
+        "issue:44683",
+        "issue:44829",
+        "issue:44898",
+        "issue:44991",
+        "issue:45020",
+        "issue:45072",
+        "issue:45081",
+        "issue:45325",
+        "issue:45357",
+        "issue:45381",
+        "issue:45479",
+        "issue:45491",
+        "issue:45538",
+        "issue:45663",
+        "issue:45812"
+      ],
+      "soft_pairs": [
+        "issue:44079|issue:45663",
+        "issue:43818|issue:45357",
+        "issue:43421|issue:44568",
+        "issue:36032|issue:41720",
+        "issue:44462|issue:45812",
+        "issue:43632|issue:44534",
+        "issue:43792|issue:44589",
+        "issue:44683|issue:45020",
+        "issue:44403|issue:44991",
+        "issue:33666|issue:40444",
+        "issue:43482|issue:45538",
+        "issue:44829|issue:45479",
+        "issue:40444|issue:45325",
+        "issue:43482|issue:43525",
+        "issue:44291|issue:44589",
+        "issue:43377|issue:45491",
+        "issue:41720|issue:45081",
+        "issue:43632|issue:45072",
+        "issue:44265|issue:45072",
+        "issue:42175|issue:43296",
+        "issue:43611|issue:43646",
+        "issue:43531|issue:44291",
+        "issue:44079|issue:44534",
+        "issue:43881|issue:44488",
+        "issue:43994|issue:44898",
+        "issue:43454|issue:44466",
+        "issue:43828|issue:44683",
+        "issue:42915|issue:43482",
+        "issue:43792|issue:43957",
+        "issue:44079|issue:45072",
+        "issue:40444|issue:45381",
+        "issue:43644|issue:44079",
+        "issue:45325|issue:45381"
+      ],
+      "prepared_review_unit_hash": "ae9c9d0ee673880d08295af1d16486a30457b4910278d432c8607555a1f94cbe",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 37885,
+        "estimated_input_tokens": 9472,
+        "estimated_eval_tokens": 19200
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a loose mix of mostly unrelated Transformers regression reports. The soft pairs share broad themes like model loading, tokenizer behavior, or v5 regressions, but they do not look like the same underlying bug/change, so I would not merge any of them as duplicates.",
+        "confidence": 0.54,
+        "canonical_issue_reason": "issue:43611 is the cleanest, most concrete representative here: a specific Transformers v5 model-loading regression with a well-defined failure mode. It is more actionable than the broader or more model-specific reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43611 is the best single issue to keep as a reference point because it is precise, technical, and centered on a core code path rather than a vague symptom.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44079",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "Different bugs: ModelOutput key assignment vs Gemma-4 FSDP2 shared_kv_states handling."
+          },
+          {
+            "left": "issue:43818",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different areas: Video-LLaVA tower initialization/weight sharing vs Qwen3.5 save_pretrained visual-key serialization."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both concern special tokens, but one is runtime post-processor updates and the other is BOS/EOS not being added; not the same concrete bug."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Completely unrelated: T5Tokenizer method-name conflict vs Qwen3 auto device mapping CUDA assert."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both involve AutoTokenizer, but one ignores tokenizer.json while the other yields wrong IDs for Granite models; different failure modes."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Related to v5 initialization behavior, but one is _is_hf_initialized and the other is non-persistent buffer contents; too different to treat as duplicates."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Different model/runtime failures: Whisper loading vs Float8 storage lookup error."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "One is a torch flex_attention compilation regression, the other is a broad remote_code loading breakage; not the same bug."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Generic loading noise vs tokenizer loading regression for a specific repo; no shared concrete code path."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "Both are Qwen-VL training issues, but multi-GPU training and iterable-dataset multi-image failure are distinct."
+          },
+          {
+            "left": "issue:43482",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "GGUF/v5 loading failure is unrelated to CLIPTokenizer's extreme model_max_length default."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Both affect training, but flash_attention_3 degenerate training and num_labels=1 zero-loss classification are different problems."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Different Qwen2.5-VL issues: iterable dataset multi-image failure vs still-image rope index scaling bug."
+          },
+          {
+            "left": "issue:43482",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "GGUF loading failure vs Llama4Config missing pad_token_id; unrelated."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Different v5 regressions: unexpected _is_hf_initialized argument vs missing Float8 storage object."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Both are batching/shape-related, but MIMI padding-mask mismatch and Gemma3 sliding-window NaNs are separate code paths."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping CUDA assert is unrelated to Mistral tokenizer regex patching."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Both are v5 regressions, but they concern different mechanisms: initialization flag handling vs bf16 dtype mismatches."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "torch.export/torch_compilable_check failure is unrelated to bf16 inference dtype mismatches."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43296",
+            "accept": false,
+            "reason": "Backend installation packaging issue vs PaddleOCR-VL loading failure; no common bug."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43646",
+            "accept": false,
+            "reason": "Both are model-loading/init regressions in v5, but the concrete failures differ enough that they do not look like the same bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding-window behavior is unrelated to init_empty_weights/_is_hf_initialized argument handling."
+          },
+          {
+            "left": "issue:44079",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "ModelOutput key assignment and non-persistent buffer junk are different internal behaviors."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Different models and failure modes: glm-4v-9b loading vs sleng-bert loading."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Different vision-model regressions: SigLIP2 nonsensical outputs vs Perceiver resolution/interpolation failure."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "AyaVision weight tying and lm_head serialization are related only broadly; they are not the same concrete issue."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Autocast dtype mismatch for Phi-tiny-MoE is distinct from compiled flex_attention failing on torch>=2.9."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "FineGrainedFP8Config failure for Qwen3Moe is unrelated to Qwen2.5-GGUF loading in v5."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Whisper loading failure and torch.device('meta') loading breakage are different code paths."
+          },
+          {
+            "left": "issue:44079",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "ModelOutput key assignment is unrelated to bf16 dtype mismatches in SwitchTransformers/TimmWrapperModel."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both are Qwen2.5-VL-related, but multi-image iterable dataset failure and video position-id bug are different problems."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Non-persistent buffer initialization and ModelOutput key assignment are unrelated regressions."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both involve Qwen2.5-VL position handling, but still-image temporal scaling and video vision_position_ids are distinct."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 10,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:35707",
+        "issue:39692",
+        "issue:40444",
+        "issue:40990",
+        "issue:41720",
+        "issue:43262",
+        "issue:43296",
+        "issue:43317",
+        "issue:43377",
+        "issue:43404",
+        "issue:43454",
+        "issue:43526",
+        "issue:43550",
+        "issue:43572",
+        "issue:43606",
+        "issue:43632",
+        "issue:43701",
+        "issue:43720",
+        "issue:43746",
+        "issue:43792",
+        "issue:43819",
+        "issue:43873",
+        "issue:43881",
+        "issue:43883",
+        "issue:44186",
+        "issue:44265",
+        "issue:44291",
+        "issue:44387",
+        "issue:44423",
+        "issue:44466",
+        "issue:44610",
+        "issue:44617",
+        "issue:44683",
+        "issue:44877",
+        "issue:44898",
+        "issue:44977",
+        "issue:44991",
+        "issue:45072",
+        "issue:45325",
+        "issue:45357",
+        "issue:45406",
+        "issue:45563",
+        "issue:45584",
+        "issue:45632",
+        "issue:45684",
+        "issue:45701",
+        "issue:45759",
+        "issue:45800"
+      ],
+      "soft_pairs": [
+        "issue:44387|issue:44617",
+        "issue:43296|issue:45800",
+        "issue:35707|issue:40444",
+        "issue:43526|issue:43883",
+        "issue:44291|issue:44877",
+        "issue:43606|issue:44265",
+        "issue:43317|issue:43873",
+        "issue:43881|issue:44991",
+        "issue:43701|issue:45325",
+        "issue:44186|issue:44265",
+        "issue:43632|issue:45563",
+        "issue:43606|issue:43819",
+        "issue:43262|issue:43746",
+        "issue:44610|issue:45325",
+        "issue:43454|issue:43550",
+        "issue:43377|issue:44610",
+        "issue:43454|issue:45072",
+        "issue:43377|issue:44898",
+        "issue:43572|issue:45759",
+        "issue:40990|issue:45584",
+        "issue:39692|issue:40444",
+        "issue:44466|issue:45357",
+        "issue:43296|issue:44991",
+        "issue:43792|issue:44683",
+        "issue:43454|issue:43720",
+        "issue:43873|issue:45701",
+        "issue:41720|issue:44977",
+        "issue:45632|issue:45684",
+        "issue:43404|issue:45357",
+        "issue:44423|issue:45406"
+      ],
+      "prepared_review_unit_hash": "a4ff4c3a288a2f6276d583d4b9f23c1c8e3568357c3b477e0c43db7db8834deb",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36898,
+        "estimated_input_tokens": 9225,
+        "estimated_eval_tokens": 18706
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "All candidate soft links are false positives: they point to distinct bugs spanning generation, loading, quantization, processors, and model-specific regressions. No pull requests are present. The broadest and most representative issue in this set is the open offloading/quantization report.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "issue:43873 is the best cluster representative because it is open, has the most discussion/inbound references, and covers a broadly relevant runtime problem rather than a narrow model-specific edge case.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43873 is the strongest overall issue to anchor the cluster because it is active and widely applicable; the other issues are more isolated, model-specific regressions or documentation/example problems.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44387",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Both mention OOM, but one is a Transformers v5 int4 quantization memory regression and the other is a SAM3Video model-specific memory issue; different failure modes and code paths."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "PaddleOCR-VL loading in vLLM and a torch/Transformers version incompatibility are unrelated problems despite both being load failures."
+          },
+          {
+            "left": "issue:35707",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "Progressive generation with inputs_embeds/past_key_values is a generation API bug; Qwen2.5-VL iterable dataset with multiple images is a multimodal finetuning failure."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction and Molmo tied-weights serialization are completely different subsystems."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Both affect loading in v5, but one is init_empty_weights/_is_hf_initialized handling and the other is strict config parsing for granite_speech."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44265",
+            "accept": false,
+            "reason": "CPU offload device mismatch in suno/bark-small is not the same as torch.export failing on torch_compilable_check models."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Both involve offloading, but one is a specific dequantized-model load failure with device_map=auto while the other is a broader quantization/offloading behavior report."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "glm-4v-9b model loading and tokenizer loading for EMBEDDIA/est-roberta are unrelated."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "resume_from_checkpoint key mismatch is a training/checkpoint issue, while 45325 is a Qwen2.5-VL rope/position-id scaling bug."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44265",
+            "accept": false,
+            "reason": "LayoutLMv2Tokenizer NER/padding crashes are unrelated to torch.export and torch_compilable_check."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "_is_hf_initialized breakage in v5 and a stale generate warning for num_return_sequences do not share the same bug."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "CPU offload device mismatch and DAC.from_latents vs forward mismatch are different model behaviors and code paths."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate defaults and GraniteSpeech PEFT checkpoint loading are unrelated."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "OmDet-Turbo input-size mismatch and Qwen2.5-VL temporal position-id scaling are different multimodal regressions."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "lm_head weight tying in AyaVisionConfig is unrelated to Bamba-9B-v2 torch.compile+SDPA failure."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "MIMI batched-vs-single padding-mask mismatch and OmDet-Turbo processor size mismatch are different bugs."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Weight-tying in AyaVisionConfig is not the same as dtype mismatches in SwitchTransformers/TimmWrapperModel bfloat16 inference."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "MIMI padding-mask support and Perceiver interpolation at non-default resolution are separate issues."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Missing pad_token_idx in StableLmConfig and unwraping text_config for composite Qwen multimodal configs are unrelated config/model-loading issues."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "High perplexity on GPT-OSS-20B and Whisper empty-transcription generation failure are different generation/evaluation problems."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "SigLIP2 docs example errors and Qwen2.5-VL multi-image iterable dataset failures are unrelated."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Serialization differences for lm_head.weight by device and incorrect visual-encoder keys in Qwen3.5 save_pretrained are distinct save/load regressions."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "PaddleOCR-VL/vLLM loading and tokenizer loading for est-roberta are unrelated failures."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Whisper runtime failure and torch>=2.9 compiled flex_attention failure are different issues."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "AyaVision lm_head tying and BitNet packed-weight unpacking are unrelated model-state loading bugs."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Quantization/offloading behavior and version-dependent tokenization changes do not describe the same underlying bug."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "A800 auto device mapping cudaErrorAssert and Qwen3.5 flash-attention generation failure are different Qwen regressions."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:45684",
+            "accept": false,
+            "reason": "Cache-path collision for trust_remote_code and read-only permission propagation in save_pretrained are unrelated file-handling bugs."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "lm_head tying in Mistral3ForConditionalGeneration and incorrect visual encoder keys in Qwen3.5 save_pretrained are different regressions."
+          },
+          {
+            "left": "issue:44423",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Continuous-batching multimodal crash and Gemma4Processor missing _tokenizer are separate serve/processor issues."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 11,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:34634",
+        "issue:35707",
+        "issue:36010",
+        "issue:36331",
+        "issue:40990",
+        "issue:41720",
+        "issue:42617",
+        "issue:42915",
+        "issue:43295",
+        "issue:43316",
+        "issue:43317",
+        "issue:43322",
+        "issue:43421",
+        "issue:43450",
+        "issue:43526",
+        "issue:43531",
+        "issue:43540",
+        "issue:43550",
+        "issue:43572",
+        "issue:43577",
+        "issue:43606",
+        "issue:43708",
+        "issue:43720",
+        "issue:43749",
+        "issue:43782",
+        "issue:43873",
+        "issue:43881",
+        "issue:44186",
+        "issue:44292",
+        "issue:44361",
+        "issue:44387",
+        "issue:44479",
+        "issue:44534",
+        "issue:44610",
+        "issue:44671",
+        "issue:44743",
+        "issue:44756",
+        "issue:44877",
+        "issue:44898",
+        "issue:44913",
+        "issue:44918",
+        "issue:44991",
+        "issue:45072",
+        "issue:45081",
+        "issue:45161",
+        "issue:45701",
+        "issue:45704",
+        "issue:45812"
+      ],
+      "soft_pairs": [
+        "issue:34634|issue:35707",
+        "issue:44186|issue:45072",
+        "issue:43577|issue:43708",
+        "issue:35707|issue:41720",
+        "issue:43526|issue:43881",
+        "issue:43873|issue:44756",
+        "issue:43749|issue:44534",
+        "issue:44877|issue:44991",
+        "issue:42617|issue:42915",
+        "issue:43450|issue:45701",
+        "issue:43550|issue:43720",
+        "issue:36331|issue:42915",
+        "issue:44610|issue:44898",
+        "issue:44186|issue:44610",
+        "issue:36331|issue:42617",
+        "issue:44387|issue:45704",
+        "issue:44479|issue:44918",
+        "issue:43316|issue:44913",
+        "issue:42915|issue:43317",
+        "issue:43550|issue:43606",
+        "issue:41720|issue:43322",
+        "issue:45081|issue:45812",
+        "issue:40990|issue:45701",
+        "issue:43295|issue:45701",
+        "issue:41720|issue:45161",
+        "issue:43531|issue:43572",
+        "issue:44671|issue:44743",
+        "issue:44292|issue:44479",
+        "issue:43421|issue:45812",
+        "issue:41720|issue:42617",
+        "issue:36010|issue:42617",
+        "issue:43540|issue:43782",
+        "issue:44186|issue:44361"
+      ],
+      "prepared_review_unit_hash": "7b2e6f21c039293909c84b1bbb54e3ddd2d3fe20d8e03a3f4b6ac7dccc9783c5",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 37360,
+        "estimated_input_tokens": 9340,
+        "estimated_eval_tokens": 18936
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The set is highly heterogeneous: most issues describe unrelated model-specific failures, tokenizer regressions, offloading/memory problems, or trainer/config edge cases. No soft pair looks like the same concrete bug or change, so there is no strong duplicate cluster here.",
+        "confidence": 0.34,
+        "canonical_issue_reason": "There is no true canonical duplicate for this cluster; if a representative issue is needed, issue 45812 is the best anchor because it is a concrete, recent tokenizer regression with broad impact across Granite models.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45812 is the strongest representative because it has a clear failure mode, a broad regression scope, and enough specificity to act as an anchor if this cluster must be summarized by one issue.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:34634",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "Different subsystems and failure modes: Bark voice presets vs generation with inputs_embeds/past_key_values."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Tokenizer/NER padding crash and bfloat16 dtype mismatch are unrelated bugs."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43708",
+            "accept": false,
+            "reason": "BLIP2 dtype persistence and Trainer max_steps calculation are separate training/loading issues."
+          },
+          {
+            "left": "issue:35707",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Progressive generation cache handling is unrelated to Qwen3 auto device mapping CUDA asserts."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction and glm-4v-9b loading failure are not the same bug."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Both involve memory/offload themes, but one is quantization offloading and the other is mmap on Strix Halo; different root causes."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "FSDP CPU RAM efficient loading and non-persistent buffer junk are distinct v5 regressions."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are loading-related, but one is strict config parsing and the other is tokenizer loading for a specific model."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "3d_parallel.py execution failure and Qwen3MoE FP8 config failure are unrelated."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Video batch shape bugs and tokenizer version-dependent behavior are not the same issue."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "One is a torch.compile/SDPA problem, the other is accelerate loading of packed BitNet weights."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Trainer compute_loss signature change is unrelated to Qwen3MoE FP8 loading."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Both concern image preprocessing sizes, but they affect different models and different failure points."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "LayoutLMv2 tokenizer crashes and OmDet-Turbo image size mismatch are unrelated."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "CustomTrainer API breakage and a 3d_parallel.py runtime failure do not share a bug."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Both are memory-related, but one is int4 quantization reserved-memory growth and the other is an apex RMSNorm leak."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Qwen video-input regression and Qwen3.5 embedding unpacking under TRL SFT are different code paths."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Different config regressions: Gemma3TextConfig API mismatch vs GPTNeoX rotary_pct reload behavior."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43317",
+            "accept": false,
+            "reason": "Qwen3MoE FP8 failure and dequantized model offload failure are different loading paths."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43606",
+            "accept": false,
+            "reason": "torch.compile with SDPA on Bamba and Bark CPU-offload device mismatch are unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43322",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping CUDA assert and Llava Next segfault are not the same bug."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but one is Mistral regex patching and the other is Granite token IDs; different concrete failures."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Perplexity on gpt-oss-20b and a generic tokenization change report are not the same defect."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Processor/tokenizer API regression and version-dependent tokenization changes are related only broadly, not as one concrete bug."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Both involve Qwen3/GPT-OSS MoE parallelism, but the reported failures and code paths differ."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Sliding-window behavior in Qwen3-MoE is unrelated to StableLmConfig missing pad_token_idx."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "CamemBERT masked LM predictions and Qwen modular recurrent-state resets are unrelated."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "NVFP4 execution failure and Qwen video-input regression are different issues."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Runtime special-token post-processor updates and Granite token ID regression are both tokenizer-related but not the same bug."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "Auto device mapping CUDA assert and 3d_parallel.py execution failure are unrelated."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "GenerationMixin import error and 3d_parallel.py failure do not share a root cause."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:43782",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video ValueError and Qwen3VL weight_only loading error are different multimodal loading paths."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Tokenizer crashes happen in both, but on different tokenizers/models and with different task triggers."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 12,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:30064",
+        "issue:33666",
+        "issue:34567",
+        "issue:34634",
+        "issue:35707",
+        "issue:39692",
+        "issue:41720",
+        "issue:42886",
+        "issue:42915",
+        "issue:43322",
+        "issue:43381",
+        "issue:43388",
+        "issue:43421",
+        "issue:43531",
+        "issue:43550",
+        "issue:43606",
+        "issue:43653",
+        "issue:43720",
+        "issue:43746",
+        "issue:43844",
+        "issue:43873",
+        "issue:43927",
+        "issue:44062",
+        "issue:44368",
+        "issue:44451",
+        "issue:44492",
+        "issue:44509",
+        "issue:44514",
+        "issue:44545",
+        "issue:44589",
+        "issue:44610",
+        "issue:44779",
+        "issue:44843",
+        "issue:44898",
+        "issue:44987",
+        "issue:44991",
+        "issue:45042",
+        "issue:45059",
+        "issue:45072",
+        "issue:45357",
+        "issue:45406",
+        "issue:45464",
+        "issue:45588",
+        "issue:45701",
+        "issue:45704",
+        "issue:45759",
+        "issue:45812"
+      ],
+      "soft_pairs": [
+        "issue:45406|issue:45464",
+        "issue:41720|issue:43873",
+        "issue:30064|issue:34634",
+        "issue:43421|issue:45701",
+        "issue:44492|issue:44509",
+        "issue:43388|issue:45059",
+        "issue:43606|issue:43720",
+        "issue:44779|issue:45812",
+        "issue:42915|issue:43322",
+        "issue:43550|issue:43653",
+        "issue:44514|issue:44545",
+        "issue:43531|issue:45759",
+        "issue:43606|issue:43653",
+        "issue:43421|issue:43927",
+        "issue:34567|issue:35707",
+        "issue:44368|issue:45357",
+        "issue:41720|issue:43746",
+        "issue:43873|issue:45704",
+        "issue:44451|issue:44991",
+        "issue:44898|issue:45072",
+        "issue:44062|issue:44451",
+        "issue:33666|issue:35707",
+        "issue:44987|issue:45042",
+        "issue:45588|issue:45759",
+        "issue:43421|issue:43653",
+        "issue:43531|issue:44589",
+        "issue:42886|issue:44843",
+        "issue:39692|issue:44610",
+        "issue:43381|issue:43844"
+      ],
+      "prepared_review_unit_hash": "7118f06e0cf11f114203211c5b3b55c068014115eb8d7a31a3bec03ee0b75800",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 35939,
+        "estimated_input_tokens": 8985,
+        "estimated_eval_tokens": 18226
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most items are unrelated. The only clear duplicate pairs are the offline-tokenizer bug (42886 \u2194 44843) and the exact Qwen2_5_VLProcessor padding crash (44514 \u2194 44545). No PRs are present.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "Issue 44843 is the strongest canonical issue: it names the exact HF_HUB_OFFLINE/model_info regression, and 42886 is the broader user-facing symptom of the same bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44843 is the best global issue candidate because it is the most specific root-cause report and subsumes the offline cache failure described in 42886.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45406",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Different components and failures: Gemma4Processor missing _tokenizer vs Qwen3.5 streaming inference crash."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Both are device/offload related, but they describe different bugs and code paths (Qwen3 auto device map assert vs quantized offloading behavior)."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:34634",
+            "accept": false,
+            "reason": "Unrelated processor bugs: void segmentation maps vs BarkProcessor voice_preset handling."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Runtime special-token post-processing bug vs a broad tokenization regression claim; not the same concrete issue."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "A typo in cache strategy docs/code wording is unrelated to removed pipeline-task documentation."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:45059",
+            "accept": false,
+            "reason": "Distributed metric gathering label truncation is unrelated to SAM3 text/box behavior."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Different failure modes and models: CPU offload device mismatch vs packed-weight unpacking during accelerate loading."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different models and symptoms; not clearly the same bug."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43322",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config failure is unrelated to Llava Next segmentation fault on load."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "torch.compile SDPA bug is unrelated to BigBirdTokenizer special-token registration."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:44545",
+            "accept": true,
+            "reason": "Same title and same crash: Qwen2_5_VLProcessor.apply_chat_template on batched input with padding=False."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Sliding-window behavior in Qwen3-MoE is a different model/runtime issue than composite config unwrapping in from_config."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "CPU offload device mismatch and tokenizer mask-token registration are unrelated."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Post-processor refresh on special-token edits is unrelated to DiaConfig losing custom token IDs on save/load."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "TrainerState token counting and progressive generation with inputs_embeds/past_key_values are different bugs."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Warning about tie_word_embeddings is unrelated to incorrect visual-encoder keys saved by Qwen3.5."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Auto device mapping cudaErrorAssert and PEFT-adapter local checkpoint loading are different code paths."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Quantization/offloading behavior is unrelated to the T5 apex RMSNorm memory leak."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are model loading failures, but the models and failure details differ; no evidence of the same root cause."
+          },
+          {
+            "left": "issue:44898",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Perceiver resolution handling and bfloat16 dtype mismatch are unrelated issues."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "AddedToken special-argument TypeError is unrelated to ScandiBERT loading failure."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "Multi-GPU training for Qwen2-VL is unrelated to progressive generation inputs_embeds/past_key_values."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "A model-loading regression for one repo is unrelated to PIL backend image processors requiring torchvision."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Flash-attention crash on s_aux=None is unrelated to from_config text_config unwrapping for multimodal Qwen configs."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Runtime special-token post-processor update and BigBirdTokenizer mask-token registration are different tokenizer bugs."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding-window behavior is unrelated to a Float8 storage type error."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:44843",
+            "accept": true,
+            "reason": "Same offline loading bug: both describe tokenizer/model_info behavior breaking HF_HUB_OFFLINE cache resolution."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "SigLIP2 doc example errors are unrelated to OmDet-Turbo processor/model image-size mismatch."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode is unrelated to ZeRO-3 gradient growth."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded and conservative overall. The only accepted duplicates are well-supported by the titles/packet (42886\u219444843 and 44514\u219444545), and the remaining soft-edge rejections are phrased cautiously enough for the evidence available. No PR pairs are present."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 13,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:30064",
+        "issue:34634",
+        "issue:36010",
+        "issue:39692",
+        "issue:41720",
+        "issue:42915",
+        "issue:43334",
+        "issue:43335",
+        "issue:43381",
+        "issue:43421",
+        "issue:43525",
+        "issue:43531",
+        "issue:43540",
+        "issue:43550",
+        "issue:43575",
+        "issue:43596",
+        "issue:43618",
+        "issue:43653",
+        "issue:43697",
+        "issue:43716",
+        "issue:43720",
+        "issue:43746",
+        "issue:43756",
+        "issue:43782",
+        "issue:43819",
+        "issue:43873",
+        "issue:43994",
+        "issue:44008",
+        "issue:44077",
+        "issue:44186",
+        "issue:44220",
+        "issue:44625",
+        "issue:44661",
+        "issue:44671",
+        "issue:44811",
+        "issue:44843",
+        "issue:44898",
+        "issue:44918",
+        "issue:44936",
+        "issue:45072",
+        "issue:45081",
+        "issue:45183",
+        "issue:45356",
+        "issue:45381",
+        "issue:45479",
+        "issue:45704",
+        "issue:45759"
+      ],
+      "soft_pairs": [
+        "issue:44918|issue:44936",
+        "issue:43421|issue:44186",
+        "issue:44625|issue:45479",
+        "issue:43525|issue:43531",
+        "issue:43782|issue:45381",
+        "issue:45081|issue:45356",
+        "issue:43994|issue:44671",
+        "issue:34634|issue:41720",
+        "issue:43381|issue:45704",
+        "issue:39692|issue:43873",
+        "issue:43596|issue:44220",
+        "issue:42915|issue:45759",
+        "issue:45183|issue:45479",
+        "issue:43540|issue:44008",
+        "issue:44220|issue:44661",
+        "issue:43618|issue:43697",
+        "issue:44077|issue:44625",
+        "issue:43720|issue:43819",
+        "issue:43575|issue:45704",
+        "issue:30064|issue:36010",
+        "issue:43335|issue:43756",
+        "issue:43653|issue:43720",
+        "issue:43334|issue:45759",
+        "issue:43716|issue:44008",
+        "issue:42915|issue:43746",
+        "issue:43746|issue:45381",
+        "issue:43746|issue:43873",
+        "issue:43550|issue:45072",
+        "issue:44186|issue:44898",
+        "issue:43550|issue:43819",
+        "issue:43421|issue:44811",
+        "issue:44843|issue:45081"
+      ],
+      "prepared_review_unit_hash": "937ba681da88e09e41552a176cfa1d46d8ad4b33d4c9b8737432306c4fc8ea37",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36979,
+        "estimated_input_tokens": 9245,
+        "estimated_eval_tokens": 18746
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a set of unrelated model/tokenizer/config/runtime bugs that only share broad Transformers vocabulary. None of the soft pairs look like the same underlying issue, so no duplicate edges should be accepted.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No single issue is a good cluster-wide canonical: the items span different models, subsystems, and failure modes, with no true duplicate pair to anchor the group.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44918",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Different problems: Qwen3.5 embedding unpacking in TRL SFT vs trainer.evaluate failing after trainer.train()."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44186",
+            "accept": false,
+            "reason": "Different tokenizer bugs: runtime special-token/post-processor updates vs LayoutLMv2 NER padding/truncation crashes."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Different config/label bugs: Qwen3.5 text_config propagation vs single-label classification zero-loss behavior."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Unrelated Qwen3-MoE sliding window issue vs Llama4Config missing pad_token_id attribute."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Different Qwen multimodal failures: weight_only load error vs wrong video vision_position_ids."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Both touch tokenizer regex logic, but they describe different failures and code paths; not the same concrete bug."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Different models and symptoms: SigLIP2 bad AutoModel/pipeline outputs vs CamemBERT masked LM regression."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "BarkProcessor voice_preset issue is unrelated to Qwen3 auto device mapping cudaErrorAssert."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Gradient checkpointing eval-mode restriction is unrelated to T5 apex.FusedRMSNorm memory leak."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Documentation/example errors are not the same bug as offloading failing under quantization."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Deepspeed zero3/BertModel index error is unrelated to _torch_extract_fbank_features()."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Different Qwen config bugs: FineGrainedFP8Config failure vs from_config not unwrapping text_config."
+          },
+          {
+            "left": "issue:45183",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Generic input validation messaging bug is unrelated to the num_labels=1 zero-loss issue."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44008",
+            "accept": false,
+            "reason": "Video input processing failure in Qwen3OmniMoe is unrelated to Gemma3n variable-name collision."
+          },
+          {
+            "left": "issue:44220",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "Audio feature extraction bug is unrelated to add-new-model-like failing inside TOKENIZER_MAPPING_NAMES."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "Missing CLIPOutput attentions field is unrelated to RTDetrV2 output differences in v5."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44625",
+            "accept": false,
+            "reason": "patchtsmixer post_init policy and Qwen3.5 num_labels propagation are unrelated issues."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "BitNet packed-weight loading and DAC latent decoding mismatch are different code paths and bugs."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Tensor-parallel OOM for Qwen2 is unrelated to the T5 apex RMSNorm leak."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "ImageProcessor segmentation-map handling and GenerationMixin import failure are unrelated."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43756",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer construction and Smollm3 RoPE-layer dropping are unrelated model-configuration bugs."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "BigBirdTokenizer mask-token registration is unrelated to BitNet accelerate loading of packed weights."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both involve Qwen multimodal configs, but one is a missing pad_token_id attribute and the other is from_config text_config unwrapping; not the same bug."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44008",
+            "accept": false,
+            "reason": "Mistral-3 image dtype mismatch is unrelated to Gemma3n audio_mel_mask attribute collision."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 loading and GraniteSpeech PEFT local-checkpoint loading are unrelated."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "GraniteSpeech PEFT checkpoint loading is unrelated to Qwen2.5-VL video position-id errors."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Local-checkpoint loading for PEFT adapters is unrelated to quantization/offloading behavior."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Bamba torch.compile SDPA failure and SwitchTransformers/TimmWrapper dtype mismatch are unrelated."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "LayoutLMv2 tokenizer crash and Perceiver interpolation failure are different model-specific bugs."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Different subsystems: torch.compile SDPA issue vs DAC.from_latents mismatch."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Runtime post-processor updates and Whisper batch_decode skip_special_tokens behavior are related only loosely to tokenization, not the same bug."
+          },
+          {
+            "left": "issue:44843",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both touch _patch_mistral_regex, but one is offline model_info() access and the other is backend_tokenizer AttributeError; separate failures."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 14,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:30064",
+        "issue:34634",
+        "issue:36010",
+        "issue:36331",
+        "issue:42175",
+        "issue:43322",
+        "issue:43404",
+        "issue:43452",
+        "issue:43531",
+        "issue:43540",
+        "issue:43577",
+        "issue:43582",
+        "issue:43606",
+        "issue:43643",
+        "issue:43650",
+        "issue:43653",
+        "issue:43720",
+        "issue:43782",
+        "issue:43818",
+        "issue:43819",
+        "issue:43825",
+        "issue:43827",
+        "issue:43906",
+        "issue:43994",
+        "issue:44062",
+        "issue:44186",
+        "issue:44442",
+        "issue:44448",
+        "issue:44485",
+        "issue:44488",
+        "issue:44509",
+        "issue:44561",
+        "issue:44610",
+        "issue:44671",
+        "issue:44805",
+        "issue:44821",
+        "issue:44877",
+        "issue:44936",
+        "issue:44977",
+        "issue:44991",
+        "issue:45020",
+        "issue:45072",
+        "issue:45081",
+        "issue:45084",
+        "issue:45381",
+        "issue:45406",
+        "issue:45685",
+        "issue:45701"
+      ],
+      "soft_pairs": [
+        "issue:43643|issue:44877",
+        "issue:43531|issue:43643",
+        "issue:43577|issue:43818",
+        "issue:44062|issue:44488",
+        "issue:43540|issue:45381",
+        "issue:43653|issue:43819",
+        "issue:44448|issue:44671",
+        "issue:43782|issue:44977",
+        "issue:44561|issue:45020",
+        "issue:43322|issue:45685",
+        "issue:43825|issue:43994",
+        "issue:44821|issue:45084",
+        "issue:30064|issue:36331",
+        "issue:43825|issue:43827",
+        "issue:44485|issue:44509",
+        "issue:43577|issue:43582",
+        "issue:44186|issue:44442",
+        "issue:43606|issue:45072",
+        "issue:34634|issue:36010",
+        "issue:44805|issue:44936",
+        "issue:43452|issue:44991",
+        "issue:44610|issue:45072",
+        "issue:42175|issue:45081",
+        "issue:43650|issue:43906",
+        "issue:43720|issue:45072",
+        "issue:43404|issue:43782",
+        "issue:45081|issue:45406",
+        "issue:44671|issue:45701"
+      ],
+      "prepared_review_unit_hash": "947cb3db463e3fc14076e8e4bfa1943330162809ea2c0d60752edd184b9b3b91",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36080,
+        "estimated_input_tokens": 9020,
+        "estimated_eval_tokens": 18296
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is highly heterogeneous: most pairs share only a broad subsystem or version-regression theme, not the same concrete bug. I would not merge any of the soft pairs. If a representative issue is needed, the broadest umbrella is the remote_code regression report, but the cluster is not a tight duplicate set.",
+        "confidence": 0.78,
+        "canonical_issue_reason": "issue:45020 is the broadest, most umbrella-like report in the set (recent-version regressions affecting remote_code models), so it is the least narrow representative if one must be chosen.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45020 is the best cluster representative by scope, but only loosely so; the rest of the items are mostly unrelated one-off bugs across different models and code paths.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43643",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Different config-loading bugs: missing fields with trust_remote_code vs strict config rejecting granite_speech."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior is unrelated to trust_remote_code config field loss."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43818",
+            "accept": false,
+            "reason": "BLIP2 dtype propagation and Video-LLaVA temporal-attention/shared-weight issues are different code paths and failures."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "AddedToken special-arg duplication is not the same as sleng-bert tokenizer loading failure."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both are video-related, but they hit different models and different bugs: ValueError in processing vs wrong vision_position_ids."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Tokenizer special-token registration and DAC latent/STE reconstruction are unrelated."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Both describe regressions in model outputs, but for different models and different mechanisms."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Qwen3VL from_pretrained weight_only loading and Qwen3.5 flash-attention generation are distinct failures."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "44561 is a specific v5 helper-removal regression; 45020 is a broader umbrella report about remote_code breakage, not the same concrete bug."
+          },
+          {
+            "left": "issue:43322",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Llava Next segfault and MPS histogram/int failure in moe are unrelated issues."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Pipeline error-message wording and SigLIP2 incorrect outputs are different bugs."
+          },
+          {
+            "left": "issue:44821",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "AutoImageProcessor URL loading and template compilation errors are unrelated."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "ImageProcessor void segmentation maps and CustomTrainer compute_loss signature mismatch are unrelated."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "Same broad v5 pipeline-removal topic, but one is runtime messaging and the other is stale docs; not the same underlying bug."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "GLM-5 RoPE implementation and docs mentioning removed pipeline tasks are unrelated."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "BLIP2 dtype handling and Apple Silicon caching_allocator_warmup TypeError are unrelated."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "LayoutLMv2 tokenizer crashes on NER/padding and FastSpeech2ConformerTokenizer loading failure are different tokenizer bugs."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "CPU offload device mismatch and bfloat16 dtype mismatch are different failures."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "BarkProcessor voice_preset and GenerationMixin import failure are unrelated."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Mask shape IndexError and trainer.evaluate-after-train failure do not point to the same bug."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "gguf_file loading breakage and est-roberta tokenizer loading regression are separate tokenizer/model-loading problems."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Processor image-size mismatch and bfloat16 dtype mismatch are unrelated."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Missing TensorFlow in the torch extra and Mistral regex patch crashing on tokenizer internals are unrelated."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "The titles/body intent are unrelated; one is a malformed issue, the other is a reproduction of a different bug."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "BitNet packed-weight unpacking and dtype mismatch in inference are different problems."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43782",
+            "accept": false,
+            "reason": "Mistral3 weight tying and Qwen3VL weight_only loading are unrelated model-specific issues."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both are loading-time AttributeErrors, but they affect different components and root causes."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "CamemBERT masked-LM regression is not the same concrete bug as a general tokenization-change report."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 15,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:34634",
+        "issue:36010",
+        "issue:36331",
+        "issue:38175",
+        "issue:39692",
+        "issue:43257",
+        "issue:43262",
+        "issue:43329",
+        "issue:43450",
+        "issue:43454",
+        "issue:43526",
+        "issue:43531",
+        "issue:43550",
+        "issue:43577",
+        "issue:43582",
+        "issue:43638",
+        "issue:43643",
+        "issue:43644",
+        "issue:43645",
+        "issue:43653",
+        "issue:43697",
+        "issue:43720",
+        "issue:43746",
+        "issue:43749",
+        "issue:43819",
+        "issue:43854",
+        "issue:43927",
+        "issue:44060",
+        "issue:44291",
+        "issue:44368",
+        "issue:44554",
+        "issue:44568",
+        "issue:44589",
+        "issue:44610",
+        "issue:44811",
+        "issue:44857",
+        "issue:44912",
+        "issue:44936",
+        "issue:44977",
+        "issue:45042",
+        "issue:45084",
+        "issue:45356",
+        "issue:45459",
+        "issue:45488",
+        "issue:45538",
+        "issue:45600",
+        "issue:45759",
+        "issue:45803"
+      ],
+      "soft_pairs": [
+        "issue:44060|issue:44368",
+        "issue:43582|issue:44857",
+        "issue:45459|issue:45803",
+        "issue:43262|issue:43819",
+        "issue:43526|issue:43697",
+        "issue:45084|issue:45538",
+        "issue:43644|issue:43749",
+        "issue:43329|issue:45600",
+        "issue:43653|issue:44912",
+        "issue:43638|issue:44936",
+        "issue:43746|issue:44811",
+        "issue:38175|issue:44610",
+        "issue:44589|issue:45042",
+        "issue:43454|issue:43577",
+        "issue:43854|issue:44977",
+        "issue:43526|issue:44554",
+        "issue:36010|issue:43645",
+        "issue:39692|issue:44912",
+        "issue:34634|issue:36331",
+        "issue:43257|issue:43329",
+        "issue:43819|issue:43927",
+        "issue:43720|issue:44610",
+        "issue:43749|issue:44291",
+        "issue:43550|issue:44912",
+        "issue:44568|issue:45488",
+        "issue:43262|issue:44811",
+        "issue:43531|issue:45042",
+        "issue:43643|issue:45759",
+        "issue:44568|issue:45356",
+        "issue:36010|issue:36331",
+        "issue:43329|issue:43450"
+      ],
+      "prepared_review_unit_hash": "32d35bffff915b5d7902500c64623a6daa3a0add68498c438b59232808000d20",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 37491,
+        "estimated_input_tokens": 9373,
+        "estimated_eval_tokens": 19002
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is highly heterogeneous: the soft edges mostly connect unrelated Transformers regressions that only share broad subsystem/version similarity. I would not merge any of the proposed soft pairs.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "Issue 43644 is the broadest shared-runtime regression in the set and is closer to the cluster\u2019s recurring v5-loading/correctness theme than the narrower model- or tokenizer-specific reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43644 is the best overall representative because it describes a general Transformers 5.0.0 correctness break that can affect many models, not just one narrow configuration.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44060",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Both mention tied-weights warnings, but one is a concrete Qwen3-Next weight-tying bug and the other is a user-facing warning during LoRA fine-tuning; not the same code path."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Completely different failures: Apple Silicon allocator warmup TypeError vs CUDA AMP loss crash in LwDetrImageLoss."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Both are about exception handling, but they affect different tokenizer/processor functions and different failure modes."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate defaulting is unrelated to DAC latent reconstruction mismatch."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "Different models and bugs: Beit reduce_labels shape/return issue vs RTDetrV2 output regression."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Unrelated areas: compilation of template nodes vs tokenizer max_length default."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Both involve loading, but one is non-persistent buffer initialization and the other is FSDP CPU RAM efficient loading; not the same defect."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Multimodal token counting bug vs stale auto-mapping references in CI; no shared underlying change."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "BigBird special-token decode bug and MXFP4 packed-weight loading bug are unrelated."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Zero3 training index error with a non-pretrained Bert model is different from evaluate() failing after train()."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "PEFT/local checkpoint loading issue vs Whisper batch_decode skip_special_tokens regression; different layers of the stack."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Unexpected zero probabilities in SigLIP2 is unrelated to OmDet-Turbo processor resizing to the wrong image size."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Storage-type lookup failure and PIL backend torchvision dependency regression are different bugs."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "AyaVision weight-tying failure and Blip2 dtype propagation issue are distinct model-loading problems."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "GLM-4.7-Flash test loading failure is not the same as Qwen3.5 flash-attention generation regression."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "Beit label reduction and MPS attention correctness when head dims differ are unrelated."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "GenerationMixin import breakage is a different issue from Jupyter custom-model initialization regression."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Documentation example errors for SigLIP2 do not match the MXFP4 loading fallback bug."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Bark voice_preset failure and CustomTrainer.compute_loss signature mismatch are unrelated API breaks."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Qwen3 MoE weight conversion during accelerate/deepspeed loading is a different path from multimodal token counting."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "DAC latent decoding mismatch is unrelated to DiaConfig losing custom token IDs on save/load."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Packed-weight unpacking during accelerate loading is not the same as an image processor size mismatch."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "These are both loading-related, but one is an FSDP RAM-efficient loading regression and the other is an init_empty_weights argument incompatibility."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "torch.compile SDPA failure in Bamba is unrelated to MXFP4 quantization fallback."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45488",
+            "accept": false,
+            "reason": "Tokenizer-specific BOS/EOS behavior for mdeberta-v3 is not the same bug as LlamaTokenizer overriding tokenizer.json pre-tokenizer settings."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Different audio-processing APIs and different default/parameter handling bugs."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior is unrelated to the PIL backend torchvision requirement regression."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Missing fields with trust_remote_code and failure to unwrap text_config in composite configs are distinct config-loading bugs."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Tokenizer special-token insertion bug and Kimi-K2.5 codec/regex warning regression are different tokenizer issues."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 16,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:33357",
+        "issue:34567",
+        "issue:36331",
+        "issue:38175",
+        "issue:40990",
+        "issue:41720",
+        "issue:42915",
+        "issue:43257",
+        "issue:43317",
+        "issue:43329",
+        "issue:43452",
+        "issue:43540",
+        "issue:43577",
+        "issue:43582",
+        "issue:43643",
+        "issue:43646",
+        "issue:43696",
+        "issue:43697",
+        "issue:43742",
+        "issue:43749",
+        "issue:43756",
+        "issue:43906",
+        "issue:43927",
+        "issue:43931",
+        "issue:44062",
+        "issue:44291",
+        "issue:44479",
+        "issue:44560",
+        "issue:44737",
+        "issue:44756",
+        "issue:44821",
+        "issue:44849",
+        "issue:44898",
+        "issue:44912",
+        "issue:44987",
+        "issue:44991",
+        "issue:45042",
+        "issue:45072",
+        "issue:45081",
+        "issue:45137",
+        "issue:45357",
+        "issue:45375",
+        "issue:45381",
+        "issue:45405",
+        "issue:45538",
+        "issue:45588",
+        "issue:45701"
+      ],
+      "soft_pairs": [
+        "issue:34567|issue:40990",
+        "issue:43329|issue:44479",
+        "issue:44849|issue:45701",
+        "issue:43257|issue:43931",
+        "issue:43697|issue:45357",
+        "issue:43452|issue:43582",
+        "issue:41720|issue:43540",
+        "issue:44291|issue:45042",
+        "issue:42915|issue:43742",
+        "issue:33357|issue:34567",
+        "issue:43317|issue:45538",
+        "issue:44479|issue:45381",
+        "issue:44987|issue:45538",
+        "issue:43540|issue:44560",
+        "issue:43540|issue:44479",
+        "issue:42915|issue:44821",
+        "issue:43927|issue:44912",
+        "issue:43257|issue:43756",
+        "issue:43577|issue:45072",
+        "issue:42915|issue:45538",
+        "issue:43257|issue:43696",
+        "issue:45137|issue:45405",
+        "issue:38175|issue:44898",
+        "issue:44737|issue:45588",
+        "issue:43742|issue:44291",
+        "issue:44062|issue:44991",
+        "issue:43906|issue:45381",
+        "issue:43646|issue:43749",
+        "issue:43749|issue:44756",
+        "issue:34567|issue:36331",
+        "issue:43643|issue:45375",
+        "issue:44062|issue:45081"
+      ],
+      "prepared_review_unit_hash": "4112a218ad3c1ca66512871b95ac5e26c89b9a89b0cd068bd2196aedbd7831e2",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36803,
+        "estimated_input_tokens": 9201,
+        "estimated_eval_tokens": 18658
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly heterogeneous. 44479 remains the best representative issue for the Qwen video regression family, but the soft links are not strong enough to confirm a duplicate chain: 43329 is a narrower undefined-helper bug, and 45381 is only title-level similar to the broader 44479 report.",
+        "confidence": 0.62,
+        "canonical_issue_reason": "44479 is the broadest and most representative Qwen video-regression report in the set, covering multiple affected model variants and the shared video-input path.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "44479 is the strongest standalone issue to keep: it has the broadest scope and best captures the shared Qwen video-input regression family.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:34567",
+            "right": "issue:40990",
+            "accept": false,
+            "reason": "Unrelated bugs: trainer token accounting vs GPT-oss perplexity."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Related area, but not the same concrete bug: 43329 is an undefined-helper failure in token counting, while 44479 is a broader video-input regression report."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Different symptoms and scope; only broad model/version overlap."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different model families and failure modes: MoE weight conversion vs VL weight-shape mismatch."
+          },
+          {
+            "left": "issue:43697",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different code paths: RTDetrV2 output changes vs Qwen3.5 visual-key serialization."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "Tokenizer/model loading vs Apple Silicon allocator warmup TypeError are unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Both are Qwen-related, but the failures are different concrete bugs."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Different issues: init_empty_weights argument handling vs image processors requiring torchvision."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Different model-loading bugs in different subsystems."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Completely unrelated: Mac bus error on CLIP vs TrainerState token counter update."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Different problems: device-map/offload loading failure vs tokenizer max_length constant."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Very similar titles, but the packet only supports this at a broad thematic level; not enough evidence for the same concrete bug."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Unrelated: model-loading failure vs tokenizer max_length issue."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Both are video-input bugs, but they affect different model families and appear to be different failures."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Qwen video processing bug vs broader video-position regression; related topic, not enough to call duplicate."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Different subsystems: Qwen3Moe FP8 load failure vs image-processor URL loading."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Saving/loading custom token IDs is unrelated to MXFP4 quantization fallback."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43756",
+            "accept": false,
+            "reason": "Different bugs in different loading paths: MoE weight conversion vs RoPE-layer regression."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Both mention dtype, but they affect different models and code paths."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Different bugs: Qwen3Moe FP8 loading vs CLIPTokenizer max_length behavior."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43696",
+            "accept": false,
+            "reason": "Different issues: MoE weight conversion vs GPT-oss CUDA OOM."
+          },
+          {
+            "left": "issue:45137",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO3 deque error is unrelated to the PEFT version bump."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Different models and symptoms: SigLIP zero probabilities vs Perceiver resolution handling."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "XLNet CPU placement bug vs flash-attention crash are unrelated."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "MobileLLM loading key error vs init_empty_weights argument handling are different bugs."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Tokenizer-loading theme only; not enough evidence of the same AddedToken bug."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Isolated reproduction of another issue vs Qwen2.5-VL video-position regression; no clear identity."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Custom model init breakage is unrelated to FSDP CPU RAM-efficient loading."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Different memory/loading problems: FSDP CPU RAM-efficient loading vs mmap OOM."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "TrainerState token counter bug vs compute_loss signature break are unrelated."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both involve missing fields, but the deserialization failures are not the same concrete bug."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Different tokenizer bugs: AddedToken construction vs Mistral regex patching."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The cluster summary is directionally plausible, but it overstates the evidence for the Qwen video cluster. In particular, the accepted soft-edge pair issue:43329 \u2194 issue:44479 is not strongly grounded by the packet alone: 43329 is about an undefined helper in `_get_num_multimodal_tokens`, while 44479 is a broader video-input regression report. The second accept (44479 \u2194 45381) is more plausible, but the overall verdict still isn\u2019t conservative enough because the only explicit support here is title-level similarity. I\u2019d keep this as a likely related cluster, not a confirmed duplicate chain."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 17,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:28282",
+        "issue:29127",
+        "issue:34689",
+        "issue:36010",
+        "issue:41762",
+        "issue:42175",
+        "issue:43452",
+        "issue:43493",
+        "issue:43526",
+        "issue:43531",
+        "issue:43577",
+        "issue:43582",
+        "issue:43643",
+        "issue:43697",
+        "issue:43720",
+        "issue:43742",
+        "issue:43756",
+        "issue:43819",
+        "issue:43824",
+        "issue:43994",
+        "issue:44291",
+        "issue:44466",
+        "issue:44484",
+        "issue:44488",
+        "issue:44514",
+        "issue:44561",
+        "issue:44589",
+        "issue:44625",
+        "issue:44704",
+        "issue:44792",
+        "issue:44811",
+        "issue:44821",
+        "issue:44843",
+        "issue:44877",
+        "issue:44898",
+        "issue:44936",
+        "issue:44964",
+        "issue:44977",
+        "issue:45005",
+        "issue:45042",
+        "issue:45081",
+        "issue:45216",
+        "issue:45290",
+        "issue:45325",
+        "issue:45381",
+        "issue:45405",
+        "issue:45479",
+        "issue:45600"
+      ],
+      "soft_pairs": [
+        "issue:36010|issue:42175",
+        "issue:43493|issue:43994",
+        "issue:42175|issue:43824",
+        "issue:43452|issue:44488",
+        "issue:43526|issue:45479",
+        "issue:44291|issue:44821",
+        "issue:44792|issue:44964",
+        "issue:28282|issue:34689",
+        "issue:45042|issue:45081",
+        "issue:43756|issue:45325",
+        "issue:45405|issue:45600",
+        "issue:43742|issue:44589",
+        "issue:44484|issue:44704",
+        "issue:43697|issue:45216",
+        "issue:43531|issue:44821",
+        "issue:44514|issue:45381",
+        "issue:28282|issue:29127",
+        "issue:44625|issue:44877",
+        "issue:43582|issue:44977",
+        "issue:43452|issue:43577",
+        "issue:43720|issue:44898",
+        "issue:43643|issue:44561",
+        "issue:44514|issue:45290",
+        "issue:43452|issue:44843",
+        "issue:43819|issue:44811",
+        "issue:41762|issue:44936",
+        "issue:42175|issue:44488",
+        "issue:44466|issue:45005"
+      ],
+      "prepared_review_unit_hash": "40ee75d690413e6d34e9284f46b56bf4473f2db8c799035c4628279f0140f9d2",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36507,
+        "estimated_input_tokens": 9127,
+        "estimated_eval_tokens": 18510
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most items are unrelated one-off bugs. The only plausible duplicate pair is the v5 tied-weights serialization/reporting regression around lm_head.weight on translation models; everything else is too broad or touches different code paths.",
+        "confidence": 0.42,
+        "canonical_issue_reason": "issue:45005 is the best umbrella report for the tied-weights regression in v5; it is broader than the more specific device-dependent lm_head.weight symptom in issue:44466.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45005 has the broadest, clearest description of the duplicated tied-weights problem and is the best issue to keep as the cluster representative.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:36010",
+            "right": "issue:42175",
+            "accept": false,
+            "reason": "Different failures: missing PyTorch backend vs TensorFlow not being included in a torch install. No shared code path."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Both mention SigLIP2, but one is an implementation-vs-JAX discrepancy and the other is bad outputs in AutoModel/pipeline; not enough to treat as the same bug."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Backend packaging issue vs a Qwen2.5-VL import error. Different subsystem and symptom."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "gguf loading breakage vs a model-specific load failure for sleng-bert; different root causes."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Image processor label reduction bug vs sequence-classification zero-loss bug; unrelated code paths."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Model loading with init_empty_weights vs loading AutoImageProcessor from URL. Different APIs and failures."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Janus test failure and Phi-4 multimodal load failure are both multimodal-related, but not the same concrete bug."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Generic PyTorch-missing ImportError vs a Llama 3.2 Vision loading regression. Different causes."
+          },
+          {
+            "left": "issue:45042",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Image processor torchvision dependency regression vs Mistral regex patch crash. Unrelated."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Different models and different RoPE/position-id logic; not the same bug."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "One is an unreleased dependency version bump; the other is stale auto_mappings references. Different maintenance issues."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "KeyError while loading one model vs missing Float8 storage support. Different failure modes."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "save_pretrained shard-size question vs AutoProcessor kwarg propagation bug. Not the same issue."
+          },
+          {
+            "left": "issue:43697",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "RTDetrV2 output regression vs Qwen3.5 save_pretrained regression. Different models and code paths."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window bug vs AutoImageProcessor URL loading bug. Unrelated."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Different Qwen2.5-VL failures: batched chat template crash vs video position-id bug. Same family, not the same concrete bug."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "PyTorch import failure vs LayoutLMv3 box-content error messaging. Different problems."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "num_labels propagation bug in Qwen3.5 vs strict config blocking granite_speech. Different config issues."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Apple Silicon allocator TypeError vs Qwen3.5 flash-attention generation issue. No shared code path."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "gguf loading failure vs dtype not propagating in Blip2. Different subsystems."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "BitNet accelerate-loading unpack issue vs Perceiver interpolation failure. Different model logic."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "Config field loss under trust_remote_code vs removal of is_torch_fx_available breaking remote-code models. Different regressions."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both are apply_chat_template crashes, but they involve different triggers and likely different branches; not clearly the same bug."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44843",
+            "accept": false,
+            "reason": "gguf loading breakage vs offline model_info call in mistral regex patch. Different failure paths."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "DAC latent/forward mismatch vs Whisper batch_decode skipping special tokens. Unrelated."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO-3 model load IndexError vs trainer.evaluate() after training. Different issues."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "TensorFlow backend packaging vs inability to load a specific BERT model. Not the same bug."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45005",
+            "accept": true,
+            "reason": "Both describe the same v5 tied-weights/ lm_head serialization regression; 45005 is the broader umbrella report and 44466 is a specific manifestation."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The cluster summary is conservative and grounded: the accepted pair (44466, 45005) is a plausible tied-weights/lm_head v5 duplicate, and the other soft-edge rejections avoid overclaiming by keeping same-family-but-different-symptom cases separate."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 18,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:24643",
+        "issue:28282",
+        "issue:29127",
+        "issue:30064",
+        "issue:32090",
+        "issue:33666",
+        "issue:34567",
+        "issue:35141",
+        "issue:36010",
+        "issue:38175",
+        "issue:41720",
+        "issue:43122",
+        "issue:43295",
+        "issue:43329",
+        "issue:43452",
+        "issue:43504",
+        "issue:43525",
+        "issue:43531",
+        "issue:43550",
+        "issue:43572",
+        "issue:43577",
+        "issue:43653",
+        "issue:43688",
+        "issue:43697",
+        "issue:43720",
+        "issue:43742",
+        "issue:43749",
+        "issue:43906",
+        "issue:43927",
+        "issue:44361",
+        "issue:44442",
+        "issue:44451",
+        "issue:44464",
+        "issue:44479",
+        "issue:44568",
+        "issue:44610",
+        "issue:44743",
+        "issue:44821",
+        "issue:44843",
+        "issue:44936",
+        "issue:44977",
+        "issue:44991",
+        "issue:45325",
+        "issue:45356",
+        "issue:45440",
+        "issue:45588",
+        "issue:45812"
+      ],
+      "soft_pairs": [
+        "issue:43295|issue:45356",
+        "issue:43653|issue:44568",
+        "issue:24643|issue:29127",
+        "issue:38175|issue:43295",
+        "issue:44361|issue:44442",
+        "issue:30064|issue:33666",
+        "issue:43653|issue:44361",
+        "issue:43577|issue:43697",
+        "issue:44464|issue:44743",
+        "issue:44843|issue:44991",
+        "issue:44479|issue:44936",
+        "issue:41720|issue:43906",
+        "issue:44977|issue:45356",
+        "issue:43688|issue:45440",
+        "issue:43577|issue:45588",
+        "issue:43531|issue:43742",
+        "issue:43653|issue:43927",
+        "issue:44464|issue:44977",
+        "issue:43550|issue:43927",
+        "issue:43550|issue:44361",
+        "issue:32090|issue:34567",
+        "issue:43122|issue:43906",
+        "issue:43504|issue:43720",
+        "issue:43504|issue:43550",
+        "issue:32090|issue:33666",
+        "issue:28282|issue:35141",
+        "issue:43504|issue:44610",
+        "issue:43452|issue:44451",
+        "issue:44568|issue:45812",
+        "issue:43742|issue:44821",
+        "issue:41720|issue:44464",
+        "issue:43688|issue:43749",
+        "issue:33666|issue:36010",
+        "issue:43329|issue:45325",
+        "issue:43525|issue:43572"
+      ],
+      "prepared_review_unit_hash": "c973c31dd9263dcab0b3991f6d03ff778f2c036b87fe08b61c53f42612348e5f",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 35,
+        "serialized_chars": 37786,
+        "estimated_input_tokens": 9447,
+        "estimated_eval_tokens": 19150
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly a grab-bag of unrelated HF issues with only loose lexical similarity. I found no soft pair that clearly describes the same underlying bug/change, so all candidate edges are rejected. As a representative issue, I\u2019d use the broad Granite tokenizer regression because it is the most general and user-impacting item in the set.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No perfect canonical issue exists for this heterogeneous cluster; issue 45812 is the best representative because it is broad, user-facing, and framed as a silent tokenizer regression across all Granite models.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45812 is the strongest cluster representative: it targets a widespread tokenizer regression, is explicit about the symptom, and is likely to subsume nearby tokenizer-related reports better than the narrower model-specific bugs.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43295",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different model families and different failure modes (processor/tokenizer API break vs codec/warning regression)."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Different tokenizer bugs: BigBird mask token registration vs mDeBERTa BOS/EOS insertion. Similar area, not the same issue."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "Unrelated problems: DeepSpeed weight dimensionality error vs LayoutLMv3 box-validation messaging."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Different subsystems and symptoms: SigLIP2 zero probabilities vs processor/tokenizer regression in custom multimodal code."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "Both involve tokenizer loading failures, but for unrelated tokenizers/models with no shared code-path bug."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "Image processor void segmentation maps and Qwen2-VL multi-GPU training are unrelated."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Different tokenizer/model failures; only broad tokenization similarity."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "Blip2 dtype casting and RTDetrV2 output divergence are unrelated model-runtime bugs."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Both touch generation behavior, but one is chunked generation with compilation and the other is recurrent-state reset in a Qwen3.5 module."
+          },
+          {
+            "left": "issue:44843",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but one is an offline hub call bug and the other is a model-specific tokenizer loading regression."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Video-input regression and trainer evaluate-after-train failure are unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "No concrete shared code-path is evident; one is Qwen3 device-mapping CUDA assert, the other is just an isolated reproduction of an older issue."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Different models and different generation/tokenizer regressions; not the same fix."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Both are model correctness issues, but they concern different architectures and different bugs."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Blip2 dtype loading and flash_attention crash on sink-less models do not share the same underlying bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding-window behavior and MobileLLM loading key error are unrelated."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Special-token registration in BigBirdTokenizer is unrelated to DiaConfig losing custom token IDs on save/load."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Both are generation issues, but they involve different mechanisms and code paths (compiled forward vs flash-attention)."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Torch.compile/SDPA failure in Bamba and DiaConfig token-ID loss are unrelated."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Different model/runtime bug vs tokenizer AttributeError; no common underlying issue."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Trainer NoneType broadcast failure and token-count tracking not updating are separate trainer issues."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "Different tokenizer regression report vs isolated reproduction; no evidence they are the same bug."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Different load-time issues: legacy-field preset loading vs packed-weight unpacking during accelerate loading."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Beit pretrained loading failure and Bamba torch.compile SDPA failure are unrelated."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "Trainer broadcast error vs Qwen2-VL multi-GPU training is only broad training-context similarity."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "PyTorch-missing import error and token-embedding reinit on resize are unrelated."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Model loading legacy-field bug vs processor output-size mismatch are different problems."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Both are model loading failures, but not the same concrete code-path or model family."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are tokenizer regressions in v5, but they affect different models and different token handling paths."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "MobileLLM loading key error and AutoImageProcessor-from-URL failure are unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44464",
+            "accept": false,
+            "reason": "Qwen3 device-mapping CUDA assert and compiled chunked generation inconsistency are different runtime bugs."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Aux-loss normalization and FSDP CPU RAM efficient loading are unrelated."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Qwen2-VL multi-GPU training and GenerationMixin import failure are unrelated."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Both are multimodal/Qwen-VL related, but one is an undefined-function bug in video token counting and the other is a rope-index scaling issue for still images."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Similar missing-config-field theme, but different models and different field names; 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": 2,
+      "component_count": 16,
+      "review_unit_index": 19,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:28282",
+        "issue:32090",
+        "issue:33357",
+        "issue:33666",
+        "issue:36331",
+        "issue:39692",
+        "issue:43257",
+        "issue:43262",
+        "issue:43278",
+        "issue:43295",
+        "issue:43299",
+        "issue:43334",
+        "issue:43493",
+        "issue:43540",
+        "issue:43572",
+        "issue:43577",
+        "issue:43606",
+        "issue:43646",
+        "issue:43688",
+        "issue:43716",
+        "issue:43720",
+        "issue:43756",
+        "issue:43824",
+        "issue:43825",
+        "issue:43844",
+        "issue:43927",
+        "issue:44361",
+        "issue:44423",
+        "issue:44466",
+        "issue:44479",
+        "issue:44488",
+        "issue:44554",
+        "issue:44589",
+        "issue:44625",
+        "issue:44734",
+        "issue:44843",
+        "issue:44849",
+        "issue:44913",
+        "issue:44964",
+        "issue:44977",
+        "issue:45083",
+        "issue:45200",
+        "issue:45229",
+        "issue:45357",
+        "issue:45405",
+        "issue:45520",
+        "issue:45701",
+        "issue:45704"
+      ],
+      "soft_pairs": [
+        "issue:43493|issue:43825",
+        "issue:43262|issue:43295",
+        "issue:43756|issue:44913",
+        "issue:44849|issue:45357",
+        "issue:43540|issue:43716",
+        "issue:43720|issue:43927",
+        "issue:43606|issue:43927",
+        "issue:28282|issue:32090",
+        "issue:28282|issue:33357",
+        "issue:43334|issue:43572",
+        "issue:43577|issue:44554",
+        "issue:43688|issue:45704",
+        "issue:33666|issue:36331",
+        "issue:39692|issue:43493",
+        "issue:43720|issue:44361",
+        "issue:44964|issue:45405",
+        "issue:44479|issue:44625",
+        "issue:43540|issue:44977",
+        "issue:43493|issue:43844",
+        "issue:44488|issue:44843",
+        "issue:43756|issue:45083",
+        "issue:45200|issue:45229",
+        "issue:43278|issue:43844",
+        "issue:43257|issue:43299",
+        "issue:43646|issue:44466",
+        "issue:43824|issue:44589",
+        "issue:39692|issue:43295",
+        "issue:43688|issue:45701",
+        "issue:44589|issue:45520",
+        "issue:33357|issue:33666",
+        "issue:44423|issue:44734"
+      ],
+      "prepared_review_unit_hash": "53d39a5428be6bc5353f353a8e8785ac590fe9fe274682895d0b6a772558d36a",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 37264,
+        "estimated_input_tokens": 9316,
+        "estimated_eval_tokens": 18888
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is heterogeneous: most pairs are only loosely similar at the model/regression theme level, not true duplicates. I would not merge any of the soft-edge pairs; 43646 is the best broad representative if a single issue must anchor the cluster.",
+        "confidence": 0.72,
+        "canonical_issue_reason": "No clear duplicate canonical exists across the cluster; issue 43646 is the broadest and most general v5 regression report, so it best fits as a representative anchor.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43646 is the most general user-facing breakage report in the set and covers a broad custom-model initialization regression, making it the least misleading cluster representative.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43493",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Different bugs: SigLIP2 implementation discrepancy vs pipeline translation-task error messaging."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Audio sampling-rate default bug and processor.tokenizer regression are unrelated code paths."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Both involve rotary settings, but one is missing RoPE layers while the other is config reload reverting a percentage value."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different Qwen3.5 issues: hidden-states behavior vs save_pretrained writing wrong visual encoder keys."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:43716",
+            "accept": false,
+            "reason": "Unrelated multimodal failures: Qwen3OmniMoe video processing vs Mistral-3 image-preprocessor dtype mismatch."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "BitNet packed-weight loading bug and DiaConfig token-id persistence bug are separate failures."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "CPU offload device mismatch is unrelated to config token ID loss."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "PyTorch import missing vs Trainer GPU broadcast TypeError are different issues."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "ImportError on missing PyTorch is not the same as a MacOS bus error on a CLIP model."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Similar symptom class, but different configs/models and different missing fields ('pad_token_id' vs 'pad_token_idx')."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "BLIP2 dtype loading issue and MPS attention correctness issue affect different subsystems and code paths."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Aux-loss normalization bug is unrelated to T5's apex RMSNorm memory leak."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Qwen2-VL multi-GPU training request and custom Trainer signature mismatch are different problems."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Documentation example errors are not the same as the SigLIP2 implementation discrepancy."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Packed-weight loading and MLukeTokenizer task AttributeError are unrelated bugs."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Phi-4 multimodal load failure is unrelated to a PEFT version pin bump."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44625",
+            "accept": false,
+            "reason": "Qwen video-input regression and Qwen3.5 num_labels propagation bug are distinct multimodal/config issues."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Different Qwen3* video/audio-generation paths; one is video input processing, the other flash-attention generation behavior."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "SigLIP2 discrepancy and ZeRO-3 gradient growth are different model/training bugs."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44843",
+            "accept": false,
+            "reason": "Loading a specific model and offline tokenizer patch behavior are separate issues."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "RoPE layer count bug and feature-length helper behavior in qwen3_omni_moe are not the same defect."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "mm_token_type_ids defaulting and multi-GPU CUDA OOM are unrelated problems."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "Embedding dtype drift during eval is not the same as gradient explosion with ZeRO-3."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43299",
+            "accept": false,
+            "reason": "Qwen3 MOE weight conversion/loading failures are related only broadly; the specific breakages differ."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Custom model initialization regression and lm_head.weight serialization inconsistency are different code paths."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Missing Qwen2.5-VL import and Float8 storage lookup error are unrelated import/runtime failures."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Doc-example mistakes and v4.57.5 processor API regression are not the same bug."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Auxiliary loss normalization and tokenization changes across versions are unrelated."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Float8 storage lookup error and Python 3.13 flash_attn KeyError are distinct compatibility issues."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "MacOS bus error with a community CLIP model is unrelated to Qwen2-VL multi-GPU training."
+          },
+          {
+            "left": "issue:44423",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Both are serve-related crashes, but one is multimodal input handling and the other is KV-cache continuation tensor indexing."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 20,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:15354",
+        "issue:28218",
+        "issue:29942",
+        "issue:30333",
+        "issue:33357",
+        "issue:34567",
+        "issue:38175",
+        "issue:43232",
+        "issue:43278",
+        "issue:43504",
+        "issue:43540",
+        "issue:43572",
+        "issue:43575",
+        "issue:43606",
+        "issue:43653",
+        "issue:43688",
+        "issue:43701",
+        "issue:43844",
+        "issue:43873",
+        "issue:43927",
+        "issue:43994",
+        "issue:44112",
+        "issue:44291",
+        "issue:44315",
+        "issue:44361",
+        "issue:44464",
+        "issue:44493",
+        "issue:44521",
+        "issue:44554",
+        "issue:44568",
+        "issue:44625",
+        "issue:44792",
+        "issue:44849",
+        "issue:44964",
+        "issue:45072",
+        "issue:45084",
+        "issue:45198",
+        "issue:45200",
+        "issue:45290",
+        "issue:45325",
+        "issue:45356",
+        "issue:45405",
+        "issue:45446",
+        "issue:45520",
+        "issue:45584",
+        "issue:45701",
+        "issue:45759"
+      ],
+      "soft_pairs": [
+        "issue:43504|issue:43606",
+        "issue:43572|issue:44291",
+        "issue:43927|issue:44361",
+        "issue:43504|issue:43653",
+        "issue:44625|issue:45759",
+        "issue:43575|issue:45405",
+        "issue:43606|issue:44112",
+        "issue:44792|issue:45405",
+        "issue:44521|issue:45290",
+        "issue:43994|issue:44493",
+        "issue:44464|issue:45584",
+        "issue:29942|issue:30333",
+        "issue:28218|issue:34567",
+        "issue:43844|issue:44964",
+        "issue:43540|issue:45200",
+        "issue:43688|issue:43873",
+        "issue:45084|issue:45520",
+        "issue:44554|issue:45446",
+        "issue:15354|issue:38175",
+        "issue:44568|issue:45701",
+        "issue:45072|issue:45198",
+        "issue:43232|issue:44849",
+        "issue:43278|issue:43701",
+        "issue:28218|issue:33357",
+        "issue:44315|issue:44849",
+        "issue:45200|issue:45325",
+        "issue:44554|issue:45356"
+      ],
+      "prepared_review_unit_hash": "46ae78ffcdf2bb45cd26d2f19ee02361e16fa2fd7b5d1df21a8b1c764a54c367",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 27,
+        "serialized_chars": 35194,
+        "estimated_input_tokens": 8799,
+        "estimated_eval_tokens": 17854
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is highly heterogeneous: the issues span unrelated tokenizer, generation, multimodal, config, training, CI, and backend problems. None of the soft pairs look like the same underlying bug/change, so all soft edges should be rejected.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "No single canonical issue stands out because the cluster is not a coherent duplicate set; it contains many unrelated problem types across different models and code paths.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No best issue for deduplication purposes: none of these issues is representative of the whole set, and several are clearly independent bug reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43504",
+            "right": "issue:43606",
+            "accept": false,
+            "reason": "Different failures: BEiT semantic segmentation preset loading/legacy field vs Bark CPU offload device mismatch."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Both concern config/loading regressions, but one is missing pad_token_idx in StableLmConfig while the other is an init_empty_weights TypeError from an unexpected argument."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Different tokenizers and symptoms: DiaConfig save/load loses custom token IDs, while MLukeTokenizer raises AttributeError on tasks."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Unrelated models and bugs: BEiT legacy-field loading vs BigBirdTokenizer mask token registration/empty decode output."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Related family, but not the same bug: Qwen3.5 num_labels propagation vs AutoModelForCausalLM unwrapping text_config for composite multimodal configs."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Different problem classes: Qwen2-57B-A14B-Instruct tensor-parallel OOM vs an unreleased MIN_PEFT_VERSION bump."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Both are CI/runtime device issues, but they affect different models and code paths: Bark CPU offload mismatch vs stale GraniteSpeech device override test."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "A janus model test failure is unrelated to the PEFT version bump issue."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both touch chat/template handling, but one is all-zero assistant masks for multimodal inputs and the other crashes on tool-call assistant messages with no content."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Different symptoms: Siglip2 nonsensical outputs with AutoModel/pipeline vs widespread unexpected key handling for position ids."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Different generation bugs: inconsistent chunked generation with compiled forward vs Whisper empty-transcription failure after align_special_tokens."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:30333",
+            "accept": false,
+            "reason": "Flash Attention 2 test failures are not the same as MLflow not reporting failed training jobs."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Tokenizer spacing after added token is unrelated to TrainerState.num_input_tokens_seen not updating."
+          },
+          {
+            "left": "issue:43844",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different model/load paths: ZeRO-3 gradient abnormal increase vs Phi-4 multimodal model load failure."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Different multimodal failures: Qwen3OmniMoe video processing ValueError vs Gemma 4 mm_token_type_ids defaulting issue."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Auxiliary loss normalization in MoE models is unrelated to generic offloading/quantization behavior."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Different import/parser problems: template node compilation TypeError vs flash_attn KeyError on Python 3.13."
+          },
+          {
+            "left": "issue:44554",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "MPS attention correctness with mismatched head dims is unrelated to the PyTorch version check for AuxRequest import in flex_attention."
+          },
+          {
+            "left": "issue:15354",
+            "right": "issue:38175",
+            "accept": false,
+            "reason": "GeneratorExp/TorchScript export failure is unrelated to Siglip2 zero probability outputs."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "mdeberta-v3-base add_special_tokens regression is not the same as broader tokenization changes across transformers versions."
+          },
+          {
+            "left": "issue:45072",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Different dtype/load issues: SwitchTransformers/TimmWrapperModel bfloat16 mismatch vs Wav2Vec2 save_pretrained/tokenization failure."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44849",
+            "accept": false,
+            "reason": "Different generation/config bugs: sync_gpus update_model_kwargs issue vs Qwen3.5 output_hidden_states bug."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Embedding dtype drift in eval is unrelated to resume_from_checkpoint key mismatch."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Tokenizer added-space behavior is unrelated to a MacOS bus error in a CLIP community model."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44849",
+            "accept": false,
+            "reason": "Liger Kernel not applied with model_init is unrelated to Qwen3.5 output_hidden_states behavior."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Gemma 4 mm_token_type_ids defaulting is unrelated to Qwen2.5-VL rope index scaling for still images."
+          },
+          {
+            "left": "issue:44554",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "MPS attention head-dim correctness is unrelated to Kimi-K2.5 tokenizer codec handling and warning regression."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 21,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:37428",
+        "issue:38175",
+        "issue:42915",
+        "issue:43262",
+        "issue:43278",
+        "issue:43295",
+        "issue:43334",
+        "issue:43425",
+        "issue:43450",
+        "issue:43493",
+        "issue:43525",
+        "issue:43572",
+        "issue:43575",
+        "issue:43653",
+        "issue:43701",
+        "issue:43819",
+        "issue:43824",
+        "issue:43827",
+        "issue:43844",
+        "issue:43856",
+        "issue:43873",
+        "issue:43901",
+        "issue:43957",
+        "issue:43975",
+        "issue:43994",
+        "issue:44112",
+        "issue:44206",
+        "issue:44360",
+        "issue:44442",
+        "issue:44464",
+        "issue:44466",
+        "issue:44514",
+        "issue:44521",
+        "issue:44561",
+        "issue:44610",
+        "issue:44792",
+        "issue:44857",
+        "issue:44898",
+        "issue:44936",
+        "issue:44987",
+        "issue:45083",
+        "issue:45127",
+        "issue:45198",
+        "issue:45245",
+        "issue:45412",
+        "issue:45563",
+        "issue:45584"
+      ],
+      "soft_pairs": [
+        "issue:43295|issue:44610",
+        "issue:43334|issue:43525",
+        "issue:44466|issue:45127",
+        "issue:44898|issue:45412",
+        "issue:43493|issue:43701",
+        "issue:37428|issue:43957",
+        "issue:43572|issue:44987",
+        "issue:43572|issue:43957",
+        "issue:43873|issue:44464",
+        "issue:43295|issue:43450",
+        "issue:43824|issue:43975",
+        "issue:43844|issue:44936",
+        "issue:43295|issue:43873",
+        "issue:44514|issue:44521",
+        "issue:38175|issue:45412",
+        "issue:43295|issue:43994",
+        "issue:43278|issue:43856",
+        "issue:37428|issue:43824",
+        "issue:44442|issue:45198",
+        "issue:38175|issue:43493",
+        "issue:42915|issue:43824",
+        "issue:43827|issue:43901",
+        "issue:43653|issue:44112",
+        "issue:43450|issue:43493",
+        "issue:43450|issue:43575",
+        "issue:44792|issue:45563",
+        "issue:44360|issue:44561",
+        "issue:45127|issue:45245",
+        "issue:43819|issue:45083",
+        "issue:44206|issue:45584",
+        "issue:43262|issue:43450",
+        "issue:43819|issue:44112",
+        "issue:44610|issue:44857",
+        "issue:37428|issue:43425"
+      ],
+      "prepared_review_unit_hash": "8ba3a3f95f079eedc83f052a69ce56d841af5ac22eb40e4dc5b507e6360aeae7",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 34,
+        "serialized_chars": 37294,
+        "estimated_input_tokens": 9324,
+        "estimated_eval_tokens": 18904
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a mixed bag of distinct transformer regressions, with only a couple of true duplicate-like pairs: one around SigLIP2 output correctness and one around missing `pad_token_id` on model configs. The rest are related only at a broad subsystem level (processors, tokenizers, multimodal loading) and should stay separate.",
+        "confidence": 0.69,
+        "canonical_issue_reason": "Issue 43295 is the broadest representative of the processor/tokenizer regression subset: it describes a v4.57.5 breakage in multimodal processor access and is the most central umbrella title among the processor-related reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43295 is the best overall issue representative because it captures a broad, high-impact regression in processor/tokenizer behavior and best summarizes the most recurring theme in this cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43295",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Different bugs: custom model code accessing `processor.tokenizer` vs OmDet-Turbo image size mismatch."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43525",
+            "accept": true,
+            "reason": "Same underlying regression class: both report missing `pad_token_id` on model configs during load."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45127",
+            "accept": false,
+            "reason": "Different code paths: serialization of tied `lm_head.weight` vs LoRA merge collapse with extended vocab."
+          },
+          {
+            "left": "issue:44898",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Unrelated failures: Perceiver positional interpolation vs RT-DETR memory release after GC."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "SigLIP2 output discrepancy vs training checkpoint key mismatch; not the same bug."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Import error for `_flash_supports_window_size` vs meta-device model loading breakage."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Both are load/regression reports, but for different models and failure modes; not one fix."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Missing `pad_token_idx` on StableLmConfig is separate from meta-device loading failures."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44464",
+            "accept": false,
+            "reason": "Quantization/offloading behavior vs compiled chunked generation inconsistency; distinct bugs."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Processor/tokenizer regression is not the same as batched video shape output bug."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43975",
+            "accept": false,
+            "reason": "Different issues: Qwen2.5-VL import failure vs DeepSeek detokenization regression."
+          },
+          {
+            "left": "issue:43844",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "ZeRO-3 gradient growth vs trainer evaluate-after-train failure are unrelated."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Processor/tokenizer regression is separate from quantization/offloading problems."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Both involve `apply_chat_template`, but one is a batching crash and the other is incorrect assistant masks; too different to merge."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "SigLIP2 zero probabilities and RT-DETR memory leak are unrelated."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Shared model/tooling area only; the titles point to different SigLIP2 vs processor-tokenizer problems."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "Embedding dtype drift during evaluate is unrelated to Qwen3 MoE memory efficiency."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but they concern different symbols/models and different fixes."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Tokenizer loading failure and save_pretrained/tokenization failure are different model-specific bugs."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43493",
+            "accept": true,
+            "reason": "Both describe SigLIP2 output correctness discrepancies; likely the same underlying implementation bug."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config failure is unrelated to a Qwen2.5-VL import error."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Docs regressions around `pipeline()` and `return_all_scores` are separate documentation issues."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Different tokenizer/CI issues: BigBird special-token registration vs GraniteSpeech device override test."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Batched video shape bug and SigLIP2 implementation discrepancy are different subsystems."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Video processor shape handling and Qwen2-57B tensor-parallel OOM are unrelated."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Janus image-generation test failure vs stale paged-generate warning; not the same issue."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "DSA indexer discussion vs removal of `is_torch_fx_available` breaking remote-code models."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "LoRA merge collapse and category-cardinality runtime error have no shared code path."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "DAC latent reconstruction mismatch and Qwen3 Omni feature-length helper behavior are distinct."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Feature extractor center-arg crash and Whisper empty-transcription failure are separate bugs."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Audio chat-template sample-rate default bug is unrelated to batched video shape output."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "DAC forward mismatch and GraniteSpeech CI test failure do not share the same bug."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "OmDet-Turbo processor resolution bug is unrelated to LwDetrImageLoss AMP/CUDA crashing."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43425",
+            "accept": false,
+            "reason": "Flash-attention import failure and Torch 2.10 compatibility are different breakages."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary stays within the packet and is appropriately cautious. The two accepted soft-edge pairs are plausibly mergeable at the bug-class level, while the rejected pairs are framed conservatively as distinct failure modes. No clear overstatement or unsupported canonicalization stands out."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 22,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:28218",
+        "issue:29942",
+        "issue:32090",
+        "issue:35141",
+        "issue:37428",
+        "issue:38175",
+        "issue:39692",
+        "issue:43278",
+        "issue:43335",
+        "issue:43388",
+        "issue:43404",
+        "issue:43504",
+        "issue:43526",
+        "issue:43550",
+        "issue:43606",
+        "issue:43611",
+        "issue:43653",
+        "issue:43720",
+        "issue:43819",
+        "issue:43825",
+        "issue:43931",
+        "issue:44112",
+        "issue:44292",
+        "issue:44351",
+        "issue:44448",
+        "issue:44462",
+        "issue:44493",
+        "issue:44561",
+        "issue:44743",
+        "issue:44779",
+        "issue:44855",
+        "issue:44857",
+        "issue:44977",
+        "issue:45072",
+        "issue:45083",
+        "issue:45198",
+        "issue:45200",
+        "issue:45229",
+        "issue:45290",
+        "issue:45310",
+        "issue:45381",
+        "issue:45459",
+        "issue:45478",
+        "issue:45520",
+        "issue:45588",
+        "issue:45593",
+        "issue:45712",
+        "issue:45803"
+      ],
+      "soft_pairs": [
+        "issue:45290|issue:45381",
+        "issue:44857|issue:45593",
+        "issue:43278|issue:43825",
+        "issue:45200|issue:45588",
+        "issue:43606|issue:45083",
+        "issue:44448|issue:44743",
+        "issue:43335|issue:43550",
+        "issue:37428|issue:44351",
+        "issue:43611|issue:43931",
+        "issue:43388|issue:43526",
+        "issue:39692|issue:43526",
+        "issue:44561|issue:44779",
+        "issue:44462|issue:44561",
+        "issue:43819|issue:44493",
+        "issue:43335|issue:43653",
+        "issue:43404|issue:45200",
+        "issue:45310|issue:45478",
+        "issue:44855|issue:45083",
+        "issue:39692|issue:44493",
+        "issue:29942|issue:32090",
+        "issue:43550|issue:44112",
+        "issue:43335|issue:45083",
+        "issue:43335|issue:43504",
+        "issue:44292|issue:45520",
+        "issue:38175|issue:44493",
+        "issue:44448|issue:44779",
+        "issue:28218|issue:35141",
+        "issue:43335|issue:45072",
+        "issue:44977|issue:45229",
+        "issue:43335|issue:45593",
+        "issue:43720|issue:44112",
+        "issue:44448|issue:44977",
+        "issue:45459|issue:45712",
+        "issue:45198|issue:45803"
+      ],
+      "prepared_review_unit_hash": "6996b83567290e63caec9ce5d1eeba360881fa7a95d2744ee08fcd4a76d5ac82",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 38161,
+        "estimated_input_tokens": 9541,
+        "estimated_eval_tokens": 19338
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The only clear duplicate pair is the repeated Qwen3.5 MoE from_pretrained error report (issues 45310 and 45478). The rest are separate bugs touching different models, APIs, or failure modes.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "Issue 45310 is the earlier report of the same Qwen3.5 MoE from_pretrained failure and has slightly more context/activity, so it is the better canonical issue.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45310 is the strongest representative overall because it is the earlier, fuller report of the duplicated Qwen3.5 MoE loading bug.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45290",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Different features and failure modes: chat template/tool-call handling vs Qwen2.5-VL video position IDs."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Unrelated model bugs: AMP/CUDA loss crash vs missing auxiliary losses when denoising is off."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Different areas: dtype regression in evaluate vs pipeline error-message wording for translation tasks."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Different code paths: Gemma 4 text-only mm_token_type_ids vs flash_attention integration crashing on sink-less models."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Different subsystems and symptoms: CPU offload device mismatch vs feature-length helper behavior."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Different model behaviors: Pegasus output drift vs Qwen3.5 recurrent-state reset bug."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Different problems: SwitchTransformers config layer creation vs Bamba SDPA/torch.compile failure."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but they reference different missing symbols and likely different breakages."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different loading failures: base_model_prefix regression vs Qwen3-VL weight shape mismatch."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Different utilities and bugs: metrics gathering on tuple labels vs Beit fast processor reduce_labels behavior."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "SigLIP2 docs/example issues are unrelated to BeitImageProcessorFast label reduction."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Different regressions: removed helper breaking trust_remote_code vs Deepseek tokenizer output change."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "Tokenizer loading behavior vs v5 helper removal; not the same underlying bug."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Different symptoms and code paths: DAC latent reconstruction vs unexpected position-id keys."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Different models and bugs: SwitchTransformers sparse-layer config vs BigBirdTokenizer special-token registration."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Different model classes and issues: lm_head tying in Mistral3 vs Gemma 4 token-type defaults."
+          },
+          {
+            "left": "issue:45310",
+            "right": "issue:45478",
+            "accept": true,
+            "reason": "Same reported bug: identical Qwen3.5 MoE from_pretrained error on transformers>=5.4.0."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Different failures: Python 3.13 JIT parsing/indentation issue vs qwen3_omni_moe feature-length helper."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Different problems: SigLIP2 documentation errors vs position-id key warnings across models."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Unrelated CI failures: Flash Attention 2 tests vs trainer GPU broadcast NoneType error."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Different CI/runtime issues: Bamba torch.compile SDPA failure vs GraniteSpeech stale device override test."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Different model/config bug vs helper-function behavior in qwen3_omni_moe."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43504",
+            "accept": false,
+            "reason": "Different models and symptoms: SwitchTransformers config vs BEiT semantic segmentation pretrained loading."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Different import/runtime issues: Qwen-3-8B-NVFP4 execution vs flash_attn import KeyError on Python 3.13."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Different model-specific output issue vs broad unexpected position-id key warnings."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Different models and bugs: Pegasus output differences vs Deepseek tokenizer regression."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Different tokenizer spacing bug vs embedding reinitialization after resize."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Different failure modes: sparse-layer config bug vs dtype mismatch in bfloat16 inference."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Different issues: flash-attention generation correctness vs multi-GPU CUDA OOM."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Different model behavior: SwitchTransformers config bug vs missing auxiliary losses in D-FINE."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Different CI/runtime concerns: BitNet packed-weight loading vs GraniteSpeech test flake."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Different models and bugs: Pegasus version output drift vs Qwen3.5 flash-attention generation issue."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Different scope: tokenizer protobuf error handling vs leftover dummy classes leaking into the package namespace."
+          },
+          {
+            "left": "issue:45198",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Different tokenizer/save_pretrained failure vs bare-except error swallowing in FuyuBatchFeature."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The conclusion is grounded in the packet: 45310 and 45478 are the same Qwen3.5 MoE from_pretrained bug, with matching titles and the strongest similarity score. The remaining soft pairs are rejected conservatively and the reasons generally track the issue titles without overstating identity or mergeability."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 23,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:25251",
+        "issue:29127",
+        "issue:29942",
+        "issue:32090",
+        "issue:34689",
+        "issue:35141",
+        "issue:36010",
+        "issue:37428",
+        "issue:40444",
+        "issue:43065",
+        "issue:43066",
+        "issue:43278",
+        "issue:43381",
+        "issue:43412",
+        "issue:43526",
+        "issue:43550",
+        "issue:43606",
+        "issue:43653",
+        "issue:43701",
+        "issue:43873",
+        "issue:43927",
+        "issue:43976",
+        "issue:43994",
+        "issue:44186",
+        "issue:44361",
+        "issue:44448",
+        "issue:44451",
+        "issue:44462",
+        "issue:44466",
+        "issue:44493",
+        "issue:44568",
+        "issue:44743",
+        "issue:44857",
+        "issue:44977",
+        "issue:45072",
+        "issue:45083",
+        "issue:45198",
+        "issue:45245",
+        "issue:45372",
+        "issue:45412",
+        "issue:45440",
+        "issue:45459",
+        "issue:45491",
+        "issue:45520",
+        "issue:45563",
+        "issue:45588",
+        "issue:45698"
+      ],
+      "soft_pairs": [
+        "issue:44568|issue:45491",
+        "issue:43606|issue:44857",
+        "issue:29127|issue:32090",
+        "issue:40444|issue:43526",
+        "issue:44857|issue:45072",
+        "issue:34689|issue:35141",
+        "issue:43927|issue:45698",
+        "issue:43701|issue:45083",
+        "issue:44977|issue:45563",
+        "issue:45372|issue:45588",
+        "issue:43550|issue:45083",
+        "issue:43412|issue:45588",
+        "issue:45245|issue:45563",
+        "issue:43873|issue:44493",
+        "issue:29942|issue:35141",
+        "issue:36010|issue:37428",
+        "issue:45372|issue:45459",
+        "issue:43066|issue:44462",
+        "issue:43065|issue:45412",
+        "issue:43278|issue:43381",
+        "issue:44857|issue:45588",
+        "issue:25251|issue:32090",
+        "issue:44448|issue:44466",
+        "issue:43653|issue:44743",
+        "issue:44451|issue:45520",
+        "issue:44186|issue:45198",
+        "issue:43976|issue:45440",
+        "issue:43994|issue:45412",
+        "issue:44361|issue:45198"
+      ],
+      "prepared_review_unit_hash": "6d6435d0e3fefb8b92dafb6b92bbbbd24d13c58e8bc9f71221db3e619f2291cf",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 35869,
+        "estimated_input_tokens": 8968,
+        "estimated_eval_tokens": 18192
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is highly heterogeneous: most pairs share only broad vocabulary (tokenization, loading, dtype, import, eval) but describe different bugs and code paths. I rejected all soft edges because none look like the same underlying issue or a mergeable PR pair.",
+        "confidence": 0.89,
+        "canonical_issue_reason": "No single issue cleanly dominates as a duplicate anchor because the cluster spans unrelated reports. If one must be chosen, issue 43873 is the best anchor: it is open, broadly phrased, and has the most follow-up activity/inbound references.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43873 is the best representative issue in this set because it is broad enough to serve as a future duplicate target and has the strongest activity signal among the open reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44568",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Different bugs: tokenizer special-token behavior vs NaN embeddings in Gemma3 sliding-window attention."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Both are runtime failures, but one is device offload mismatch and the other is AMP/CUDA loss crash; different code paths."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Unrelated problems: LayoutLMv3 error-message clarity vs Trainer GPU broadcast failing on NoneType."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Different components and failures: multimodal iterable dataset handling vs BeitImageProcessorFast label reduction."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Both mention dtype, but one is a loss crash under AMP and the other is inference mismatch with bfloat16 in CI."
+          },
+          {
+            "left": "issue:34689",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Model-loading regression vs output embedding reinitialization after resize; not the same bug."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Both involve save/load, but one loses custom token IDs and the other loads the wrong custom module; different serialization failures."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Resume checkpoint key mismatch is unrelated to the qwen3_omni_moe feature-length helper issue."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Different generation issues: flash-attention correctness vs a stale warning for num_return_sequences."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Different failures: missing ReasoningEffort import vs flash-attention crash on s_aux=None."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "torch.compile + SDPA model failure is unrelated to the feature-length helper bug."
+          },
+          {
+            "left": "issue:43412",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Executorch export/unsupported op issues are not the same as the flash_attention.py AttributeError."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Category-limit RuntimeError and a stale warning in paged generate are unrelated."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Quantization/offloading behavior is different from unexpected position-id keys during loading."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Flash Attention 2 test failures are unrelated to embedding reinitialization after resize."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:37428",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and different subsystems."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "One is a missing dependency import, the other is an overly broad exception masking tokenizer errors."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:44462",
+            "accept": false,
+            "reason": "Tokenizer decoder-type mismatch is not the same as AutoTokenizer ignoring tokenizer.json."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Dummy Conv2d in Sam3PixelDecoder is unrelated to RT-DETR memory not being released."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "BF16-to-FP32 dtype drift in evaluation is different from gradient checkpointing being disallowed in eval mode."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Different runtime paths: AMP loss crash vs flash_attention AttributeError on sink-less models."
+          },
+          {
+            "left": "issue:25251",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Pipeline top_k nesting behavior has nothing to do with Trainer broadcast errors."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Both are version regressions, but one is generation output differences and the other is state-dict serialization of tied weights."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Tokenizer special-token registration and recurrent-state reset are unrelated."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Model load failure for ScandiBERT is unrelated to the Python 3.13 flash_attn import KeyError."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Different tokenizers and different failure modes: LayoutLMv2 NER/padding crash vs Wav2Vec2 save_pretrained/tokenization failure."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Python-version incompatibility is unrelated to DeepseekV3MoE divergence from the remote implementation."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Flash-attention generation correctness and a stale warning are different problems."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Not the same tokenizer bug; one crashes on NER padding/truncation, the other on save/load and tokenization."
+          },
+          {
+            "left": "issue:44493",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Unexpected position-id keys and flash_attention s_aux=None are different issues."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "MLukeTokenizer AttributeError is unrelated to Wav2Vec2 save_pretrained/tokenization failure."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 24,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:29127",
+        "issue:29942",
+        "issue:30064",
+        "issue:32090",
+        "issue:33357",
+        "issue:34567",
+        "issue:36010",
+        "issue:39692",
+        "issue:41720",
+        "issue:42915",
+        "issue:42994",
+        "issue:43065",
+        "issue:43278",
+        "issue:43329",
+        "issue:43334",
+        "issue:43381",
+        "issue:43421",
+        "issue:43526",
+        "issue:43531",
+        "issue:43575",
+        "issue:43653",
+        "issue:43701",
+        "issue:43723",
+        "issue:43756",
+        "issue:43824",
+        "issue:43873",
+        "issue:43994",
+        "issue:44186",
+        "issue:44206",
+        "issue:44265",
+        "issue:44279",
+        "issue:44442",
+        "issue:44479",
+        "issue:44554",
+        "issue:44610",
+        "issue:44734",
+        "issue:44871",
+        "issue:44977",
+        "issue:45103",
+        "issue:45198",
+        "issue:45229",
+        "issue:45245",
+        "issue:45372",
+        "issue:45406",
+        "issue:45412",
+        "issue:45440",
+        "issue:45491",
+        "issue:45593"
+      ],
+      "soft_pairs": [
+        "issue:42915|issue:42994",
+        "issue:44265|issue:45198",
+        "issue:44871|issue:44977",
+        "issue:43421|issue:45491",
+        "issue:44206|issue:44479",
+        "issue:43065|issue:43994",
+        "issue:44734|issue:45406",
+        "issue:29942|issue:30064",
+        "issue:33357|issue:39692",
+        "issue:29127|issue:34567",
+        "issue:44610|issue:45412",
+        "issue:45103|issue:45372",
+        "issue:43873|issue:45229",
+        "issue:43756|issue:45440",
+        "issue:43723|issue:44554",
+        "issue:45372|issue:45406",
+        "issue:43278|issue:43575",
+        "issue:43381|issue:43701",
+        "issue:43278|issue:43329",
+        "issue:43653|issue:44442",
+        "issue:41720|issue:45229",
+        "issue:45440|issue:45593",
+        "issue:29942|issue:36010",
+        "issue:32090|issue:33357",
+        "issue:43334|issue:43531",
+        "issue:43526|issue:44186",
+        "issue:44977|issue:45245",
+        "issue:43824|issue:44279"
+      ],
+      "prepared_review_unit_hash": "233f207ffccb456899eb129b9abfba5cfa6510d79b77c81e61d62b1d67b5f5c1",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36166,
+        "estimated_input_tokens": 9042,
+        "estimated_eval_tokens": 18340
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are a mixed set of mostly unrelated Transformers issues; none of the soft pairs look like the same underlying bug/change, so no duplicate edges should be accepted.",
+        "confidence": 0.92,
+        "canonical_issue_reason": "No single issue is a true canonical duplicate target here; if an anchor is needed, issue 43873 is the broadest and most active report in the set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43873 is the best representative overall because it has the broadest scope, strong discussion activity, and inbound references, making it the most central candidate in this otherwise unrelated set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42915",
+            "right": "issue:42994",
+            "accept": false,
+            "reason": "Both involve quantization, but one is a FineGrainedFP8 model failure and the other is quantized model saving; different code paths and symptoms."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Unrelated failures: torch.export with torch_compilable_check versus Wav2Vec2 save_pretrained/tokenization."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Gemma eos_token config mismatch is separate from Qwen3.5 flash-attention generation failure."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Runtime special-token post-processor updates are unrelated to Gemma3 NaN embeddings from sliding-window attention."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Unsupported feature-processor arg in LasrFeatureExtractor is not the same regression as Qwen VL video-input handling."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "A dummy Conv2d in Sam3PixelDecoder is unrelated to SigLIP2 AutoModel/pipeline nonsense outputs."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both touch transformers serve, but one is KV-cache continuation indexing and the other is Gemma4Processor initialization."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "Flash Attention 2 test failures and void segmentation-map processing are different subsystems and bugs."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:39692",
+            "accept": false,
+            "reason": "MacOS bus error with a community CLIP model is unrelated to SigLIP2 documentation example mistakes."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "LayoutLMv3 box error messaging and TrainerState token counting are unrelated issues."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "OmDet-Turbo processor size mismatch is unrelated to RT-DETR memory release behavior."
+          },
+          {
+            "left": "issue:45103",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "auto_docstring annotation parsing and Gemma4 processor import failure are different bugs."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Quantization offloading problems and Gemma4 multi-GPU CUDA OOM are not the same failure mode."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Smollm3 RoPE-layer mismatch and DeepseekV3 divergence are distinct model-implementation issues."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "Tokenizer loading in v5 is unrelated to MPS attention correctness with mismatched value head dims."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Gemma 4 processor loading import failure is separate from transformers serve missing _tokenizer."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Embedding dtype drift between train/eval is unrelated to tensor-parallel OOM."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode is a different issue from resume_from_checkpoint key mismatch."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Embedding dtype changes and undefined video-token-count helpers are unrelated."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration and FastSpeech2Conformer tokenizer loading are different tokenizer bugs."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping cudaErrorAssert is not the same as Gemma4 multi-GPU OOM."
+          },
+          {
+            "left": "issue:45440",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "DeepseekV3 implementation divergence is unrelated to D-FINE auxiliary-loss handling."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Flash Attention 2 tests and GenerationMixin import errors are unrelated."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Trainer broadcast NoneType and MacOS CLIP bus error are different failures."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Qwen3-VL pad_token_id loading failure is unrelated to Qwen3-MoE sliding_window behavior."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44186",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction bug and LayoutLMv2Tokenizer NER/padding crash are unrelated."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "Qwen3.5 flash-attention generation and 2^24 category overflow are different problems."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "A specific Qwen2.5 VL import error is not the same as a generic dependency issue with transformers."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 25,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:28218",
+        "issue:29127",
+        "issue:32090",
+        "issue:33290",
+        "issue:33357",
+        "issue:34567",
+        "issue:34689",
+        "issue:35141",
+        "issue:38175",
+        "issue:42907",
+        "issue:42915",
+        "issue:43262",
+        "issue:43316",
+        "issue:43441",
+        "issue:43452",
+        "issue:43525",
+        "issue:43526",
+        "issue:43575",
+        "issue:43611",
+        "issue:43646",
+        "issue:43701",
+        "issue:43761",
+        "issue:43931",
+        "issue:44077",
+        "issue:44079",
+        "issue:44188",
+        "issue:44206",
+        "issue:44387",
+        "issue:44442",
+        "issue:44464",
+        "issue:44479",
+        "issue:44610",
+        "issue:44743",
+        "issue:44811",
+        "issue:44857",
+        "issue:44964",
+        "issue:44977",
+        "issue:45072",
+        "issue:45216",
+        "issue:45229",
+        "issue:45290",
+        "issue:45440",
+        "issue:45491",
+        "issue:45584",
+        "issue:45588",
+        "issue:45600",
+        "issue:45698",
+        "issue:45803"
+      ],
+      "soft_pairs": [
+        "issue:43526|issue:44857",
+        "issue:43526|issue:44610",
+        "issue:43262|issue:44206",
+        "issue:42915|issue:43931",
+        "issue:43441|issue:45588",
+        "issue:44442|issue:45072",
+        "issue:44464|issue:45491",
+        "issue:43452|issue:45588",
+        "issue:44743|issue:44811",
+        "issue:43575|issue:43701",
+        "issue:44442|issue:45803",
+        "issue:43646|issue:44077",
+        "issue:29127|issue:35141",
+        "issue:43575|issue:44387",
+        "issue:45072|issue:45229",
+        "issue:44977|issue:45584",
+        "issue:33357|issue:38175",
+        "issue:45588|issue:45600",
+        "issue:44479|issue:45290",
+        "issue:43316|issue:43525",
+        "issue:44964|issue:45440",
+        "issue:43761|issue:44079",
+        "issue:28218|issue:29127",
+        "issue:43931|issue:45216",
+        "issue:42907|issue:45698",
+        "issue:28218|issue:32090",
+        "issue:33290|issue:34567",
+        "issue:34689|issue:43611",
+        "issue:43761|issue:44188"
+      ],
+      "prepared_review_unit_hash": "60e80a58c3eb4a893e7d8691e7a19cf2d610c31df2e66a6a97bcd286ed48d978",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36404,
+        "estimated_input_tokens": 9101,
+        "estimated_eval_tokens": 18458
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This is a loose, heterogeneous cluster of Transformers issues; the soft pairs mostly share broad keywords or subsystems rather than the same underlying bug. I would not merge any of the proposed pairs. If a single issue must anchor the cluster, the clearest representative is the save/load regression around `from_pretrained`/`save_pretrained` behavior.",
+        "confidence": 0.39,
+        "canonical_issue_reason": "issue:45698 is the broadest representative of the repeated save/load serialization failures in this cluster, with a concrete repro and active discussion. The cluster is not tight, but this issue best matches the dominant model-loading theme.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45698 is the best single issue to keep as the cluster representative because it is clear, user-facing, and closest to the recurring serialization/loading regressions seen here.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43526",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Different bugs: BeitImageProcessorFast label reduction vs LwDetrImageLoss AMP/CUDA crash."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Unrelated processor issues; one is label reduction, the other is an OmDet-Turbo size mismatch."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44206",
+            "accept": false,
+            "reason": "Different audio/feature-extractor APIs and failures; sampling-rate default vs unsupported `center` arg."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different model families and failure modes: FineGrainedFP8 config vs Qwen3-VL shape mismatch."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Both mention flash attention, but one is a model-specific generation failure and the other is an `AttributeError` in integration code."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Tokenizer loading failure vs bfloat16 dtype mismatch in CI; not the same underlying bug."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Different code paths: compiled chunked generation inconsistency vs Gemma3 NaN embeddings with sliding-window attention."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "GGUF/AutoTokenizer loading breakage is unrelated to the flash-attention `s_aux=None` crash."
+          },
+          {
+            "left": "issue:44743",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Recurrent-state reset in modular Qwen3.5 vs Whisper batch_decode ignoring `skip_special_tokens`."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "OOM during tensor-parallel model load vs checkpoint key mismatch on resume; different problems."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Tokenizer load failure vs bare `except` swallowing interrupts in `convert_to_tensors()`; unrelated."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "Both touch initialization, but one is custom model init regression and the other is a `patchtsmixer` `post_init` policy issue."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "LayoutLMV3 error-message clarity is unrelated to embedding reinitialization after resize."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Both are OOMs, but one is tensor-parallel load OOM and the other is increased reserved memory under int4 quantization."
+          },
+          {
+            "left": "issue:45072",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "CI dtype mismatch vs multi-GPU CUDA OOM; different failure modes."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Qwen3.5 flash-attention generation issue vs Whisper empty-transcription/chat-template bug; unrelated."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:38175",
+            "accept": false,
+            "reason": "MacOS bus error on CLIP model vs SigLIP2 zero probabilities; different models and symptoms."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Flash-attention `AttributeError` vs stale `auto_mappings.py` references; no shared bug."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Video-input regression in Qwen models vs chat-template crash on tool-call assistant messages; different paths."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Both are config/API issues, but they concern different models and different missing attributes."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Phi-4 multimodal loading error vs DeepseekV3MoE divergence; unrelated model families and bugs."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "CLIPVisionModel hidden-states regression vs `ModelOutput` key assignment bug; different abstractions."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "Tokenizer spacing bug vs LayoutLMV3 error clarity issue; not the same underlying defect."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both involve loading/saving Qwen models, but one is a weight-shape mismatch and the other is a bad checkpoint saved by `save_pretrained`."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Both are save/load related, but one is dequantized model saving and the other is wrong custom-module loading after save/load; too different to merge."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Tokenizer added-space bug vs Trainer `_gpu_broadcast_one` NoneType error; unrelated."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Deepspeed/Adafactor OOM vs token counter not updating; different components and symptoms."
+          },
+          {
+            "left": "issue:34689",
+            "right": "issue:43611",
+            "accept": false,
+            "reason": "Both are model-loading regressions, but the Llama 3.2 vision failure and `base_model_prefix` loading bug are distinct root causes."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44188",
+            "accept": false,
+            "reason": "CLIPVisionModel hidden-state regression vs attention-kernel divergence under `torch.compile`; unrelated."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 26,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:29127",
+        "issue:32090",
+        "issue:33290",
+        "issue:33357",
+        "issue:34689",
+        "issue:36010",
+        "issue:39692",
+        "issue:41628",
+        "issue:42175",
+        "issue:43295",
+        "issue:43316",
+        "issue:43381",
+        "issue:43404",
+        "issue:43441",
+        "issue:43493",
+        "issue:43540",
+        "issue:43550",
+        "issue:43643",
+        "issue:43720",
+        "issue:43742",
+        "issue:43827",
+        "issue:43856",
+        "issue:43873",
+        "issue:43994",
+        "issue:44077",
+        "issue:44162",
+        "issue:44190",
+        "issue:44206",
+        "issue:44442",
+        "issue:44479",
+        "issue:44496",
+        "issue:44514",
+        "issue:44568",
+        "issue:44743",
+        "issue:44792",
+        "issue:44855",
+        "issue:44871",
+        "issue:44877",
+        "issue:44898",
+        "issue:44913",
+        "issue:45137",
+        "issue:45372",
+        "issue:45479",
+        "issue:45491",
+        "issue:45584",
+        "issue:45656",
+        "issue:45759"
+      ],
+      "soft_pairs": [
+        "issue:44442|issue:44898",
+        "issue:39692|issue:45491",
+        "issue:43381|issue:43873",
+        "issue:43994|issue:44206",
+        "issue:43493|issue:43827",
+        "issue:43550|issue:44743",
+        "issue:43295|issue:45584",
+        "issue:43316|issue:45759",
+        "issue:44568|issue:45479",
+        "issue:41628|issue:45372",
+        "issue:43720|issue:44442",
+        "issue:43643|issue:44913",
+        "issue:32090|issue:33290",
+        "issue:44190|issue:44496",
+        "issue:34689|issue:36010",
+        "issue:43873|issue:44871",
+        "issue:43381|issue:43856",
+        "issue:44077|issue:44479",
+        "issue:33290|issue:45656",
+        "issue:44855|issue:45137",
+        "issue:43540|issue:44514",
+        "issue:29127|issue:33357",
+        "issue:43381|issue:43742",
+        "issue:43316|issue:44877",
+        "issue:33290|issue:34689",
+        "issue:42175|issue:43316",
+        "issue:44162|issue:44792",
+        "issue:43404|issue:43441"
+      ],
+      "prepared_review_unit_hash": "2ca42eb2b637b6f97380fb1786165e27326a5d71f98977cdf38a49b88cc855ae",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 28,
+        "serialized_chars": 35839,
+        "estimated_input_tokens": 8960,
+        "estimated_eval_tokens": 18176
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is not a duplicate set; it is a heterogeneous grab-bag of unrelated Transformers bugs spanning config loading, multimodal processors, deepspeed, tokenizers, and model-specific regressions. None of the soft pairs look like the same underlying bug or change.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "No single canonical issue fits: the items share only a broad Transformers theme, but their concrete failure modes, code paths, and affected components are mostly unrelated.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No strong best issue; even the broadest-looking reports are still distinct bugs, so any single issue would misrepresent the cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44442",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Tokenizer loading failure vs Perceiver image-classification resolution bug; different models and code paths."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "SigLIP2 doc/example errors and quantization mismatch are unrelated to Gemma3 NaN embeddings from mixed-length batching."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Eval-mode gradient checkpointing and quantization/offloading behavior are different bugs."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44206",
+            "accept": false,
+            "reason": "SigLIP2 output quality issue vs LastFeatureExtractor crash on an unsupported argument; unrelated failures."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "SigLIP2 implementation discrepancy is not the same as stale docs still using pipeline()."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "torch.compile+SDPA failure in Bamba is unrelated to Qwen recurrent-state reset logic."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Both involve processors/tokenizers, but one is a processor.tokenizer regression and the other is empty-transcription handling in Whisper."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Gemma3TextConfig API inconsistency and Qwen composite-config unwrapping are different config bugs."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Special-token insertion bug is unrelated to the zero-loss sequence-classification issue."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and different root causes."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "BitNet packed-weight loading and FastSpeech2Conformer tokenizer loading are unrelated."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Config field loss on remote-code load and GPTNeoX rotary_pct not round-tripping are distinct config persistence bugs."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "Trainer broadcast type error and AdaFactor/deepspeed OOM are unrelated distributed-training issues."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Local dataset loading in a training script is unrelated to an unrecognized model/config.json issue."
+          },
+          {
+            "left": "issue:34689",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Model-loading regression and GenerationMixin import failure are different breakages, even if both relate to version upgrades."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Quantization/offloading problems are unrelated to an eos_token_id config mismatch."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode is not the same as Qwen3 MoE memory inefficiency."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "patchtsmixer post_init policy and Qwen video-input regression are unrelated."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "AdaFactor OOM and duplicate optimizer.step() with deepspeed are different bugs."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Python 3.13 JIT parsing/import failure and DeepSpeed ZeRO3 deque IndexError are unrelated."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "Both are multimodal/video-related, but they fail in different Qwen processor paths and are not the same bug."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "LayoutLMv3 error-message clarity and MacOS CLIP bus error are unrelated."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Eval checkpointing limitation is unrelated to the MobileLLM key error during loading."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Gemma3TextConfig discrepancy and Granite Speech strict-config loading are separate config-handling issues."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Deepspeed/AdaFactor memory pressure does not match a Llama 3.2 model-loading regression."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43316",
+            "accept": false,
+            "reason": "Packaging/backend installation issue is unrelated to Gemma3TextConfig API mismatch."
+          },
+          {
+            "left": "issue:44162",
+            "right": "issue:44792",
+            "accept": false,
+            "reason": "ESM2 breakage and a Janus image-generation test failure affect different models and code paths."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Mistral3 lm_head weight tying and Ministral FlashAttention failure are different model-specific bugs."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 27,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:25251",
+        "issue:28282",
+        "issue:33290",
+        "issue:35141",
+        "issue:36331",
+        "issue:41720",
+        "issue:41950",
+        "issue:42617",
+        "issue:43065",
+        "issue:43232",
+        "issue:43295",
+        "issue:43388",
+        "issue:43452",
+        "issue:43475",
+        "issue:43550",
+        "issue:43575",
+        "issue:43606",
+        "issue:43646",
+        "issue:43673",
+        "issue:43701",
+        "issue:43742",
+        "issue:43761",
+        "issue:43825",
+        "issue:43827",
+        "issue:43906",
+        "issue:43931",
+        "issue:43994",
+        "issue:44060",
+        "issue:44077",
+        "issue:44242",
+        "issue:44464",
+        "issue:44479",
+        "issue:44496",
+        "issue:44534",
+        "issue:44610",
+        "issue:44617",
+        "issue:44734",
+        "issue:44743",
+        "issue:44792",
+        "issue:44871",
+        "issue:44913",
+        "issue:44936",
+        "issue:44977",
+        "issue:45127",
+        "issue:45200",
+        "issue:45290",
+        "issue:45584",
+        "issue:45698"
+      ],
+      "soft_pairs": [
+        "issue:25251|issue:28282",
+        "issue:43232|issue:43388",
+        "issue:43295|issue:43825",
+        "issue:41950|issue:43475",
+        "issue:44871|issue:45584",
+        "issue:43550|issue:43701",
+        "issue:45127|issue:45200",
+        "issue:43606|issue:43701",
+        "issue:43295|issue:44871",
+        "issue:43065|issue:44617",
+        "issue:43761|issue:44479",
+        "issue:33290|issue:35141",
+        "issue:33290|issue:36331",
+        "issue:43575|issue:43742",
+        "issue:43452|issue:44496",
+        "issue:42617|issue:43575",
+        "issue:43646|issue:44913",
+        "issue:43673|issue:44743",
+        "issue:44936|issue:45698",
+        "issue:43761|issue:44242",
+        "issue:41720|issue:43575",
+        "issue:44077|issue:44534",
+        "issue:44479|issue:44734",
+        "issue:43931|issue:44060",
+        "issue:44464|issue:44871",
+        "issue:44617|issue:45584",
+        "issue:43906|issue:45290",
+        "issue:43065|issue:44610",
+        "issue:43827|issue:43994",
+        "issue:44792|issue:44977",
+        "issue:44534|issue:44913"
+      ],
+      "prepared_review_unit_hash": "e783555b7420e84049f2f1f2e30497e481fe1a2a32eaf8057d7cd0554db1315e",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 37010,
+        "estimated_input_tokens": 9253,
+        "estimated_eval_tokens": 18762
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is mostly a grab-bag of unrelated bug reports. None of the soft pairs look like true duplicates of the same underlying bug/change, so I would not merge any of them.",
+        "confidence": 0.81,
+        "canonical_issue_reason": "No single canonical issue stands out: the issues span unrelated subsystems and symptoms, and the soft pairs do not show a shared root cause.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43825 is the cleanest, most self-contained user-facing regression in the set, so it is the best representative issue even though it is not a duplicate hub.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:25251",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Pipeline output-shape bug vs missing PyTorch ImportError; completely different failure modes."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:43388",
+            "accept": false,
+            "reason": "Generation kwargs/update logic is unrelated to gather_for_metrics label truncation."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Custom processor/tokenizer regression is different from a pipeline error-message/docs regression."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Video-classification processor lookup issue does not match the SAM3 video attribute error."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "eos_token_id config mismatch is unrelated to Whisper empty-transcription generation failure."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure is unrelated to resume_from_checkpoint key mismatches."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "LoRA merge with extended vocab is a different code path than Gemma 4 mm_token_type_ids defaults."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "CPU offload device mismatch is unrelated to checkpoint resume key handling."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Processor/tokenizer API regression is not the same as eos_token_id configuration inconsistency."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Dummy Conv2d in Sam3PixelDecoder is unrelated to Sam3Video CUDA OOM."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "CLIPVision hidden_states regression is different from Qwen/VL video-input regression."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Adafactor+DeepSpeed OOM and post_init reinitialization are separate training/model-init bugs."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Optimizer OOM is unrelated to CustomTrainer.compute_loss signature breakage."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "TP OOM on a large Qwen model is not the same as a MobileLLM loading KeyError."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "gguf_file handling and missing model_type are both loading issues, but with different root causes and fixes."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "3d_parallel execution failure is unrelated to tensor-parallel OOM on Qwen2-57B."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Custom model initialization breakage is not the same as GPTNeoX rotary_pct reload reset."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Missing GenerationMixin cache in chunked_prefill differs from recurrent-state reset in modular_qwen3_5."
+          },
+          {
+            "left": "issue:44936",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "evaluate-after-train behavior and wrong custom module loading after save/load are different problems."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44242",
+            "accept": false,
+            "reason": "CLIPVision hidden_states regression is unrelated to load-balancing loss missing when router logits are off."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping CUDA assert is a different failure than TP OOM on Qwen2-57B."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "patchtsmixer post_init policy and non-persistent buffer junk are unrelated initialization changes."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Video-input regression in Qwen VL models is not the same as serve KV-cache continuation indexing."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44060",
+            "accept": false,
+            "reason": "Weight-shape mismatch during load is unrelated to a tied-weights warning bug."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Chunked generation inconsistency with compiled forward is unrelated to eos_token_id config mismatch."
+          },
+          {
+            "left": "issue:44617",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Sam3Video OOM does not share the same code path as Whisper empty-transcription generation."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "A reproduction of a prior issue is unrelated to chat template crashes on tool-call assistant messages."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Sam3PixelDecoder dummy layer issue is unrelated to OmDet-Turbo processor/model input-size mismatch."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Docs referencing removed pipeline() APIs are different from a SigLIP2 AutoModel/pipeline misbehavior."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Janus image-generation test failure is unrelated to Qwen3.5 flash-attention generation bugs."
+          },
+          {
+            "left": "issue:44534",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Buffer initialization corruption is not the same as GPTNeoX rotary_pct not persisting on reload."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 28,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:36246",
+        "issue:38175",
+        "issue:41628",
+        "issue:41720",
+        "issue:43065",
+        "issue:43116",
+        "issue:43295",
+        "issue:43335",
+        "issue:43404",
+        "issue:43479",
+        "issue:43540",
+        "issue:43550",
+        "issue:43644",
+        "issue:43650",
+        "issue:43761",
+        "issue:43976",
+        "issue:44077",
+        "issue:44112",
+        "issue:44186",
+        "issue:44190",
+        "issue:44220",
+        "issue:44464",
+        "issue:44496",
+        "issue:44521",
+        "issue:44734",
+        "issue:44737",
+        "issue:44743",
+        "issue:44792",
+        "issue:44855",
+        "issue:44877",
+        "issue:44913",
+        "issue:44933",
+        "issue:44938",
+        "issue:44964",
+        "issue:45127",
+        "issue:45216",
+        "issue:45245",
+        "issue:45276",
+        "issue:45357",
+        "issue:45362",
+        "issue:45375",
+        "issue:45381",
+        "issue:45542",
+        "issue:45563",
+        "issue:45584",
+        "issue:45588",
+        "issue:45600",
+        "issue:45803"
+      ],
+      "soft_pairs": [
+        "issue:41720|issue:45245",
+        "issue:44190|issue:44964",
+        "issue:44792|issue:45600",
+        "issue:44737|issue:45803",
+        "issue:44855|issue:45588",
+        "issue:44521|issue:45381",
+        "issue:45127|issue:45357",
+        "issue:43335|issue:44855",
+        "issue:44077|issue:44877",
+        "issue:44938|issue:45542",
+        "issue:43644|issue:44913",
+        "issue:43976|issue:44855",
+        "issue:43550|issue:44190",
+        "issue:43540|issue:44734",
+        "issue:43761|issue:44743",
+        "issue:44496|issue:45803",
+        "issue:45276|issue:45584",
+        "issue:44112|issue:45600",
+        "issue:43650|issue:44220",
+        "issue:43335|issue:44190",
+        "issue:43116|issue:44190",
+        "issue:43295|issue:45276",
+        "issue:45245|issue:45584",
+        "issue:43404|issue:45216",
+        "issue:43479|issue:45375",
+        "issue:36246|issue:41628",
+        "issue:44933|issue:45362",
+        "issue:44186|issue:44734",
+        "issue:44464|issue:45563",
+        "issue:38175|issue:43065"
+      ],
+      "prepared_review_unit_hash": "52e1213b2319a874ecdec3cb7f8ee5c4785682e3839c451569c7ba4822201637",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36850,
+        "estimated_input_tokens": 9213,
+        "estimated_eval_tokens": 18682
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is highly heterogeneous and does not look like a real duplicate group. The issues span unrelated models, configs, generation paths, and packaging/import regressions, so no soft pair should be merged.",
+        "confidence": 0.11,
+        "canonical_issue_reason": "There is no strong canonical issue here; if one must be chosen, issue 44964 is the broadest user-facing loading regression, but it does not actually represent the cluster well.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44964 is the most generally framed and broadly user-impacting item in the set, so it is the least-bad representative choice despite the cluster\u2019s lack of cohesion.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41720",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "Qwen3 device-mapping CUDA assert and a category-limit runtime error are different failures in different code paths."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Local dataset loading in an example script is unrelated to a model-load regression in latest transformers."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "A janus generation test failure and removed auto-mapping config references are different problems."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "XLNet CPU tensor creation and a bare-except bug in Fuyu tensor conversion are unrelated."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Python 3.13 import parsing failure and flash-attention handling of s_aux=None are unrelated."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both are multimodal, but one is assistant mask generation and the other is video position-id handling; different bugs."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "LoRA merge/tied-embedding collapse and incorrect visual encoder keys on save_pretrained are separate serialization/model-weight issues."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer config and DebertaV2 Python 3.13 import parsing are unrelated."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Both involve config behavior, but they target different models and different validations."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Python 3.14 import/load failure and a missing TensorFlow backend detection issue are not the same bug."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Non-persistent buffer initialization and GPTNeoX config reload losing rotary_pct are distinct state/serialization issues."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "Both are version-compatibility reports, but they describe different breakages and different Python versions."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure in Bamba is unrelated to a dataset-loading example script issue."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video input validation and serve KV-cache continuation indexing are separate code paths."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "CLIPVision hidden_states regression and Qwen3 recurrent-state reset are unrelated model-specific bugs."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Model auto-loading/config recognition and Fuyu batch feature exception handling are different failures."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Token embedding resize propagation and Whisper empty-transcription generation are unrelated."
+          },
+          {
+            "left": "issue:44112",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "GraniteSpeech CI stale-device test and Sam3LiteText auto-mapping references are different CI/config problems."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "The issue titled 'ADD THE DATA' is not the same underlying bug as the fbank feature extraction problem."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "SwitchTransformers config initialization and image-classification dataset loading are unrelated."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Multi-label classification empty outputs and local dataset loading are different example-script issues."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "Processor/tokenizer regression with images and token embedding resize behavior are unrelated."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Category-count overflow and Whisper empty transcription are different runtime failures."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "A tied lm_head weight bug in Mistral3 and an incorrect save_pretrained checkpoint for Qwen3.5 are distinct model persistence issues."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both are config-field problems, but they concern different models and different defaulting/strictness behavior."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "A missing model-specific image processor import is not the same as a missing top-level AutoImageProcessor export."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "An image_utils import problem and a Qwen3.5 chat crash are unrelated."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "LayoutLMv2Tokenizer padding/NER crashes and serve KV-cache tensor indexing are separate bugs."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Compiled-forward chunked generation inconsistency and a stale warning in paged generate are different generation issues."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43065",
+            "accept": false,
+            "reason": "Zero probabilities in SigLIP2 and a dummy Conv2d in Sam3PixelDecoder are unrelated model implementations."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 29,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:36246",
+        "issue:38617",
+        "issue:41950",
+        "issue:42175",
+        "issue:42907",
+        "issue:43295",
+        "issue:43299",
+        "issue:43381",
+        "issue:43388",
+        "issue:43441",
+        "issue:43526",
+        "issue:43531",
+        "issue:43577",
+        "issue:43643",
+        "issue:43673",
+        "issue:43688",
+        "issue:43854",
+        "issue:43901",
+        "issue:44016",
+        "issue:44075",
+        "issue:44112",
+        "issue:44220",
+        "issue:44242",
+        "issue:44246",
+        "issue:44279",
+        "issue:44336",
+        "issue:44368",
+        "issue:44466",
+        "issue:44496",
+        "issue:44561",
+        "issue:44683",
+        "issue:44734",
+        "issue:44913",
+        "issue:44964",
+        "issue:45003",
+        "issue:45042",
+        "issue:45200",
+        "issue:45216",
+        "issue:45335",
+        "issue:45357",
+        "issue:45406",
+        "issue:45446",
+        "issue:45563",
+        "issue:45584",
+        "issue:45593",
+        "issue:45600",
+        "issue:45632",
+        "issue:45759"
+      ],
+      "soft_pairs": [
+        "issue:44683|issue:45446",
+        "issue:44466|issue:44913",
+        "issue:44561|issue:45632",
+        "issue:44734|issue:45632",
+        "issue:44246|issue:45003",
+        "issue:36246|issue:38617",
+        "issue:43643|issue:44496",
+        "issue:45406|issue:45632",
+        "issue:45563|issue:45584",
+        "issue:41950|issue:44016",
+        "issue:44964|issue:45600",
+        "issue:43388|issue:43688",
+        "issue:42907|issue:43441",
+        "issue:44112|issue:44964",
+        "issue:43295|issue:45335",
+        "issue:43854|issue:44964",
+        "issue:43673|issue:44336",
+        "issue:44016|issue:44220",
+        "issue:44496|issue:45042",
+        "issue:43299|issue:45216",
+        "issue:43531|issue:45600",
+        "issue:42175|issue:44279",
+        "issue:43526|issue:43901",
+        "issue:45335|issue:45584",
+        "issue:44075|issue:44368",
+        "issue:43381|issue:45593",
+        "issue:43299|issue:45357",
+        "issue:44242|issue:45200",
+        "issue:44496|issue:45759",
+        "issue:43901|issue:44496",
+        "issue:43577|issue:45216"
+      ],
+      "prepared_review_unit_hash": "04bf0559fd79a14d628f93aa6b7835898707ca32199351a04a811193b7997bd8",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 37212,
+        "estimated_input_tokens": 9303,
+        "estimated_eval_tokens": 18862
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly a set of loosely related, but not duplicate, Transformers issues spanning model loading, generation, config, and tooling. I did not find any soft pair that clearly describes the same underlying bug/change, so all soft edges are rejected.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "No exact duplicate stands out; issue 44964 is the most representative user-facing regression because it reports a concrete model-loading failure on a named checkpoint with broad impact.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44964 is the strongest anchor for the cluster: it is a clear, reproducible loading breakage affecting a specific multimodal model and is more representative than the narrower or more incidental reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44683",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Both mention flex_attention, but one is a torch-compatibility failure and the other is a bad version check for a different import path; not the same bug."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "One is tied-weights serialization/device-dependent save behavior; the other is a config field not persisting on reload. Different code paths and symptoms."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:45632",
+            "accept": false,
+            "reason": "Both involve trust_remote_code, but one is a missing internal API after v5 removal and the other is a cache-path collision for local models; unrelated failures."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:45632",
+            "accept": false,
+            "reason": "KV-cache continuation indexing in serve is unrelated to trust_remote_code cache-path collisions."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Import slowness and unsafe sys.modules access are not the same underlying defect."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Both are ImportErrors, but they are for different missing symbols and different modules."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Both concern config/model loading, but one is missing fields with trust_remote_code and the other is an unrecognized model lacking model_type; not the same bug."
+          },
+          {
+            "left": "issue:45406",
+            "right": "issue:45632",
+            "accept": false,
+            "reason": "Gemma4Processor _tokenizer crash in serve is unrelated to trust_remote_code cache directory collisions."
+          },
+          {
+            "left": "issue:45563",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "A stale warning in paged generate() and a Whisper failure after align_special_tokens are different generation issues."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:44016",
+            "accept": false,
+            "reason": "Video-classification pipeline processor lookup and a notebook syntax error are unrelated."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Phi-4 multimodal loading failure and stale auto_mappings references to removed configs are different model/config maintenance issues."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43688",
+            "accept": false,
+            "reason": "Metric gathering dropping tuple labels and auxiliary-loss normalization are unrelated training/eval bugs."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Saving dequantized Ministrals/Devstrals and FlashAttention loading failures for Ministral-3 are different behaviors in different code paths."
+          },
+          {
+            "left": "issue:44112",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "A stale CI test failure in GraniteSpeech is not the same as a real model-loading regression."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "Processor.tokenizer passthrough regression and resize_token_embeddings not updating decoder.embed_tokens are distinct model API bugs."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different failing models and different failure modes; both are loading issues but not duplicates."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "Chunked-prefill cache behavior and ANSI output formatting have no substantive overlap."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Notebook syntax error vs. feature-extraction bug; unrelated."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Unrecognized model config and PIL image processor torchvision dependency are unrelated."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "One is Qwen3VL model loading in v5.0.0.dev0; the other is incorrect save_pretrained output for Qwen3.5 in v5.4.0. Same family, different bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior and Sam3LiteText auto_mappings cleanup are unrelated."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Missing TensorFlow in a torch extra install and a generic dependency issue are not the same concrete problem."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction bug and documentation mismatch for return_all_scores are unrelated."
+          },
+          {
+            "left": "issue:45335",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Token embedding resize behavior and Whisper empty-transcription generation failure do not share a code path."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Unused SGD args and a tied-embeddings warning are different optimizer/model-config issues."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode and D-FINE auxiliary-loss handling are unrelated."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Load-balancing loss omission and Gemma4 mm_token_type_ids defaults are different model-training concerns."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Unrecognized model type in config and AutoModelForCausalLM not unwrapping text_config are distinct config-loading failures."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "A docs/behavior mismatch for pipelines is unrelated to model recognition/config parsing."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "BLIP2 dtype propagation and Qwen3.5 save_pretrained serialization are unrelated bugs."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 30,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:32090",
+        "issue:35141",
+        "issue:38175",
+        "issue:39692",
+        "issue:41950",
+        "issue:42994",
+        "issue:43065",
+        "issue:43295",
+        "issue:43299",
+        "issue:43317",
+        "issue:43388",
+        "issue:43441",
+        "issue:43450",
+        "issue:43479",
+        "issue:43525",
+        "issue:43650",
+        "issue:43824",
+        "issue:43827",
+        "issue:43854",
+        "issue:43901",
+        "issue:43976",
+        "issue:44016",
+        "issue:44062",
+        "issue:44261",
+        "issue:44263",
+        "issue:44279",
+        "issue:44297",
+        "issue:44351",
+        "issue:44462",
+        "issue:44466",
+        "issue:44521",
+        "issue:44617",
+        "issue:44683",
+        "issue:44779",
+        "issue:44898",
+        "issue:44933",
+        "issue:44987",
+        "issue:45020",
+        "issue:45081",
+        "issue:45276",
+        "issue:45278",
+        "issue:45290",
+        "issue:45325",
+        "issue:45405",
+        "issue:45542",
+        "issue:45600",
+        "issue:45712"
+      ],
+      "soft_pairs": [
+        "issue:41950|issue:45020",
+        "issue:44016|issue:44062",
+        "issue:44351|issue:45278",
+        "issue:43976|issue:45405",
+        "issue:43388|issue:43827",
+        "issue:43479|issue:43525",
+        "issue:43824|issue:44933",
+        "issue:44683|issue:45542",
+        "issue:44279|issue:45405",
+        "issue:43299|issue:43854",
+        "issue:38175|issue:44617",
+        "issue:32090|issue:35141",
+        "issue:39692|issue:43901",
+        "issue:43295|issue:44779",
+        "issue:41950|issue:43650",
+        "issue:44462|issue:45081",
+        "issue:44297|issue:44462",
+        "issue:44521|issue:45325",
+        "issue:43976|issue:44279",
+        "issue:43976|issue:44987",
+        "issue:45600|issue:45712",
+        "issue:44462|issue:44779",
+        "issue:43450|issue:44279",
+        "issue:42994|issue:43317",
+        "issue:43065|issue:44898",
+        "issue:44466|issue:45276",
+        "issue:44261|issue:44263",
+        "issue:43650|issue:45290",
+        "issue:43441|issue:44683"
+      ],
+      "prepared_review_unit_hash": "8b0d249a1a2c58ff7482dadab0b5ad23c2d8dea12a36e951b176c25a7357dc93",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 35832,
+        "estimated_input_tokens": 8958,
+        "estimated_eval_tokens": 18172
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "All soft edges look like broad topical similarity rather than the same underlying bug/change. The set is heterogeneous across unrelated models, tokenizer, config, docs, quantization, and pipeline issues, so I would not merge any pair as duplicates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "No canonical issue stands out: the items span many distinct bug reports and documentation regressions with different code paths and symptoms.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No single best issue candidate for a duplicate cluster; the apparent matches are only broad subsystem overlap, not the same concrete failure.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41950",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Video-classification pipeline processor lookup and remote_code loading failures are different bugs in different paths."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44062",
+            "accept": false,
+            "reason": "Notebook syntax error and AddedToken keyword-argument TypeError are unrelated failure modes."
+          },
+          {
+            "left": "issue:44351",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Missing HybridCache import and broad post-upgrade import errors are not the same concrete issue."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Python version compatibility and an unreleased PEFT version pin are separate dependency problems."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "Metric-gathering label truncation and docs still mentioning pipeline() are unrelated."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Default multimodal config initialization and missing pad_token_id are different config bugs."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "These are both import errors, but for different missing symbols and different root causes."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "flex_attention failures on torch>=2.9 and TensorBoard/TensorFlow backend issues are unrelated."
+          },
+          {
+            "left": "issue:44279",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Generic dependency trouble does not match the specific PEFT version bump regression."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Qwen3VL MoE loading and GLM-4.7 test loading failures are model-specific and not the same bug."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Unexpected zero probabilities in SigLIP2 and SAM3 video OOM are different symptoms and code paths."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Trainer NoneType broadcast error and embedding reinitialization after resize are unrelated training/model-init bugs."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Both are docs-related, but one is a SigLIP2 example/model mismatch and the other is a text-classification API docs mismatch."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Custom-model processor/tokenizer regression and Deepseek tokenizer output corruption are distinct tokenizer issues."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:43650",
+            "accept": false,
+            "reason": "A pipeline processor bug and a low-signal placeholder issue are clearly unrelated."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "AutoTokenizer ignoring tokenizer.json and Mistral regex patch crashes are different loading/patching failures."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:44462",
+            "accept": false,
+            "reason": "Tokenizer class serialization mismatch and repository tokenizer.json selection are related only broadly, not the same bug."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "All-zero assistant masks and Qwen2.5-VL rope-index scaling are different multimodal behaviors."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Python support regression and a generic dependency issue are not the same concrete problem."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Python-version compatibility and loading physical-intelligence/fast are unrelated regressions."
+          },
+          {
+            "left": "issue:45600",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Removed Sam3LiteText mappings and leftover dummy classes are separate repo-maintenance issues."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "AutoTokenizer preferring tokenizer.json and Deepseek tokenizer incorrect results are different tokenizer bugs."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Video processor batch-shape bug and generic dependency issue have no shared code path."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43317",
+            "accept": false,
+            "reason": "Quantized model saving and dequantized model loading/offload are different stages of the model lifecycle."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Dummy Conv2d in Sam3PixelDecoder and Perceiver interpolation failures are unrelated vision-model issues."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "lm_head serialization depending on device and Gemma4 resize_token_embeddings not updating outputs are not the same bug."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "MLA layernorm epsilon precision discussion and GlmMoeDsaIndexer torch.split return values are unrelated."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "A placeholder issue and a chat-template tool-call crash are clearly unrelated."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Both mention attention backends, but one is a model-specific FlashAttention failure and the other is a compiled flex_attention torch-version regression."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 31,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:16998",
+        "issue:22355",
+        "issue:41628",
+        "issue:42491",
+        "issue:43122",
+        "issue:43262",
+        "issue:43299",
+        "issue:43352",
+        "issue:43450",
+        "issue:43475",
+        "issue:43525",
+        "issue:43650",
+        "issue:43653",
+        "issue:43673",
+        "issue:43723",
+        "issue:43761",
+        "issue:43784",
+        "issue:43881",
+        "issue:43901",
+        "issue:43976",
+        "issue:43994",
+        "issue:44016",
+        "issue:44162",
+        "issue:44188",
+        "issue:44242",
+        "issue:44246",
+        "issue:44291",
+        "issue:44295",
+        "issue:44297",
+        "issue:44336",
+        "issue:44361",
+        "issue:44373",
+        "issue:44496",
+        "issue:44521",
+        "issue:44589",
+        "issue:44617",
+        "issue:44623",
+        "issue:44671",
+        "issue:44749",
+        "issue:44779",
+        "issue:44987",
+        "issue:45250",
+        "issue:45362",
+        "issue:45397",
+        "issue:45542",
+        "issue:45593",
+        "issue:45759"
+      ],
+      "soft_pairs": [
+        "issue:44623|issue:45397",
+        "issue:43723|issue:44162",
+        "issue:44623|issue:44749",
+        "issue:43653|issue:45593",
+        "issue:43784|issue:45542",
+        "issue:44671|issue:44779",
+        "issue:43122|issue:43650",
+        "issue:43901|issue:43994",
+        "issue:44361|issue:45593",
+        "issue:44987|issue:45542",
+        "issue:44016|issue:44291",
+        "issue:43262|issue:44521",
+        "issue:44297|issue:45397",
+        "issue:43475|issue:44016",
+        "issue:44188|issue:44336",
+        "issue:43673|issue:43761",
+        "issue:43525|issue:45759",
+        "issue:43525|issue:43881",
+        "issue:43450|issue:44617",
+        "issue:44589|issue:45362",
+        "issue:42491|issue:43299",
+        "issue:43901|issue:44246",
+        "issue:41628|issue:43976",
+        "issue:44336|issue:44373",
+        "issue:16998|issue:22355",
+        "issue:44242|issue:45593",
+        "issue:43352|issue:45250",
+        "issue:44295|issue:45397",
+        "issue:44246|issue:44496"
+      ],
+      "prepared_review_unit_hash": "26310a68113ff8f3725a4ea32eb3c40e6be11162669adc894ea15b41f864cae8",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 35505,
+        "estimated_input_tokens": 8877,
+        "estimated_eval_tokens": 18010
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a mixed bag of largely unrelated Transformers issues, with no strong duplicate pairs. The closest theme is v5 loading/tokenizer regressions, but the soft edges still don\u2019t line up to the same concrete bug.",
+        "confidence": 0.92,
+        "canonical_issue_reason": "issue:43723 is the best representative because it is a broad v5 tokenizer-loading regression, which is more central and general than the other highly model-specific or docs-only issues.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43723 is the strongest cluster anchor for the same reason: it captures the widest, most generic failure mode in the set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44623",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "Processor save_pretrained file omissions and gemma-4 zero3 from_pretrained failures are different code paths and different models."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "Tokenizer loading in v5 and ESM2 breakage are unrelated subsystems with no shared concrete bug."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "A missing processor file bug and a data-filtering slowdown are unrelated."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token decoding and D-FINE auxiliary-loss behavior are unrelated model bugs."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Both are import-time failures, but one is a missing nn symbol and the other is a missing tf backend; different causes and paths."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "CamemBERT MLM prediction errors and Deepseek tokenizer regressions are different model/tokenizer issues."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:43650",
+            "accept": false,
+            "reason": "Tokenizer output changes across versions and an unrelated placeholder issue are not the same bug."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Docs mentioning return_all_scores and a bad SigLIP2 inference result are unrelated."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "MLukeTokenizer AttributeError and D-FINE loss omission are unrelated."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Model loading failure and TensorFlow backend import error are different problems."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "A notebook syntax error and an init_empty_weights TypeError are unrelated."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Both involve chat templates, but one is audio sampling-rate defaulting and the other is assistant-mask generation; not the same bug."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "Tokenizer metadata saved incorrectly and gemma-4 loading failure are unrelated."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:44016",
+            "accept": false,
+            "reason": "SAM3 video encoder attribute failure and a notebook syntax error are unrelated."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "Torch.compile attention-kernel divergence and ANSI code output formatting are unrelated."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:43761",
+            "accept": false,
+            "reason": "Generation cache missing and CLIP hidden_states regression are different code paths and models."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Missing pad_token_id on Llama4Config and text_config unwrapping for Qwen multimodal configs are different config-loading bugs."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "Llama4Config pad_token_id and glm-4v-9b loading failures are unrelated model-specific issues."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "A batched video shape bug and CUDA OOM are not the same underlying issue."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "Float8 storage lookup failure and Qwen3.5 chat crash are unrelated."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43299",
+            "accept": false,
+            "reason": "Both mention Qwen3 MoE compatibility, but one is LoRA cross-version incompatibility and the other is a model-loading regression; not the same concrete bug."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "A documentation mismatch and slow import are unrelated."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "AutoImageProcessor import failure and Python version incompatibility are different issues."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "ANSI codes in loading_report and a wrong docstring for position_ids are unrelated."
+          },
+          {
+            "left": "issue:16998",
+            "right": "issue:22355",
+            "accept": false,
+            "reason": "A model_max_length question and a missing transformers.onnx module are unrelated."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Both are missing-loss complaints, but they affect different models and different flags, so they are not the same bug."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:45250",
+            "accept": false,
+            "reason": "A specific model not supporting Flash Attention 2 and a generic Flash Attention topic are not duplicate bugs."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "position_ids buffer access errors and gemma-4 loading failures are unrelated."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Slow import and unrecognized model config 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": 2,
+      "component_count": 16,
+      "review_unit_index": 32,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:36246",
+        "issue:41950",
+        "issue:42757",
+        "issue:43065",
+        "issue:43232",
+        "issue:43257",
+        "issue:43262",
+        "issue:43408",
+        "issue:43441",
+        "issue:43526",
+        "issue:43540",
+        "issue:43596",
+        "issue:43673",
+        "issue:43824",
+        "issue:43901",
+        "issue:43976",
+        "issue:44162",
+        "issue:44220",
+        "issue:44242",
+        "issue:44246",
+        "issue:44295",
+        "issue:44336",
+        "issue:44351",
+        "issue:44451",
+        "issue:44485",
+        "issue:44496",
+        "issue:44561",
+        "issue:44749",
+        "issue:44779",
+        "issue:44857",
+        "issue:44871",
+        "issue:44877",
+        "issue:44964",
+        "issue:44987",
+        "issue:44995",
+        "issue:45003",
+        "issue:45137",
+        "issue:45200",
+        "issue:45292",
+        "issue:45362",
+        "issue:45375",
+        "issue:45459",
+        "issue:45561",
+        "issue:45563",
+        "issue:45584",
+        "issue:45701",
+        "issue:45759"
+      ],
+      "soft_pairs": [
+        "issue:41950|issue:44220",
+        "issue:36246|issue:42757",
+        "issue:44336|issue:44485",
+        "issue:36246|issue:44351",
+        "issue:43673|issue:44561",
+        "issue:43824|issue:43976",
+        "issue:43976|issue:44964",
+        "issue:44242|issue:44857",
+        "issue:44246|issue:44871",
+        "issue:44162|issue:45200",
+        "issue:45362|issue:45375",
+        "issue:44295|issue:44749",
+        "issue:43441|issue:43540",
+        "issue:44779|issue:45292",
+        "issue:44877|issue:45375",
+        "issue:45292|issue:45584",
+        "issue:45292|issue:45701",
+        "issue:44987|issue:45459",
+        "issue:43901|issue:45003",
+        "issue:36246|issue:43596",
+        "issue:41950|issue:44451",
+        "issue:45561|issue:45563",
+        "issue:43232|issue:44336",
+        "issue:45375|issue:45759",
+        "issue:43526|issue:45003",
+        "issue:43065|issue:43408",
+        "issue:43257|issue:45137",
+        "issue:44995|issue:45200",
+        "issue:44496|issue:45003",
+        "issue:43262|issue:44246"
+      ],
+      "prepared_review_unit_hash": "bdc3258bef6598236cfe97fb9576a559bae27a57c668d50f5fea982738e81f85",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 30,
+        "serialized_chars": 36008,
+        "estimated_input_tokens": 9002,
+        "estimated_eval_tokens": 18260
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a mixed cluster of distinct Transformers bugs (imports, configs, generation, model-specific regressions). None of the soft pairs look like the same underlying bug/change, so I would not merge any of them; 43824 is the strongest representative issue by signal and impact.",
+        "confidence": 0.38,
+        "canonical_issue_reason": "issue:43824 has the strongest discussion/inbound-reference signal and is a clear, user-facing regression in model import/export behavior, making it the best anchor among this mixed set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43824 is the clearest, highest-signal issue in the cluster and most suitable as the representative duplicate target.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41950",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Different subsystems: video-classification pipeline processor selection vs audio feature extraction."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are ImportError reports, but for different missing symbols and unrelated code paths."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "ANSI color handling in loading_report is unrelated to GLM-5 RoPE implementation."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Different missing imports: Qwen2.5-VL image processor vs HybridCache."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "Generation cache/chunked_prefill issue is unrelated to trust_remote_code breakage from a removed utility."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "Qwen2.5-VL import failure is unrelated to the Python version compatibility regression."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Python version support problem vs model-loading failure for a specific checkpoint."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Different bugs in different model losses; no shared failing code path."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Import slowdown is unrelated to eos_token_id configuration mismatch."
+          },
+          {
+            "left": "issue:44162",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "ESM2 breakage and Gemma 4 mm_token_type_ids defaulting are separate model-specific issues."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Qwen3.5 chat crash and missing config field in Qwen3_5MoeVisionConfig are not the same defect."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "Position_ids buffer error is unrelated to the reported slow filtering after upgrade."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Ministral FlashAttention failure and Qwen3OmniMoe video-input ValueError are different code paths."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:45292",
+            "accept": false,
+            "reason": "Tokenizer regression and resize_token_embeddings/output_embeddings mismatch are unrelated."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Strict config loading failure for granite_speech is not the same as a silently dropped Qwen3_5MoeVisionConfig field."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Embedding resize bug is unrelated to Whisper empty-transcription generation failure."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Output embedding resize bug is not the same as version-dependent tokenization changes."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "Model load failure for physical-intelligence/fast is unrelated to protobuf error masking."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Docs/behavior mismatch in text classification is unrelated to unsafe sys.modules access."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:43596",
+            "accept": false,
+            "reason": "Qwen2.5-VL import error and DeepSpeed ZeRO3 BertModel IndexError are unrelated."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Video-classification processor lookup and ScandiBERT loading failure do not share the same code path."
+          },
+          {
+            "left": "issue:45561",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "xdist file-race bug is unrelated to the stale num_return_sequences warning."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "Generation kwargs update bug is unrelated to terminal ANSI output handling."
+          },
+          {
+            "left": "issue:45375",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Missing config field in Qwen3_5MoeVisionConfig and text_config unwrapping in AutoModelForCausalLM are different composite-model loading bugs."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction bug is unrelated to sys.modules access."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:43408",
+            "accept": false,
+            "reason": "Dummy Conv2d in Sam3PixelDecoder and sam3_video vs sam3_tracker warning are distinct SAM3 issues."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Qwen3 MoE weight conversion under accelerate/deepspeed is unrelated to the ZeRO3 deque pop error."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "GlmMoeDsa stale indexer cache and Gemma 4 token-type defaults are unrelated."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Unrecognized model loading and unsafe sys.modules access are different failures."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "Audio chat template sampling-rate default bug is unrelated to import-time slowdown."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 33,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:28282",
+        "issue:30990",
+        "issue:41628",
+        "issue:42371",
+        "issue:42886",
+        "issue:42994",
+        "issue:43232",
+        "issue:43452",
+        "issue:43475",
+        "issue:43482",
+        "issue:43540",
+        "issue:43643",
+        "issue:43646",
+        "issue:43653",
+        "issue:43673",
+        "issue:43827",
+        "issue:43931",
+        "issue:44038",
+        "issue:44060",
+        "issue:44077",
+        "issue:44162",
+        "issue:44297",
+        "issue:44361",
+        "issue:44368",
+        "issue:44466",
+        "issue:44623",
+        "issue:44734",
+        "issue:44749",
+        "issue:44869",
+        "issue:44933",
+        "issue:44977",
+        "issue:44995",
+        "issue:45003",
+        "issue:45092",
+        "issue:45137",
+        "issue:45229",
+        "issue:45230",
+        "issue:45310",
+        "issue:45335",
+        "issue:45356",
+        "issue:45381",
+        "issue:45397",
+        "issue:45459",
+        "issue:45561",
+        "issue:45563",
+        "issue:45588",
+        "issue:45600",
+        "issue:45656"
+      ],
+      "soft_pairs": [
+        "issue:42886|issue:45229",
+        "issue:43475|issue:43540",
+        "issue:44466|issue:45335",
+        "issue:44995|issue:45588",
+        "issue:43653|issue:44162",
+        "issue:45397|issue:45656",
+        "issue:44368|issue:44977",
+        "issue:44368|issue:45563",
+        "issue:44060|issue:44077",
+        "issue:45137|issue:45600",
+        "issue:43673|issue:44734",
+        "issue:45003|issue:45561",
+        "issue:43931|issue:44368",
+        "issue:44933|issue:45459",
+        "issue:44869|issue:45356",
+        "issue:44297|issue:44623",
+        "issue:43646|issue:44162",
+        "issue:44361|issue:44749",
+        "issue:43232|issue:43673",
+        "issue:43452|issue:45003",
+        "issue:44623|issue:45656",
+        "issue:43673|issue:43827",
+        "issue:28282|issue:30990",
+        "issue:42994|issue:43482",
+        "issue:44749|issue:45656",
+        "issue:44749|issue:45310",
+        "issue:43475|issue:43643",
+        "issue:45230|issue:45397",
+        "issue:44038|issue:44623",
+        "issue:42371|issue:45092",
+        "issue:43475|issue:45381",
+        "issue:41628|issue:44933"
+      ],
+      "prepared_review_unit_hash": "6cd56be695765442c277eab541594aee39ea136de53a3acedb19218a4d191a22",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37130,
+        "estimated_input_tokens": 9283,
+        "estimated_eval_tokens": 18822
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a loose grab-bag of Transformers regressions and incompatibilities, mostly unrelated concrete bugs. The broadest and most representative issue is the v5 custom model initialization regression.",
+        "confidence": 0.62,
+        "canonical_issue_reason": "issue:43646 is the most general, high-level regression in the set and best matches the common theme of Transformers v5 breaking model initialization paths.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43646 is the best cluster representative because it is broad, user-facing, and covers a core v5 breakage rather than a narrow model-specific symptom.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42886",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Offline tokenizer cache loading and Gemma4 multi-GPU CUDA OOM are unrelated failure modes."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Different models and different code paths: SAM3 vision encoder attribute error vs Qwen3OmniMoe video-input validation."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "Both mention embeddings/tied weights, but one is serialization inconsistency and the other is resize_token_embeddings not updating decoder embeddings; not the same bug."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "A stale indexer cache crash in GlmMoeDsa is unrelated to flash_attention.py crashing on sink-less models."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration and ESM2 breakage are different tokenizer/model bugs."
+          },
+          {
+            "left": "issue:45397",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Gemma-4 ZeRO3 loading and a double optimizer step under deepspeed are different DeepSpeed problems."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "A tied-embeddings warning in Qwen3.5 and a flash-attention generation bug are not the same underlying issue."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Both involve warnings, but one is a tied-weights config warning and the other is a stale num_return_sequences warning in paged generate."
+          },
+          {
+            "left": "issue:44060",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "Qwen3-Next tied-weights warning vs patchtsmixer post_init policy are unrelated."
+          },
+          {
+            "left": "issue:45137",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO3 deque underflow is unrelated to removed Sam3LiteText config references in auto_mappings.py."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Both are generation/cache related, but chunked_prefill cache missing and /v1/responses KV indexing are distinct concrete bugs."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Unsafe sys.modules access is unrelated to the xdist captured_info.txt race in testing utils."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Qwen3-VL weight-shape loading failure is not the same as a tied-embeddings warning in Qwen3.5."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "A missing image_utils import and tokenizer protobuf error handling are different import/tokenizer issues."
+          },
+          {
+            "left": "issue:44869",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Whisper timestamp decoding crash and Kimi-K2.5 tokenizer codec regression are separate tokenizer bugs."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Tokenizer save_pretrained writing the wrong tokenizer_class is distinct from processor.save_pretrained missing files."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "Both are broad breakage reports, but ESM2 likely has its own model-specific failure mode rather than the same custom-init regression."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "MLukeTokenizer task failures and Chinese preprocessing slowdown after upgrade are unrelated."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "Both touch generation internals, but sync_gpus model-kwarg updating and chunked_prefill cache handling are different code paths."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "gguf_file loading regressions and sys.modules access safety are unrelated loader issues."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Processor serialization and double optimizer stepping are unrelated subsystems."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "A runtime cache bug and a docs-only pipeline reference issue are not the same change or bug."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:30990",
+            "accept": false,
+            "reason": "PyTorch missing ImportError and Sentence Transformers hanging during load are different environment/runtime problems."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Quantized model saving failure and GGUF loading failure in v5 are separate model I/O bugs."
+          },
+          {
+            "left": "issue:44749",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Data filtering slowdown after upgrade is unrelated to deepspeed optimizer stepping twice."
+          },
+          {
+            "left": "issue:44749",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Preprocessing slowdown and Qwen3.5 MoE from_pretrained error are different regressions."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "SAM3 video attribute error and trust_remote_code AutoConfig field loss are unrelated."
+          },
+          {
+            "left": "issue:45230",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "A generic bug report is not the same as the Gemma-4 ZeRO3 loading issue."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe v5 bug and processor.save_pretrained missing files are unrelated."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "TF32 API-settings documentation and InternVL2 meta-init incompatibility are unrelated."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Different video-related bugs in different models; no shared underlying code-path is evident."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Cannot import AutoImageProcessor and missing image_utils import are separate import-path breakages, not one bug."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 34,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:30333",
+        "issue:30990",
+        "issue:33290",
+        "issue:35141",
+        "issue:36010",
+        "issue:41628",
+        "issue:42175",
+        "issue:42222",
+        "issue:42673",
+        "issue:42831",
+        "issue:42898",
+        "issue:42994",
+        "issue:43232",
+        "issue:43278",
+        "issue:43377",
+        "issue:43475",
+        "issue:43525",
+        "issue:43704",
+        "issue:43723",
+        "issue:43784",
+        "issue:43824",
+        "issue:43901",
+        "issue:44038",
+        "issue:44188",
+        "issue:44222",
+        "issue:44242",
+        "issue:44351",
+        "issue:44373",
+        "issue:44496",
+        "issue:44568",
+        "issue:44623",
+        "issue:44661",
+        "issue:44734",
+        "issue:44792",
+        "issue:44964",
+        "issue:45081",
+        "issue:45103",
+        "issue:45137",
+        "issue:45200",
+        "issue:45310",
+        "issue:45356",
+        "issue:45375",
+        "issue:45478",
+        "issue:45561",
+        "issue:45600",
+        "issue:45712",
+        "issue:45812"
+      ],
+      "soft_pairs": [
+        "issue:45081|issue:45103",
+        "issue:42994|issue:43525",
+        "issue:42673|issue:43704",
+        "issue:43475|issue:44661",
+        "issue:45375|issue:45478",
+        "issue:45310|issue:45375",
+        "issue:42831|issue:44222",
+        "issue:42222|issue:45200",
+        "issue:41628|issue:44351",
+        "issue:44222|issue:45310",
+        "issue:42994|issue:43824",
+        "issue:43377|issue:44242",
+        "issue:42175|issue:45712",
+        "issue:36010|issue:41628",
+        "issue:44038|issue:45310",
+        "issue:44496|issue:45561",
+        "issue:44964|issue:45137",
+        "issue:45561|issue:45600",
+        "issue:30990|issue:35141",
+        "issue:43901|issue:45561",
+        "issue:42898|issue:44568",
+        "issue:43232|issue:44792",
+        "issue:43232|issue:44188",
+        "issue:42898|issue:45812",
+        "issue:44188|issue:44373",
+        "issue:43784|issue:43824",
+        "issue:43278|issue:45137",
+        "issue:44623|issue:45310",
+        "issue:30333|issue:33290",
+        "issue:44734|issue:45103",
+        "issue:43723|issue:44568",
+        "issue:42898|issue:45356"
+      ],
+      "prepared_review_unit_hash": "6f3c1c1ae2dd4f60db6c41413d5bbc322ba4df79d5a615a44baf5e34146c0b04",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36753,
+        "estimated_input_tokens": 9189,
+        "estimated_eval_tokens": 18634
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a heterogeneous issue cluster with two clear duplicate subclusters: Qwen3ForCausalLM VRAM leak and Qwen3.5 MoE from_pretrained errors. The rest are similar-looking but distinct regressions/import errors/docs issues, so they should not be merged.",
+        "confidence": 0.7,
+        "canonical_issue_reason": "issue:42673 is the strongest canonical candidate: it is the earlier, more active copy of the exact same VRAM-leak report as issue:43704, with a clearer duplicate signal than the rest of the cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:42673 is the best single issue in this set because it is an exact duplicate pair with a later clone (issue:43704) and has the strongest standalone discussion/reference activity. It is still only a partial representative of the broader heterogeneous cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45081",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Different bugs: Mistral tokenizer regex crash vs auto-docstring annotation handling crash."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Different failures: quantized save_pretrained vs missing pad_token_id on Llama4Config."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:43704",
+            "accept": true,
+            "reason": "Same exact bug/title: Qwen3ForCausalLM VRAM leak in multi-threaded dataloader use."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "Different code paths: missing Sam3VisionEncoderOutput attribute vs add-new-model-like tokenizer mapping failure."
+          },
+          {
+            "left": "issue:45375",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Same model family, but the reported problems differ: missing config field vs from_pretrained error; not enough to treat as one bug."
+          },
+          {
+            "left": "issue:45310",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Different underlying issues: from_pretrained regression vs config-field omission/silent drop."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Different bugs: accuracy drift with FineGrainedFP8 vs save_pretrained failure for FP8 MoE."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Different models and symptoms; not the same underlying defect."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and likely different breakages."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Saving FP8 MoE vs loading Qwen3.5 MoE are different concrete code paths."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Unrelated issues: quantized saving failure vs missing Qwen2_5_VL import."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44242",
+            "accept": false,
+            "reason": "Different model behavior bugs: batched padding-mask mismatch vs router loss omission."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Different areas: backend dependency packaging vs leftover dummy class cleanup."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "Both are import regressions, but for different APIs and failure modes."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Related Qwen model/version regressions, but not the same concrete bug."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Unrecognized model/config issue vs xdist file race; unrelated."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Different subsystems entirely: multimodal model loading vs DeepSpeed ZeRO3 deque error."
+          },
+          {
+            "left": "issue:45561",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Different failures: xdist capture-file race vs stale auto_mappings references."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Completely different bugs: loading hang vs token-embedding reinit on resize."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Docs mismatch vs test utility race; not the same issue."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both are tokenizer-related regressions, but they affect different behaviors and models."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44792",
+            "accept": false,
+            "reason": "Generation kwargs bug vs a Janus model test failure; no concrete match."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44188",
+            "accept": false,
+            "reason": "Different generation/attention code paths and symptoms."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Different tokenizer regressions affecting different models and behaviors."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "Attention-kernel divergence vs docstring typo; unrelated."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but for different packages/symbols and likely unrelated."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Different issues: dtype change during eval vs DeepSpeed deque underflow."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Processor save_pretrained missing files vs Qwen3.5 MoE from_pretrained error."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "MLFlow reporting failure is unrelated to an Adafactor OOM in DeepSpeed."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "KV cache continuation crash vs auto-docstring annotation crash; unrelated."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both involve tokenizer changes in v5, but they are different concrete bugs."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Different tokenizer regressions; not the same underlying defect."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded overall and appropriately conservative. The canonical choice of issue:42673 is supported by the exact title match and higher discussion activity relative to issue:43704. The soft-edge verdicts consistently stay on the cautious side, using the packet\u2019s titles to distinguish likely different bugs, and the only accepted pair is the clearly identical VRAM-leak report. The summary also matches the packet\u2019s mixed cluster structure without overclaiming mergeability."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 35,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:28282",
+        "issue:30333",
+        "issue:34689",
+        "issue:35141",
+        "issue:36331",
+        "issue:41628",
+        "issue:42175",
+        "issue:42757",
+        "issue:42831",
+        "issue:42898",
+        "issue:43479",
+        "issue:43644",
+        "issue:43723",
+        "issue:43873",
+        "issue:43874",
+        "issue:43976",
+        "issue:44060",
+        "issue:44079",
+        "issue:44188",
+        "issue:44206",
+        "issue:44222",
+        "issue:44261",
+        "issue:44263",
+        "issue:44295",
+        "issue:44297",
+        "issue:44360",
+        "issue:44373",
+        "issue:44448",
+        "issue:44483",
+        "issue:44484",
+        "issue:44485",
+        "issue:44514",
+        "issue:44623",
+        "issue:44779",
+        "issue:44869",
+        "issue:44871",
+        "issue:44933",
+        "issue:44945",
+        "issue:44964",
+        "issue:44977",
+        "issue:45092",
+        "issue:45103",
+        "issue:45125",
+        "issue:45310",
+        "issue:45405",
+        "issue:45406",
+        "issue:45656",
+        "issue:45712"
+      ],
+      "soft_pairs": [
+        "issue:44060|issue:45125",
+        "issue:43874|issue:45092",
+        "issue:45405|issue:45712",
+        "issue:42898|issue:44779",
+        "issue:42898|issue:44448",
+        "issue:43479|issue:44079",
+        "issue:44188|issue:44964",
+        "issue:44295|issue:45656",
+        "issue:44933|issue:45712",
+        "issue:42175|issue:44933",
+        "issue:44222|issue:44297",
+        "issue:43976|issue:44933",
+        "issue:44297|issue:45656",
+        "issue:43479|issue:44869",
+        "issue:44206|issue:45103",
+        "issue:44222|issue:44623",
+        "issue:42831|issue:44623",
+        "issue:30333|issue:36331",
+        "issue:41628|issue:42757",
+        "issue:30333|issue:34689",
+        "issue:44297|issue:45310",
+        "issue:43723|issue:44977",
+        "issue:28282|issue:30333",
+        "issue:30333|issue:35141",
+        "issue:44222|issue:44484",
+        "issue:44483|issue:45406",
+        "issue:43873|issue:44945",
+        "issue:44263|issue:45092",
+        "issue:44263|issue:44360",
+        "issue:44514|issue:45103",
+        "issue:44373|issue:44485",
+        "issue:43479|issue:43644",
+        "issue:44295|issue:44623",
+        "issue:44261|issue:45092",
+        "issue:43723|issue:44871"
+      ],
+      "prepared_review_unit_hash": "e1a45c55f31d8987eb0e07458212ced58b8bfb33cbebe23956295f7224212bbb",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 35,
+        "serialized_chars": 38122,
+        "estimated_input_tokens": 9531,
+        "estimated_eval_tokens": 19318
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are not duplicates of one another. They span many unrelated Transformers regressions and support issues: imports/install extras, trainer behavior, tokenizer/save_pretrained bugs, multimodal processor/config issues, quantization/offloading, and version-specific v5 regressions. Similarity is mostly superficial because many titles mention the same package, version, or API surface.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44297 is the cleanest anchor: it describes a concrete, self-contained regression in tokenizer.save_pretrained with a specific observable mismatch and likely a focused fix.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44060",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "Both are Qwen-related, but one is a tied-weights warning and the other is a missing _tp_plan for tensor parallelism; different code paths and fixes."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Missing image-processor method vs remote-code checkpoint/meta-init incompatibility; not the same underlying bug."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "One is a dependency version bump problem, the other is cleanup of leftover dummy classes; unrelated."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both involve tokenization in v5, but one is a clean-up flag behavior change and the other is a DeepSeek tokenizer correctness bug."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44448",
+            "accept": false,
+            "reason": "Both report v5 output changes, but the symptoms and likely causes are model- and tokenizer-specific, not one shared defect."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Multimodal config default initialization vs ModelOutput key assignment; completely different components."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "torch.compile attention-kernel divergence is unrelated to a Phi-4 multimodal load failure."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Reading a registered buffer and a double optimizer-step under deepspeed are unrelated failures."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "One is a missing import/export from image_utils; the other is stale dummy classes in dummy_pt_objects.py."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Package extra/backend inclusion issue vs missing image_utils import; different problem and fix."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "Both touch save_pretrained, but one is FP8 MoE serialization and the other is tokenizer class metadata mismatch; not the same bug."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Python-version compatibility regression is unrelated to a missing import from image_utils."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Tokenizer serialization metadata vs deepspeed optimizer stepping; no shared underlying code path."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44869",
+            "accept": false,
+            "reason": "Phi4MultimodalConfig initialization issue vs Whisper word-timestamp decode crash; unrelated."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Unsupported feature-extractor argument is unrelated to the auto_docstring annotations crash."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "FP8 MoE save_pretrained issue vs processor save_pretrained missing files; both serialization-related but different objects and symptoms."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Accuracy regression in FineGrainedFP8 is not the same as missing files in processor.save_pretrained."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "MLFlow reporting of failed jobs and an unexpected Trainer compute_loss kwarg are unrelated."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "AutoImageProcessor import failure vs is_offline_mode import failure; different missing symbols and dependency breakage."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "MLFlow status reporting and Llama 3.2 Vision model loading regression do not match."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Tokenizer save metadata bug vs Qwen3.5 MoE from_pretrained error; different subsystem and failure mode."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Tokenizer loading regression in v5 is unrelated to flash-attention generation issues."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:30333",
+            "accept": false,
+            "reason": "Missing PyTorch for AutoModel import is unrelated to MLFlow job reporting."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Training-job reporting and embedding-resize reinitialization are different bugs."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "FP8 serialization bug vs save_pretrained shard-size default question; not duplicates."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Chat-completions request rejection and Gemma4Processor missing _tokenizer are different failures."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Quantization offloading issue is not the same as pipeline-parallelism output corruption."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "GlmMoeDsaIndexer torch.split issue vs remote-code checkpoint/meta-init incompatibility; unrelated."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Both mention DSA indexer, but one is about split return values and the other about a missing ReLU; distinct bugs."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Batched chat-template crash and auto_docstring annotations crash are unrelated."
+          },
+          {
+            "left": "issue:44373",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Docstring correction for position_ids is not the same as GLM-5 RoPE implementation behavior."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:43644",
+            "accept": false,
+            "reason": "Default multimodal config initialization vs non-persistent buffer junk fill; different code paths."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "position_ids buffer read error and missing processor save files are different bugs."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "MLA q_a_layernorm epsilon precision issue is unrelated to remote-code checkpoint meta-init incompatibility."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Tokenizer loading regression in v5 is not the same as inconsistent eos_token_id configuration."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 36,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:22355",
+        "issue:28282",
+        "issue:30990",
+        "issue:33357",
+        "issue:38175",
+        "issue:39401",
+        "issue:41628",
+        "issue:42371",
+        "issue:42757",
+        "issue:42831",
+        "issue:42886",
+        "issue:42907",
+        "issue:42994",
+        "issue:43066",
+        "issue:43122",
+        "issue:43232",
+        "issue:43479",
+        "issue:43576",
+        "issue:43723",
+        "issue:43756",
+        "issue:43784",
+        "issue:43824",
+        "issue:43937",
+        "issue:44038",
+        "issue:44060",
+        "issue:44117",
+        "issue:44263",
+        "issue:44295",
+        "issue:44297",
+        "issue:44373",
+        "issue:44442",
+        "issue:44514",
+        "issue:44521",
+        "issue:44568",
+        "issue:44704",
+        "issue:44779",
+        "issue:44913",
+        "issue:44945",
+        "issue:45092",
+        "issue:45245",
+        "issue:45278",
+        "issue:45290",
+        "issue:45310",
+        "issue:45357",
+        "issue:45593",
+        "issue:45600",
+        "issue:45753"
+      ],
+      "soft_pairs": [
+        "issue:45092|issue:45357",
+        "issue:43479|issue:44913",
+        "issue:30990|issue:38175",
+        "issue:22355|issue:28282",
+        "issue:42886|issue:42994",
+        "issue:41628|issue:43824",
+        "issue:42831|issue:45310",
+        "issue:39401|issue:44568",
+        "issue:44263|issue:44373",
+        "issue:44945|issue:45753",
+        "issue:42831|issue:44038",
+        "issue:42757|issue:45278",
+        "issue:43576|issue:43784",
+        "issue:42886|issue:43066",
+        "issue:43232|issue:45593",
+        "issue:44117|issue:44704",
+        "issue:42371|issue:44263",
+        "issue:43756|issue:44060",
+        "issue:30990|issue:33357",
+        "issue:44704|issue:45290",
+        "issue:42907|issue:44945",
+        "issue:43122|issue:44514",
+        "issue:44295|issue:44297",
+        "issue:43232|issue:45245",
+        "issue:43784|issue:45600",
+        "issue:43723|issue:44779",
+        "issue:43723|issue:44442",
+        "issue:42907|issue:43937",
+        "issue:43122|issue:44521"
+      ],
+      "prepared_review_unit_hash": "79e720d994e3b8a1e8e4a1c36af3f0b426b89b3fbc6755d36dc66c23b04976da",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 35335,
+        "estimated_input_tokens": 8834,
+        "estimated_eval_tokens": 17924
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a loose set of unrelated Transformers issues, with only weak topical similarity in a few tokenizer/import-regression reports. I do not see any true duplicate pairs among the soft edges.",
+        "confidence": 0.48,
+        "canonical_issue_reason": "Issue 45278 is the broadest umbrella-style report in the set, covering multiple import regressions after upgrading to v5; it is the closest thing to a cluster representative, though the cluster is still heterogeneous.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45278 best represents the only recurring high-level theme here (v5 import failures), but it should be treated as a broad representative rather than a true duplicate canonical.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45092",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Both are Qwen3.5-related, but one is a v5 meta-init incompatibility and the other is a save_pretrained key-mapping regression; different bugs."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Different models and code paths: default multimodal config initialization vs GPTNeoX rotary_pct reload behavior."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:38175",
+            "accept": false,
+            "reason": "Unrelated symptoms and subsystems: Sentence Transformers loading hang vs zero probabilities for a SigLIP2 model."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Different import failures: missing transformers.onnx module vs AutoModel complaining PyTorch is absent."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:42994",
+            "accept": false,
+            "reason": "Offline cache loading for tokenizers is unrelated to quantized model saving failures."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but for different missing symbols and different model families; not the same bug."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "FineGrainedFP8 accuracy regression is unrelated to Qwen3.5-Moe from_pretrained errors."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both involve tokenizers, but offset_mapping is a different failure mode from add_special_tokens/BOS-EOS behavior."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "A runtime torch.split issue in a model component is unrelated to a docstring typo."
+          },
+          {
+            "left": "issue:44945",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Pipeline-parallel output corruption and a specific-image infinite loop are distinct inference bugs."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44038",
+            "accept": false,
+            "reason": "Different models and problems: FP8 accuracy vs Qwen3-VL-Moe v5 incompatibility."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "42757 is a single missing import from huggingface_hub; 45278 is a broad report about many import errors, not the same concrete bug."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:43784",
+            "accept": false,
+            "reason": "A broken CLI env command and a NameError during sentence-transformers import are different regressions."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:43066",
+            "accept": false,
+            "reason": "Offline tokenizer cache loading does not match a tokenizer decoder-type mismatch in v5."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Generation kwargs synchronization and D-FINE auxiliary-loss behavior are unrelated."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "TOKENIZER_MAPPING_NAMES handling and AutoProcessor cached_file kwargs are different loader bugs."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "TF32 settings guidance and a torch.split return-value issue have no overlap in code path."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44060",
+            "accept": false,
+            "reason": "Smollm3 RoPE-layer truncation and Qwen3-Next tied-weights warnings are unrelated model/config issues."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "A loading hang and a macOS bus error on a CLIP community model are different failures."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "AutoProcessor argument forwarding and chat-template crashes with tool calls are separate code paths."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Quantized/dequantized save issues are unrelated to incorrect outputs under pipeline parallelism."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "Tokenizer-version output drift and batched apply_chat_template crashes are distinct problems."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "Reading a registered buffer and save_pretrained writing the wrong tokenizer_class are different bugs."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "Generation after sync_gpus is unrelated to the 2^24 category-limit runtime error."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Missing nn during import and stale auto_mappings references are different breakages."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both concern tokenizer loading/regressions, but one is a load failure and the other is incorrect tokenization output; not the same concrete bug."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "These are separate AutoTokenizer failures affecting different tokenizer classes and symptoms."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43937",
+            "accept": false,
+            "reason": "Model save failures and invalid GenerationConfig are unrelated."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Tokenization drift and all-zero assistant masks in multimodal chat templates are different issues."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 37,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:22355",
+        "issue:28218",
+        "issue:29127",
+        "issue:30990",
+        "issue:36296",
+        "issue:39401",
+        "issue:41628",
+        "issue:42222",
+        "issue:42371",
+        "issue:42757",
+        "issue:42907",
+        "issue:43066",
+        "issue:43116",
+        "issue:43122",
+        "issue:43232",
+        "issue:43408",
+        "issue:43441",
+        "issue:43452",
+        "issue:43475",
+        "issue:43784",
+        "issue:43824",
+        "issue:43874",
+        "issue:43937",
+        "issue:43976",
+        "issue:43992",
+        "issue:44038",
+        "issue:44222",
+        "issue:44261",
+        "issue:44295",
+        "issue:44297",
+        "issue:44351",
+        "issue:44360",
+        "issue:44373",
+        "issue:44485",
+        "issue:44488",
+        "issue:44521",
+        "issue:44625",
+        "issue:44704",
+        "issue:44743",
+        "issue:44779",
+        "issue:44829",
+        "issue:44964",
+        "issue:44977",
+        "issue:45200",
+        "issue:45216",
+        "issue:45278",
+        "issue:45446",
+        "issue:45478"
+      ],
+      "soft_pairs": [
+        "issue:43408|issue:43475",
+        "issue:41628|issue:45278",
+        "issue:44261|issue:44373",
+        "issue:44360|issue:44485",
+        "issue:43232|issue:44743",
+        "issue:43116|issue:45200",
+        "issue:44625|issue:44829",
+        "issue:29127|issue:30990",
+        "issue:42907|issue:44964",
+        "issue:43122|issue:44779",
+        "issue:44038|issue:44297",
+        "issue:44038|issue:44295",
+        "issue:42907|issue:43452",
+        "issue:39401|issue:44488",
+        "issue:36296|issue:43116",
+        "issue:43824|issue:44351",
+        "issue:44038|issue:44222",
+        "issue:39401|issue:44521",
+        "issue:45216|issue:45478",
+        "issue:42757|issue:44351",
+        "issue:22355|issue:28218",
+        "issue:43066|issue:44779",
+        "issue:43122|issue:44977",
+        "issue:42907|issue:44222",
+        "issue:43784|issue:45446",
+        "issue:43992|issue:44704",
+        "issue:43937|issue:44485",
+        "issue:42222|issue:43441",
+        "issue:43976|issue:45446",
+        "issue:42371|issue:43874"
+      ],
+      "prepared_review_unit_hash": "9672fe689458ff9cc908972b0592d52008cf1580e4d3d2eb5b6ebf8dcca241a9",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36250,
+        "estimated_input_tokens": 9063,
+        "estimated_eval_tokens": 18382
+      },
+      "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 across tokenization, model loading/saving, generation, and import regressions. The soft-similarity pairs usually share a subsystem or model family, but not the same concrete underlying bug, so I would not merge them as duplicates.",
+        "confidence": 0.61,
+        "canonical_issue_reason": "Issue 45278 is the broadest umbrella-style report in the set: it describes multiple import failures after upgrading across versions, which makes it the best fit as a representative cluster anchor among these mostly import/model-regression reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45278 is the strongest global representative because it frames a broad, user-visible regression with upgrade impact and can subsume nearby import-related reports better than the more narrow model-specific or example-script issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43408",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Both are SAM 3 video related, but one is a model-type/config warning and the other is a missing attribute on encoder output; different code paths and fixes."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Both concern imports after upgrade, but 41628 is a specific AutoImageProcessor import failure while 45278 is a broad multi-import regression; not the same bug."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "A precision/config bug in MLA q_a_layernorm is unrelated to a docstring issue for position_ids."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "DSA indexer/ReLU discussion is unrelated to GLM-5 RoPE implementation."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Both involve generation state handling, but one is sync_gpus/_update_model_kwargs_for_generation and the other is recurrent-state reset with cache; different concrete defects."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Example-script multi-label classification output is unrelated to Gemma 4 mm_token_type_ids defaults."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Qwen3.5 num_labels propagation is a config bug, while flash_attention_3 degenerate training is a training-kernel behavior bug."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:30990",
+            "accept": false,
+            "reason": "LayoutLMv3 box-clarity/error-message issue is unrelated to Sentence Transformers loading stalling."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Saving dequantized Ministral/Devstral checkpoints and loading Phi-4 multimodal-instruct are different model-specific failures."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both are tokenizer regressions in v5, but one is generic tokenization drift and the other is Deepseek-specific incorrect results; not clearly the same bug."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe runtime bug and tokenizer_class metadata mismatch in save_pretrained are different issues."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44295",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe behavior and position_ids buffer reading are unrelated code paths."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "Saving dequantized models and gguf_file loading via from_pretrained are both I/O-related, but they are not the same underlying failure."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both are tokenizer-related, but one is Qwen3 offset_mapping and the other is loading cjvt/sleng-bert; different bugs."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43116",
+            "accept": false,
+            "reason": "Tensor-parallel training and example-script classification output are unrelated."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but they concern different symbols from different packages and different breakages."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Both mention Moe/Qwen3-VL-Moe, but one is a general model/version bug and the other is FP8 save_pretrained behavior; not the same change."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Offset mapping for Qwen3 tokenizer and assistant_mask generation for multimodal chat templates are distinct tokenizer features."
+          },
+          {
+            "left": "issue:45216",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both involve Qwen3.5 and Transformers v5.x, but one is save_pretrained checkpoint corruption and the other is a from_pretrained loading error; different concrete bugs."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Missing is_offline_mode and missing HybridCache are separate import regressions."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:28218",
+            "accept": false,
+            "reason": "Missing transformers.onnx and extra space after added token are unrelated."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Wrong decoder type in v5 and Deepseek tokenization regression are both tokenizer-related but not the same underlying issue."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Tokenization drift and Qwen3.5 flash-attention generation failure are unrelated."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Saving dequantized checkpoints and FP8 save_pretrained for moe are both saving-related, but they affect different models and failure modes."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Sentence-Transformers import NameError and flex_attention PyTorch version check are unrelated."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "Missing embed_tokens.weight in UMT5Encoder loading and AutoProcessor kwargs not forwarded to cached_file are different API bugs."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "GLM-5 generation-config validation and GLM-5 RoPE implementation are separate defects."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "VitPose model breakage and Ministral-3 FlashAttention failure are unrelated."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Python version compatibility and a PyTorch import/version check bug are different release/environment issues."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:43874",
+            "accept": false,
+            "reason": "TF32 API guidance and GLM46V image-patch counting 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": 2,
+      "component_count": 16,
+      "review_unit_index": 38,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:22355",
+        "issue:29127",
+        "issue:33453",
+        "issue:38617",
+        "issue:39401",
+        "issue:41628",
+        "issue:42175",
+        "issue:42673",
+        "issue:42831",
+        "issue:43064",
+        "issue:43116",
+        "issue:43381",
+        "issue:43408",
+        "issue:43421",
+        "issue:43450",
+        "issue:43493",
+        "issue:43673",
+        "issue:43742",
+        "issue:43825",
+        "issue:43827",
+        "issue:43937",
+        "issue:43986",
+        "issue:44038",
+        "issue:44075",
+        "issue:44117",
+        "issue:44222",
+        "issue:44291",
+        "issue:44410",
+        "issue:44779",
+        "issue:44829",
+        "issue:44936",
+        "issue:44938",
+        "issue:44945",
+        "issue:44977",
+        "issue:45071",
+        "issue:45072",
+        "issue:45081",
+        "issue:45083",
+        "issue:45216",
+        "issue:45278",
+        "issue:45310",
+        "issue:45405",
+        "issue:45446",
+        "issue:45478",
+        "issue:45563",
+        "issue:45588",
+        "issue:45753"
+      ],
+      "soft_pairs": [
+        "issue:44075|issue:44936",
+        "issue:43493|issue:45753",
+        "issue:42831|issue:45753",
+        "issue:43116|issue:43825",
+        "issue:44038|issue:45478",
+        "issue:44222|issue:45478",
+        "issue:43116|issue:43493",
+        "issue:43064|issue:43673",
+        "issue:45083|issue:45446",
+        "issue:43937|issue:45563",
+        "issue:42175|issue:44938",
+        "issue:38617|issue:45278",
+        "issue:33453|issue:39401",
+        "issue:39401|issue:44779",
+        "issue:22355|issue:29127",
+        "issue:44117|issue:45588",
+        "issue:43827|issue:45446",
+        "issue:43381|issue:44945",
+        "issue:43937|issue:44977",
+        "issue:45405|issue:45446",
+        "issue:43450|issue:43986",
+        "issue:42673|issue:45072",
+        "issue:44117|issue:45081",
+        "issue:39401|issue:43421",
+        "issue:38617|issue:41628",
+        "issue:44829|issue:45216",
+        "issue:44410|issue:45310",
+        "issue:44291|issue:45071",
+        "issue:43408|issue:43742"
+      ],
+      "prepared_review_unit_hash": "b48511a64e9e44ca6b56674f0f7e1db092650ad39e90b2a5298ccb22565eb8de",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 35579,
+        "estimated_input_tokens": 8895,
+        "estimated_eval_tokens": 18046
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These issues are mostly unrelated Transformers bugs across different subsystems; none of the soft pairs look like the same underlying defect, so I would not merge any of them as duplicates.",
+        "confidence": 0.31,
+        "canonical_issue_reason": "issue:45278 is the broadest representative of a regression-style import breakage after a version upgrade, and it is the most umbrella-like issue in the set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the most general issue and best fits as a cluster representative, even though the surrounding items are not true duplicates.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44075",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Different bugs: optimizer argument handling vs. train/eval lifecycle failure."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Different model families and symptoms; one is SigLIP2 implementation drift, the other is a Qwen3.5 image-specific generation loop."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "FP8 accuracy regression is unrelated to a Qwen3.5 infinite-loop generation bug."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "One is a multi-label classification example bug; the other is a v5 pipeline error-message/docs issue."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both mention Qwen3.5/MoE, but one is a broad v5 bug and the other is a specific from_pretrained error; not the same code-path."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "FP8 save_pretrained behavior and Qwen3.5 from_pretrained loading are different failures."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Classification example output bug and SigLIP2 model discrepancy are unrelated."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "Training optimizer-state corruption is unrelated to GenerationMixin cache handling during chunked prefill."
+          },
+          {
+            "left": "issue:45083",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Different subsystems: qwen3_omni_moe feature-length helper vs. flex_attention import/version check."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "One is invalid GenerationConfig, the other is a stale warning in paged generate()."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "Backend dependency packaging issue is unrelated to a Python 3.14 import/load failure."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "45278 is a broad import regression report, but 38617 is a specific missing symbol; too generic to treat as the same bug."
+          },
+          {
+            "left": "issue:33453",
+            "right": "issue:39401",
+            "accept": false,
+            "reason": "Tokenizer loading regression and Qwen3 offset_mapping bug are different tokenizer failures."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but for different models and different symptoms."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "Module import failure and LayoutLMv3 error-message clarity issue are unrelated."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Tokenizer mapping fallback bug is unrelated to flash_attention handling of s_aux=None."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Docs/pipeline deprecation issue and flex_attention version-check bug are different changes."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Gradient checkpointing eval-mode constraint is not the same as incorrect pipeline-parallel output."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Both affect generation, but one is config validation and the other is flash-attention generation failure."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Missing PEFT release version and PyTorch import gating are unrelated packaging/version issues."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:43986",
+            "accept": false,
+            "reason": "Batched video processor shape bug and torchvision-missing crash are different problems."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "VRAM leak under multithreaded dataloading is unrelated to dtype mismatch in inference CI."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Different tokenizer-related failures: mapping lookup vs. Mistral regex patch crashing on backend_tokenizer."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:43421",
+            "accept": false,
+            "reason": "Offset_mapping regression and runtime post-processor update bug are different tokenizer behaviors."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "Missing layer_type_validation import and missing AutoImageProcessor import are distinct import regressions."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Flash-attention-3 training degeneration is unrelated to Qwen3.5 save_pretrained checkpoint corruption."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Missing projections in qwen3next layer 0 is not the same as Qwen3.5 MoE from_pretrained failure."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "init_empty_weights constructor argument breakage and PretrainedConfig type-checking regression are different code paths."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "SAM model-type warning and MobileLLM key error are unrelated loading issues."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 39,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:38617",
+        "issue:41628",
+        "issue:42175",
+        "issue:42222",
+        "issue:42371",
+        "issue:42673",
+        "issue:42831",
+        "issue:43012",
+        "issue:43066",
+        "issue:43097",
+        "issue:43122",
+        "issue:43335",
+        "issue:43408",
+        "issue:43441",
+        "issue:43452",
+        "issue:43577",
+        "issue:43874",
+        "issue:43906",
+        "issue:43931",
+        "issue:43937",
+        "issue:43976",
+        "issue:43986",
+        "issue:44038",
+        "issue:44075",
+        "issue:44117",
+        "issue:44220",
+        "issue:44261",
+        "issue:44263",
+        "issue:44368",
+        "issue:44410",
+        "issue:44483",
+        "issue:44484",
+        "issue:44485",
+        "issue:44493",
+        "issue:44496",
+        "issue:44829",
+        "issue:44945",
+        "issue:44964",
+        "issue:44977",
+        "issue:44987",
+        "issue:45071",
+        "issue:45278",
+        "issue:45310",
+        "issue:45459",
+        "issue:45464",
+        "issue:45478",
+        "issue:45542",
+        "issue:45656"
+      ],
+      "soft_pairs": [
+        "issue:42222|issue:45542",
+        "issue:44484|issue:45310",
+        "issue:42175|issue:42222",
+        "issue:43937|issue:44496",
+        "issue:44483|issue:45464",
+        "issue:44075|issue:45656",
+        "issue:43012|issue:43577",
+        "issue:44964|issue:45278",
+        "issue:43066|issue:43122",
+        "issue:42371|issue:44261",
+        "issue:38617|issue:44220",
+        "issue:43874|issue:44261",
+        "issue:43097|issue:44493",
+        "issue:44410|issue:45478",
+        "issue:44484|issue:45478",
+        "issue:42222|issue:45459",
+        "issue:44263|issue:44485",
+        "issue:43441|issue:44829",
+        "issue:43012|issue:43408",
+        "issue:41628|issue:45071",
+        "issue:45071|issue:45542",
+        "issue:42673|issue:44977",
+        "issue:42371|issue:42831",
+        "issue:42673|issue:43906",
+        "issue:44117|issue:44987",
+        "issue:43986|issue:44484",
+        "issue:43097|issue:43452",
+        "issue:43931|issue:44038",
+        "issue:43335|issue:45071",
+        "issue:43976|issue:44945",
+        "issue:43408|issue:44368",
+        "issue:43874|issue:44263"
+      ],
+      "prepared_review_unit_hash": "f1ed28b3f7d8eb312e02435492145f0c4c63a98cb0ccf9921d2ad018cf939728",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37125,
+        "estimated_input_tokens": 9282,
+        "estimated_eval_tokens": 18820
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly heterogeneous Transformers regressions and API/import breakages; there is no strong duplicate pair, so I reject all soft-edge merges. The most representative issue is the broad v5.5.0 import-regression report.",
+        "confidence": 0.61,
+        "canonical_issue_reason": "issue:45278 is the broadest umbrella report here: \"Many import errors after update from 4.57.0 to 5.5.0,\" which best matches the cluster\u2019s general upgrade/import-breakage theme.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 has the clearest high-level symptom, widest scope, and strongest potential to represent the cluster\u2019s recurring v5 import regression pattern.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42222",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Different failures: VitPose model loading vs missing TensorFlow backend when only tensorboard is installed."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Unrelated problems: shard-size configuration question vs Qwen3.5 MoE from_pretrained loading error."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:42222",
+            "accept": false,
+            "reason": "Backend packaging issue vs a broken VitPose model path; not the same bug."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "GLM-5 GenerationConfig validation failure is distinct from an unknown-model/config.json model_type error."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Both hit chat/completions, but one is a general request-format rejection and the other is a Qwen3.5-0.8B streaming failure."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Different optimizer bugs: SGD args ignored vs DeepSpeed causing optimizer.step() to run twice."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "Both involve dtype/precision, but one is a PyTorch warning during compile and the other is incorrect dtype loading for Blip2."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Both are load/import regressions, but one is Phi-4 multimodal model loading and the other is a broad import-breakage report."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:43122",
+            "accept": false,
+            "reason": "Both concern tokenizer behavior in v5, but one is decoder type selection and the other is changed tokenization output."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44261",
+            "accept": false,
+            "reason": "TF32 settings guidance is unrelated to MLA layernorm epsilon precision drift."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Different import-time failure versus audio feature extraction bug."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:44261",
+            "accept": false,
+            "reason": "Missing image-patch method in GLM46V is unrelated to MLA RMS norm epsilon precision issues."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Different regressions: removed tie_embeddings flag vs unexpected position-id keys across models."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Different Qwen-family loading bugs; one is missing projections in qwen3next, the other is Qwen3.5 MoE loading failure."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Shard-size default discussion is unrelated to Qwen3.5 MoE from_pretrained errors."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "VitPose model breakage is unrelated to tokenizer error masking when protobuf is absent."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Different GLM internals: torch.split return handling vs RoPE implementation."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Both mention flash attention, but one is Ministral-3 inference failure and the other is degenerate training in sequence classification."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43408",
+            "accept": false,
+            "reason": "Precision warning on bf16 compile is unrelated to a model-type mismatch warning."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "AutoImageProcessor import failure and PreTrainedConfig type-checking regression are different code paths."
+          },
+          {
+            "left": "issue:45071",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Type-checking regression is unrelated to undefined TensorFlow backend handling."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "VRAM leak in multithreaded Qwen3ForCausalLM usage is not the same as flash-attention generation corruption in Qwen3.5."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:42831",
+            "accept": false,
+            "reason": "TF32 control API discussion is unrelated to FineGrainedFP8 accuracy drift."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "No evidence they describe the same underlying bug; one is a VRAM leak, the other a reproduction of another issue."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Tokenizer mapping None handling is not the same as loading a specific model package failing in v5.1.0."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Video-model crash without torchvision is unrelated to a save_pretrained shard-size question."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "Removed tie_embeddings behavior does not match gguf_file breaking AutoTokenizer/AutoModel loading."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44038",
+            "accept": false,
+            "reason": "Both are Qwen3-VL related, but weight-shape mismatch and Moe loading failure are distinct defects."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer config bug is unrelated to PreTrainedConfig type-checking breakage."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Python version compatibility issue is unrelated to incorrect outputs under pipeline parallelism."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Two different config warnings; one is SAM3 model-type mismatch, the other is tie_word_embeddings guidance."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Layer projection missing in qwen3next is not the same as Qwen3.5 MoE from_pretrained loading error."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "No overlap beyond both being Transformers issues; the symptoms and code paths differ completely."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "Different failure modes: removed config flag vs gguf_file loading breakage."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44038",
+            "accept": false,
+            "reason": "Shape mismatch on one model is not the same as a broader Qwen3-VL-Moe bug."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Sparse-layer creation bug and type-checking regression are unrelated."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Python support regression is unrelated to pipeline-parallel output corruption."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "These are distinct warnings with different triggers and fixes."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "Different model families and different internals; image-patch counting vs torch.split indexing."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 40,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:38617",
+        "issue:42617",
+        "issue:42673",
+        "issue:42757",
+        "issue:42886",
+        "issue:42907",
+        "issue:43012",
+        "issue:43064",
+        "issue:43066",
+        "issue:43097",
+        "issue:43208",
+        "issue:43381",
+        "issue:43502",
+        "issue:43531",
+        "issue:43550",
+        "issue:43577",
+        "issue:43825",
+        "issue:43873",
+        "issue:43901",
+        "issue:43906",
+        "issue:43931",
+        "issue:43986",
+        "issue:43992",
+        "issue:44038",
+        "issue:44062",
+        "issue:44117",
+        "issue:44261",
+        "issue:44315",
+        "issue:44351",
+        "issue:44368",
+        "issue:44410",
+        "issue:44464",
+        "issue:44484",
+        "issue:44485",
+        "issue:44655",
+        "issue:44704",
+        "issue:44829",
+        "issue:44964",
+        "issue:45071",
+        "issue:45103",
+        "issue:45245",
+        "issue:45290",
+        "issue:45341",
+        "issue:45372",
+        "issue:45375",
+        "issue:45561",
+        "issue:45588",
+        "issue:45753"
+      ],
+      "soft_pairs": [
+        "issue:44829|issue:45245",
+        "issue:38617|issue:42757",
+        "issue:43064|issue:44315",
+        "issue:44464|issue:45071",
+        "issue:44410|issue:45375",
+        "issue:44062|issue:44117",
+        "issue:43097|issue:45588",
+        "issue:38617|issue:44351",
+        "issue:42673|issue:43577",
+        "issue:44704|issue:45103",
+        "issue:44038|issue:44410",
+        "issue:44261|issue:44485",
+        "issue:43931|issue:44410",
+        "issue:44964|issue:45341",
+        "issue:43986|issue:44117",
+        "issue:43502|issue:43992",
+        "issue:43097|issue:43873",
+        "issue:43066|issue:43906",
+        "issue:43986|issue:45290",
+        "issue:43986|issue:45372",
+        "issue:45341|issue:45561",
+        "issue:43986|issue:45341",
+        "issue:43097|issue:43901",
+        "issue:43825|issue:44655",
+        "issue:42617|issue:45341",
+        "issue:43992|issue:44484",
+        "issue:45341|issue:45588",
+        "issue:43012|issue:43550",
+        "issue:43208|issue:43550",
+        "issue:42886|issue:43381",
+        "issue:43012|issue:44368",
+        "issue:42907|issue:45753",
+        "issue:42907|issue:44655",
+        "issue:43531|issue:45341"
+      ],
+      "prepared_review_unit_hash": "62e9f4c12a443b13d5650bf2579e24c8e650f09f90ebab3269798dc3f4719bc7",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 37768,
+        "estimated_input_tokens": 9442,
+        "estimated_eval_tokens": 19140
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is very heterogeneous; the soft pairs are mostly same-subsystem but not the same underlying bug. No soft edge looks safe to merge as a duplicate, so I\u2019d anchor on the most broadly described user-facing regression rather than treating the set as one duplicate family.",
+        "confidence": 0.81,
+        "canonical_issue_reason": "Issue 43986 is the strongest standalone anchor here: it has a clear reproducible crash, broader user impact, and more discussion than the other narrowly scoped reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43986 is the best issue candidate if one must be chosen, because it is concrete, well-scoped, and more representative than the one-off bug reports in this mixed cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44829",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "Different failures: SDPA/torch.compile training regression vs category-cardinality runtime error."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and different breakages."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44315",
+            "accept": false,
+            "reason": "FSDP2 optimizer-state corruption and Liger application with model_init are unrelated code paths."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Compiled chunked generation output mismatch is unrelated to PreTrainedConfig type-checking changes."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Model projection bug and missing config field are different issues, not the same defect."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44117",
+            "accept": false,
+            "reason": "AddedToken constructor misuse is separate from TOKENIZER_MAPPING_NAMES returning None."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Deprecated embedding-tie API removal is unrelated to flash-attention sink-less crash."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Different missing imports from different modules; same general symptom only."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "VRAM leak in Qwen3ForCausalLM threads is unrelated to Blip2 dtype loading."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "AutoProcessor kwargs passthrough bug is unrelated to auto_docstring annotation handling."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44410",
+            "accept": false,
+            "reason": "Different Qwen model families and different failure modes."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "RMS norm epsilon config bug and GLM-5 RoPE implementation issue are not the same bug."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44410",
+            "accept": false,
+            "reason": "Weight-shape mismatch on Qwen3-VL-30B is distinct from missing Qwen3next projections."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Model loading regression and a testing_utils bug are unrelated."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44117",
+            "accept": false,
+            "reason": "Video-model loading without torchvision is unrelated to tokenizer mapping handling."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:43992",
+            "accept": false,
+            "reason": "Offline-mode API requests and UMT5Encoder missing embed_tokens are different loader bugs."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "API removal warning and quantization offloading behavior are unrelated."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "Wrong decoder type and an isolated reproduction of another issue do not match concretely."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "AutoProcessor/torchvision crash is unrelated to chat-template tool-call handling."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "Different import/dependency failures in different processor code paths."
+          },
+          {
+            "left": "issue:45341",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "The titles suggest related testing utilities, but not the same specific defect."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "A processor-loading crash is not the same as a generic testing_utils bug."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Embedding-tie API removal and return_all_scores documentation mismatch are unrelated."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "pipeline translation-task messaging and save_pretrained pipeline serialization are separate topics."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "3d_parallel.py execution failure is unrelated to testing_utils."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Missing encoder weights on load is unrelated to the default shard-size question."
+          },
+          {
+            "left": "issue:45341",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Testing-utils race conditions and flash_attention sink handling are different code paths."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Compilation warning and degenerate training under SDPA are not the same bug."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "xLSTM training blockers and Bamba SDPA compile failure are unrelated."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "Offline cache loading and gradient-checkpointing-in-eval are separate failures."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Both involve warnings, but they stem from different mechanisms and fixes."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Saving dequantized models and Qwen3_5 image looping are unrelated."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Both mention saving, but model dequantization and Pipeline serialization are different problems."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior is unrelated to testing_utils."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 41,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:36010",
+        "issue:36683",
+        "issue:38617",
+        "issue:41669",
+        "issue:42757",
+        "issue:42994",
+        "issue:43012",
+        "issue:43064",
+        "issue:43066",
+        "issue:43208",
+        "issue:43299",
+        "issue:43352",
+        "issue:43441",
+        "issue:43475",
+        "issue:43502",
+        "issue:43525",
+        "issue:43526",
+        "issue:43618",
+        "issue:43756",
+        "issue:43761",
+        "issue:43824",
+        "issue:43867",
+        "issue:43881",
+        "issue:43901",
+        "issue:43906",
+        "issue:43937",
+        "issue:43976",
+        "issue:43992",
+        "issue:44038",
+        "issue:44220",
+        "issue:44230",
+        "issue:44242",
+        "issue:44246",
+        "issue:44263",
+        "issue:44380",
+        "issue:44483",
+        "issue:44485",
+        "issue:44514",
+        "issue:44655",
+        "issue:44683",
+        "issue:45216",
+        "issue:45290",
+        "issue:45310",
+        "issue:45375",
+        "issue:45381",
+        "issue:45440",
+        "issue:45479",
+        "issue:45753"
+      ],
+      "soft_pairs": [
+        "issue:44483|issue:44514",
+        "issue:43208|issue:43526",
+        "issue:42757|issue:43824",
+        "issue:43012|issue:44683",
+        "issue:44483|issue:45290",
+        "issue:44038|issue:44230",
+        "issue:43618|issue:43761",
+        "issue:38617|issue:43502",
+        "issue:44483|issue:45381",
+        "issue:43064|issue:44380",
+        "issue:43299|issue:43867",
+        "issue:43901|issue:44655",
+        "issue:36683|issue:43475",
+        "issue:43867|issue:45216",
+        "issue:42994|issue:44655",
+        "issue:43352|issue:43937",
+        "issue:43066|issue:45381",
+        "issue:36010|issue:42757",
+        "issue:44220|issue:44655",
+        "issue:43756|issue:44485",
+        "issue:44263|issue:45440",
+        "issue:36683|issue:43525",
+        "issue:43618|issue:45216",
+        "issue:43906|issue:45753",
+        "issue:36683|issue:44242",
+        "issue:36683|issue:43881",
+        "issue:43299|issue:45753",
+        "issue:43352|issue:43976",
+        "issue:43441|issue:44380",
+        "issue:43352|issue:44380",
+        "issue:43992|issue:45310",
+        "issue:44380|issue:45479",
+        "issue:41669|issue:44246",
+        "issue:43867|issue:45375"
+      ],
+      "prepared_review_unit_hash": "e15974d79016ce190cb5009f1650cc04f0d9e50b4cf0d398091a3c86fc125daf",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 37522,
+        "estimated_input_tokens": 9381,
+        "estimated_eval_tokens": 19018
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous transformer bugs, with only a couple of plausible duplicate pairs. The strongest match is the import-time slowdown pair; the CLIP vision output regressions also look related. Most other pairs are broad-similarity matches only and should stay separate.",
+        "confidence": 0.78,
+        "canonical_issue_reason": "Issue 41669 is the most concrete root-cause report in the cluster and cleanly anchors the import-performance duplicate pair with 44246.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 41669 is the best overall representative because it describes a specific, actionable regression rather than a vague symptom.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44483",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "Both are chat-related failures, but one is API request validation and the other is batched chat templating; different code paths."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Unrelated models/subsystems: xLSTM training bugs vs BeitImageProcessorFast label reduction."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and different causes."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "BF16 warning behavior and compiled flex_attention failure are different bugs."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both touch chat templating, but one is /v1/chat/completions request acceptance and the other is assistant tool-call templating crashes."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "Same model family, but one is a loading bug and the other is fp8 support/inference request; not the same underlying issue."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:43761",
+            "accept": true,
+            "reason": "Both point to the same CLIP v5 regression class where model outputs are not populated correctly in forward return paths."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:43502",
+            "accept": false,
+            "reason": "Import error from configuration_utils and unwanted network access with local_files_only=True are unrelated."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Chat API request rejection and Qwen2.5-VL video position-id mismatch are unrelated bugs."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "FSDP/optimizer-state corruption and GPT2 attention-scaling behavior are different problems."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43867",
+            "accept": false,
+            "reason": "Qwen3VL-Moe loading failure and sorted-state_dict load error do not appear to share the same root cause."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Pipeline docs mismatch is documentation/behavior drift; save_pretrained pipeline persistence is a separate bug."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Gemma3Config missing vocab_size and Sam3VisionEncoderOutput missing fpn_position_embeddings are unrelated attribute regressions."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both are checkpoint-related, but one is load failure from state_dict ordering and the other is incorrect save_pretrained output for Qwen3.5."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Quantized model saving failure and Pipeline.save_pretrained failure are separate serialization issues."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:43937",
+            "accept": false,
+            "reason": "FlashAttention support issue for Nemotron and GLM-5 generation-config validation are unrelated."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Tokenizer decoder-type regression and Qwen2.5-VL video position-id bug are different code paths."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are import errors, but for different missing symbols and different modules."
+          },
+          {
+            "left": "issue:44220",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Audio feature extraction and pipeline persistence are unrelated."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "RoPE-layer count mismatch in Smollm3 and GLM-5 RoPE implementation discussion are not the same concrete bug."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Both concern MoE internals, but they are different models and different failure modes."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Missing config attributes in different model families are not duplicates."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "CLIP output regression and Qwen3.5 checkpoint-saving regression are unrelated."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "A reproduction of an earlier issue and a Qwen3_5 infinite-loop report are not the same bug."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:44242",
+            "accept": false,
+            "reason": "Gemma3Config attribute absence and MoE load-balancing loss behavior are unrelated."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "Different model load/attribute issues with no clear shared root cause."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Qwen3VL-Moe load failure and Qwen3_5 image infinite loop are separate model bugs."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "FlashAttention support and Python version compatibility are unrelated."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Both mention attention backends, but one is unsupported FlashAttention for a specific model and the other is GPT2 attention scaling being ignored."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Different attention-backend issues without a shared concrete code-path fix."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Missing embed_tokens.weight in UMT5Encoder and Qwen3.5 Moe from_pretrained failure are separate loading bugs."
+          },
+          {
+            "left": "issue:44380",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Attention scaling and zero-loss classification behavior are unrelated."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44246",
+            "accept": true,
+            "reason": "Both describe the same import-performance regression: import * in models makes importing transformers slow."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Sorted state_dict load failure and missing deepstack_visual_indexes config field are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded overall. The import-time slowdown pair is clearly supported by the packet, and the CLIP pair is a reasonable soft duplicate call without overclaiming. The rejected pairs stay conservative and mostly distinguish different subsystems or code paths."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 2,
+      "component_count": 16,
+      "review_unit_index": 42,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:28282",
+        "issue:30990",
+        "issue:31795",
+        "issue:33453",
+        "issue:36296",
+        "issue:36683",
+        "issue:42548",
+        "issue:42890",
+        "issue:43010",
+        "issue:43064",
+        "issue:43097",
+        "issue:43352",
+        "issue:43502",
+        "issue:43575",
+        "issue:43618",
+        "issue:43867",
+        "issue:43901",
+        "issue:43976",
+        "issue:43992",
+        "issue:44062",
+        "issue:44075",
+        "issue:44077",
+        "issue:44112",
+        "issue:44117",
+        "issue:44222",
+        "issue:44351",
+        "issue:44355",
+        "issue:44360",
+        "issue:44380",
+        "issue:44484",
+        "issue:44485",
+        "issue:44492",
+        "issue:44655",
+        "issue:44749",
+        "issue:44855",
+        "issue:44857",
+        "issue:44938",
+        "issue:44964",
+        "issue:44977",
+        "issue:45125",
+        "issue:45341",
+        "issue:45357",
+        "issue:45375",
+        "issue:45440",
+        "issue:45478",
+        "issue:45561",
+        "issue:45588",
+        "issue:45600"
+      ],
+      "soft_pairs": [
+        "issue:43064|issue:44492",
+        "issue:43618|issue:44380",
+        "issue:44355|issue:44855",
+        "issue:43352|issue:45478",
+        "issue:36683|issue:44117",
+        "issue:42890|issue:45561",
+        "issue:43097|issue:43352",
+        "issue:43867|issue:45357",
+        "issue:43976|issue:44938",
+        "issue:43992|issue:45478",
+        "issue:43064|issue:44655",
+        "issue:44360|issue:45440",
+        "issue:36296|issue:43575",
+        "issue:44077|issue:44380",
+        "issue:28282|issue:33453",
+        "issue:43867|issue:44484",
+        "issue:43618|issue:45588",
+        "issue:43618|issue:44857",
+        "issue:43618|issue:44977",
+        "issue:43901|issue:44938",
+        "issue:44655|issue:44938",
+        "issue:42890|issue:44964",
+        "issue:42890|issue:44112",
+        "issue:42890|issue:45600",
+        "issue:43502|issue:44062",
+        "issue:43010|issue:44075",
+        "issue:42890|issue:45341",
+        "issue:42548|issue:44351",
+        "issue:44222|issue:45125",
+        "issue:43502|issue:44485",
+        "issue:44355|issue:44749",
+        "issue:30990|issue:31795",
+        "issue:45125|issue:45478",
+        "issue:45125|issue:45375"
+      ],
+      "prepared_review_unit_hash": "26044eec5228f76ec484aba5e3b0ef8653b61b58e8c3fac59d30064ed48c48db",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 37434,
+        "estimated_input_tokens": 9359,
+        "estimated_eval_tokens": 18974
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous: it mixes unrelated model-loading, tensor-parallel, flash-attention, test-flakiness, documentation, and Python-version issues. None of the soft pairs look like the same concrete bug or change.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "No reliable canonical issue: the items span unrelated failures across different models, subsystems, and versions, so there is no clear duplicate anchor.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No strong representative issue for the cluster; the issues are too diverse to treat one as the global duplicate target.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43064",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "Different problems: FSDP2/PEFT optimizer-state corruption vs a cache-strategy typo."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Both touch attention code, but one is a missing output field and the other is GPT-2 attention scaling under SDPA/FlashAttention."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "Separate import failures with different causes: compiled-file loading vs torch.jit parsing on Python 3.13."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Different model-specific issues: Nemotron Flash Attention 2 support vs Qwen3.5 MoE from_pretrained failure."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:44117",
+            "accept": false,
+            "reason": "One is a missing config attribute; the other is TOKENIZER_MAPPING_NAMES returning None during from_pretrained."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Both are test-related, but one is missing set_seed in model tests and the other is an xdist race on a captured-info file."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43352",
+            "accept": false,
+            "reason": "Unrelated API removal vs flash-attention support error."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different serialization/load bugs: state_dict sorting error vs incorrect visual encoder keys in save_pretrained."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "Both involve Python compatibility, but they affect different versions and failure modes."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "UMT5Encoder missing embed_tokens.weight is a different load-path bug from the Qwen3.5 MoE from_pretrained error."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Optimizer-state corruption during training is unrelated to saving Pipeline objects."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Different code paths: DSA indexer ReLU omission vs DeepseekV3MoE implementation divergence."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Tensor-parallel training bug vs OOM when loading a large Qwen2 model with tp; not the same defect."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "patchtsmixer post_init validation is unrelated to GPT-2 attention scaling under backend changes."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Different regressions: missing PyTorch dependency vs tokenizer loading regression."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Load error with sorted state_dict is unrelated to max_shard_size behavior in save_pretrained."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Both mention attention, but one is CLIPOutput attentions missing and the other is a flash-attention crash on s_aux=None."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Missing attentions output and a float16 AMP/CUDA crash are different bugs."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Attention-output issue is unrelated to Qwen3.5 generation problems with flash-attention."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "Docs mismatch for return_all_scores is unrelated to a Python 3.14 import/load failure."
+          },
+          {
+            "left": "issue:44655",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "Saving Pipeline objects and importing transformers on Python 3.14 are different issues."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Test-seed flakiness is unrelated to the Phi-4 multimodal model loading error."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Both are CI/test problems, but one is missing set_seed and the other is a stale GraniteSpeech device-override test."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Missing set_seed in a model test is unrelated to auto_mappings referencing removed configs."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44062",
+            "accept": false,
+            "reason": "Local-files-only network leakage is unrelated to the AddedToken constructor argument error."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:44075",
+            "accept": false,
+            "reason": "Cache update no_grad decoration and unused SGD arguments are unrelated code paths."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "A missing set_seed in SamHQ tests is not the same as a generic testing_utils bug."
+          },
+          {
+            "left": "issue:42548",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import-name errors, but they concern different symbols and likely different export paths."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "FP8 save_pretrained for MoE and missing _tp_plan for tensor parallelism are different bugs."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "local_files_only network requests have nothing to do with max_shard_size defaults."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "Compiled-file import errors are unrelated to the data-filtering slowdown regression."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:31795",
+            "accept": false,
+            "reason": "Sentence-transformers loading hang is unrelated to model.forward documentation confusion."
+          },
+          {
+            "left": "issue:45125",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both involve Qwen3.5 MoE, but one is a missing tensor-parallel plan and the other is a from_pretrained load error."
+          },
+          {
+            "left": "issue:45125",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Missing _tp_plan and a missing deepstack_visual_indexes field are different configuration/parallelism issues."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 43,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:30333",
+        "issue:31795",
+        "issue:33453",
+        "issue:34689",
+        "issue:36296",
+        "issue:38617",
+        "issue:41628",
+        "issue:42548",
+        "issue:42757",
+        "issue:42831",
+        "issue:43010",
+        "issue:43232",
+        "issue:43299",
+        "issue:43502",
+        "issue:43526",
+        "issue:43576",
+        "issue:43723",
+        "issue:43824",
+        "issue:43827",
+        "issue:43874",
+        "issue:43901",
+        "issue:43976",
+        "issue:44038",
+        "issue:44222",
+        "issue:44230",
+        "issue:44261",
+        "issue:44273",
+        "issue:44315",
+        "issue:44351",
+        "issue:44355",
+        "issue:44360",
+        "issue:44368",
+        "issue:44387",
+        "issue:44393",
+        "issue:44556",
+        "issue:44623",
+        "issue:44655",
+        "issue:44908",
+        "issue:44961",
+        "issue:45003",
+        "issue:45071",
+        "issue:45125",
+        "issue:45230",
+        "issue:45278",
+        "issue:45468",
+        "issue:45542",
+        "issue:45656",
+        "issue:45753"
+      ],
+      "soft_pairs": [
+        "issue:44387|issue:44556",
+        "issue:43010|issue:45278",
+        "issue:30333|issue:33453",
+        "issue:43723|issue:44556",
+        "issue:43010|issue:44368",
+        "issue:43502|issue:44222",
+        "issue:43502|issue:44961",
+        "issue:44315|issue:44360",
+        "issue:42831|issue:43874",
+        "issue:30333|issue:36296",
+        "issue:44355|issue:44655",
+        "issue:44355|issue:45071",
+        "issue:44222|issue:44230",
+        "issue:43827|issue:44961",
+        "issue:43010|issue:45656",
+        "issue:33453|issue:34689",
+        "issue:42757|issue:43576",
+        "issue:43299|issue:44556",
+        "issue:33453|issue:36296",
+        "issue:43576|issue:45071",
+        "issue:43576|issue:43824",
+        "issue:44961|issue:45468",
+        "issue:43010|issue:43976",
+        "issue:31795|issue:38617",
+        "issue:41628|issue:42548",
+        "issue:38617|issue:43576",
+        "issue:43576|issue:44351",
+        "issue:44908|issue:44961",
+        "issue:43010|issue:43232",
+        "issue:36296|issue:43901",
+        "issue:41628|issue:44355",
+        "issue:36296|issue:43526",
+        "issue:44261|issue:44360",
+        "issue:44273|issue:45003",
+        "issue:44623|issue:45230",
+        "issue:44273|issue:45542",
+        "issue:44393|issue:45753",
+        "issue:43576|issue:44556",
+        "issue:44393|issue:45071",
+        "issue:44038|issue:45125"
+      ],
+      "prepared_review_unit_hash": "3c9e8f1c6d014bbe350d2401f72e1d3f38aa849299208d33552111cc5e465578",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 40,
+        "serialized_chars": 39001,
+        "estimated_input_tokens": 9751,
+        "estimated_eval_tokens": 19758
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a loose collection of Transformers v5 regression reports, especially import/load/runtime breakages, but the individual issues are too varied to treat as duplicates of one another. No PRs are present.",
+        "confidence": 0.63,
+        "canonical_issue_reason": "issue:45278 is the broadest representative of the dominant theme: many import/runtime breakages after upgrading to v5. It best captures the cluster\u2019s shared regression pattern without being tied to a single narrow subsystem.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the strongest single issue to anchor the cluster because it is open, version-upgrade focused, and broad enough to subsume several of the nearby import/loading complaints.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44387",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Different problems: int4 quantization memory regression vs checkpoint reload compatibility after upgrading."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Cache update no_grad behavior is unrelated to the broad import-error regression report."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "MLflow reporting on failed training jobs is unrelated to tokenizer loading."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Both involve loading, but one is tokenizer loading in v5 and the other is checkpoint reload after version upgrade; not the same concrete bug."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Cache update decoration and a tie_word_embeddings warning are different code paths and symptoms."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "local_files_only network leakage is unrelated to FP8 save_pretrained behavior."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "The second item is not a real matching bug report; no shared underlying issue."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Liger Kernel application in model_init is unrelated to the DSA indexer ReLU complaint."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:43874",
+            "accept": false,
+            "reason": "FP8 accuracy drift and a missing image-patch helper are different model/image-processing bugs."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "MLflow failed-job reporting does not match a tensor-parallel training bug."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Compiled-file import/runtime errors and save_pretrained behavior are distinct failures."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Compiled Python file errors are unrelated to PretrainedConfig type checking."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "FP8 save_pretrained for MoE and FP8 inference support for Qwen3-VL/Qwen3.5 are separate changes."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "Docs mentioning pipeline() and the unrelated racoon issue do not describe the same bug."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "A no_grad cache/update fix is not the same as double optimizer stepping under deepspeed."
+          },
+          {
+            "left": "issue:33453",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Tokenizer loading regression and Llama 3.2 Vision model loading are different model-load failures."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:43576",
+            "accept": false,
+            "reason": "Missing is_offline_mode import and a broken env command are related to v5 packaging, but not the same concrete defect."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe loading and reloading old checkpoints after upgrade are distinct model-loading issues."
+          },
+          {
+            "left": "issue:33453",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "Tokenizer loading regression is unrelated to tensor-parallel training behavior."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Broken env command and PretrainedConfig type checking are separate regressions."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "A broken env command is not the same as a missing Qwen2_5_VLForConditionalGeneration import."
+          },
+          {
+            "left": "issue:44961",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "Racoon is not a meaningful issue match for Gemma4AudioRelPositionalEncoding."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "Cache no_grad behavior is unrelated to Python-version compatibility."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Documentation confusion around forward args is unrelated to an import error for layer_type_validation."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:42548",
+            "accept": false,
+            "reason": "AutoImageProcessor and PreTrainedModel import failures are different missing symbols."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:43576",
+            "accept": false,
+            "reason": "A missing utility import and a broken env command are not the same bug."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Broken env command and missing HybridCache import are separate import/export issues."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "Scheduler kwargs handling is unrelated to the racoon issue."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43232",
+            "accept": false,
+            "reason": "Both touch generation internals, but they describe different code paths and symptoms."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Tensor-parallel training bug and a docs mismatch for return_all_scores are unrelated."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44355",
+            "accept": false,
+            "reason": "Missing AutoImageProcessor import is not the same as compiled Python file errors."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Tensor-parallel training and BeitImageProcessorFast label reduction are unrelated."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "MLA layernorm precision issue and DSA indexer ReLU omission are different bugs."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Lazy loading failure and unsafe sys.modules access are related at a high level but not the same concrete defect."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "Processor.save_pretrained missing files is specific; the other item is too generic to match."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Lazy loading behavior and an undefined tf backend error are unrelated."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "2D bbox hallucination and infinite-loop generation are different Qwen3-VL symptoms."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Broken env command and checkpoint reload incompatibility are distinct regression reports."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Qwen3-VL output quality issue and PretrainedConfig type checking are unrelated."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe loading and missing _tp_plan for tensor parallelism are not 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": 2,
+      "component_count": 16,
+      "review_unit_index": 44,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:31795",
+        "issue:33453",
+        "issue:34689",
+        "issue:37428",
+        "issue:41669",
+        "issue:42371",
+        "issue:42890",
+        "issue:42913",
+        "issue:43066",
+        "issue:43352",
+        "issue:43408",
+        "issue:43502",
+        "issue:43519",
+        "issue:43576",
+        "issue:43606",
+        "issue:43673",
+        "issue:43704",
+        "issue:43827",
+        "issue:43976",
+        "issue:44077",
+        "issue:44246",
+        "issue:44261",
+        "issue:44263",
+        "issue:44273",
+        "issue:44295",
+        "issue:44297",
+        "issue:44315",
+        "issue:44351",
+        "issue:44360",
+        "issue:44373",
+        "issue:44393",
+        "issue:44485",
+        "issue:44492",
+        "issue:44556",
+        "issue:44623",
+        "issue:44655",
+        "issue:44704",
+        "issue:44829",
+        "issue:44861",
+        "issue:44908",
+        "issue:44961",
+        "issue:45071",
+        "issue:45092",
+        "issue:45230",
+        "issue:45310",
+        "issue:45341",
+        "issue:45468",
+        "issue:45478"
+      ],
+      "soft_pairs": [
+        "issue:41669|issue:43673",
+        "issue:44373|issue:44961",
+        "issue:44556|issue:45092",
+        "issue:44704|issue:44908",
+        "issue:42890|issue:44908",
+        "issue:44246|issue:44273",
+        "issue:43408|issue:44315",
+        "issue:43976|issue:44393",
+        "issue:43352|issue:44393",
+        "issue:44297|issue:45230",
+        "issue:43576|issue:44908",
+        "issue:44295|issue:45230",
+        "issue:44908|issue:45071",
+        "issue:43673|issue:44492",
+        "issue:41669|issue:44351",
+        "issue:41669|issue:45341",
+        "issue:41669|issue:44492",
+        "issue:44393|issue:45478",
+        "issue:44315|issue:44829",
+        "issue:44077|issue:44861",
+        "issue:43827|issue:44908",
+        "issue:31795|issue:33453",
+        "issue:41669|issue:44908",
+        "issue:42371|issue:44492",
+        "issue:44655|issue:44861",
+        "issue:43576|issue:44273",
+        "issue:44360|issue:45468",
+        "issue:45230|issue:45310",
+        "issue:43827|issue:44492",
+        "issue:44273|issue:44908",
+        "issue:43519|issue:44485",
+        "issue:41669|issue:43704",
+        "issue:43519|issue:44908",
+        "issue:31795|issue:34689",
+        "issue:31795|issue:37428",
+        "issue:44315|issue:45092",
+        "issue:43704|issue:44485",
+        "issue:44485|issue:45468",
+        "issue:42371|issue:43704",
+        "issue:43502|issue:43519",
+        "issue:43519|issue:44623",
+        "issue:43519|issue:43704",
+        "issue:44263|issue:45468",
+        "issue:43606|issue:45468",
+        "issue:44261|issue:45468",
+        "issue:43704|issue:45341",
+        "issue:42913|issue:43066"
+      ],
+      "prepared_review_unit_hash": "7389be6608bd87627a5852c867fdfaac9843b305252fb1e347d940798d5c8b2b",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 47,
+        "serialized_chars": 41100,
+        "estimated_input_tokens": 10275,
+        "estimated_eval_tokens": 20806
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is thematically related around Transformers v5 regressions, tokenizer/model loading, and import/lazy-loading behavior, but the soft pairs do not look like true duplicates of the same concrete bug/change. No PRs are present.",
+        "confidence": 0.35,
+        "canonical_issue_reason": "issue:41669 is the broadest and most central issue here: it targets import-time/lazy-loading behavior and has the strongest linkage to several other candidates in the set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:41669 is the best representative issue because it is the clearest hub for the cluster\u2019s import/lazy-loading performance theme and is more central than the narrower model-specific bugs.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41669",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "Import-time/lazy-loading performance vs chunked-prefill cache availability; different bug paths."
+          },
+          {
+            "left": "issue:44373",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "Docstring typo/clarification vs unrelated placeholder issue."
+          },
+          {
+            "left": "issue:44556",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Both are v5 loading regressions, but one is generic checkpoint reload and the other is remote-code/meta-init incompatibility; not the same bug."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Processor kwargs forwarding vs scheduler keyword handling; unrelated subsystems."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Test flakiness from missing seed vs scheduler kwarg bug; no shared code path."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44273",
+            "accept": false,
+            "reason": "Import slowdowns vs lazy loading malfunction; related theme, but not the same concrete defect."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:44315",
+            "accept": false,
+            "reason": "Model-type warning mismatch vs Liger Kernel/model_init application; unrelated."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44393",
+            "accept": false,
+            "reason": "Python version compatibility vs Qwen3-VL bounding-box behavior; unrelated."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:44393",
+            "accept": false,
+            "reason": "Flash Attention support guard vs multimodal bbox output issue; different model paths."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "Tokenizer save metadata bug vs generic bug report; insufficient and unrelated."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "CLI env command breakage vs scheduler kwargs bug; unrelated."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "Position_ids buffer read error vs generic bug report; not the same issue."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Scheduler kwargs ignored vs PretrainedConfig type checking; different code paths."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "Cache missing in generation vs cache-strategy typo; unrelated."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Import-star cleanup/performance vs missing HybridCache export; related import surface but different defects."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Import performance vs testing_utils bug; unrelated."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "Import-time performance vs cache-strategy typo; not duplicate."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Different Qwen3-VL/Qwen3.5 MoE problems and different symptoms."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Liger Kernel model-init integration vs flash-attention training degeneration; unrelated."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Optional post_init policy vs tied-weights-key AttributeError; different implementation issues."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Docs still mention pipeline() vs scheduler kwargs handling; unrelated."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Forward-doc confusion vs tokenizer loading regression; unrelated."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Import/lazy-loading performance vs scheduler kwarg bug; no shared bug."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "TF32 API settings vs cache-strategy typo; unrelated."
+          },
+          {
+            "left": "issue:44655",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Pipeline save_pretrained missing files vs tied-weights-key crash; different save/load paths."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44273",
+            "accept": false,
+            "reason": "env command breakage vs lazy loading problem; only broad v5 theme overlaps."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "DSA indexer behavior vs Gemma-4 audio positional encoding; unrelated model/code path."
+          },
+          {
+            "left": "issue:45230",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Generic bug report vs specific Qwen3.5 MoE from_pretrained error; not enough to be same bug."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "Docs referencing pipeline() vs cache-strategy typo; unrelated."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Lazy loading malfunction vs scheduler kwargs bug; different areas."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Qwen3VL timestamp math vs GLM-5 RoPE implementation; different model-specific logic."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:43704",
+            "accept": false,
+            "reason": "Import performance vs VRAM leak in dataloader threads; unrelated."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Processor timestamp calculation vs scheduler kwargs; unrelated."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Doc confusion vs model loading regression; unrelated."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:37428",
+            "accept": false,
+            "reason": "Doc confusion vs flash-attention import error; unrelated."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Liger Kernel/model_init integration vs v5 checkpoint incompatibility; different bugs."
+          },
+          {
+            "left": "issue:43704",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "VRAM leak vs RoPE implementation; unrelated."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "Different model implementation bugs with no shared code path."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "MLA q_a_layernorm epsilon precision issue vs Gemma-4 audio positional encoding; unrelated."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "torch.split return semantics in DSA indexer vs Gemma-4 audio encoding; unrelated."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "CPU offload device mismatch vs Gemma-4 audio positional encoding; unrelated."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "Same as above: not the same concrete defect."
+          }
+        ]
+      },
+      "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": 16,
+      "review_unit_index": 45,
+      "review_unit_count": 45,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-15354-388",
+      "nodes": [
+        "issue:30333",
+        "issue:31515",
+        "issue:31795",
+        "issue:33453",
+        "issue:36296",
+        "issue:38617",
+        "issue:41669",
+        "issue:43066",
+        "issue:43298",
+        "issue:43525",
+        "issue:43572",
+        "issue:43704",
+        "issue:44230",
+        "issue:44360",
+        "issue:44623",
+        "issue:44861",
+        "issue:44908",
+        "issue:44998",
+        "issue:45290",
+        "issue:45310"
+      ],
+      "soft_pairs": [
+        "issue:44908|issue:44998",
+        "issue:44861|issue:45290",
+        "issue:43525|issue:44861",
+        "issue:43704|issue:44998",
+        "issue:41669|issue:44998",
+        "issue:43572|issue:44861",
+        "issue:43066|issue:44998",
+        "issue:31515|issue:38617",
+        "issue:44360|issue:44998",
+        "issue:44230|issue:44623",
+        "issue:30333|issue:31515",
+        "issue:31515|issue:36296",
+        "issue:31515|issue:31795",
+        "issue:44230|issue:45310",
+        "issue:31515|issue:33453",
+        "issue:43298|issue:44230"
+      ],
+      "prepared_review_unit_hash": "cc0023c3561622a78a0f644ea6f469f4a40889b14a7d1772197afb953edf98e8",
+      "component_budget": {
+        "node_count": 388,
+        "item_count": 388,
+        "soft_pair_count": 1459,
+        "serialized_chars": 751559,
+        "estimated_input_tokens": 187890,
+        "estimated_eval_tokens": 376036
+      },
+      "budget": {
+        "node_count": 20,
+        "item_count": 20,
+        "soft_pair_count": 16,
+        "serialized_chars": 16034,
+        "estimated_input_tokens": 4009,
+        "estimated_eval_tokens": 8274
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These issues are mostly unrelated standalone reports; none of the proposed soft edges look like true duplicates. The clearest representative issue is the tokenizer-loading regression report.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "issue:33453 is a concrete, high-activity regression report with a well-scoped failure mode, making it the best issue representative.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:33453 has the strongest signal of a real, actionable bug and the broadest representative value among the set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44908",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Inverse-sqrt scheduler kwargs bug is unrelated to the unrelated 'Unemployment' issue title and body."
+          },
+          {
+            "left": "issue:44861",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Different code paths: tied-weight key handling versus chat template tool-call serialization."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Config attribute missing versus tied-weight key crash; not the same underlying defect."
+          },
+          {
+            "left": "issue:43704",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "VRAM leak in Qwen3ForCausalLM is unrelated to the unrelated 'Unemployment' issue."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Import-star performance cleanup is unrelated to the unrelated 'Unemployment' issue."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Missing pad_token_idx in StableLmConfig is a different config bug from _get_tied_weight_keys crashing."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Wrong tokenizer decoder type is unrelated to the unrelated 'Unemployment' issue."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Slow checkpoint loading and ImportError for layer_type_validation are different failure modes and code paths."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "DSA indexer/ReLU discussion is unrelated to the unrelated 'Unemployment' issue."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "FP8 model support and processor.save_pretrained missing files are different problems."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:31515",
+            "accept": false,
+            "reason": "MLFlow job-status reporting and slow checkpoint loading are unrelated bugs."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "Checkpoint loading slowness is unrelated to tensor parallel training behavior."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:31795",
+            "accept": false,
+            "reason": "Loading speed issue and documentation confusion are different issues."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Both mention Qwen3.5 MoE, but one is FP8 inference support and the other is a from_pretrained error; not the same concrete bug."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Slow from_pretrained checkpoint loading is not the same as a tokenizer-loading regression."
+          },
+          {
+            "left": "issue:43298",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "Continuous batching audio-model support and FP8 Qwen3-VL/MoE support are distinct feature gaps."
+          }
+        ]
+      },
+      "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": 16,
+      "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": "f20b8e9877cd8ae04064e9d87ee43f27a13c4cd1899c9bc5220c8c0c86f895cf",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1589,
+        "estimated_input_tokens": 398,
+        "estimated_eval_tokens": 1052
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1589,
+        "estimated_input_tokens": 398,
+        "estimated_eval_tokens": 1052
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are both broad community-contribution tracker issues, but they target different features and code paths: Flash Attention 2 support vs GGUF support. They are not duplicates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "No single canonical issue: the two issues track different architecture-support efforts with different implementation work.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Neither issue is a better canonical duplicate of the other; both are separate umbrella requests with distinct scope.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:26350",
+            "right": "issue:33260",
+            "accept": false,
+            "reason": "Same general phrasing, but the underlying requests are different (Flash Attention 2 support vs GGUF support) and would not be merged into one issue/PR."
+          }
+        ]
+      },
+      "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": 16,
+      "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-27088-4",
+      "nodes": [
+        "issue:27088",
+        "issue:43834",
+        "issue:44365",
+        "issue:44370"
+      ],
+      "soft_pairs": [
+        "issue:27088|issue:44370",
+        "issue:27088|issue:44365",
+        "issue:27088|issue:43834",
+        "issue:43834|issue:44370",
+        "issue:43834|issue:44365",
+        "issue:44365|issue:44370"
+      ],
+      "prepared_review_unit_hash": "3797dcb49cf83fa760cabfe063568eafc58f5560eaf607756a0aae5c97e0b70d",
+      "component_budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4454,
+        "estimated_input_tokens": 1114,
+        "estimated_eval_tokens": 2484
+      },
+      "budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4454,
+        "estimated_input_tokens": 1114,
+        "estimated_eval_tokens": 2484
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are not duplicates: they are separate docs-translation requests for different language targets, and several titles are just placeholder variants of the same template. The only concrete request is the Turkish one, so it is the best representative issue if a canonical item is needed.",
+        "confidence": 0.95,
+        "canonical_issue_reason": "Issue 27088 is the only concrete, language-specific request ([i18n-TR] Turkish) and is more representative than the placeholder-language template issues.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 27088 is the most actionable and specific item in the cluster; the others are generic placeholder variants rather than distinct, mergeable duplicates.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:27088",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Different translation targets: Turkish vs a placeholder language template. Same topic area, but not the same underlying issue."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Different translation targets: Turkish vs a placeholder language template. Not a duplicate request."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:43834",
+            "accept": false,
+            "reason": "Different translation targets: Turkish vs a placeholder language template. Same theme, different issue."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Both are placeholder/template-style translation issues, but they are not the same concrete language request."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Template-like translation issues for different language values; not the same underlying task."
+          },
+          {
+            "left": "issue:44365",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Template-like translation issues for different language values; 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": 5,
+      "component_count": 16,
+      "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-36817-14",
+      "nodes": [
+        "issue:36817",
+        "issue:37015",
+        "issue:41115",
+        "issue:41211",
+        "issue:42503",
+        "issue:42738",
+        "issue:43671",
+        "issue:43837",
+        "issue:43909",
+        "issue:44121",
+        "issue:44205",
+        "issue:45027",
+        "issue:45177",
+        "issue:45322"
+      ],
+      "soft_pairs": [
+        "issue:36817|issue:41211",
+        "issue:36817|issue:43671",
+        "issue:37015|issue:43837",
+        "issue:37015|issue:44121",
+        "issue:37015|issue:43909",
+        "issue:37015|issue:42503",
+        "issue:37015|issue:45027",
+        "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:42738",
+        "issue:42738|issue:45027",
+        "issue:43671|issue:43837",
+        "issue:43671|issue:45027",
+        "issue:43671|issue:45177",
+        "issue:43671|issue:44121",
+        "issue:43837|issue:45322",
+        "issue:43837|issue:43909",
+        "issue:43909|issue:44205",
+        "issue:43671|issue:43909",
+        "issue:44121|issue:45322",
+        "issue:43837|issue:44121",
+        "issue:43837|issue:45027",
+        "issue:43837|issue:45177"
+      ],
+      "prepared_review_unit_hash": "8ed9e51206fc7f368e6d45558aaf38da253fc8d223df39563d6890c834705fe3",
+      "component_budget": {
+        "node_count": 14,
+        "item_count": 14,
+        "soft_pair_count": 28,
+        "serialized_chars": 17650,
+        "estimated_input_tokens": 4413,
+        "estimated_eval_tokens": 9082
+      },
+      "budget": {
+        "node_count": 14,
+        "item_count": 14,
+        "soft_pair_count": 28,
+        "serialized_chars": 17650,
+        "estimated_input_tokens": 4413,
+        "estimated_eval_tokens": 9082
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster is a heterogeneous set of model-request issues across unrelated model families and modalities (BERT variants, audio/ASR/TTS, OCR, vision encoders, sparse transformers). The soft similarities are mostly superficial naming overlap, not duplicate bug/change reports.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:36817",
+            "right": "issue:41211",
+            "accept": false,
+            "reason": "EuroBert config support and DEIMv2 are different model requests with no shared underlying change."
+          },
+          {
+            "left": "issue:36817",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "EuroBert configuration support and Qwen3-TTS support target different model families and modalities."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "NeoBERT and Qwen3-ASR are unrelated model additions: text encoder vs speech recognition."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "NeoBERT and the weight-sparse transformer request concern different architectures and code paths."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "NeoBERT and LFM2.5 Audio are distinct model additions, not the same implementation work."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:42503",
+            "accept": false,
+            "reason": "NeoBERT and ModernVBERT are separate model families; no common duplicate bug/change."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "NeoBERT and Voxtral-4B-TTS are different model requests in different audio/text paths."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "MiniCPM3 and Qwen3-ASR are different model additions with different interfaces."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "MiniCPM3 support and the sparse-transformer request are unrelated model-architecture additions."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:42503",
+            "accept": false,
+            "reason": "MiniCPM3 and ModernVBERT are distinct models; the requests do not describe the same change."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "MiniCPM3 and Voxtral-4B-TTS are separate model requests, not one duplicate."
+          },
+          {
+            "left": "issue:41211",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "DEIMv2 and Qwen3-TTS are different models and likely different code surfaces."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "ModernVBERT and Qwen3-ASR are unrelated requests across different modalities."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "ModernVBERT and Qwen3-TTS are separate model additions with no shared underlying fix."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:42738",
+            "accept": false,
+            "reason": "ModernVBERT and BERT-like models with RoPE overlap only broadly on BERT naming, not on the same concrete implementation."
+          },
+          {
+            "left": "issue:42738",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "BERT-like models with RoPE and Voxtral-4B-TTS are unrelated model families and behaviors."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "Qwen3-TTS and Qwen3-ASR are different speech tasks; not the same model support."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "Qwen3-TTS and Voxtral-4B-TTS are different TTS model requests."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:45177",
+            "accept": false,
+            "reason": "Qwen3-TTS and DeepSeek-OCR2 target different modalities and code paths."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "Qwen3-TTS support and the sparse-transformer request are not the same underlying change."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45322",
+            "accept": false,
+            "reason": "Qwen3-ASR and EUPE are unrelated speech-recognition vs encoder-model requests."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "Qwen3-ASR and LFM2.5 Audio are different audio models with different support needs."
+          },
+          {
+            "left": "issue:43909",
+            "right": "issue:44205",
+            "accept": false,
+            "reason": "LFM2.5 Audio and SAM3-LiteText are distinct model additions in different domains."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "Qwen3-TTS and LFM2.5 Audio are separate speech-model requests, not duplicates."
+          },
+          {
+            "left": "issue:44121",
+            "right": "issue:45322",
+            "accept": false,
+            "reason": "The sparse-transformer request and EUPE request are different architectures and changes."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "Qwen3-ASR and the sparse-transformer request do not share the same concrete implementation problem."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "Qwen3-ASR and Voxtral-4B-TTS are different speech tasks and model families."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45177",
+            "accept": false,
+            "reason": "Qwen3-ASR and DeepSeek-OCR2 are unrelated speech vs OCR model requests."
+          }
+        ]
+      },
+      "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": 16,
+      "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-37934-6",
+      "nodes": [
+        "issue:37934",
+        "issue:42936",
+        "issue:43240",
+        "issue:43668",
+        "issue:43813",
+        "issue:44031"
+      ],
+      "soft_pairs": [
+        "issue:37934|issue:43668",
+        "issue:42936|issue:43240",
+        "issue:42936|issue:44031",
+        "issue:42936|issue:43813",
+        "issue:42936|issue:43668",
+        "issue:43240|issue:43668",
+        "issue:43668|issue:44031",
+        "issue:43668|issue:43813"
+      ],
+      "prepared_review_unit_hash": "ea7f69455ee3cbb178a470b5c47e36308cda7338aec09ebe34489c434e21706a",
+      "component_budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 8,
+        "serialized_chars": 6290,
+        "estimated_input_tokens": 1573,
+        "estimated_eval_tokens": 3402
+      },
+      "budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 8,
+        "serialized_chars": 6290,
+        "estimated_input_tokens": 1573,
+        "estimated_eval_tokens": 3402
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These items are not duplicates of one another; they span unrelated bugs in different models/configs/utilities, with only superficial textual similarity. No PRs are present.",
+        "confidence": 0.95,
+        "canonical_issue_reason": "There is no true canonical issue for this cluster because the reports concern different subsystems and different failure modes. If a representative must be chosen, issue 43668 is the cleanest self-contained bug report.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43668 is the most precise and narrowly scoped of the set: a specific type-hint bug in ModernBERTConfig with a clear, isolated fix path.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:37934",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Different concerns: Llama4TextL2Norm normalization semantics vs. ModernBERTConfig type annotation. No shared code-path bug."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43240",
+            "accept": false,
+            "reason": "Mask2Former ignore_value handling and missing kwargs propagation are unrelated model-loss issues."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:44031",
+            "accept": false,
+            "reason": "A Mask2Former loss-parameter bug is unrelated to tokenizer regex warning behavior."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43813",
+            "accept": false,
+            "reason": "Different modules and bug types: model loss parameter usage vs. a typo in quantization naming."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "One is a model runtime bug; the other is a config type-hint mistake. Not the same underlying defect."
+          },
+          {
+            "left": "issue:43240",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Passing kwargs to a loss function and correcting a config annotation are unrelated changes."
+          },
+          {
+            "left": "issue:43668",
+            "right": "issue:44031",
+            "accept": false,
+            "reason": "ModernBERTConfig type hint issue does not overlap with tokenizer regex warning regression."
+          },
+          {
+            "left": "issue:43668",
+            "right": "issue:43813",
+            "accept": false,
+            "reason": "A type-hint typo and a quantization attribute typo are different files/features and not one fix."
+          }
+        ]
+      },
+      "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": 16,
+      "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": "3bc951fc9ad5cfb3a17180bdd4eacd1aa04fe738e49514f1fb17f0117826f30b",
+      "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 do not look like duplicates: one is about block decoding mechanics, while the other is a request for DeepSeek v3.2 model support. They may both touch model/inference functionality, but they describe different underlying changes.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43037 is the more specific, user-facing request and is better defined as a standalone tracking issue; issue 41084 is broader and less clearly tied to the DeepSeek support request.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41084",
+            "right": "issue:43037",
+            "accept": false,
+            "reason": "Different underlying requests: block decoding implementation vs adding DeepSeek v3.2 support. Shared broad inference context is not enough to treat them as duplicates."
+          }
+        ]
+      },
+      "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": 16,
+      "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": "8f35b320ccc397c40a666b4c6dfd01c84821345b2cb1b400fb42911d87c6f105",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1583,
+        "estimated_input_tokens": 396,
+        "estimated_eval_tokens": 1048
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1583,
+        "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 to Trainer customization, but they describe different problems: one is about `remove_unused_column` behavior, the other about estimating FLOPS for custom trainers. I do not consider them duplicates.",
+        "confidence": 0.92,
+        "canonical_issue_reason": "Issue 43598 is the clearer representative of the cluster because it focuses on a concrete Trainer customization limitation, while 43600 is a separate FLOPS-estimation concern.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43598 is the best single issue to keep if one must be chosen, since it has the more direct and actionable customization request.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43598",
+            "right": "issue:43600",
+            "accept": false,
+            "reason": "They involve different code paths and outcomes: `remove_unused_column` customization versus FLOPS estimation for custom trainers. Shared Trainer context is too broad to treat as 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": 9,
+      "component_count": 16,
+      "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": "f75b90561a47ccfed928cff0a69d05d4ef78538a3397584fade29e8487ff9bf0",
+      "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": "The two issues are about different Qwen model families and different AutoModel capabilities, so they do not look like duplicates.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "Issue #43684 is the more general registration request and is the best representative of the cluster, but it is not the same underlying task as #44405.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#43684 is the clearest standalone issue in the set; #44405 targets sequence classification support for a different config/model family.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43684",
+            "right": "issue:44405",
+            "accept": false,
+            "reason": "Different model families (Qwen3-Omni vs Qwen3.5) and different API targets (AutoModel/AutoModelForConditionalGeneration vs AutoModelForSequenceClassification); 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": 10,
+      "component_count": 16,
+      "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": "166e7e4b6a44f738df76e34f8b91d124a4286a158959e69239aa02e89bfcd6d8",
+      "component_budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4263,
+        "estimated_input_tokens": 1066,
+        "estimated_eval_tokens": 2388
+      },
+      "budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4263,
+        "estimated_input_tokens": 1066,
+        "estimated_eval_tokens": 2388
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Three issues share the same near-empty title/body pattern around \u201chuggingface \u201d and are likely duplicates of the same report; \u201cFett\u201d looks unrelated. No pull requests are present.",
+        "confidence": 0.83,
+        "canonical_issue_reason": "Issue 43845 is the earliest of the repeated \u201chuggingface \u201d reports and is the best representative of that duplicate set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43845 is the strongest global representative because it matches the other two \u201chuggingface \u201d issues closely and was created first.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43835",
+            "right": "issue:43845",
+            "accept": false,
+            "reason": "Different titles and no evidence they describe the same underlying problem."
+          },
+          {
+            "left": "issue:43835",
+            "right": "issue:43846",
+            "accept": false,
+            "reason": "Different titles and no concrete signal of the same bug or request."
+          },
+          {
+            "left": "issue:43835",
+            "right": "issue:43859",
+            "accept": false,
+            "reason": "Different title and no clear shared underlying issue."
+          },
+          {
+            "left": "issue:43845",
+            "right": "issue:43846",
+            "accept": true,
+            "reason": "Same title pattern and very similar sparse issue content; likely duplicate reports."
+          },
+          {
+            "left": "issue:43845",
+            "right": "issue:43859",
+            "accept": true,
+            "reason": "Same near-identical title/body pattern; likely the same underlying issue."
+          },
+          {
+            "left": "issue:43846",
+            "right": "issue:43859",
+            "accept": true,
+            "reason": "All three \u201chuggingface \u201d issues appear to be the same minimal report, so this pair is likely duplicate."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: the three 'huggingface ' issues have identical titles and similar sparse metadata, while 'Fett' is distinct by title. The soft-edge verdicts are conservative overall, and choosing 43845 as the earliest representative is supported by the timestamps. No PRs are present."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 11,
+      "component_count": 16,
+      "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-44371-5",
+      "nodes": [
+        "issue:44371",
+        "issue:44418",
+        "issue:44572",
+        "issue:44573",
+        "issue:44937"
+      ],
+      "soft_pairs": [
+        "issue:44371|issue:44572",
+        "issue:44371|issue:44573",
+        "issue:44371|issue:44418",
+        "issue:44371|issue:44937",
+        "issue:44418|issue:44572",
+        "issue:44418|issue:44573",
+        "issue:44572|issue:44573",
+        "issue:44572|issue:44937",
+        "issue:44573|issue:44937"
+      ],
+      "prepared_review_unit_hash": "b570d69da9e7946ac886837edc223689ce4490ff45d3ab5c12f89178e1fa0fae",
+      "component_budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 9,
+        "serialized_chars": 5791,
+        "estimated_input_tokens": 1448,
+        "estimated_eval_tokens": 3152
+      },
+      "budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 9,
+        "serialized_chars": 5791,
+        "estimated_input_tokens": 1448,
+        "estimated_eval_tokens": 3152
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are all low-information/spam-like closed issues with placeholder or unrelated titles; there is no evidence they describe the same underlying bug or change. The cluster is best treated as noise rather than a duplicate set.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "issue:44371 is the cleanest canonical placeholder because it is the earliest obvious spam/placeholder report and is representative of the cluster's low-signal content.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44371 is the best issue to anchor the cluster if one must be chosen, because it is an archetypal spam placeholder and the earliest of the identical <spam> entries.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44371",
+            "right": "issue:44572",
+            "accept": false,
+            "reason": "Both are generic spam placeholders, but there is no concrete shared bug or change to indicate duplication."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "Same placeholder title class, but no substantive description or target makes them distinct noisy issues rather than duplicates."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44418",
+            "accept": false,
+            "reason": "One is a spam placeholder and the other is a generic documentation-suggestion title; they do not appear to be the same issue."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "The titles are unrelated and there is no supporting detail tying them to one underlying report."
+          },
+          {
+            "left": "issue:44418",
+            "right": "issue:44572",
+            "accept": false,
+            "reason": "A documentation suggestion is not the same as a spam placeholder; no shared problem statement is evident."
+          },
+          {
+            "left": "issue:44418",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "The titles and apparent intent differ, with no concrete overlap beyond low-signal content."
+          },
+          {
+            "left": "issue:44572",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "Both are spam placeholders, but there is still no evidence of one underlying bug or change; they look like separate noise reports."
+          },
+          {
+            "left": "issue:44572",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "A spam placeholder and a specific quoted-name title are unrelated without any shared technical context."
+          },
+          {
+            "left": "issue:44573",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "No common concrete issue is visible; the reports are different low-information entries."
+          }
+        ]
+      },
+      "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": 16,
+      "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": "e38b9e28c62f4279f508c9e5b35880186a814df554adb59a8b2f31380118142e",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1645,
+        "estimated_input_tokens": 412,
+        "estimated_eval_tokens": 1080
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1645,
+        "estimated_input_tokens": 412,
+        "estimated_eval_tokens": 1080
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two closed issues are likely duplicates: they have the same title, very similar body lengths, and #45145 is earlier with one inbound reference. The evidence is suggestive rather than definitive because the bodies are not included and the only pair signal is soft similarity.",
+        "confidence": 0.77,
+        "canonical_issue_reason": "#45145 is the better provisional canonical issue because it was created earlier and already has an inbound reference.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#45145 is the best issue candidate on the available evidence: it appears to be the earlier copy of the same report and has slightly more traceability.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45145",
+            "right": "issue:45151",
+            "accept": true,
+            "reason": "Accept as a likely duplicate issue pair: the titles are identical, the body lengths are very close, and there is no evidence of different targets or scope. The evidence is not definitive, so this is a conservative duplicate call."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The summary is a bit overconfident for the evidence in the packet. What is grounded: the two issues have the same title, similar body lengths, and #45145 is earlier with one inbound reference. What is not fully supported: claims like \u201cessentially the same report content,\u201d \u201cno differentiating targets or filenames,\u201d and a 0.98 confidence level, since the packet does not include the bodies and the pair only has soft_similarity evidence with deterministic_accept=false. A more conservative conclusion would be that they are likely duplicates, with #45145 as the provisional canonical issue."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 13,
+      "component_count": 16,
+      "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": 14,
+      "component_count": 16,
+      "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-45672-2",
+      "nodes": [
+        "issue:45672",
+        "issue:45674"
+      ],
+      "soft_pairs": [
+        "issue:45672|issue:45674"
+      ],
+      "prepared_review_unit_hash": "3c0c4627d22a3787327662766fb0a577b5be701bc471675d4ce3bb41afecf085",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1668,
+        "estimated_input_tokens": 417,
+        "estimated_eval_tokens": 1090
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1668,
+        "estimated_input_tokens": 417,
+        "estimated_eval_tokens": 1090
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are two distinct 4-bit inference bugs: one specific to Gemma4 audio modules hitting a torch.finfo() TypeError on uint8 weights, and one in BitsAndBytesConfig where llm_int8_skip_modules removes the default lm_head exclusion and triggers an AssertionError. They are related only at a broad subsystem level, not the same underlying defect.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "issue #45674 is the better representative because it describes a clearer, more general configuration regression with a concrete failing path and broader reuse across 4-bit inference setups.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue #45674 is the strongest standalone report: it has a precise root-cause description (default lm_head exclusion being cleared) and a direct, actionable failure mode.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45672",
+            "right": "issue:45674",
+            "accept": false,
+            "reason": "Both concern 4-bit inference, but the failure mechanisms differ: Gemma4 audio uint8/torch.finfo TypeError vs. BitsAndBytesConfig lm_head exclusion AssertionError. 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": 15,
+      "component_count": 16,
+      "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": "59562361351efadc15cb4488a7f61f51d51af8c1bec75eb1bb9f952af88819b2",
+      "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": "The two closed issues are just person-name/date entries with no body or links and no evidence of a shared bug/report. They do not look like duplicates or related issue reports.",
+        "confidence": 0.96,
+        "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": "Different names and birth dates, no description, no shared targets, and no issue-specific evidence of 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": 16,
+      "component_count": 16,
+      "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": "5a8e2b645ef367042a7b0468c5866b26d779b7bd4e5890377c064b0c3269e18d",
+      "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": "The two issues are not duplicates: they only share a surname and date-like text, with no body, context, or shared problem statement. They should remain separate.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "No canonical issue: the items do not describe the same underlying bug or request.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Neither issue is a good canonical choice; each appears to be a distinct, unrelated ticket.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45764",
+            "right": "issue:45765",
+            "accept": false,
+            "reason": "Same surname is not enough to infer duplication; the titles refer to different people and different dates, with no shared problem context."
+          }
+        ]
+      },
+      "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": 26,
+      "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-39895-13",
+      "nodes": [
+        "pull_request:39895",
+        "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:45795"
+      ],
+      "soft_pairs": [
+        "pull_request:39895|pull_request:43424",
+        "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:45795",
+        "pull_request:44827|pull_request:45655",
+        "pull_request:44827|pull_request:45213",
+        "pull_request:43823|pull_request:45213",
+        "pull_request:43823|pull_request:45795",
+        "pull_request:43251|pull_request:43254"
+      ],
+      "prepared_review_unit_hash": "bfa0cdeac41d327395bac30a546e9bab851b3bf07c915fdb65ed2f1efd942968",
+      "component_budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 19,
+        "serialized_chars": 35198,
+        "estimated_input_tokens": 8800,
+        "estimated_eval_tokens": 17856
+      },
+      "budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 19,
+        "serialized_chars": 35198,
+        "estimated_input_tokens": 8800,
+        "estimated_eval_tokens": 17856
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The cluster is mostly heterogeneous and not a true duplicate set. The only clear duplicate pair is the two loss-utils PRs that both extend `fixed_cross_entropy` to pass supported `cross_entropy` kwargs; everything else is an unrelated model addition, docs change, refactor, or test-only PR.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "If a canonical PR must be chosen, `pull_request:43251` is the best representative of the only real duplicate pair because it has the explicit issue target (`43240`) and a fuller implementation of the same `fixed_cross_entropy` kwargs fix.",
+        "best_issue_reason": null,
+        "best_pr_reason": "`pull_request:43251` is the strongest PR in the cluster to anchor on: it directly addresses the tracked loss bug, targets the same code path as `43254`, and is the only clear duplicate-worthy change here.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:39895",
+            "right": "pull_request:43424",
+            "accept": false,
+            "reason": "Unrelated changes: VideoPrism model integration vs. an ExecuTorch dynamic-shapes test. Same repository area, but not the same bug or change."
+          },
+          {
+            "left": "pull_request:41855",
+            "right": "pull_request:43424",
+            "accept": false,
+            "reason": "Tokenizer method additions are unrelated to ExecuTorch export tests; no shared underlying defect."
+          },
+          {
+            "left": "pull_request:41855",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Mistral tokenizer missing methods and adding MobileLLM are separate model/tokenizer workstreams."
+          },
+          {
+            "left": "pull_request:42668",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Processor auto-loading robustness changes are unrelated to adding a new MobileLLM model."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Loss-function kwargs fix is unrelated to introducing MobileLLM model files and tests."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "`fixed_cross_entropy` kwargs support and Mistral4 test fixes are different code paths and different bugs."
+          },
+          {
+            "left": "pull_request:43424",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "ExecuTorch dynamic-shape export test and Mistral4 fixes do not address the same underlying change."
+          },
+          {
+            "left": "pull_request:43424",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Dynamic-shape export testing is unrelated to MobileLLM model addition."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:45655",
+            "accept": false,
+            "reason": "New model support and a `cls.config` resolution-order fix are unrelated."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:43995",
+            "accept": false,
+            "reason": "MobileLLM model addition and Falcon output-interface refactor are distinct changes."
+          },
+          {
+            "left": "pull_request:43995",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "Falcon refactor and the model-creation skill draft are not the same concrete bug/change."
+          },
+          {
+            "left": "pull_request:43995",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "Falcon standardized-output refactor is unrelated to Mistral4 test fixes."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45170",
+            "accept": false,
+            "reason": "Mistral4 test fixes and `layrnorm` renaming across CLIP-like models are separate changes."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45795",
+            "accept": false,
+            "reason": "Mistral4 test fixes are unrelated to docs for audio/video processors."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45655",
+            "accept": false,
+            "reason": "Mistral4 test work and `cls.config` resolution-order fix are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "A Mistral4 test fix and the model-creation skill draft do not represent the same underlying change."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "Adding MobileLLM and adding the model-creation skill are different concerns; one is product code, the other is tooling/docs."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:45795",
+            "accept": false,
+            "reason": "MobileLLM model implementation and processor docs are unrelated."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:43254",
+            "accept": true,
+            "reason": "Both PRs make the same `fixed_cross_entropy` change in `loss_utils.py` to pass supported kwargs (`weight`/`label_smoothing`) to `nn.functional.cross_entropy`; they fix the same concrete bug and could be merged into one."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative: it identifies the only clearly mergeable duplicate pair (43251/43254) based on the shared loss_utils diff and shared issue target, while treating the rest as heterogeneous. The soft-edge rejections are also cautious and consistent with the provided diffs. No overstatement detected."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 2,
+      "component_count": 26,
+      "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-40546-2",
+      "nodes": [
+        "pull_request:40546",
+        "pull_request:42781"
+      ],
+      "soft_pairs": [
+        "pull_request:40546|pull_request:42781"
+      ],
+      "prepared_review_unit_hash": "14418f4f708943ab930c8f0f2cc658605cfd7402cfa03812e2cfaffbb3cb0c3c",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 7225,
+        "estimated_input_tokens": 1807,
+        "estimated_eval_tokens": 3870
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 7225,
+        "estimated_input_tokens": 1807,
+        "estimated_eval_tokens": 3870
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Related VibeVoice PRs, but not duplicates: 40546 is the initial VibeVoice implementation/docs integration, while 42781 expands into VibeVoice Realtime with the acoustic tokenizer and additional generation/model changes.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 40546 is the foundational VibeVoice integration and the clearest representative of the base feature work; it is narrower and more complete as the original model addition.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 40546 is the best single representative of the cluster because it covers the core VibeVoice import and registration work, while 42781 is a later follow-on for realtime/tokenizer functionality.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:40546",
+            "right": "pull_request:42781",
+            "accept": false,
+            "reason": "They share the VibeVoice namespace and some files, but the scope differs: 40546 adds the base model/docs/auto-mapping integration, whereas 42781 adds realtime support plus an acoustic tokenizer and related generation/path updates. These are adjacent feature PRs, 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": 3,
+      "component_count": 26,
+      "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": "9ac404caba0955a15011b005c361a639100d06174e09f08a930691f24a2e7d1b",
+      "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": "These two PRs are unrelated feature additions in different model families: TimesFM covariate forecasting vs DINOv3 image classification support. The soft similarity is only superficial.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #40755 is the better cluster representative because it is the broader, more substantial functional change with core implementation across model code, utils, tests, and docs.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #40755 is the strongest standalone PR in the set: it introduces a new forecasting capability with implementation work rather than only model/classification enablement and example updates.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:40755",
+            "right": "pull_request:41224",
+            "accept": false,
+            "reason": "Different subsystems and problems: TimesFM time-series covariate forecasting vs DINOv3 image classification support. No shared filenames, targets, or code path, so they are 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": 4,
+      "component_count": 26,
+      "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": 5,
+      "component_count": 26,
+      "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-41797-3",
+      "nodes": [
+        "pull_request:41797",
+        "pull_request:45186",
+        "pull_request:45630"
+      ],
+      "soft_pairs": [
+        "pull_request:41797|pull_request:45186",
+        "pull_request:45186|pull_request:45630"
+      ],
+      "prepared_review_unit_hash": "1abe801664b2f08cf6713e8f215b32b6e7a9eacac07616c57599545bd27281f2",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 10426,
+        "estimated_input_tokens": 2607,
+        "estimated_eval_tokens": 5470
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 10426,
+        "estimated_input_tokens": 2607,
+        "estimated_eval_tokens": 5470
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Three unrelated PRs adding support/docs for different models (DeepSeek OCR, Isaac, and Kimi2-6). They share template-like integration files, but each targets a distinct model family and change set, so they are not 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:41797",
+            "right": "pull_request:45186",
+            "accept": false,
+            "reason": "Different model families and codepaths: DeepSeek OCR vs Isaac. Shared framework integration files are generic onboarding scaffolding, not the same fix/change."
+          },
+          {
+            "left": "pull_request:45186",
+            "right": "pull_request:45630",
+            "accept": false,
+            "reason": "Different model additions (Isaac vs Kimi2-6) with distinct filenames and model implementations. Similar auto-registration/docs updates are too broad to treat 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": 6,
+      "component_count": 26,
+      "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": "61cc767ab4c9064a76539b77498a9d44cb60ab97e9d72a2b357471f2df5633c2",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5935,
+        "estimated_input_tokens": 1484,
+        "estimated_eval_tokens": 3224
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5935,
+        "estimated_input_tokens": 1484,
+        "estimated_eval_tokens": 3224
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both pull requests are GitHub Actions maintenance upgrades, but they target different actions and different underlying changes: one is mostly checkout/cache/upload-artifact for Node 24 compatibility, while the other bumps docker/build-push-action. They overlap in workflow files but are not the same fix and should not be merged as duplicates.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No canonical duplicate PR. If one must be chosen as the broader representative, 42975 is more umbrella-like because it updates several core workflow actions for Node 24 compatibility, whereas 42976 is a narrower docker/build-push-action version bump.",
+        "best_issue_reason": null,
+        "best_pr_reason": "42975 is the better representative of the cluster because it spans more workflow actions and is centered on a specific compatibility goal. 42976 is related maintenance, but it is a separate change focused on docker/build-push-action.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:42975",
+            "right": "pull_request:42976",
+            "accept": false,
+            "reason": "Both are workflow action version bumps, but they change different third-party actions and address different maintenance concerns. They do not appear to fix the same concrete code-path problem or be safely mergeable into one PR 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": 7,
+      "component_count": 26,
+      "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:43451",
+        "pull_request:45490"
+      ],
+      "soft_pairs": [
+        "pull_request:42978|pull_request:43448",
+        "pull_request:42978|pull_request:43451",
+        "pull_request:42978|pull_request:45490",
+        "pull_request:43448|pull_request:43451",
+        "pull_request:43448|pull_request:45490",
+        "pull_request:43451|pull_request:45490"
+      ],
+      "prepared_review_unit_hash": "8085010eb1f29e9977e50251b11ec2d580297cf79d80faf93ba2923ac66877b5",
+      "component_budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 14079,
+        "estimated_input_tokens": 3520,
+        "estimated_eval_tokens": 7296
+      },
+      "budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 14079,
+        "estimated_input_tokens": 3520,
+        "estimated_eval_tokens": 7296
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "All four pull requests are separate model additions (ViT NEPA, Molmo, Molmo2, CTSM) and do not fix the same bug or implement the same change. The overlap is limited to common docs/auto-registration plumbing, which is not enough to treat them as duplicates.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No true duplicate PR exists in this cluster; if a representative must be chosen, 43451 is the most substantial and most discussed PR, but it addresses a distinct model add from the others.",
+        "best_issue_reason": null,
+        "best_pr_reason": "43451 is the strongest representative by scope and review activity, but it should not be considered a duplicate of the others because it adds Molmo2 rather than fixing the same concrete code path as any other PR.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:43448",
+            "accept": false,
+            "reason": "Different model additions: ViT NEPA vs Molmo. Shared auto/docs files are generic registration changes, not the same underlying feature or bug fix."
+          },
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:43451",
+            "accept": false,
+            "reason": "Different model additions with different architectures and files. Both touch auto-mappings, but they implement separate model integrations, not one merged change."
+          },
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "ViT NEPA and CTSM are unrelated model integrations. The common toctree/auto-model edits are incidental and do not indicate the same change."
+          },
+          {
+            "left": "pull_request:43448",
+            "right": "pull_request:43451",
+            "accept": false,
+            "reason": "Molmo and Molmo2 are related by naming, but they are distinct model families with different code paths, tests, and docs; not the same concrete change."
+          },
+          {
+            "left": "pull_request:43448",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "Molmo and CTSM are unrelated additions. Shared repo plumbing does not make them duplicates."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "Molmo2 and CTSM are separate model implementations. They overlap only in generic registration/docs updates, so they 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": 8,
+      "component_count": 26,
+      "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-43291-3",
+      "nodes": [
+        "pull_request:43291",
+        "pull_request:43665",
+        "pull_request:44815"
+      ],
+      "soft_pairs": [
+        "pull_request:43291|pull_request:43665",
+        "pull_request:43291|pull_request:44815",
+        "pull_request:43665|pull_request:44815"
+      ],
+      "prepared_review_unit_hash": "91af6d51b249076d97dcbdb493b8216942ac481dfd657765de91f4ff07a0e228",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 7093,
+        "estimated_input_tokens": 1774,
+        "estimated_eval_tokens": 3804
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 7093,
+        "estimated_input_tokens": 1774,
+        "estimated_eval_tokens": 3804
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The three PRs are not duplicates: one is a Whisper tokenizer fix, one is a test-only change across CLIP/ViT, and one is a broader FP8/dequant loading fix. They touch different code paths and would not plausibly be merged into one PR.",
+        "confidence": 0.92,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #43291 is the clearest self-contained bug fix: it changes Whisper decoding logic in the tokenizer and updates related tests, with a focused concrete code path.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #43291 is the best representative of a real change because it addresses an actual runtime behavior bug rather than only modifying tests or spanning multiple unrelated subsystems.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:43665",
+            "accept": false,
+            "reason": "Different scope and code paths: Whisper tokenization/decoding versus CLIP/ViT test edits. The second PR appears test-only and does not fix the same underlying bug."
+          },
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:44815",
+            "accept": false,
+            "reason": "Unrelated fixes: Whisper decoding versus FP8/dequant/model-loading changes for Mistral/quantization. No shared concrete bug or mergeable common patch."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:44815",
+            "accept": false,
+            "reason": "Test-only edits for CLIP/ViT are not the same underlying change as FP8 dequant/model-loading work. They are too far apart to treat as one 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": 9,
+      "component_count": 26,
+      "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-43488-2",
+      "nodes": [
+        "pull_request:43488",
+        "pull_request:43532"
+      ],
+      "soft_pairs": [
+        "pull_request:43488|pull_request:43532"
+      ],
+      "prepared_review_unit_hash": "4bc701c8231180896bc58682d82274a2d08ec63dc543fdef4ea1fbc1b3bde073",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4563,
+        "estimated_input_tokens": 1141,
+        "estimated_eval_tokens": 2538
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4563,
+        "estimated_input_tokens": 1141,
+        "estimated_eval_tokens": 2538
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two pull requests are not duplicates: one is a small repo-bot/check related change around setup/dependency/test metadata, while the other is a broader CI/workflow and testing-utilities update. They only overlap on a Vit test file and do not appear to fix the same concrete problem.",
+        "confidence": 0.78,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #43488 is the more coherent representative of the cluster because it is a small, self-contained change centered on repository setup/versioning and a test hook, whereas #43532 spans multiple workflows and utilities with a much broader scope.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #43488 is the best single PR candidate in this set: it has a narrow, specific purpose and a consistent change surface. PR #43532 is broader and touches unrelated CI plumbing, so it is less suitable as the cluster representative.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43488",
+            "right": "pull_request:43532",
+            "accept": false,
+            "reason": "Reject: they do not target the same underlying bug or change. The shared Vit test file is incidental; the rest of the diffs are about different areas (setup/versioning vs CI/workflow and notification plumbing)."
+          }
+        ]
+      },
+      "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": 26,
+      "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": "344e611ebb5d894530ad2ebb7db0bc1b3ed6bd86e2e773f03810d838f68a59bc",
+      "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": "These are two closely related but distinct pull requests: one adds a Promptable Concept Segmentation pipeline, the other adds a Promptable Visual Segmentation pipeline. They overlap in framework/docs/test scaffolding, but they implement different pipeline classes and different underlying model paths, so they are not duplicates.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No clear canonical PR: the two changes target different features and different concrete code paths, even though they share a lot of boilerplate and integration files.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #43613 is the broader of the two and touches more model integrations, so it is the better representative of the pair if one must be chosen.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43612",
+            "right": "pull_request:43613",
+            "accept": false,
+            "reason": "Reject: they add different pipelines ('Promptable Concept Segmentation' vs 'Promptable Visual Segmentation') and different model integrations. Shared docs/test infrastructure is not enough to treat them as the same underlying change or safely merge them 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": 11,
+      "component_count": 26,
+      "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": "4d01db88fa26ff17763b27eeb298d5b71df3f4aa30324176401793332962ce6a",
+      "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": "Both items touch `src/transformers/trainer.py`, but they address different trainer extensions: one changes gradient-accumulation loss scaling, the other changes signature-column filtering for custom trainers. They do not look like duplicates or the same fix.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": "No single PR is a clear canonical representative here; the two PRs are about separate trainer customization points and should be triaged independently.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43651",
+            "right": "pull_request:43663",
+            "accept": false,
+            "reason": "Different code paths and behaviors: `_loss_is_scaled_for_ga` changes loss normalization in `training_step`, while `_get_signature_columns` changes dataset column filtering. Shared filename alone is not enough to treat them as the same 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": 12,
+      "component_count": 26,
+      "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-43656-3",
+      "nodes": [
+        "pull_request:43656",
+        "pull_request:43836",
+        "pull_request:43842"
+      ],
+      "soft_pairs": [
+        "pull_request:43656|pull_request:43842",
+        "pull_request:43836|pull_request:43842"
+      ],
+      "prepared_review_unit_hash": "146523d779b736cfb6c0e35f0b38e4497b13e11b85d2516274ffe68789cadcf9",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 5993,
+        "estimated_input_tokens": 1499,
+        "estimated_eval_tokens": 3254
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 5993,
+        "estimated_input_tokens": 1499,
+        "estimated_eval_tokens": 3254
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "All three PRs are trying to fix the same import-time TypeAdapter annotation problem in `src/transformers/cli/serve.py`; two are near-direct variants of the same fix, and one broader annotation cleanup appears to cover the same underlying bug path.",
+        "confidence": 0.84,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #43842 is the most focused and explicit fix for the TypeAdapter NameError, with a narrow change set and a clear description of the runtime failure it addresses.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #43842 is the best representative because it directly targets the TypeAdapter NameError, keeps the scope tight, and is less of a general typing cleanup than #43656 or the broader refactor in #43836.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43656",
+            "right": "pull_request:43842",
+            "accept": true,
+            "reason": "Both target the same TypeAdapter NameError in `serve.py` and change annotation/runtime typing behavior in the same code path; they look like alternative fixes for the same bug."
+          },
+          {
+            "left": "pull_request:43836",
+            "right": "pull_request:43842",
+            "accept": true,
+            "reason": "Both explicitly fix the TypeAdapter annotation failure in the CLI server module, and the diffs are compatible as variants of the same underlying import-time bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: all three PRs touch `src/transformers/cli/serve.py`, their titles and issue targets point to the same TypeAdapter/annotation NameError family, and the soft-edge pairs look like mergeable alternative fixes rather than unrelated changes. The verdicts are conservative enough."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 13,
+      "component_count": 26,
+      "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": 14,
+      "component_count": 26,
+      "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:44068|pull_request:44116",
+        "pull_request:44068|pull_request:44722",
+        "pull_request:44071|pull_request:44722",
+        "pull_request:44018|pull_request:44116",
+        "pull_request:44018|pull_request:44086",
+        "pull_request:44018|pull_request:44085",
+        "pull_request:44068|pull_request:44086",
+        "pull_request:44071|pull_request:44085",
+        "pull_request:44059|pull_request:44086",
+        "pull_request:44085|pull_request:44086",
+        "pull_request:44059|pull_request:44722",
+        "pull_request:44098|pull_request:44101",
+        "pull_request:44018|pull_request:44071",
+        "pull_request:44017|pull_request:44722",
+        "pull_request:44017|pull_request:44161",
+        "pull_request:44017|pull_request:44073",
+        "pull_request:44068|pull_request:44071",
+        "pull_request:44071|pull_request:44086",
+        "pull_request:43997|pull_request:44073",
+        "pull_request:44017|pull_request:44071",
+        "pull_request:44073|pull_request:44101",
+        "pull_request:44059|pull_request:44071",
+        "pull_request:44059|pull_request:44068",
+        "pull_request:44018|pull_request:44059",
+        "pull_request:44056|pull_request:44101",
+        "pull_request:44073|pull_request:44161",
+        "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:43997|pull_request:44018",
+        "pull_request:44056|pull_request:44161",
+        "pull_request:43997|pull_request:44161",
+        "pull_request:44074|pull_request:44161",
+        "pull_request:43997|pull_request:44071",
+        "pull_request:44010|pull_request:44071",
+        "pull_request:44010|pull_request:44073",
+        "pull_request:44018|pull_request:44129",
+        "pull_request:43997|pull_request:44074",
+        "pull_request:43997|pull_request:44129",
+        "pull_request:44010|pull_request:44018",
+        "pull_request:44071|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"
+      ],
+      "prepared_review_unit_hash": "23b2c7af0a19e20f1883cf03d70eeb623466c397e466faffed77c1431a3160d5",
+      "component_budget": {
+        "node_count": 28,
+        "item_count": 28,
+        "soft_pair_count": 77,
+        "serialized_chars": 84498,
+        "estimated_input_tokens": 21125,
+        "estimated_eval_tokens": 42506
+      },
+      "budget": {
+        "node_count": 28,
+        "item_count": 28,
+        "soft_pair_count": 72,
+        "serialized_chars": 76931,
+        "estimated_input_tokens": 19233,
+        "estimated_eval_tokens": 38722
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster splits into two real duplicate/superset chains: GPT-J/CodeGen (44066 \u2194 44722, with 44085 as a narrower GPT-J slice) and GPT-Neo (44018 \u2194 44068). Everything else is the same output-tracing refactor theme, but across different model code paths, so it should stay separate.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "43979 is the shared tracking issue for the standardized output-tracing refactor work across the cluster.",
+        "canonical_pr_reason": "44722 is the best canonical PR for the GPT-J/CodeGen chain because it is the later, broader version of that same concrete refactor on the same code path; it should not be taken as covering the separate GPT-Neo chain.",
+        "best_issue_reason": "43979 is the central issue anchor for the refactor family in this cluster and the cleanest issue-level representative.",
+        "best_pr_reason": "44722 is the best single PR anchor for the cluster because it is the most complete later version of the largest concrete duplicate chain, but the GPT-Neo chain still needs its own representative PR.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "Different model families; same output-tracing theme only."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "FNet/CVT vs DeBERTaV2; not the same code path."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44085",
+            "accept": true,
+            "reason": "Same GPT-J output-tracing code path; 44085 is a narrower GPT-J slice of the broader 44066 refactor."
+          },
+          {
+            "left": "pull_request:44007",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "ResNet/RT-DETR ResNet vs EfficientNet; different concrete implementations."
+          },
+          {
+            "left": "pull_request:44072",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "EfficientNet vs GPT-J/CodeGen; unrelated model code."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "GPT-J/CodeGen vs EfficientNet; only a shared refactor pattern."
+          },
+          {
+            "left": "pull_request:44013",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "MobileNetV2 and DeBERTaV2 are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-J/CodeGen vs MGP-STR; different model internals."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-J/CodeGen vs MPT; same theme, different implementation."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44068",
+            "accept": true,
+            "reason": "Same GPT-Neo output-tracing refactor on the same file/code path; 44068 looks like the fuller follow-up version."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "GPT-J/CodeGen vs GPT-Neo; different models."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44066",
+            "accept": false,
+            "reason": "GPT-Neo vs GPT-J/CodeGen; different model code."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "ResNet vs GPT-J/CodeGen; different code paths."
+          },
+          {
+            "left": "pull_request:44086",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "MGP-STR vs GPT-J/CodeGen; unrelated models."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "ResNet vs GPT-J; only broad output-tracing similarity."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44019",
+            "accept": false,
+            "reason": "GPT-Neo vs ResNet; different model families."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "ResNet vs MPT; unrelated code paths."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44116",
+            "accept": false,
+            "reason": "GPT-Neo vs Flaubert; different models."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "GPT-Neo vs GPT-J/CodeGen; not the same refactor."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "MPT vs GPT-J/CodeGen; different model internals."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44116",
+            "accept": false,
+            "reason": "GPT-Neo vs Flaubert; same pattern only."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo vs MGP-STR; unrelated code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "GPT-Neo vs GPT-J; different model code."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo vs MGP-STR; not a duplicate."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "MPT vs GPT-J; different concrete changes."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-2 vs MGP-STR; unrelated models."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-J vs MGP-STR; only shared refactor style."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "GPT-2 vs GPT-J/CodeGen; different code paths."
+          },
+          {
+            "left": "pull_request:44098",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "ViLT vs XLM; different model families."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-Neo vs MPT; not the same bug/change."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "Segformer vs GPT-J/CodeGen; unrelated implementations."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "Segformer vs LongT5; different code paths."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "Segformer vs VisualBert; same pattern only."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-Neo vs MPT; different models."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "MPT vs MGP-STR; unrelated refactors."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "RegNet vs VisualBert; different code paths."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "MPNet vs XLM; not the same underlying change."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "VisualBert vs LongT5; different implementations."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "Segformer vs MPT; no shared concrete change."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "VisualBert vs XLM; same broad refactor theme only."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-2 vs MPT; different model internals."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "GPT-2 vs GPT-Neo; not duplicates."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "GPT-2 vs GPT-J; different concrete code."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44017",
+            "accept": false,
+            "reason": "MPNet vs Segformer; unrelated model paths."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "SqueezeBERT vs MPT; different code paths."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "SqueezeBERT vs VisualBert; not the same change."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "DeBERTaV2 vs SpeechT5; different model internals."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44154",
+            "accept": false,
+            "reason": "DeBERTaV2 vs VITS; unrelated model code."
+          },
+          {
+            "left": "pull_request:44024",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "FocalNet vs ImageGPT; same refactor pattern only."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44722",
+            "accept": true,
+            "reason": "Same GPT-J/CodeGen output-tracing refactor; 44722 is the broader later version."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44722",
+            "accept": true,
+            "reason": "Both target GPT-J output tracing on the same code path; they are mergeable as one PR."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "GPT-J vs SpeechT5; different models."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44007",
+            "accept": false,
+            "reason": "FNet/CVT vs ResNet/RT-DETR ResNet; unrelated code paths."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "FNet/CVT vs SpeechT5; same refactor theme only."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "DeBERTaV2 vs SpeechT5; different concrete change."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44154",
+            "accept": false,
+            "reason": "DeBERTaV2 vs VITS; unrelated model code."
+          },
+          {
+            "left": "pull_request:44024",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "FocalNet vs ImageGPT; not the same underlying bug/change."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The cluster summary is mostly grounded, but it overstates uniqueness: the GPT-Neo pair (44018/44068) is also a high-confidence duplicate/superset chain, so 44722 is not the canonical PR for the 'only' strong duplicate chain. The GPT-J/CodeGen chain is a valid anchor, but the writeup should explicitly acknowledge the separate GPT-Neo duplicate chain and avoid implying the rest is only loose theme similarity."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 14,
+      "component_count": 26,
+      "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:44073|pull_request:44074",
+        "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": "f895268fb71b9c3ea4408b6600e063db97e221bd328e3b3f643cd4f527ac5b13",
+      "component_budget": {
+        "node_count": 28,
+        "item_count": 28,
+        "soft_pair_count": 77,
+        "serialized_chars": 84498,
+        "estimated_input_tokens": 21125,
+        "estimated_eval_tokens": 42506
+      },
+      "budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 5,
+        "serialized_chars": 13067,
+        "estimated_input_tokens": 3267,
+        "estimated_eval_tokens": 6790
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "All items are model-specific PRs for the same broad refactor theme\u2014migrating to standardized output tracing/capture\u2014but they touch different architectures and code paths, so they should not be deduped into one change.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 44073 is a clean, representative example of the cluster: it clearly migrates one model to standardized output tracing with a focused diff and no extra cross-cutting changes.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 44129 is the strongest standalone PR in the group because it is explicitly tied to an issue target and has a concrete, well-scoped output-tracing refactor for SpeechT5.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "Same refactor pattern, but different models and files (VisualBert vs TextNet). The implementations are not the same concrete code-path and would not plausibly merge into one PR."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "Both are standardized output-tracing migrations, but they target different architectures (MPNet vs VisualBert) with model-specific hooks and metadata. Similar theme is not enough for a duplicate."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "These are separate model migrations with different forward/model wiring. They share a refactor style, not the same underlying bug or change."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44010",
+            "accept": false,
+            "reason": "RegNet and SqueezeBert are unrelated model implementations. The diffs reflect parallel modernization work, not one shared fix."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "SqueezeBert and SpeechT5 are different model code paths. Even though both migrate output tracing, they are independent PRs and should remain separate."
+          }
+        ]
+      },
+      "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": 26,
+      "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": "167f2d1d2a5d5ce6c5d380f21e549561f2e2d73ab2c0036c0fec215e95711c68",
+      "component_budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 40,
+        "serialized_chars": 40176,
+        "estimated_input_tokens": 10044,
+        "estimated_eval_tokens": 20344
+      },
+      "budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 40,
+        "serialized_chars": 40176,
+        "estimated_input_tokens": 10044,
+        "estimated_eval_tokens": 20344
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These PRs are all variations of the same output-tracing/refactor pattern, but they are model-specific changes to different code paths, not duplicates of one underlying fix.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "44025 is the broadest representative of the set: it applies the refactor across two related depth-anything model files and most clearly matches the cluster theme.",
+        "best_issue_reason": null,
+        "best_pr_reason": "44025 is the best anchor for the cluster because it is a multi-file refactor and cleanly exemplifies the common output-tracing migration pattern.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Same refactor theme, but different model families and forward paths; not the same concrete change."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44000",
+            "accept": false,
+            "reason": "Both are output-tracing refactors, but they target different model implementations and are not mergeable as one fix."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Different model code paths; shared helper migration alone is too broad to treat as a duplicate."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "These touch different architectures and distinct output-tracing implementations, so they are not the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Same broad subsystem, but separate model-specific refactors with different forward code paths."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Similar tracing cleanup, but unrelated model paths and no shared concrete defect."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Both are refactors, but they affect different models and should not be merged as duplicates."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "Different model families and changes; only the refactor style matches."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44026",
+            "accept": false,
+            "reason": "No evidence of a shared bug or identical code path; just a similar output-tracing migration."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "Same broad cleanup pattern, but distinct wrapper/model implementations."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Different models and forward logic; the similarity is only at the refactor pattern level."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Not the same code path; they are separate model-specific output-tracing updates."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "These are separate model implementations with different refactor scopes, so they are not duplicates."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "Shared helper migration is not enough; these change different model code paths."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Different models with no shared concrete fix, only the same refactor theme."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "Distinct architectures and output handling; not one underlying bug/change."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "Separate model-specific refactors, not the same code path."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Similar output-tracing cleanup, but unrelated models and implementation details."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "These are different model families and should not be collapsed into one PR."
+          },
+          {
+            "left": "pull_request:44003",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Same general refactor family, but different architectures and forward paths."
+          },
+          {
+            "left": "pull_request:44003",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Different model subsystems and changes; not a single concrete bug or fix."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Only the broad output-tracing theme overlaps; the code paths are unrelated."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "These are separate model implementations, so the refactor similarity is too weak for duplication."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Different target models and code paths; no same underlying change."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Shared output-tracing language, but distinct model code and no common concrete bug."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44026",
+            "accept": false,
+            "reason": "Both are wrapper-style refactors, but they are different model wrappers and not the same change."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "Different encoder-decoder wrapper implementations; similar refactor, not duplicate."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Separate model code paths with only superficial refactor similarity."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Same tracing migration pattern, but distinct models and fixes."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Not the same underlying issue; just a similar output-tracing refactor."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Different model classes and code paths; too broad to merge as duplicates."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "Both are wrapper refactors, but they are separate encoder-decoder models and not one concrete fix."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Distinct model implementations; shared theme only."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Different models and changes; no shared underlying bug or change."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Separate model-specific refactors, not a single code-path problem."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "No evidence of the same concrete change beyond the refactor pattern."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Different model code paths; not mergeable as one duplicate fix."
+          },
+          {
+            "left": "pull_request:44028",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Same broad output-tracing cleanup, but unrelated model implementations."
+          },
+          {
+            "left": "pull_request:44028",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Distinct models and code paths; not the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:44029",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Different models; only the general refactor theme overlaps."
+          }
+        ]
+      },
+      "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": 26,
+      "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-44662-3",
+      "nodes": [
+        "pull_request:44662",
+        "pull_request:45075",
+        "pull_request:45401"
+      ],
+      "soft_pairs": [
+        "pull_request:44662|pull_request:45401",
+        "pull_request:44662|pull_request:45075",
+        "pull_request:45075|pull_request:45401"
+      ],
+      "prepared_review_unit_hash": "9be26e50c650fa64cad6acfa55495980c3366f546bb4e2f1ebb1908bc24f67ee",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 10914,
+        "estimated_input_tokens": 2729,
+        "estimated_eval_tokens": 5714
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 10914,
+        "estimated_input_tokens": 2729,
+        "estimated_eval_tokens": 5714
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are three separate model-integration pull requests for different models (PenguinVL, DeepSeek-OCR-2, and Voxtral TTS). They share Transformers boilerplate and Auto/Docs wiring, but they do not target the same code-path problem or the same change.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45075 is the best representative because it is a mature, non-draft model integration with substantial implementation coverage and the highest review activity among the three.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45075 is the strongest overall representative of the cluster: it is a full-featured integration, not a draft, and has the most review activity.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44662",
+            "right": "pull_request:45401",
+            "accept": false,
+            "reason": "Different model integrations for unrelated architectures/tasks (PenguinVL vs Voxtral TTS); shared docs/auto-registration files are just boilerplate, not evidence of the same fix."
+          },
+          {
+            "left": "pull_request:44662",
+            "right": "pull_request:45075",
+            "accept": false,
+            "reason": "PenguinVL and DeepSeek-OCR-2 are distinct model additions with different model code, configs, and converters; the overlap is only in common Transformers scaffolding."
+          },
+          {
+            "left": "pull_request:45075",
+            "right": "pull_request:45401",
+            "accept": false,
+            "reason": "DeepSeek-OCR-2 and Voxtral TTS are unrelated model integrations; they touch similar registry/docs areas but do not fix or add the same underlying functionality."
+          }
+        ]
+      },
+      "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": 26,
+      "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": "deb61c381b92e4ceba5a3960f9949310d19ff5b52dbac4719b19dced7343dc45",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4408,
+        "estimated_input_tokens": 1102,
+        "estimated_eval_tokens": 2460
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4408,
+        "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 feature-area PRs about fusion, but they implement different abstractions and code paths: one adds a generic Module Fusion API, the other adds n-to-1 kernel fusion via KernelConfig and model-loading integration. They are not duplicates.",
+        "confidence": 0.92,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45363 is the broader and more concrete implementation of kernel-backed fusion, touching the loader/integration path and KernelConfig plumbing rather than just introducing a standalone API scaffold.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45363 is the best representative of the cluster because it implements the end-to-end user-facing fusion behavior, while PR 44979 is a separate API foundation.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44979",
+            "right": "pull_request:45363",
+            "accept": false,
+            "reason": "They target different layers: 44979 adds a new generic Module Fusion API and tests, while 45363 wires n-to-1 kernel fusion through KernelConfig, integrations, and model loading. Related theme, but not the same underlying change or 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": 18,
+      "component_count": 26,
+      "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": "dc3a2b77fef08234750a5a94b21b4fff98b4ab68146220ab473401b64d8d9a0a",
+      "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 patch the same AttributeError in `_patch_mistral_regex` by switching from `tokenizer.backend_tokenizer.pre_tokenizer` to `tokenizer.pre_tokenizer` in the same file; the later PR adds tests and is the more complete representative.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45317 is the better canonical PR because it fixes the same concrete bug as #45086 but also includes test coverage and a clearer title describing the exact failing path (`fix_mistral_regex=True`).",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45317 is the best overall PR for the cluster: same code change, same issue target, and it adds regression tests, making it the most complete and merge-ready representation.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45086",
+            "right": "pull_request:45317",
+            "accept": true,
+            "reason": "Accept: both PRs address the same AttributeError in `_patch_mistral_regex` with the same line-level fix in `tokenization_utils_tokenizers.py`; #45317 is essentially the same patch plus tests, so they are duplicate-change candidates."
+          }
+        ]
+      },
+      "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": 26,
+      "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": "41070347c6f8aadb9f4bec1ca58fd65b9742ff0652482c83fe08540ff48c1c8c",
+      "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": "These two PRs are related to Nemotron H, but they are not the same underlying change: one is a narrow conversion/refactor PR, while the other is a broader architecture split introducing dedicated dense/sparse model variants and docs/tests.",
+        "confidence": 0.9,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45569 is the better representative for this cluster because it is the broader, more complete Nemotron H change set (new dense/sparse variants, auto-mapping, docs, and tests), whereas #45115 is a narrower conversion-mapping/refactor PR.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45569 is the strongest standalone PR here due to its broader scope and clearer user-facing impact; #45115 looks like a smaller preparatory/internal refactor rather than the main architecture change.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45115",
+            "right": "pull_request:45569",
+            "accept": false,
+            "reason": "Both touch Nemotron H files, but they target different changes: #45115 adds checkpoint conversion/key renaming for GraniteMoeHybrid compatibility, while #45569 restructures Nemotron H into dedicated dense/sparse model variants with docs and tests. They are not mergeable as one 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": 20,
+      "component_count": 26,
+      "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": "c55cac4f5bef2da99d6cb51bcf00b279d623ffce9bd542515a6d05d318d9793d",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5397,
+        "estimated_input_tokens": 1350,
+        "estimated_eval_tokens": 2956
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5397,
+        "estimated_input_tokens": 1350,
+        "estimated_eval_tokens": 2956
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two related but distinct PRs: one adds heterogeneous per-layer config support, the other extends that foundation into broader heterogeneous model/runtime support. They overlap in shared config utilities but are not the same concrete change.",
+        "confidence": 0.84,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45332 is the broader, more representative change: it includes heterogeneous config handling plus model/cache/masking/runtime integration and tests, making it the best canonical PR for the cluster.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45332 is the best representative because it covers the larger end-to-end heterogeneous model support rather than only the config layer plumbing.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45332",
+            "right": "pull_request:45333",
+            "accept": false,
+            "reason": "They share heterogeneous config utilities, but #45333 is only per-layer configuration while #45332 adds broader model/runtime support. This is related foundation work, not the same fix/change that should 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": 21,
+      "component_count": 26,
+      "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": "7721d2b2c5d12f2f10ffbdae79837ba5b2183949a7ccb61f9b3dec78324f16e4",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5513,
+        "estimated_input_tokens": 1379,
+        "estimated_eval_tokens": 3014
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5513,
+        "estimated_input_tokens": 1379,
+        "estimated_eval_tokens": 3014
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs are typing-related, but they are not the same change: #45415 is a broad type-checking sweep across many `src/transformers/*.py` files, while #45425 adds specific typing protocol/alias support for a smaller set of code paths. Shared filenames are incidental, not evidence of the same fix.",
+        "confidence": 0.89,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "#45425 is the more focused and concrete typing change; it introduces specific protocol/alias definitions to satisfy a narrower set of typing needs, whereas #45415 is a wide cleanup across many files.",
+        "best_issue_reason": null,
+        "best_pr_reason": "#45425 is the best representative for the cluster because it is a self-contained typing improvement with a clear purpose and smaller scope.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45415",
+            "right": "pull_request:45425",
+            "accept": false,
+            "reason": "They are related only at a high level (typing work) and touch some overlapping files, but they address different problems: a broad type-checking pass versus targeted typing protocol/alias additions. They would not plausibly be merged into one PR as 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": 22,
+      "component_count": 26,
+      "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-45487-2",
+      "nodes": [
+        "pull_request:45487",
+        "pull_request:45615"
+      ],
+      "soft_pairs": [
+        "pull_request:45487|pull_request:45615"
+      ],
+      "prepared_review_unit_hash": "4d452c0889d073b90531568d25de57b9ddccd272cb56d24a2d268f234eadbf09",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5159,
+        "estimated_input_tokens": 1290,
+        "estimated_eval_tokens": 2836
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5159,
+        "estimated_input_tokens": 1290,
+        "estimated_eval_tokens": 2836
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are unrelated PRs: one fixes a model-parallel device placement bug in AltCLIP/ChineseCLIP code, while the other only adds XPU test expectations for Qianfan OCR. They should not be clustered as duplicates.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45487 is the substantive code fix, addressing a concrete device mismatch in model-parallel forward logic across AltCLIP/ChineseCLIP and related models.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45487 is the best representative because it changes runtime model code to fix a real bug; PR 45615 is only updating expected test outputs for another model.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45487",
+            "right": "pull_request:45615",
+            "accept": false,
+            "reason": "Different models and different change types: 45487 fixes a device-placement bug in model code, while 45615 only updates XPU test expectations for Qianfan OCR. They do not appear to address 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": 23,
+      "component_count": 26,
+      "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": "49fbc7772e9f05e2978776cd59a1c60531632be67281b39adf9aba42154a01ed",
+      "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 pull requests are related to expert parallelism, but they target different failure modes and code paths: one focuses on EP/ZeRO-3 loading and config serialization, while the other fixes experts being overwritten by rank-0 broadcast during distributed loading. They should not be treated as duplicates.",
+        "confidence": 0.91,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45662 is the better canonical representative because it targets a concrete distributed-loading bug with a clear failure mode (experts being overwritten) and a focused fix in the loading path; PR #45548 is narrower and addresses EP/ZeRO-3 launch/serialization behavior instead.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45662 is the strongest standalone candidate: it fixes a specific correctness issue in distributed model loading and has the most direct, concrete bug description. PR #45548 is useful but more ancillary, centered on launch compatibility and config handling.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45548",
+            "right": "pull_request:45662",
+            "accept": false,
+            "reason": "Both involve expert parallelism, but they fix different bugs: #45548 adds EP config/compatibility handling for ZeRO-3 loading, while #45662 prevents experts from being overwritten during rank-0 broadcast in the loading path. 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": 24,
+      "component_count": 26,
+      "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": "719c76c451bec8aae35576e448d0eea85ef44d6884da0d83fcf90965be09f395",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4589,
+        "estimated_input_tokens": 1148,
+        "estimated_eval_tokens": 2552
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4589,
+        "estimated_input_tokens": 1148,
+        "estimated_eval_tokens": 2552
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs address the same xdist-related captured_info logging problem in testing_utils; PR 45645 is the broader, more complete version because it also updates CI handling for multiple captured_info files.",
+        "confidence": 0.92,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45645 covers the same core xdist-safe captured_info fix as 45639 and additionally adjusts the CI workflow to consume the new per-worker artifact naming, making it the more complete implementation.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45645 is the best representative because it includes the core code-path fix plus the downstream CI/log collection changes needed to use it reliably.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45639",
+            "right": "pull_request:45645",
+            "accept": true,
+            "reason": "Same underlying captured_info xdist collision fix; 45645 is a superset that includes the testing_utils change and the related CI workflow update."
+          }
+        ]
+      },
+      "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": 26,
+      "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-45683-2",
+      "nodes": [
+        "pull_request:45683",
+        "pull_request:45734"
+      ],
+      "soft_pairs": [
+        "pull_request:45683|pull_request:45734"
+      ],
+      "prepared_review_unit_hash": "eba4ca0f26b2d3be0c00e316270a9341d31af2208b44dbab31b1d3a8e81e7538",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4469,
+        "estimated_input_tokens": 1118,
+        "estimated_eval_tokens": 2492
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4469,
+        "estimated_input_tokens": 1118,
+        "estimated_eval_tokens": 2492
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open PRs touch the same quantization helper, but they fix different bugs: one special-cases audio submodules, the other changes how default skip lists are combined with user-supplied skips. They should not be clustered as duplicates.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "45734 is the better representative PR because it fixes the core skip-module composition logic in `get_modules_to_not_convert`, which is broader and more central than the narrow audio-module special case in 45683.",
+        "best_issue_reason": null,
+        "best_pr_reason": "45734 is the strongest standalone PR in the set: it addresses a concrete quantization failure mode in the main code path and its fix is directly about the helper\u2019s behavior rather than a model-specific workaround.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45683",
+            "right": "pull_request:45734",
+            "accept": false,
+            "reason": "Both touch quantization skip logic, but 45683 adds audio-tower/embed_audio exclusions while 45734 changes additive handling of `skip_modules`; these are different underlying bugs and would not naturally merge 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": 26,
+      "component_count": 26,
+      "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
+    }
+  ]
+}
diff --git a/snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/manifest.json b/snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..8b293e7239475b64375343a1ccaacde6249114b9
--- /dev/null
+++ b/snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/manifest.json
@@ -0,0 +1,14 @@
+{
+  "analysis_id": "hybrid-model-20260507t060024z",
+  "artifacts": {
+    "hybrid": "snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.json",
+    "hybrid_reviews": "snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.llm-reviews.json"
+  },
+  "channel": "canonical",
+  "model": null,
+  "published_at": "2026-05-07T06:05:26Z",
+  "repo": "huggingface/transformers",
+  "schema_version": 1,
+  "snapshot_id": "20260507T060024Z",
+  "variant": "hybrid"
+}
diff --git a/snapshots/20260507T060024Z/manifest.json b/snapshots/20260507T060024Z/manifest.json
index 2c3f1aa4daffbc7bf62b3d6f79a2fb2a06e7a1a7..dae50b1f4285ab009aa7f458b0bc29ba24b03ecb 100644
--- a/snapshots/20260507T060024Z/manifest.json
+++ b/snapshots/20260507T060024Z/manifest.json
@@ -32,6 +32,24 @@
     "timeline_events": 559
   },
   "extracted_at": "2026-05-07T06:00:24Z",
+  "published_analysis": {
+    "canonical_analysis_id": "hybrid-model-20260507t060024z",
+    "runs": {
+      "hybrid-model-20260507t060024z": {
+        "analysis_id": "hybrid-model-20260507t060024z",
+        "artifacts": {
+          "hybrid": "snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.json",
+          "hybrid_reviews": "snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/analysis-report-hybrid.llm-reviews.json"
+        },
+        "channel": "canonical",
+        "manifest_path": "snapshots/20260507T060024Z/analysis-runs/hybrid-model-20260507t060024z/manifest.json",
+        "model": null,
+        "published_at": "2026-05-07T06:05:26Z",
+        "variant": "hybrid"
+      }
+    },
+    "schema_version": 1
+  },
   "repo": "huggingface/transformers",
   "snapshot_id": "20260507T060024Z",
   "watermark": {