diff --git a/analysis/current/analysis-report-hybrid.json b/analysis/current/analysis-report-hybrid.json
index b60994f359abdbf91130032383a9ef4b41d2f57a..58b9255c33d57bf53fe4b434dc1a9ba6a2a7b721 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": "20260515T000050Z",
-  "generated_at": "2026-05-15T00:09:04Z",
+  "snapshot_id": "20260515T060029Z",
+  "generated_at": "2026-05-15T06:04:51Z",
   "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 ce5b46977a13bbeb4218ad9f701c11c547530659..fe13419bf490f489af924c38ba39cd1442ff3d4d 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": "20260515T000050Z",
-  "generated_at": "2026-05-15T00:09:04Z",
+  "snapshot_id": "20260515T060029Z",
+  "generated_at": "2026-05-15T06:04:51Z",
   "model": "gpt-5.4-mini?service_tier=flex",
   "reviews": [
     {
@@ -13,8 +13,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -402,8 +402,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -773,8 +773,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -1108,8 +1108,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -1444,8 +1444,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -1773,8 +1773,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -2130,8 +2130,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -2455,8 +2455,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -2781,8 +2781,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -3102,8 +3102,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -3417,8 +3417,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -3735,8 +3735,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -4053,8 +4053,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -4360,8 +4360,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -4667,8 +4667,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -4958,8 +4958,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -5249,8 +5249,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -5564,8 +5564,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -5879,8 +5879,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -6169,8 +6169,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -6508,8 +6508,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -6834,8 +6834,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -7125,8 +7125,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -7429,8 +7429,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -7779,8 +7779,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -8090,8 +8090,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -8381,8 +8381,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -8679,8 +8679,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -8977,8 +8977,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -9289,8 +9289,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -9601,8 +9601,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -9913,8 +9913,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -10218,8 +10218,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -10516,8 +10516,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -10849,8 +10849,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -11180,8 +11180,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -11508,8 +11508,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -11855,8 +11855,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -12174,8 +12174,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -12493,8 +12493,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -12634,8 +12634,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -12967,8 +12967,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -13268,8 +13268,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -13566,8 +13566,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -13885,8 +13885,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -14246,8 +14246,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -14663,8 +14663,8 @@
       "review_unit_count": 47,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10000-400",
       "nodes": [
@@ -14907,8 +14907,8 @@
       "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": [
@@ -14978,8 +14978,8 @@
       "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": [
@@ -15041,8 +15041,8 @@
       "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-27088-5",
       "nodes": [
@@ -15170,8 +15170,8 @@
       "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-36817-15",
       "nodes": [
@@ -15484,8 +15484,8 @@
       "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": [
@@ -15600,8 +15600,8 @@
       "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": [
@@ -15663,8 +15663,8 @@
       "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": [
@@ -15726,8 +15726,8 @@
       "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": [
@@ -15789,8 +15789,8 @@
       "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": [
@@ -15889,8 +15889,8 @@
       "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": [
@@ -16011,8 +16011,8 @@
       "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": [
@@ -16131,8 +16131,8 @@
       "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": [
@@ -16194,8 +16194,8 @@
       "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": [
@@ -16257,8 +16257,8 @@
       "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": [
@@ -16320,8 +16320,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-45869-5",
       "nodes": [
@@ -16483,175 +16483,163 @@
         "pull_request:44827|pull_request:45922",
         "pull_request:43251|pull_request:43254"
       ],
-      "prepared_review_unit_hash": "99631cbf12ce2689134cc870710aef6de6eef92202ad087ae1bc19a4b3152376",
+      "prepared_review_unit_hash": "96ea4833ea1cb2ce921e5550a0c00cd4c21fe491dcdb3813e0833889bd189ec9",
       "component_budget": {
         "node_count": 15,
         "item_count": 15,
         "soft_pair_count": 22,
-        "serialized_chars": 42212,
-        "estimated_input_tokens": 10553,
-        "estimated_eval_tokens": 21362
+        "serialized_chars": 42217,
+        "estimated_input_tokens": 10555,
+        "estimated_eval_tokens": 21366
       },
       "budget": {
         "node_count": 15,
         "item_count": 15,
         "soft_pair_count": 22,
-        "serialized_chars": 42212,
-        "estimated_input_tokens": 10553,
-        "estimated_eval_tokens": 21362
+        "serialized_chars": 42217,
+        "estimated_input_tokens": 10555,
+        "estimated_eval_tokens": 21366
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "The cluster contains one true duplicate pair: the two `fixed_cross_entropy` PRs (#43251 and #43254). The rest are unrelated model additions, docs, or refactors that only share broad subsystem similarity.",
-        "confidence": 0.98,
+        "summary": "The cluster is mostly unrelated PRs that only share broad Transformer-maintenance themes. The only strong duplicate pair is the two `fixed_cross_entropy` PRs, which address the same loss-helper change for the same underlying `cross_entropy` kwargs support.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #43254 is a solid canonical representative: it is explicitly linked to issue #43240, it is the matched duplicate partner of #43251, and both PRs make the same `fixed_cross_entropy` kwargs/label_smoothing change in the same file.",
+        "canonical_pr_reason": "`pull_request:43251` is the better canonical PR because it directly targets issue 43240, clearly states the bug fix, and is the more complete implementation of the shared `fixed_cross_entropy` change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #43254 is the best overall representative in the packet for the duplicate code-fix pair because it is explicitly tied to #43240 and is the matched duplicate partner of #43251; the two PRs address the same concrete `fixed_cross_entropy` change.",
+        "best_pr_reason": "`pull_request:43251` is the best representative of the duplicate pair: same file, same bug, same functional goal, and it has the explicit tracking issue link. The other PR (`43254`) is a near-duplicate variant of the same fix.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:39895",
             "right": "pull_request:43424",
             "accept": false,
-            "reason": "VideoPrism model addition vs ExecuTorch dynamic-shape export test; unrelated changes."
+            "reason": "Unrelated PRs: one adds VideoPrism, the other adds an Executorch dynamic-shape test. No shared code-path or concrete bug fix."
           },
           {
             "left": "pull_request:41855",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Tokenizer missing-methods fix vs adding MobileLLM; different code paths and goals."
+            "reason": "Different model/task areas: Mistral tokenizer helper methods vs adding MobileLLM. Shared tokenization context is too broad to be the same change."
           },
           {
             "left": "pull_request:41855",
             "right": "pull_request:43424",
             "accept": false,
-            "reason": "Tokenizer API fix vs export test; no shared underlying bug or change."
+            "reason": "Tokenizer utility additions and an Executorch export test are unrelated changes."
           },
           {
             "left": "pull_request:42668",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Processor robustness refactor vs new model implementation; not the same fix."
+            "reason": "Both touch processor/model infrastructure, but they target different models and different problems. Not the same underlying bug or change."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Loss-function kwargs fix vs model addition; unrelated."
+            "reason": "Loss-helper kwargs support and adding MobileLLM are unrelated; no shared code-path."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "Loss-function kwargs fix vs Mistral4 test/refactor work; different problems."
+            "reason": "Both are small fixes, but one is cross-entropy kwargs handling and the other is Mistral4 test cleanup. Different code paths and goals."
           },
           {
             "left": "pull_request:43424",
             "right": "pull_request:45919",
             "accept": false,
-            "reason": "Dynamic-shape export test vs Sapiens2 model addition; unrelated."
+            "reason": "Executorch dynamic-shape export test vs adding Sapiens2 model. Not a duplicate change."
           },
           {
             "left": "pull_request:43424",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "Export test vs Mistral4 test/refactor; different concrete issues."
+            "reason": "Different concerns: dynamic-shape export testing versus Mistral4 test fixes. No shared underlying bug."
           },
           {
             "left": "pull_request:43424",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Export test vs MobileLLM model addition; no shared bug."
+            "reason": "Executorch export test and MobileLLM model addition are unrelated."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:45655",
             "accept": false,
-            "reason": "New model addition vs config-resolution order fix; unrelated."
+            "reason": "New model addition vs config resolution order fix. Too broad and unrelated."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:45922",
             "accept": false,
-            "reason": "New model addition vs lru-cache memory leak fix; different issues."
+            "reason": "MobileLLM model addition and vision-model memory-leak fix are unrelated changes."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:43995",
             "accept": false,
-            "reason": "New model implementation vs Falcon output-interface refactor; not the same change."
+            "reason": "Adding MobileLLM and refactoring Falcon output collection are different model-specific changes."
           },
           {
             "left": "pull_request:43995",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "Falcon refactor vs model-creation skill docs; unrelated."
+            "reason": "Falcon output-interface refactor and a model-creation skill are unrelated."
           },
           {
             "left": "pull_request:43995",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "Falcon interface refactor vs Mistral4 test fix; different code paths."
+            "reason": "Falcon model refactor and Mistral4 test fixes do not address the same bug or code path."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45170",
             "accept": false,
-            "reason": "Mistral4 tests vs CLIP-like layernorm rename; unrelated."
+            "reason": "Mistral4 test fixes and CLIP-like `layrnorm`\u2192`layernorm` renaming are unrelated."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45795",
             "accept": false,
-            "reason": "Mistral4 tests vs audio/video processor docs; only broad ecosystem overlap."
+            "reason": "Mistral4 test fixes and docs for audio/video processors are not the same change."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45655",
             "accept": false,
-            "reason": "Mistral4 tests vs `cls.config` resolution fix; not the same bug."
+            "reason": "Mistral4 test fixes and config resolution ordering are unrelated maintenance changes."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "Mistral4 tests vs model-creation skill; completely different intent."
-          },
-          {
-            "left": "pull_request:43823",
-            "right": "pull_request:45213",
-            "accept": false,
-            "reason": "Model implementation vs skill/documentation PR; unrelated."
-          },
-          {
-            "left": "pull_request:43823",
-            "right": "pull_request:45795",
-            "accept": false,
-            "reason": "Model implementation vs processor docs; different scope."
+            "reason": "Mistral4 test fixes and a model-creation skill PR are unrelated."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45922",
             "accept": false,
-            "reason": "Mistral4 tests vs vision-model memory-leak fix; unrelated."
+            "reason": "Mistral4 test fixes and vision-model memory-leak fixes are different issues in different subsystems."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:43254",
             "accept": true,
-            "reason": "Both PRs modify `fixed_cross_entropy` in the same way to forward supported `cross_entropy` kwargs; same concrete fix and plausible to merge as one."
+            "reason": "Same underlying fix in the same file: both update `fixed_cross_entropy` to support extra `nn.functional.cross_entropy` kwargs such as `weight` and `label_smoothing`. They could plausibly be merged into one PR."
           }
         ]
       },
       "evaluator_result": {
-        "accept": false,
-        "feedback": "The overall duplicate-pair call is correct, but the analysis overstates evidence in `best_pr_reason`: PR #43251 is not the only explicitly issue-linked code fix in the set, since PR #43254 is also explicitly linked to issue #43240 and is the matched duplicate partner. Please make the representative selection more conservative and grounded in the packet."
+        "accept": true,
+        "feedback": "The duplicate judgment is well grounded in the packet. The only accepted pair, `pull_request:43251` and `pull_request:43254`, clearly targets the same `fixed_cross_entropy` change in the same file and the same issue (`43240`), so mergeability as one bugfix is plausible. The other soft pairs are conservatively rejected and the summary does not overstate the evidence."
       },
       "evaluator_used": true,
-      "retried": true,
+      "retried": false,
       "accepted_nontrivial_soft_edge": true,
       "error_kind": null,
       "error_message": null
@@ -16675,7 +16663,7 @@
       "soft_pairs": [
         "pull_request:40546|pull_request:42781"
       ],
-      "prepared_review_unit_hash": "e87f3f2811ffe53c8a98822c13a0e19cb6e4152e05500e235111b2ab90872d77",
+      "prepared_review_unit_hash": "75cb1bd96934f1d1dddaa97af0355bbf0daf8bdcf072e0e20ba3f047fc1fa9de",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -16697,18 +16685,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related VibeVoice PRs, but they are not the same change: one adds the base VibeVoice model package, while the other adds a realtime variant plus an acoustic tokenizer and additional generation/runtime changes.",
-        "confidence": 0.84,
+        "summary": "The two PRs overlap heavily on VibeVoice model scaffolding, but they are not duplicates: 40546 implements the base VibeVoice model, while 42781 extends it with realtime/acoustic-tokenizer work and extra generation changes.",
+        "confidence": 0.88,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #40546 is the cleaner canonical representative of the cluster because it is the original base VibeVoice implementation and has much more review activity, while #42781 is a later, broader realtime follow-up.",
+        "canonical_pr_reason": "PR 40546 is the core standalone VibeVoice implementation and has the stronger review history; 42781 is a later realtime extension rather than the same change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #40546 is the best single PR to anchor the cluster: it introduces the core VibeVoice integration without the extra realtime/tokenizer scope, making it the more stable and representative change.",
+        "best_pr_reason": "PR 40546 is the best representative of the cluster because it covers the primary VibeVoice feature without the additional realtime follow-on scope.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:40546",
             "right": "pull_request:42781",
             "accept": false,
-            "reason": "Do not merge: #42781 is not a duplicate of #40546. It extends VibeVoice into a realtime path and adds an acoustic tokenizer plus generation/util changes, so the concrete change set and code path differ materially from the base implementation."
+            "reason": "They share the VibeVoice subsystem, but 42781 adds realtime/acoustic-tokenizer functionality and extra generation changes, so it is an extension rather than the same concrete PR."
           }
         ]
       },
@@ -16738,7 +16726,7 @@
       "soft_pairs": [
         "pull_request:40755|pull_request:41224"
       ],
-      "prepared_review_unit_hash": "5667d59e50c9c41c632b2bbb32f5d55b5929ff511bed31ee77938813dc860661",
+      "prepared_review_unit_hash": "e4b37e2b939feb7dd563df270c79ee68a22e6817d18608b68d582897629bc888",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -16760,18 +16748,18 @@
       "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.",
+        "summary": "False positive cluster: one PR adds TimesFM covariate forecasting support, while the other adds DINOv3 image-classification support. They touch different models, code paths, and user-facing features.",
         "confidence": 0.99,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No clear canonical PR for deduplication: the PRs target different models, code paths, and user-facing features.",
+        "canonical_pr_reason": "No canonical PR here: these are unrelated feature PRs for different model families and should not be deduplicated into one cluster.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #40755 is the broader representative of the cluster only in the sense that it adds a substantial new forecasting capability across multiple files; however, it is not a duplicate of PR #41224.",
+        "best_pr_reason": "No global best PR for the cluster; each PR is independently valid but they solve different problems and belong in separate clusters.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:40755",
             "right": "pull_request:41224",
             "accept": false,
-            "reason": "Different underlying changes: TimesFM covariate forecasting support versus DINOv3 image classification support. They modify different model implementations, docs, and tests, so they should not be merged as duplicates."
+            "reason": "Different model families and changesets: TimesFM covariate forecasting vs DINOv3 image classification. Not the same underlying bug or mergeable into one PR."
           }
         ]
       },
@@ -16866,7 +16854,7 @@
         "pull_request:41797|pull_request:45186",
         "pull_request:45186|pull_request:45630"
       ],
-      "prepared_review_unit_hash": "807f8d97220b5abc87753bdb72e9005457a92959e8272553bcd476e19008449c",
+      "prepared_review_unit_hash": "2ec9a732cb7a092f672c42e2ea003f88ca1d890bacfc8aa0b90aa6746e991cdf",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
@@ -16888,24 +16876,24 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Three independent model-addition PRs were grouped by broad template similarity, but they target different models and code paths: DeepSeek OCR, Isaac, and Kimi2-6. They are not duplicates and should stay separate.",
+        "summary": "Three pull requests add different model integrations (DeepSeek OCR, Isaac, Kimi2-6). They overlap in boilerplate registration/docs, but they do not fix the same underlying code-path problem, so they should not be merged as duplicates.",
         "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No single canonical PR fits this cluster because the PRs add different models with different files, configs, and tests; the similarity is only structural (new-model scaffold).",
+        "canonical_pr_reason": "PR 45630 is the strongest representative of the cluster: it has the broadest integration footprint and the highest review activity, but it is still a distinct Kimi2-6 model addition rather than the same change as the others.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45630 is the most substantial and recent standalone model-addition PR in the set, but it is not a duplicate of the others and should not be treated as a merge target for them.",
+        "best_pr_reason": "PR 45630 is the best single representative because it is the most developed and actively reviewed of the three, making it the most useful anchor if this cluster needs one canonical PR.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:41797",
             "right": "pull_request:45186",
             "accept": false,
-            "reason": "Both are new-model PRs, but they add different models (DeepSeek OCR vs Isaac) with different implementations and docs. Shared auto-registration files reflect framework scaffolding, not the same change."
+            "reason": "Both are new model additions, but they target different models and different implementation stacks (DeepSeek OCR vs Isaac). Shared auto-registration/docs files are boilerplate and not evidence of the same fix."
           },
           {
             "left": "pull_request:45186",
             "right": "pull_request:45630",
             "accept": false,
-            "reason": "Isaac and Kimi2-6 are unrelated model additions. They share broad integration files, but the concrete code paths and model-specific modules are distinct, so they should not be merged as duplicates."
+            "reason": "These are separate model integrations for Isaac and Kimi2-6. They overlap in model registry plumbing, but they do not address the same concrete bug or change and would not plausibly be merged into one PR."
           }
         ]
       },
@@ -16935,7 +16923,7 @@
       "soft_pairs": [
         "pull_request:42975|pull_request:42976"
       ],
-      "prepared_review_unit_hash": "9d55fbaafab87324b839847a1270c4fce815f932afffaff0c7c8f95139e9e076",
+      "prepared_review_unit_hash": "cdb9eaaf4d1c58320dcd8eb575f4b0c13ad8cb483009f504d858f3c7297b7a7c",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -16957,18 +16945,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Two separate maintenance PRs: one upgrades core GitHub Actions for Node 24 compatibility, the other bumps docker/build-push-action versions in Docker build workflows. They overlap in workflow files but do not address the same underlying change.",
-        "confidence": 0.91,
+        "summary": "Two open PRs both update GitHub Actions, but they target different dependency families and change sets: one is a broad Node 24 compatibility sweep for workflow actions, while the other upgrades Docker/build-related actions to newer major versions. They overlap in a few workflow files, but they do not fix the same concrete problem.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 42975 is the better representative of the cluster because it has a clear, unified goal: migrating core GitHub Actions to Node 24-compatible versions across CI workflows.",
+        "canonical_pr_reason": "PR 42975 is the better cluster representative because it is the broader, explicitly scoped workflow-actions compatibility upgrade and touches many of the same CI workflows.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 42975 is the strongest standalone choice: it is broader, more cohesive, and tied to a specific compatibility objective rather than a narrower action-version bump.",
+        "best_pr_reason": "PR 42975 is the best single representative of this cluster since it captures the general GitHub Actions upgrade theme more comprehensively than the narrower Docker-action bump in 42976.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:42975",
             "right": "pull_request:42976",
             "accept": false,
-            "reason": "Reject: these are different dependency upgrades in GitHub Actions workflows, not the same concrete bug or change. They touch some of the same files, but one updates checkout/cache/upload-artifact for Node 24 compatibility while the other updates docker/build-push-action."
+            "reason": "Do not merge: these are separate maintenance changes. 42975 upgrades multiple GitHub Actions for Node 24 compatibility, while 42976 primarily bumps docker/build-push-action and a few other actions to latest versions. Shared workflow filenames alone are not enough to treat them as the same underlying change."
           }
         ]
       },
@@ -17011,7 +16999,7 @@
         "pull_request:44601|pull_request:45838",
         "pull_request:45490|pull_request:45838"
       ],
-      "prepared_review_unit_hash": "047918875e66d67615be850070d9f4a48be701a26b4ad15bbeb12a6a084cebfe",
+      "prepared_review_unit_hash": "f23405e4678881cb0d7623cf445a7666a13ad9bd6356a79ca848daaa7b7ecaf2",
       "component_budget": {
         "node_count": 7,
         "item_count": 7,
@@ -17033,66 +17021,66 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "This cluster is a set of unrelated additions: several distinct model-introduction PRs plus one distributed loading feature. The shared files are mostly boilerplate registry/docs changes, not evidence of the same underlying bug or change.",
-        "confidence": 0.96,
+        "summary": "This cluster is mostly a set of unrelated PRs that happen to share Transformers boilerplate (docs TOC, auto registries, model package wiring). They cover different model families and one separate distributed-loading feature, so the soft similarities are false positives rather than duplicates.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
         "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45838 is the broadest model-addition PR in the set, but it is still only a single model-family support change and not a duplicate of the others; there is no strong canonical PR for the cluster.",
+        "best_pr_reason": "PR 45838 is the largest self-contained model-support addition in the set, but it is not actually a duplicate of the others; the cluster does not have a true canonical PR.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:42978",
             "right": "pull_request:43448",
             "accept": false,
-            "reason": "Different model additions (ViT NEPA vs Molmo). Similar docs/auto-registry edits are boilerplate, not the same code-path fix."
+            "reason": "ViT NEPA vs Molmo are different model-addition PRs; overlap is only shared docs/auto wiring."
           },
           {
             "left": "pull_request:42978",
             "right": "pull_request:43451",
             "accept": false,
-            "reason": "Different model families (ViT NEPA vs Molmo2) with separate implementation files and tests; not one underlying change."
+            "reason": "ViT NEPA vs Molmo2 are separate model implementations, not the same underlying change."
           },
           {
             "left": "pull_request:42978",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "ViT NEPA and CTSM are distinct models; overlap is only in generic model-registration/docs work."
+            "reason": "ViT NEPA and CTSM are distinct model-support PRs with unrelated codepaths."
           },
           {
             "left": "pull_request:43448",
             "right": "pull_request:43451",
             "accept": false,
-            "reason": "Molmo and Molmo2 are related in name but are separate model support PRs with different code paths and artifacts."
+            "reason": "Molmo and Molmo2 are related in name only; they add different model families and different support code."
           },
           {
             "left": "pull_request:43448",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "Molmo vs CTSM are unrelated model additions; shared framework wiring is not enough to treat them as duplicates."
+            "reason": "Molmo and CTSM are unrelated model additions; shared registry/docs edits are boilerplate."
           },
           {
             "left": "pull_request:43451",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "Molmo2 and CTSM are separate model families and do not appear to fix the same underlying problem."
+            "reason": "Molmo2 and CTSM are different models with different modeling/configuration changes."
           },
           {
             "left": "pull_request:43973",
             "right": "pull_request:45838",
             "accept": false,
-            "reason": "Audio model support (lfm2_audio) and OCR text-recognition support (PP-OCRv6/PPLCNetV4) are different modalities and changes."
+            "reason": "Lfm2 Audio and PP-OCRv6 text recognition are different modalities and different feature/model stacks."
           },
           {
             "left": "pull_request:44601",
             "right": "pull_request:45838",
             "accept": false,
-            "reason": "Native pipeline-parallel loading support is infrastructure work, not the same underlying change as adding OCR models."
+            "reason": "Pipeline-parallel loading support is a distributed-systems change, not the same bug/change as PP-OCRv6 model support."
           },
           {
             "left": "pull_request:45490",
             "right": "pull_request:45838",
             "accept": false,
-            "reason": "CTSM and PP-OCRv6/PPLCNetV4 are distinct model additions; the overlap is only in shared registry/docs scaffolding."
+            "reason": "CTSM and PP-OCRv6 are separate model additions; only generic docs/auto files overlap."
           }
         ]
       },
@@ -17129,70 +17117,70 @@
         "pull_request:43665|pull_request:45981",
         "pull_request:44815|pull_request:45981"
       ],
-      "prepared_review_unit_hash": "1abf9af43e5becc1ea7abfa1b79fbd986d37d5acf2933a6da109d3a8ef6fdfe1",
+      "prepared_review_unit_hash": "bf9a4f5f20e7a18001cb024aa147875a5af1f87dbd7f4889f1876000048ba476",
       "component_budget": {
         "node_count": 4,
         "item_count": 4,
         "soft_pair_count": 6,
-        "serialized_chars": 9639,
-        "estimated_input_tokens": 2410,
-        "estimated_eval_tokens": 5076
+        "serialized_chars": 9635,
+        "estimated_input_tokens": 2409,
+        "estimated_eval_tokens": 5074
       },
       "budget": {
         "node_count": 4,
         "item_count": 4,
         "soft_pair_count": 6,
-        "serialized_chars": 9639,
-        "estimated_input_tokens": 2410,
-        "estimated_eval_tokens": 5076
+        "serialized_chars": 9635,
+        "estimated_input_tokens": 2409,
+        "estimated_eval_tokens": 5074
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These PRs are not duplicates: they target different models and different code paths, with two being test-only changes and two being unrelated production fixes.",
-        "confidence": 0.96,
+        "summary": "All four PRs are only loosely similar at the title level; they do not appear to fix the same underlying bug or change. The only substantive code fix is the FP8 dequant/loading PR, while the others are unrelated test-only changes or Whisper tokenizer work.",
+        "confidence": 0.9,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #44815 is the strongest canonical candidate because it is a substantive production fix in the FP8/model-loading path, with coordinated code changes across loading, quantization, and model support rather than a test-only tweak.",
+        "canonical_pr_reason": "pull_request:44815 is the strongest representative because it contains the only concrete production code fix, centered on FP8 dequant/loading behavior, with related tests.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #44815 is the best overall representative because it addresses a concrete runtime bug in a real code path and has the clearest end-to-end fix among the set.",
+        "best_pr_reason": "pull_request:44815 is the best global representative of the cluster\u2019s substantive work; the other PRs are either unrelated model tests or Whisper-specific test/tokenization changes.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43291",
             "right": "pull_request:43665",
             "accept": false,
-            "reason": "Different subsystems and bugs: Whisper tokenizer decoding vs. CLIP/ViT test edits. They do not fix the same underlying issue."
+            "reason": "Different targets and symptoms: Whisper tokenizer decoding vs CLIP/ViT test edits. Same broad \"tests\" theme only."
           },
           {
             "left": "pull_request:43291",
             "right": "pull_request:44815",
             "accept": false,
-            "reason": "Whisper decode/timestamp handling is unrelated to FP8 dequant/model-loading behavior."
+            "reason": "Unrelated domains and code paths: Whisper decode logic vs FP8 dequant/loading internals."
           },
           {
             "left": "pull_request:43291",
             "right": "pull_request:45981",
             "accept": false,
-            "reason": "Whisper tokenizer logic and ColQwen2 test device placement are unrelated."
+            "reason": "Whisper tokenizer changes are unrelated to a ColQwen2 test device fix."
           },
           {
             "left": "pull_request:43665",
             "right": "pull_request:44815",
             "accept": false,
-            "reason": "#43665 is test-only noise across CLIP/ViT, while #44815 fixes production FP8 dequant/loading logic; not the same change."
+            "reason": "43665 is test-only and spans CLIP/ViT; 44815 is a production FP8 fix. They do not address the same bug."
           },
           {
             "left": "pull_request:43665",
             "right": "pull_request:45981",
             "accept": false,
-            "reason": "Both touch tests, but for different models and different purposes; no shared underlying bug/change."
+            "reason": "Both are test-related, but they touch different model test suites and different failure modes."
           },
           {
             "left": "pull_request:44815",
             "right": "pull_request:45981",
             "accept": false,
-            "reason": "FP8 dequant/model-loading fix and ColQwen2 test-device adjustment are unrelated."
+            "reason": "FP8 loading/dequantization code changes are unrelated to a ColQwen2 test tensor-device adjustment."
           }
         ]
       },
@@ -17222,12 +17210,12 @@
       "soft_pairs": [
         "pull_request:43488|pull_request:43532"
       ],
-      "prepared_review_unit_hash": "25f8768e9bf6e176ac4f4bbe71fead58b165041b224db08fd1a0f19a71b06972",
+      "prepared_review_unit_hash": "93b9476c93395f3daf0e7392de20c081bf47dc1e4ef2205b5b3fa34a5685e685",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4563,
+        "serialized_chars": 4562,
         "estimated_input_tokens": 1141,
         "estimated_eval_tokens": 2538
       },
@@ -17235,7 +17223,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4563,
+        "serialized_chars": 4562,
         "estimated_input_tokens": 1141,
         "estimated_eval_tokens": 2538
       },
@@ -17244,18 +17232,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These two PRs are not duplicates: one is a small repo-bot/test harness change ([don't merge] bad format to check repo bot), while the other is a broader CI/testing workflow update ([don't merge] Show diff). They only share a test filename and do not fix the same concrete code-path problem.",
-        "confidence": 0.97,
+        "summary": "These two PRs are not duplicates: #43488 is a small repo-bot/check-formatting style change around setup/versioning and a Vit test, while #43532 is a broader workflow/test-harness update affecting CI automation and several files. They do not look like the same underlying change.",
+        "confidence": 0.95,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #43532 is the better representative of the cluster because it is the more substantial, general change spanning workflow and test infrastructure; PR #43488 is a narrow synthetic/repo-bot check with an unrelated version/dependency tweak.",
+        "canonical_pr_reason": "PR #43532 is the more representative standalone change in this cluster because it touches the actual CI/test automation path and spans multiple related files, whereas #43488 is a narrow experimental/check PR.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #43532 is the best standalone PR candidate since it contains the broader functional updates and is less of a throwaway validation PR than #43488.",
+        "best_pr_reason": "#43532 is the best PR candidate to represent the cluster because it is more substantive and coherent as a single code-path change; #43488 is too small and unrelated in scope to serve as the canonical representative.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43488",
             "right": "pull_request:43532",
             "accept": false,
-            "reason": "Different purposes and code paths: #43488 is a repo-bot/test-format exercise with setup/version changes, while #43532 changes CI workflows and testing utilities. Shared Vit test file alone is not enough to treat them as the same change."
+            "reason": "Different underlying purposes and code paths: #43488 is about setup/versioning and a Vit test, while #43532 modifies CI workflow logic, testing utilities, and notification paths. Shared Vit test file is not enough to make them the same change."
           }
         ]
       },
@@ -17285,12 +17273,12 @@
       "soft_pairs": [
         "pull_request:43612|pull_request:43613"
       ],
-      "prepared_review_unit_hash": "965aa8f846e2a5ea6aa39aa48669282fa48ae85a2753406f9c2a9d4a0b5e2237",
+      "prepared_review_unit_hash": "4b13690079e2dd099370f78c27d5bb8ab72845d88103288b6d99d6a23b458f2e",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 6083,
+        "serialized_chars": 6082,
         "estimated_input_tokens": 1521,
         "estimated_eval_tokens": 3298
       },
@@ -17298,7 +17286,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 6083,
+        "serialized_chars": 6082,
         "estimated_input_tokens": 1521,
         "estimated_eval_tokens": 3298
       },
@@ -17307,18 +17295,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are two separate PRs for different features: promptable concept segmentation vs promptable visual segmentation. They share some documentation and registry scaffolding, but the pipeline code, model integrations, and tests target different underlying tasks, so they should not be deduplicated.",
+        "summary": "These are two separate feature PRs with similar scaffolding but different pipeline targets: promptable concept segmentation vs promptable visual segmentation. Shared docs/registry plumbing is not enough to treat them as the same underlying change.",
         "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No canonical PR: the PRs implement different segmentation pipelines with different targets and model/test coverage, so they are not the same change.",
+        "canonical_pr_reason": "No canonical PR: the two PRs implement different pipeline features and touch different model integrations, so neither is a duplicate of the other.",
         "best_issue_reason": null,
-        "best_pr_reason": "Neither PR is a duplicate of the other; if one must be picked as a representative, there is no clear global best because each addresses a distinct feature area.",
+        "best_pr_reason": "PR 43613 is the broader, more established feature change, but it still represents a different pipeline than 43612 and should not be merged as a duplicate representative.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43612",
             "right": "pull_request:43613",
             "accept": false,
-            "reason": "Shared docs and helper files are just shared scaffolding; the PRs add different pipelines (concept vs visual segmentation) with different code paths, model files, and tests, so they are not the same underlying change."
+            "reason": "Reject: these add different pipelines and model integrations (concept segmentation vs visual segmentation), so they are not the same concrete change and would not be merged into one PR."
           }
         ]
       },
@@ -17348,12 +17336,12 @@
       "soft_pairs": [
         "pull_request:43651|pull_request:43663"
       ],
-      "prepared_review_unit_hash": "39318d9fdcbb4acc75a1711dc2df272eedb3e040e75a157a634e34d90de7b7e3",
+      "prepared_review_unit_hash": "dddd36c9880d5e7be004f176771629ef9ac40f1da379074e6fa717c9b37a0f16",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4292,
+        "serialized_chars": 4291,
         "estimated_input_tokens": 1073,
         "estimated_eval_tokens": 2402
       },
@@ -17361,7 +17349,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4292,
+        "serialized_chars": 4291,
         "estimated_input_tokens": 1073,
         "estimated_eval_tokens": 2402
       },
@@ -17370,18 +17358,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are two distinct Trainer extension PRs: one changes gradient-accumulation loss scaling, the other refactors signature-column filtering for custom/PEFT models. They share a file but not the same underlying bug or change.",
-        "confidence": 0.98,
+        "summary": "These are two separate `trainer.py` pull requests with similar extensibility themes but different code paths: one changes gradient-accumulation loss scaling, the other extracts signature-column logic for overrideability. They should not be treated as duplicates.",
+        "confidence": 0.95,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": null,
+        "canonical_pr_reason": "PR 43651 is the more concrete functional change: it alters `training_step` behavior around gradient accumulation loss scaling and adds a dedicated property to control that logic. PR 43663 is a separate refactor/API hook for signature-column filtering.",
         "best_issue_reason": null,
-        "best_pr_reason": "No single best PR for duplication purposes; both are unrelated feature PRs in trainer.py.",
+        "best_pr_reason": "PR 43651 is the best representative of the cluster because it targets a specific runtime behavior fix, while PR 43663 is an unrelated trainer extensibility change despite touching the same file.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43651",
             "right": "pull_request:43663",
             "accept": false,
-            "reason": "Different code paths and goals: 43651 adds `_loss_is_scaled_for_ga` for loss normalization during gradient accumulation, while 43663 adds `_get_signature_columns` to control dataset column filtering. Same file only, not the same change or bug."
+            "reason": "Same file and broad trainer extensibility area, but different underlying changes: GA loss scaling vs. model signature/column filtering. They do not fix the same bug and would not plausibly merge into one PR."
           }
         ]
       },
@@ -17413,7 +17401,7 @@
         "pull_request:43656|pull_request:43842",
         "pull_request:43836|pull_request:43842"
       ],
-      "prepared_review_unit_hash": "a305258014fc434c8f2109f6f770d28dab59d954b7315ff1bb43970a8bf99928",
+      "prepared_review_unit_hash": "71930c34305857e475f8b58f5d295e317c9c9027cf8b0a764d7ed78dfc8d96fd",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
@@ -17435,30 +17423,30 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "The cluster centers on TypeAdapter-related import/annotation failures in `src/transformers/cli/serve.py`. PR #43842 is the clearest, most targeted fix; PR #43836 is close enough to the same bug to treat as a duplicate candidate; PR #43656 looks like a broader, different annotation cleanup rather than the same concrete failure mode.",
-        "confidence": 0.66,
+        "summary": "These PRs cluster around the same CLI `serve.py` runtime `TypeAdapter` annotation NameError. PR 43842 is the cleanest, most targeted fix; the other two are broader or alternate implementations of the same underlying issue.",
+        "confidence": 0.79,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #43842 is the most focused fix for the actual runtime NameError path: it guards the `TypeAdapter` import behind `TYPE_CHECKING` and quotes the annotation, directly addressing the pydantic-optional import case without extra refactoring.",
+        "canonical_pr_reason": "PR 43842 is the best canonical PR: it is narrowly scoped to the reported `TypeAdapter` NameError, has an explicit issue link, and makes the smallest localized change to `serve.py`.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #43842 best represents the underlying bug because it is the narrowest and most explicit TypeAdapter NameError fix in the same file, with minimal unrelated changes.",
+        "best_pr_reason": "PR 43842 most directly fixes the concrete runtime failure with minimal surface area, making it the strongest representative of this bug cluster.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43656",
             "right": "pull_request:43842",
-            "accept": false,
-            "reason": "Both touch `serve.py`, but #43656 appears to be a broader `__future__`/typing cleanup with different issue linkage, while #43842 specifically fixes the pydantic/TypeAdapter runtime NameError. They do not look like the same concrete change."
+            "accept": true,
+            "reason": "Both address the same `serve.py` runtime `TypeAdapter` NameError caused by evaluated annotations; they are alternate fixes for the same code-path problem."
           },
           {
             "left": "pull_request:43836",
             "right": "pull_request:43842",
             "accept": true,
-            "reason": "These are two implementations of the same `TypeAdapter` NameError fix in the same file and against the same issue target. #43836 is broader, but the core bug and code path overlap closely enough that they could plausibly be merged into one PR."
+            "reason": "Both are fixing the same `serve.py` annotation-evaluation NameError around `TypeAdapter`; despite different implementation style, they target the same underlying bug."
           }
         ]
       },
       "evaluator_result": {
         "accept": true,
-        "feedback": "Grounded summary overall. The distinction between #43842 and #43656 is conservative and supported by the diff previews: #43656 looks like a broader typing cleanup, not just the TypeAdapter import failure. The overlap between #43836 and #43842 is also plausible enough to treat as a single bugfix cluster, since both target the same file and issue and could be merged into one PR."
+        "feedback": "The summary is grounded in the packet: all three PRs touch `src/transformers/cli/serve.py`, two explicitly target the same `TypeAdapter` NameError issue, and PR 43842 is the smallest localized fix with an explicit issue link. The pairwise soft-edge acceptances are conservative enough given the titles, file overlap, and shared TypeAdapter/annotation-evaluation context."
       },
       "evaluator_used": true,
       "retried": false,
@@ -17622,8 +17610,8 @@
         "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:44018|pull_request:44161",
@@ -17647,7 +17635,7 @@
         "pull_request:44044|pull_request:44129",
         "pull_request:44044|pull_request:44154"
       ],
-      "prepared_review_unit_hash": "3a18843dbdcbf679afe80e4c75659fb8e8c1fc6a55c2ae41f584380b85582712",
+      "prepared_review_unit_hash": "bd7d6a06897f5bf88e0cd97b9af0a031b28b5fdbf55607e20ed3f38fbad1a23a",
       "component_budget": {
         "node_count": 28,
         "item_count": 28,
@@ -17660,7 +17648,7 @@
         "node_count": 27,
         "item_count": 27,
         "soft_pair_count": 72,
-        "serialized_chars": 75093,
+        "serialized_chars": 75095,
         "estimated_input_tokens": 18774,
         "estimated_eval_tokens": 37804
       },
@@ -17669,348 +17657,348 @@
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "This cluster is mostly a set of separate per-model output-tracing refactors tied to the same umbrella issue, not one broad duplicate. The only clear duplicate groups are the GPT-Neo pair and the GPT-J/CodeGen variants; the rest touch different models and should stay separate.",
-        "confidence": 0.79,
-        "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 44722 is the clearest representative of the shared GPT-J/CodeGen output-tracing refactor: it covers the same concrete code path as the other GPT-J submissions, touches both affected files, and is the most complete version of that change.",
-        "best_issue_reason": null,
-        "best_pr_reason": "PR 44722 is the strongest global representative because it is the most complete GPT-J/CodeGen standardized-tracing change among the overlapping submissions; the remaining PRs are mostly unrelated model-specific refactors.",
+        "summary": "This cluster is mostly a wave of standardized output-tracing refactors across many different model families. Most pairings are not duplicates because they touch different code paths/models; only the GPT-Neo and GPT-J/CodeGen variants clearly overlap as the same underlying refactor.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "Issue #43979 is the shared tracking issue explicitly referenced by the output-tracing refactor PRs, so it is the natural canonical issue for the cluster.",
+        "canonical_pr_reason": "PR #44722 is the strongest representative of the duplicate GPT-J/CodeGen branch: it targets the same output-tracing refactor, shares the key files, and appears to be the most complete version of that concrete change.",
+        "best_issue_reason": "Issue #43979 best represents the cluster because it is the common umbrella for the standardized output-tracing work across models.",
+        "best_pr_reason": "PR #44722 is the best single PR representative because it covers the core GPT-J/CodeGen output-tracing change and most directly reflects the shared refactor direction in the cluster.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43996",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "Different model implementations; same refactor theme, but not the same concrete code path."
+            "reason": "Different model families (CVT/FNet vs GPT-J); shared tracking theme is too broad to be a duplicate."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "CVT/FNet vs DeBERTa are unrelated model implementations."
+            "reason": "Different model families and different concrete code paths; only the output-tracing theme overlaps."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44085",
             "accept": true,
-            "reason": "Both are GPT-J output-tracing refactors on the same block/forward path; 44085 looks like a narrower overlap of the same change."
+            "reason": "Both are GPT-J output-tracing refactors on the same file and same issue; one is just a broader variant of the same underlying change."
           },
           {
             "left": "pull_request:44007",
             "right": "pull_request:44072",
             "accept": false,
-            "reason": "ResNet vs EfficientNet are different model code paths."
+            "reason": "RegNet/ResNet-style refactors are separate concrete model implementations, not the same bug or change."
           },
           {
             "left": "pull_request:44072",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "Different models and different forward paths; only the broad theme overlaps."
+            "reason": "EfficientNet and GPT-J/CodeGen are unrelated code paths despite sharing the standardized-tracing theme."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44072",
             "accept": false,
-            "reason": "GPT-J/CodeGen vs EfficientNet; same pattern, different code paths."
+            "reason": "GPT-J/CodeGen vs EfficientNet is a broad architectural similarity, not the same underlying change."
           },
           {
             "left": "pull_request:44013",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "MobileNetV2 vs DeBERTa; not the same fix."
+            "reason": "MobileNetV2 and DeBERTa are unrelated implementations; this is only a thematic similarity."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-J/CodeGen vs MGP-STR; unrelated concrete changes."
+            "reason": "GPT-J/CodeGen and MGP-STR affect different model code paths and are not the same fix."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-J/CodeGen vs MPT; same refactor style, different models."
+            "reason": "GPT-J/CodeGen and MPT are different model-specific refactors, not one duplicate change."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44068",
             "accept": true,
-            "reason": "Both are GPT-Neo output-tracing refactors in the same file and target the same change surface."
+            "reason": "These are two GPT-Neo output-tracing refactors on the same model/file set and same underlying change."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44068",
             "accept": false,
-            "reason": "GPT-J/CodeGen vs GPT-Neo; different model code paths."
+            "reason": "GPT-J/CodeGen and GPT-Neo are different model families; same refactor pattern is not enough."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44066",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J/CodeGen are separate model implementations, so this is not a duplicate fix."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "ResNet vs GPT-J/CodeGen; not the same underlying change."
+            "reason": "ResNet and GPT-J/CodeGen are different concrete model paths; only the refactor theme overlaps."
           },
           {
             "left": "pull_request:44086",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "MGP-STR vs GPT-J/CodeGen; different model implementations."
+            "reason": "MGP-STR and GPT-J/CodeGen are unrelated model-specific changes."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "ResNet and GPT-J are unrelated concrete implementations."
+            "reason": "ResNet and GPT-J are different code paths; shared issue linkage is not enough."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44019",
             "accept": false,
-            "reason": "GPT-Neo vs ResNet; only the output-tracing theme matches."
+            "reason": "GPT-Neo and ResNet are different model families and changes."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "ResNet vs MPT; different code paths."
+            "reason": "ResNet and MPT are unrelated concrete implementations."
           },
           {
             "left": "pull_request:44068",
             "right": "pull_request:44116",
             "accept": false,
-            "reason": "GPT-Neo vs Flaubert/XLM; different model families."
+            "reason": "GPT-Neo and Flaubert are different model-specific output-tracing refactors."
           },
           {
             "left": "pull_request:44068",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "GPT-Neo vs GPT-J/CodeGen; same theme, not the same fix."
+            "reason": "GPT-Neo and GPT-J/CodeGen are different code paths despite similar refactor language."
           },
           {
             "left": "pull_request:44071",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "MPT vs GPT-J/CodeGen; unrelated concrete changes."
+            "reason": "MPT and GPT-J/CodeGen are unrelated model implementations."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44116",
             "accept": false,
-            "reason": "GPT-Neo vs Flaubert/XLM; not the same underlying bug or change."
+            "reason": "GPT-Neo and Flaubert are different models and separate changes."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-Neo vs MGP-STR; different model paths."
+            "reason": "GPT-Neo and MGP-STR are different concrete implementations."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "GPT-Neo vs GPT-J; same refactor style, different code."
+            "reason": "GPT-Neo and GPT-J are different model families; not the same underlying bug."
           },
           {
-            "left": "pull_request:44018",
-            "right": "pull_request:44071",
-            "accept": false,
-            "reason": "GPT-Neo vs MPT; different model implementations."
-          },
-          {
-            "left": "pull_request:44068",
+            "left": "pull_request:44071",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-Neo vs MGP-STR; unrelated code paths."
+            "reason": "MPT and MGP-STR are unrelated code paths."
           },
           {
             "left": "pull_request:44059",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT2 vs MGP-STR; different models and fixes."
+            "reason": "GPT2 and MGP-STR are different model-specific refactors."
           },
           {
             "left": "pull_request:44085",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-J vs MGP-STR; same tracing theme, different concrete code."
+            "reason": "GPT-J and MGP-STR are unrelated concrete changes."
           },
           {
             "left": "pull_request:44098",
             "right": "pull_request:44101",
             "accept": false,
-            "reason": "ViLT vs Flaubert/XLM; different model families."
+            "reason": "ViLT and XLM/Flaubert are distinct model implementations; same theme only."
           },
           {
             "left": "pull_request:44059",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "GPT-2 vs GPT-J/CodeGen; different concrete code paths."
+            "reason": "GPT2 and GPT-J/CodeGen are different code paths; theme overlap is too broad."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-Neo vs MPT; not the same change."
+            "reason": "GPT-Neo and MPT are separate model-specific refactors."
           },
           {
             "left": "pull_request:44017",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "SegFormer vs GPT-J/CodeGen; same umbrella task, different implementations."
+            "reason": "SegFormer and GPT-J/CodeGen are unrelated model implementations."
           },
           {
             "left": "pull_request:44017",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "SegFormer vs LongT5; different model code paths."
-          },
-          {
-            "left": "pull_request:44017",
-            "right": "pull_request:44073",
-            "accept": false,
-            "reason": "SegFormer vs VisualBert; not the same fix."
+            "reason": "SegFormer and LongT5 are different models and different concrete changes."
           },
           {
             "left": "pull_request:44073",
             "right": "pull_request:44101",
             "accept": false,
-            "reason": "VisualBert vs Flaubert/XLM; different model code."
-          },
-          {
-            "left": "pull_request:44059",
-            "right": "pull_request:44086",
-            "accept": false,
-            "reason": "GPT-2 vs MGP-STR; unrelated implementations."
+            "reason": "VisualBert and Flaubert are different model code paths."
           },
           {
             "left": "pull_request:44056",
-            "right": "pull_request:44101",
+            "right": "pull_request:44161",
             "accept": false,
-            "reason": "MPNet vs Flaubert/XLM; different code paths."
+            "reason": "MPNet and LongT5 are unrelated concrete refactors."
           },
           {
-            "left": "pull_request:44073",
-            "right": "pull_request:44161",
+            "left": "pull_request:43997",
+            "right": "pull_request:44017",
             "accept": false,
-            "reason": "VisualBert vs LongT5; only the refactor pattern overlaps."
+            "reason": "RegNet and SegFormer are different model families; only the output-tracing theme overlaps."
           },
           {
             "left": "pull_request:43997",
-            "right": "pull_request:44017",
+            "right": "pull_request:44056",
             "accept": false,
-            "reason": "RegNet vs SegFormer; different model families."
+            "reason": "RegNet and MPNet are different code paths, not the same bug or change."
           },
           {
             "left": "pull_request:44071",
-            "right": "pull_request:44129",
+            "right": "pull_request:44076",
             "accept": false,
-            "reason": "MPT vs SpeechT5; unrelated concrete changes."
+            "reason": "MPT and ImageGPT are different model-specific changes."
           },
           {
             "left": "pull_request:44074",
-            "right": "pull_request:44161",
+            "right": "pull_request:44101",
             "accept": false,
-            "reason": "TextNet vs LongT5; different model code paths."
+            "reason": "TextNet and Flaubert are unrelated implementations."
           },
           {
-            "left": "pull_request:44056",
-            "right": "pull_request:44073",
+            "left": "pull_request:44018",
+            "right": "pull_request:44059",
             "accept": false,
-            "reason": "MPNet vs VisualBert; different models."
+            "reason": "GPT-Neo and GPT2 are separate model families."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44074",
+            "left": "pull_request:44017",
+            "right": "pull_request:44071",
             "accept": false,
-            "reason": "RegNet vs TextNet; not the same underlying change."
+            "reason": "SegFormer and MPT are different concrete changes."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44018",
+            "left": "pull_request:44017",
+            "right": "pull_request:44074",
             "accept": false,
-            "reason": "RegNet vs GPT-Neo; different concrete code paths."
+            "reason": "SegFormer and TextNet are unrelated model implementations."
           },
           {
-            "left": "pull_request:44071",
-            "right": "pull_request:44101",
+            "left": "pull_request:44059",
+            "right": "pull_request:44068",
             "accept": false,
-            "reason": "MPT vs Flaubert/XLM; same tracing style, different implementations."
+            "reason": "GPT2 and GPT-Neo are different model families and code paths."
           },
           {
-            "left": "pull_request:44010",
+            "left": "pull_request:44059",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "SqueezeBert vs MPT; different model families."
+            "reason": "GPT2 and MPT are unrelated concrete changes."
           },
           {
-            "left": "pull_request:44018",
-            "right": "pull_request:44129",
+            "left": "pull_request:44101",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "Flaubert/XLM and LongT5 are different model implementations."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "RegNet and TextNet are unrelated model-specific refactors."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44018",
             "accept": false,
-            "reason": "GPT-Neo vs SpeechT5; not the same fix."
+            "reason": "RegNet and GPT-Neo are different code paths."
           },
           {
             "left": "pull_request:44010",
-            "right": "pull_request:44073",
+            "right": "pull_request:44044",
             "accept": false,
-            "reason": "SqueezeBert vs VisualBert; different code paths."
+            "reason": "SqueezeBert and DeBERTa are different model families and refactors."
           },
           {
-            "left": "pull_request:44074",
-            "right": "pull_request:44161",
+            "left": "pull_request:44044",
+            "right": "pull_request:44074",
             "accept": false,
-            "reason": "TextNet vs LongT5; unrelated implementations."
+            "reason": "DeBERTa and TextNet are unrelated concrete changes."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44161",
+            "left": "pull_request:44018",
+            "right": "pull_request:44129",
             "accept": false,
-            "reason": "RegNet vs LongT5; different model code."
+            "reason": "GPT-Neo and SpeechT5 are different model code paths."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44071",
+            "left": "pull_request:44010",
+            "right": "pull_request:44073",
             "accept": false,
-            "reason": "RegNet vs MPT; not a duplicate change."
+            "reason": "SqueezeBert and VisualBert are different implementations."
           },
           {
             "left": "pull_request:44010",
             "right": "pull_request:44018",
             "accept": false,
-            "reason": "SqueezeBert vs GPT-Neo; different underlying code paths."
+            "reason": "SqueezeBert and GPT-Neo are different model families."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "CVT/FNet vs LongT5; only the output-tracing theme is shared."
+            "reason": "CVT/FNet and LongT5 are unrelated concrete changes."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44017",
             "accept": false,
-            "reason": "CVT/FNet vs SegFormer; different model implementations."
+            "reason": "CVT/FNet and SegFormer are different code paths."
           },
           {
             "left": "pull_request:44010",
-            "right": "pull_request:44044",
+            "right": "pull_request:44161",
             "accept": false,
-            "reason": "SqueezeBert vs DeBERTa; unrelated concrete fixes."
+            "reason": "SqueezeBert and LongT5 are unrelated model-specific changes."
           },
           {
             "left": "pull_request:44044",
-            "right": "pull_request:44074",
+            "right": "pull_request:44129",
             "accept": false,
-            "reason": "DeBERTa vs TextNet; different model code paths."
+            "reason": "DeBERTa and SpeechT5 are different model implementations."
           },
           {
             "left": "pull_request:44044",
             "right": "pull_request:44154",
             "accept": false,
-            "reason": "DeBERTa vs VITS; different implementations and changes."
+            "reason": "DeBERTa and ViTS are unrelated concrete refactors."
           }
         ]
       },
       "evaluator_result": {
         "accept": true,
-        "feedback": "The analysis is mostly grounded in the packet and stays conservative on the soft-edge pairs. The GPT-Neo and GPT-J/CodeGen overlaps are supported by shared files/model paths, and the remaining refusals are reasonable given different model families. The only slightly stronger claim is that PR 44722 is the \u201cmost complete\u201d version, but that is still broadly consistent with the packet and not obviously overstated."
+        "feedback": "The summary is grounded in the packet, and the duplicate/near-duplicate calls stay conservative. The selected canonical issue/PR are supported by shared issue linkage and overlapping GPT-J/CodeGen files, and the false edges are mostly justified by distinct model families or code paths."
       },
       "evaluator_used": true,
       "retried": false,
@@ -18048,7 +18036,7 @@
         "pull_request:43997|pull_request:44010",
         "pull_request:44010|pull_request:44129"
       ],
-      "prepared_review_unit_hash": "42f1a8f490d927747386a934078b67a2c62c2da27d7e2e3ec228190001036a96",
+      "prepared_review_unit_hash": "57e47d7c258906794f23d431761dea9f72023560b4f726b2dd09d2dca4d80dfd",
       "component_budget": {
         "node_count": 28,
         "item_count": 28,
@@ -18061,7 +18049,7 @@
         "node_count": 8,
         "item_count": 8,
         "soft_pair_count": 6,
-        "serialized_chars": 17124,
+        "serialized_chars": 17122,
         "estimated_input_tokens": 4281,
         "estimated_eval_tokens": 8818
       },
@@ -18070,48 +18058,48 @@
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "A cluster of model-specific PRs that all migrate different architectures to standardized output tracing/capture decorators, but they touch distinct code paths and should not be merged as duplicates.",
-        "confidence": 0.84,
+        "summary": "These are all model-specific PRs modernizing output tracing to use capture_outputs/can_return_tuple patterns. They share a theme, but each changes a different model implementation, so they are not duplicates of the same underlying fix/change.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 44129 is the strongest representative: it has an explicit issue target, a clear refactor title, and a concrete implementation of the standardized output capture pattern in SpeechT5.",
+        "canonical_pr_reason": "PR 44129 is the best representative of the cluster because it has an explicit issue target and shows a concrete, self-contained output-tracing refactor for one model; the others are similar but are separate per-model variants.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 44129 is the best single representative of the cluster\u2019s theme, but the other PRs are separate model migrations rather than the same fix.",
+        "best_pr_reason": "PR 44129 is the strongest standalone PR in the set: it is clearly scoped, tied to a target issue, and implements the standardized output-capture migration in a concrete code path.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44024",
             "right": "pull_request:44076",
             "accept": false,
-            "reason": "Both are output-capture refactors, but they are for different model families (FocalNet vs ImageGPT) with different forward paths and hook mappings; not the same concrete change."
+            "reason": "Same refactor theme, but different models and code paths (FocalNet vs ImageGPT); not the same underlying change."
           },
           {
             "left": "pull_request:44073",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "VisualBert and TextNet are separate model implementations with different output structures; this is the same refactor pattern, not the same underlying bug or change."
+            "reason": "Both modernize output tracing, but they touch different model implementations (VisualBert vs TextNet) and are not mergeable as one PR."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44073",
             "accept": false,
-            "reason": "MPNet and VisualBert each migrate their own model-specific output tracing logic; they are not the same code-path fix."
+            "reason": "Related output-capture refactors, but MPNet and VisualBert are separate code paths with distinct implementation details."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "MPNet and TextNet touch different model forward implementations and output-recording definitions, so they are not duplicates."
+            "reason": "Same broad standardization effort, but MPNet and TextNet are different model-specific changes."
           },
           {
             "left": "pull_request:43997",
             "right": "pull_request:44010",
             "accept": false,
-            "reason": "RegNet and SqueezeBert are unrelated model code paths; both use the new decorator pattern, but they do not fix the same underlying problem."
+            "reason": "Both are output-tracing migrations, but RegNet and SqueezeBert are unrelated model implementations and not the same fix."
           },
           {
             "left": "pull_request:44010",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "SqueezeBert and SpeechT5 are separate architectures with different attention/output behavior; these are distinct migrations, not one mergeable PR."
+            "reason": "Different architectures and files (SqueezeBert vs SpeechT5); they share a modernization pattern but not the same concrete bug/change."
           }
         ]
       },
@@ -18191,274 +18179,268 @@
         "pull_request:44028|pull_request:44030",
         "pull_request:44029|pull_request:44030"
       ],
-      "prepared_review_unit_hash": "ff89e64d74767482cb2ef3f78cfd3e86d8964ef5fd342a9dadaec3fbcb8a371a",
+      "prepared_review_unit_hash": "e0fc149cba7cdefdaf54a1a3d2e515a6b8cb5e2a8bf1bbc1c7ac5233c6d053aa",
       "component_budget": {
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 40,
-        "serialized_chars": 40170,
-        "estimated_input_tokens": 10043,
-        "estimated_eval_tokens": 20342
+        "serialized_chars": 40178,
+        "estimated_input_tokens": 10045,
+        "estimated_eval_tokens": 20346
       },
       "budget": {
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 40,
-        "serialized_chars": 40170,
-        "estimated_input_tokens": 10043,
-        "estimated_eval_tokens": 20342
+        "serialized_chars": 40178,
+        "estimated_input_tokens": 10045,
+        "estimated_eval_tokens": 20346
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "This cluster is a bundle of similar but distinct per-model output-tracing refactors. They share helper usage and wording, but they target different model families and code paths, so they are not duplicates.",
-        "confidence": 0.91,
+        "summary": "These PRs are all variants of the same broad refactor theme ('output tracing'), but they target different model implementations and forward paths. I do not see any pair that looks like the same concrete bug/change, so none should be merged as duplicates.",
+        "confidence": 0.87,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "44003 is the broadest single refactor in the set, spanning both Mamba and FalconMamba and adding the tracing helpers/capture path, so it best represents the cluster.",
+        "canonical_pr_reason": "PR 44025 is the cleanest representative of the cluster: it applies the tracing decorator pattern to two closely related depth-anything model files and stays narrowly scoped.",
         "best_issue_reason": null,
-        "best_pr_reason": "44003 is the most substantive and representative example of this cluster\u2019s output-tracing refactor pattern.",
+        "best_pr_reason": "44025 is the best overall representative because it is self-contained, model-specific, and shows the core output-tracing refactor with minimal extra scope.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43998",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different model families and forward paths; both refactor output handling, but not the same concrete change."
+            "reason": "Different model families and code paths (timm_backbone vs upernet); only the refactor pattern is shared."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44000",
             "accept": false,
-            "reason": "Separate architectures (`timm_backbone` vs `vision_text_dual_encoder`) with different refactor scopes."
+            "reason": "Different architectures and forward implementations; same tracing cleanup theme only."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Same general theme, but `timm_backbone` and `univnet` are unrelated code paths."
+            "reason": "Separate model code paths; not the same underlying bug/change."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Both are output-tracing refactors, but they affect different model implementations and are not mergeable as one bug fix."
+            "reason": "timm_backbone and mamba/falcon_mamba are unrelated concrete code paths despite similar refactor wording."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "`timm_backbone` and `rwkv` are distinct model families; this is only thematic similarity."
+            "reason": "Different model families; no evidence of one shared fix."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Different models and tests; same refactor pattern, not the same underlying change."
+            "reason": "MobileNetV1 and UnivNet are unrelated implementations; the similarity is only at the refactor level."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "`mobilenet_v1` and `upernet` use different forward signatures and output paths."
+            "reason": "Different model code paths and behavior; not the same change."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Unrelated model families; same tracing cleanup style, but no shared code-path bug."
+            "reason": "MobileNetV1 vs DepthAnything are distinct models; same output-tracing pattern is too broad to duplicate."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44026",
             "accept": false,
-            "reason": "Different wrapper models; both adjust output tracing, but independently."
+            "reason": "Different encoder-decoder targets and code paths; only the wrapper refactor theme matches."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "`mobilenet_v1` and `speech_encoder_decoder` are separate implementations with separate refactors."
+            "reason": "MobileNetV1 and SpeechEncoderDecoder are separate model implementations with no shared concrete bug."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "`vision_text_dual_encoder` and `univnet` are unrelated architectures; not the same fix."
+            "reason": "Vision-text dual encoder and UnivNet are unrelated code paths; shared wording is not enough."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different model code paths and output contracts; only the refactor pattern matches."
+            "reason": "Different model families and distinct forward logic."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different model families and helper usage; no shared concrete bug or change."
+            "reason": "The PRs touch different architectures and fix different model-specific tracing paths."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Separate model families; both touch output tracing, but the underlying implementations differ."
+            "reason": "Different models and output pathways; not the same bug/change."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "`univnet` and `upernet` are unrelated models with distinct forward logic."
+            "reason": "UnivNet and UperNet are unrelated; only the refactor pattern is common."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different code paths in different model families; not one merged change."
+            "reason": "Different model families and implementation details; not mergeable as one fix."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Same broad refactor idea, but `univnet` and `depth_anything` are separate implementations."
+            "reason": "Different architectures; same tracing cleanup pattern is too generic."
           },
           {
             "left": "pull_request:44002",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "`upernet` segmentation and `rwkv` language modeling are unrelated concrete changes."
+            "reason": "UperNet and RWKV are separate code paths with no concrete shared bug."
           },
           {
             "left": "pull_request:44002",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different model families and helper plumbing; no single bug fix spans both."
+            "reason": "Different model subsystems and different forward behavior."
           },
           {
             "left": "pull_request:44003",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "`mamba`/`falcon_mamba` output capturing is unrelated to `superpoint` feature extraction."
+            "reason": "Falcon/Mamba vs SuperPoint are unrelated model paths; no same underlying change."
           },
           {
             "left": "pull_request:44003",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Different internals: `mamba` state/capture refactor versus `codegen` cache/output signature changes."
+            "reason": "Mamba-family refactor and CodeGen refactor are different concrete code paths."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "`timm_backbone` and `codegen` are unrelated model paths; only the naming pattern is shared."
+            "reason": "TimmBackbone and CodeGen are unrelated implementations; same broad refactor theme only."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Different models and different output semantics; not the same concrete issue."
+            "reason": "UnivNet and CodeGen touch different model logic and cannot be treated as one duplicate fix."
           },
           {
             "left": "pull_request:44002",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "`upernet` and `codegen` have separate output contracts and code paths."
-          },
-          {
-            "left": "pull_request:44000",
-            "right": "pull_request:44004",
-            "accept": false,
-            "reason": "Different models and different refactor targets; not one duplicate change."
+            "reason": "UperNet and CodeGen are unrelated model paths."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44026",
             "accept": false,
-            "reason": "Both are encoder/decoder-style wrappers, but they are distinct models with separate forward implementations."
+            "reason": "DepthAnything and VisionEncoderDecoder are different model wrappers; same decorator pattern is not enough."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "`depth_anything` and `speech_encoder_decoder` are separate architectures and separate refactors."
+            "reason": "DepthAnything and SpeechEncoderDecoder are distinct code paths and model families."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different tasks and model code paths; same output-tracing theme only."
+            "reason": "DepthAnything vs SuperPoint: different architectures and different output handling details."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "`depth_anything` and `rwkv` are unrelated changes despite similar helper adoption."
+            "reason": "DepthAnything and RWKV are unrelated concrete changes."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Different model families and output contracts; not the same underlying fix."
+            "reason": "DepthAnything and DPR are separate model implementations; no shared underlying bug."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "`vision_encoder_decoder` and `rwkv` are unrelated code paths."
+            "reason": "VisionEncoderDecoder and RWKV are unrelated code paths."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Two different encoder-decoder wrappers; similar refactor style, but separate changes."
+            "reason": "VisionEncoderDecoder and SpeechEncoderDecoder are different wrapper implementations, not the same bug."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different model families and outputs; not a shared concrete bug."
+            "reason": "Different model families and forward paths; shared refactor style only."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "`vision_encoder_decoder` and `dpr` do not share the same code-path fix."
+            "reason": "VisionEncoderDecoder and DPR are unrelated concrete changes."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "`speech_encoder_decoder` and `superpoint` are unrelated models with different forward behavior."
+            "reason": "SpeechEncoderDecoder and SuperPoint are distinct model code paths."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Separate architectures and separate output-tracing refactors."
+            "reason": "Different model families and output semantics; not the same fix."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "`speech_encoder_decoder` and `dpr` differ in both model type and output handling."
+            "reason": "SpeechEncoderDecoder and DPR are unrelated implementations."
           },
           {
             "left": "pull_request:44028",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "`superpoint` feature extraction and `rwkv` sequence modeling are unrelated changes."
+            "reason": "SuperPoint and RWKV are unrelated model paths."
           },
           {
             "left": "pull_request:44028",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Different model families and code paths; same refactor motif only."
+            "reason": "SuperPoint and DPR do not share a concrete bug/change."
           },
           {
             "left": "pull_request:44029",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "`rwkv` and `dpr` are separate implementations with different output contracts."
+            "reason": "RWKV and DPR are unrelated model implementations; same broad refactor wording is insufficient."
           }
         ]
       },
@@ -18491,12 +18473,12 @@
         "pull_request:44662|pull_request:45075",
         "pull_request:45075|pull_request:45401"
       ],
-      "prepared_review_unit_hash": "84b0749d1b8a741c360f0e26ec4986c3de6bb4d6066311221653e8496a314d4f",
+      "prepared_review_unit_hash": "4b9104eea730c11d022d34a781a6ceafe7f5a03da7af449298ede5be27a17716",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 10912,
+        "serialized_chars": 10911,
         "estimated_input_tokens": 2728,
         "estimated_eval_tokens": 5712
       },
@@ -18504,7 +18486,7 @@
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 10912,
+        "serialized_chars": 10911,
         "estimated_input_tokens": 2728,
         "estimated_eval_tokens": 5712
       },
@@ -18513,30 +18495,30 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are three separate model-implementation PRs, not duplicates. They overlap on standard Transformer boilerplate (auto mappings, docs, package exports), but each adds a different model family and code path: PenguinVL, DeepSeek-OCR-2, and Voxtral TTS.",
-        "confidence": 0.96,
+        "summary": "These are three unrelated model-addition PRs (PenguinVL, DeepSeek-OCR-2, Voxtral TTS). They share only generic Transformers plumbing and docs registration, not the same underlying bug or change.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45075 is the strongest standalone candidate to anchor the cluster if one must be chosen: it is non-draft, relatively mature, and has the broadest review activity. It is not a duplicate of the others, just the most developed implementation PR.",
+        "canonical_pr_reason": "No true canonical PR for deduping here; the cluster is mixed. If one must be singled out, DeepSeek-OCR-2 is the most complete standalone implementation and not draft.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45075 appears to be the best overall standalone PR in this set because it is a complete, non-draft model addition with substantial review activity and full model/test/docs coverage. The other two are separate model additions rather than competing fixes.",
+        "best_pr_reason": "DeepSeek-OCR-2 is the strongest overall candidate: non-draft, substantial model + conversion + test coverage, and appears more merge-ready than the draft Voxtral TTS PR.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44662",
             "right": "pull_request:45401",
             "accept": false,
-            "reason": "Different model implementations (PenguinVL vs Voxtral TTS) with only shared registration/docs boilerplate; no same bug or mergeable single code-path fix."
+            "reason": "Different models and different additions (PenguinVL vs Voxtral TTS). Shared auto/docs files are generic scaffolding, not the same code-path fix."
           },
           {
             "left": "pull_request:44662",
             "right": "pull_request:45075",
             "accept": false,
-            "reason": "Both add different models and touch shared auto-mapping infrastructure, but they do not target the same underlying change or defect."
+            "reason": "Different model implementations (PenguinVL vs DeepSeek-OCR-2) with separate architecture, processing, and test code; only common Transformers registration plumbing overlaps."
           },
           {
             "left": "pull_request:45075",
             "right": "pull_request:45401",
             "accept": false,
-            "reason": "DeepSeek-OCR-2 and Voxtral TTS are unrelated model additions; overlap is limited to generic package/auto-registry updates."
+            "reason": "Different model families and unrelated functionality (DeepSeek-OCR-2 vs Voxtral TTS). Similar filenames are broad integration points, not evidence of one mergeable change."
           }
         ]
       },
@@ -18566,12 +18548,12 @@
       "soft_pairs": [
         "pull_request:44979|pull_request:45363"
       ],
-      "prepared_review_unit_hash": "9487f2e3cff1d15315b36643f0e4554d685bdd0d7e9ddefffbed1296a490ba34",
+      "prepared_review_unit_hash": "353f85c8de0b7494a87319632d23f91a56181bd3fd07362d0755208ebfc4018f",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4408,
+        "serialized_chars": 4407,
         "estimated_input_tokens": 1102,
         "estimated_eval_tokens": 2460
       },
@@ -18579,7 +18561,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4408,
+        "serialized_chars": 4407,
         "estimated_input_tokens": 1102,
         "estimated_eval_tokens": 2460
       },
@@ -18588,18 +18570,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These PRs are related in the broad area of module/kernel fusion, but they are not the same change: one introduces a general Module Fusion API, while the other wires up n-to-1 kernel fusion through KernelConfig and integration points.",
-        "confidence": 0.78,
+        "summary": "These are related around fusion, but they target different mechanisms: one introduces a generic Module Fusion API, while the other implements n-to-1 kernel fusion through KernelConfig and integration changes. They are not the same underlying change and should not be merged as duplicates.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #45363 is the more specific, user-facing feature implementation and spans the actual KernelConfig fusion path; PR #44979 is a lower-level foundational API that could support multiple future uses.",
+        "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45363 is the best representative of the cluster because it concretely implements a kernel-fusion workflow, whereas #44979 is a generic API scaffold rather than the full feature.",
+        "best_pr_reason": "PR 45363 is the more concrete end-user feature: it wires kernel fusion into model loading/integration and includes an example. PR 44979 is a separate API layer for module chaining/fusion, so 45363 is the better representative if one PR must be selected.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44979",
             "right": "pull_request:45363",
             "accept": false,
-            "reason": "They share the fusion theme, but #44979 adds a generic Module Fusion API and tests, while #45363 implements a specific KernelConfig-based n-to-1 kernel fusion integration. Different scope and code paths, so not a duplicate PR."
+            "reason": "Both involve fusion, but they address different code paths and abstractions. 44979 adds a generic module-fusion API in a new module, while 45363 adds KernelConfig-driven kernel fusion and integration hooks. They could coexist, but they are not the same bug/change."
           }
         ]
       },
@@ -18629,12 +18611,12 @@
       "soft_pairs": [
         "pull_request:45086|pull_request:45317"
       ],
-      "prepared_review_unit_hash": "faf827e13db9e2badbd0e008d7e2cd70404de91b71cea681db63954609702329",
+      "prepared_review_unit_hash": "2bb7317ae619e4e6dbe73d4f9ebccf463c69e4a2add67e84e2a1117d56b78e00",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4413,
+        "serialized_chars": 4414,
         "estimated_input_tokens": 1104,
         "estimated_eval_tokens": 2464
       },
@@ -18642,7 +18624,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4413,
+        "serialized_chars": 4414,
         "estimated_input_tokens": 1104,
         "estimated_eval_tokens": 2464
       },
@@ -18651,18 +18633,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both pull requests address the same Mistral regex AttributeError in `src/transformers/tokenization_utils_tokenizers.py`, with nearly identical code changes; 45317 is the more complete version because it adds tests as well.",
-        "confidence": 0.98,
+        "summary": "Both PRs target the same Mistral regex patch bug in the same file; 45317 is the stronger candidate because it includes the same fix plus tests and a clearer scope.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45317 is the better canonical PR: it implements the same fix as 45086 but also includes test coverage and appears to be the more complete patch.",
+        "canonical_pr_reason": "PR 45317 is the best canonical PR: it fixes the same AttributeError as 45086, has the same code change, and adds test coverage for the behavior.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45317 is the best overall representative because it contains the same core fix plus validation/tests, making it the stronger standalone change.",
+        "best_pr_reason": "PR 45317 is the best overall PR because it is the same fix as 45086 but more complete, with accompanying tests and clearer intent.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45086",
             "right": "pull_request:45317",
             "accept": true,
-            "reason": "Same underlying AttributeError fix in the same code path, with matching file and nearly identical diff; 45317 is essentially a superset via added tests."
+            "reason": "Same underlying AttributeError fix in _patch_mistral_regex on the same code path; 45317 is effectively the more complete version of 45086 and they could plausibly be merged as one PR."
           }
         ]
       },
@@ -18692,12 +18674,12 @@
       "soft_pairs": [
         "pull_request:45115|pull_request:45569"
       ],
-      "prepared_review_unit_hash": "302e5608d0f2d900c77e0ed5388ed9a88d7a93da2bd1a205a62fd181eaaecb18",
+      "prepared_review_unit_hash": "0fc5e678f311b76f1b7f3e3403137bd96f580f7492baaa76dc099ada7873cecb",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 6065,
+        "serialized_chars": 6066,
         "estimated_input_tokens": 1517,
         "estimated_eval_tokens": 3290
       },
@@ -18705,7 +18687,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 6065,
+        "serialized_chars": 6066,
         "estimated_input_tokens": 1517,
         "estimated_eval_tokens": 3290
       },
@@ -18714,18 +18696,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both pull requests touch NemotronH internals, but they are different changes: one is a smaller inheritance/conversion refactor, the other is a broader model split and documentation/test update. They are related, not duplicates.",
-        "confidence": 0.87,
+        "summary": "Both PRs touch NemotronH internals, but they are different changes: #45115 is a refactor/conversion-mapping inheritance cleanup, while #45569 is a broader model split into NemotronHDense/NemotronHSparse with docs and tests. They are related but not the same underlying change.",
+        "confidence": 0.91,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45569 is the broader, more complete NemotronH architecture change and is the better representative of the cluster.",
+        "canonical_pr_reason": "PR #45569 is the broader, more representative NemotronH change: it introduces the dense/sparse model split, updates auto-mapping/docs, and includes tests. It better captures the main architectural change in this cluster.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45569 is the strongest standalone PR here because it covers the full model split and accompanying API/docs/test updates, whereas 45115 is a narrower refactor.",
+        "best_pr_reason": "PR #45569 is the best single representative because it is the more complete and user-visible NemotronH update, whereas #45115 is a narrower refactor focused on conversion mapping and inheritance details.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45115",
             "right": "pull_request:45569",
             "accept": false,
-            "reason": "Shared files are in the same model family, but the PRs target different scopes: 45115 is a converter/inheritance refactor, while 45569 introduces the fuller NemotronH Dense/Sparse split plus docs/tests. They do not look like the same concrete fix and should not be merged as duplicates."
+            "reason": "They share NemotronH files, but the concrete changes differ: #45115 refactors inheritance/conversion mappings, while #45569 splits NemotronH into dense/sparse variants with docs and tests. Not the same code-path fix and not plausibly one merged PR."
           }
         ]
       },
@@ -18755,12 +18737,12 @@
       "soft_pairs": [
         "pull_request:45332|pull_request:45333"
       ],
-      "prepared_review_unit_hash": "ae4a3c1efef146351d2cba86c8ed59d2704c523edfd4611b8a59542215cc9b59",
+      "prepared_review_unit_hash": "5069797d27ef4cf3ab18fdd602a819f9ce97aeda6762560bc7483f40b572ed1b",
       "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
       },
@@ -18768,7 +18750,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5398,
+        "serialized_chars": 5397,
         "estimated_input_tokens": 1350,
         "estimated_eval_tokens": 2956
       },
@@ -18777,18 +18759,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related heterogeneous-config/modeling feature PRs, but they are not duplicates: one adds per-layer configuration plumbing, the other extends that into model/cache behavior and integrations.",
-        "confidence": 0.9,
+        "summary": "Two open PRs for heterogeneous model support: one broad end-to-end implementation and one smaller config-only foundation. They are related, but not the same concrete change.",
+        "confidence": 0.91,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45332 is the broader umbrella change and the better cluster representative: it includes the heterogeneous config machinery plus downstream modeling/cache integration, while 45333 is narrower config-only plumbing.",
+        "canonical_pr_reason": "PR #45332 is the broader umbrella change: it includes heterogeneous config handling plus model/cache/masking and GPT-OSS integration, so it best represents the overall feature effort.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45332 best represents the underlying feature work because it covers the full end-to-end heterogeneous model support, not just config serialization and accessors.",
+        "best_pr_reason": "PR #45332 is the best representative PR because it covers the full heterogeneous-model support stack and subsumes the narrower config-only work in #45333.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45332",
             "right": "pull_request:45333",
             "accept": false,
-            "reason": "Related feature work, but not the same concrete change. 45333 adds heterogeneous per-layer config support, while 45332 builds on that with model/cache/masking integration and GPT-OSS changes; they could not plausibly be merged as a single duplicate PR."
+            "reason": "Related feature slices, but not the same concrete change. #45333 is config-only per-layer configuration support, while #45332 extends into modeling/cache/masking and model integration; they could ship independently and are not safe to treat as duplicates."
           }
         ]
       },
@@ -18818,12 +18800,12 @@
       "soft_pairs": [
         "pull_request:45415|pull_request:45425"
       ],
-      "prepared_review_unit_hash": "fa945d7cabcb26e6211726d03e09d08a31d73f3035c5483cf6cacd5ceefdfc41",
+      "prepared_review_unit_hash": "44f2fb48ed7693da5d93ccff5fc9274216dc7986da13657706bfbe31082d895f",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5513,
+        "serialized_chars": 5514,
         "estimated_input_tokens": 1379,
         "estimated_eval_tokens": 3014
       },
@@ -18831,7 +18813,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5513,
+        "serialized_chars": 5514,
         "estimated_input_tokens": 1379,
         "estimated_eval_tokens": 3014
       },
@@ -18840,18 +18822,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are both typing-related PRs, but they address different scopes: one is a broad repo-wide type-checking sweep, the other adds specific internal typing protocols/aliases for modeling and PEFT integration. They should not be treated as duplicates.",
+        "summary": "These are related typing-focused PRs, but they are not duplicates: #45415 is a broad repo-wide type-checking cleanup and config change, while #45425 adds specific typing protocols/aliases and updates a few concrete call sites. They could be merged independently.",
         "confidence": 0.89,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No issue is present; if a PR must represent the cluster, #45425 is the better canonical choice because it is the more concrete, narrowly-scoped typing change, while #45415 is a broad type-checking sweep across many files.",
+        "canonical_pr_reason": "#45425 is the more focused and concrete change: it adds missing typing abstractions and updates the affected implementation sites, rather than making a broad type-checking sweep.",
         "best_issue_reason": null,
-        "best_pr_reason": "#45425 is the strongest representative of the cluster because it makes targeted typing infrastructure changes in a few related modules. #45415 is much broader and mostly about enabling type checking/ignores across many files, so it is less suitable as the cluster anchor.",
+        "best_pr_reason": "#45425 is the better representative PR for this cluster because it targets a specific typing gap and is less sprawling than the large type-checking enablement in #45415.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45415",
             "right": "pull_request:45425",
             "accept": false,
-            "reason": "Different underlying changes: #45415 is a broad type-checking pass with pyproject ignore adjustments and many file edits, while #45425 adds specific typing protocols/type aliases for modeling_utils and PEFT-related code. Shared typing files are not enough to make them the same PR."
+            "reason": "Shared typing-related files are incidental; #45415 is a broad type-checking/config cleanup, while #45425 introduces specific typing protocols and aliases for modeling_utils/peft-related code. They do not look like the same underlying change and should not be merged as duplicates."
           }
         ]
       },
@@ -18890,12 +18872,12 @@
         "pull_request:45858|pull_request:45926",
         "pull_request:45890|pull_request:45926"
       ],
-      "prepared_review_unit_hash": "57025ce7afbbd73636dab7988c5e89573a7808e8934167d18beb8a74467d9699",
+      "prepared_review_unit_hash": "df0d4c7d66cd44a67337ae5153d24069ca9e38a7a7ea64bac87e0b938c7e8512",
       "component_budget": {
         "node_count": 6,
         "item_count": 6,
         "soft_pair_count": 6,
-        "serialized_chars": 14720,
+        "serialized_chars": 14717,
         "estimated_input_tokens": 3680,
         "estimated_eval_tokens": 7616
       },
@@ -18903,7 +18885,7 @@
         "node_count": 6,
         "item_count": 6,
         "soft_pair_count": 6,
-        "serialized_chars": 14720,
+        "serialized_chars": 14717,
         "estimated_input_tokens": 3680,
         "estimated_eval_tokens": 7616
       },
@@ -18912,48 +18894,48 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These PRs are mostly unrelated: three are real model-parallel fixes for different models, and three are separate XPU test-expectation updates for different models. No soft pair looks like the same underlying change.",
-        "confidence": 0.89,
+        "summary": "No true duplicates here: the cluster only shares a broad theme of model-parallel or XPU-related changes, but the PRs target different models and different code paths. The most representative runtime fix is the AltCLIP/ChineseCLIP device-placement PR.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "pull_request:45487 is the strongest representative because it contains an actual code fix for model-parallel device placement across several related text/vision models, rather than just test-data updates.",
+        "canonical_pr_reason": "PR 45487 is the clearest substantive bug fix in model code, addressing device placement for token_type_ids in AltCLIP/ChineseCLIP rather than only adjusting tests or touching unrelated model internals.",
         "best_issue_reason": null,
-        "best_pr_reason": "pull_request:45487 is the best standalone PR in this set: it addresses a concrete bug, has substantive code changes, and is broader than the other code-fix PRs while still being a single coherent fix.",
+        "best_pr_reason": "PR 45487 is the best representative of the cluster because it fixes an actual model-parallel runtime issue and includes focused tests; the other PRs are either different models, unrelated refactors, or test-only XPU expectation updates.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45487",
             "right": "pull_request:45843",
             "accept": false,
-            "reason": "Both are model-parallel fixes, but for different model families and different code paths; they are not the same underlying bug and would not plausibly be merged into one PR."
+            "reason": "Different model families and different fixes; shared 'model parallel' theme is too broad to treat as the same change."
           },
           {
             "left": "pull_request:45487",
             "right": "pull_request:45926",
             "accept": false,
-            "reason": "45487 fixes runtime device-placement issues; 45926 only adds XPU test expectations for a different model. Same broad area, but not the same change."
+            "reason": "AltCLIP/ChineseCLIP runtime device-placement fix vs HyperCLOVAX XPU test expectations; not the same underlying bug."
           },
           {
             "left": "pull_request:45817",
             "right": "pull_request:45843",
             "accept": false,
-            "reason": "Distinct model-specific parallelism bugs in Gemma4 vs Deimv2; no shared concrete code path or single combined fix."
+            "reason": "Gemma4 model-parallel helper/device-placement fix is unrelated to the DeIMv2 changes shown here."
           },
           {
             "left": "pull_request:45858",
             "right": "pull_request:45890",
             "accept": false,
-            "reason": "Both are test-only XPU expectation updates, but they target different models and different expected outputs, so they are not the same bug/change."
+            "reason": "Both are test-only XPU expectation updates, but they target different models and different logits, so they are not the same change."
           },
           {
             "left": "pull_request:45858",
             "right": "pull_request:45926",
             "accept": false,
-            "reason": "Different test files for different models; these are independent CI expectation additions, not one unified fix."
+            "reason": "Same broad XPU-testing motif, but different model/test files and different expected outputs; not a duplicate bug fix."
           },
           {
             "left": "pull_request:45890",
             "right": "pull_request:45926",
             "accept": false,
-            "reason": "Separate XPU expectation changes for different models with no shared concrete issue beyond the generic testing theme."
+            "reason": "These are separate model-specific test expectation updates, not one shared code-path change."
           }
         ]
       },
@@ -18983,7 +18965,7 @@
       "soft_pairs": [
         "pull_request:45548|pull_request:45662"
       ],
-      "prepared_review_unit_hash": "09b75ae59449bdfca5472e81772786d4f68d8273f5b64eb4c2f3306a8d3b001a",
+      "prepared_review_unit_hash": "a421a3ba3b9725aabbe735eb169bbcfdb1084c88ff77b9b8b7b43c514da36698",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -19005,18 +18987,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related EP/distributed-loading fixes, but they address different failure modes and code paths, so they should not be deduplicated.",
-        "confidence": 0.94,
+        "summary": "Both PRs are EP-related loading fixes, but they address different failure modes: one serializes distributed_config and adds has_ep for accelerate+ZeRO-3 loading, while the other fixes expert parameter handling so rank-0 broadcast does not overwrite weights under FSDP2.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No canonical PR for dedupe: 45548 fixes EP + DeepSpeed ZeRO-3 loading/serialization via accelerate launch, while 45662 fixes a separate EP + FSDP2 bug where rank-0 broadcast overwrites experts.",
+        "canonical_pr_reason": "PR 45662 is the better canonical representative: it targets a specific runtime corruption bug in distributed loading and spans the concrete loading/MoE code path.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45662 is the stronger representative of the cluster because it targets a concrete runtime corruption bug in a distributed loading path, with a more specific and testable code change.",
+        "best_pr_reason": "PR 45662 is the strongest standalone fix in the cluster because it addresses a concrete expert-weight overwrite bug with a clear code-path change.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45548",
             "right": "pull_request:45662",
             "accept": false,
-            "reason": "Different underlying bugs: 45548 is about preserving distributed_config/EP state during loading, while 45662 is about preventing expert parameters from being overwritten after sharding/broadcast. They share EP context but not the same fix or code path."
+            "reason": "Not the same underlying bug. 45548 is about serializing distributed_config and detecting EP for loading via accelerate launch; 45662 is about expert parameters being overwritten by rank-0 broadcast during FSDP2 loading. Shared EP context only."
           }
         ]
       },
@@ -19046,40 +19028,40 @@
       "soft_pairs": [
         "pull_request:45639|pull_request:45645"
       ],
-      "prepared_review_unit_hash": "f6e8c3c3eb1cdd252dfd2f7dc2671203523ca2a7c0f2b13be6ccb0bd287df3cc",
+      "prepared_review_unit_hash": "b0784b4571d29a4173e94337cb1dd5e5f6ad14e8110ed976f8ae7ee4f81d5c7a",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4588,
-        "estimated_input_tokens": 1147,
-        "estimated_eval_tokens": 2550
+        "serialized_chars": 4590,
+        "estimated_input_tokens": 1148,
+        "estimated_eval_tokens": 2552
       },
       "budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4588,
-        "estimated_input_tokens": 1147,
-        "estimated_eval_tokens": 2550
+        "serialized_chars": 4590,
+        "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 captured_info/xdist collision problem in `testing_utils`; PR #45645 is the broader, more complete fix and subsumes #45639 by also updating CI log collection.",
-        "confidence": 0.88,
+        "summary": "Both PRs target the same xdist/captured_info logging collision path in `testing_utils.py`; PR 45645 is the broader superset because it includes the worker-safe log file fix plus CI workflow changes to read multiple log files.",
+        "confidence": 0.91,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "#45645 is the better canonical PR because it contains the core xdist-safe `captured_info` fix plus the downstream CI workflow update needed to read multiple worker logs; #45639 is just the narrower core change.",
+        "canonical_pr_reason": "PR 45645 is the better canonical PR because it covers the same core xdist-safe captured_info fix as 45639 and also updates CI to preserve and aggregate the per-worker logs.",
         "best_issue_reason": null,
-        "best_pr_reason": "#45645 best represents the cluster because it fixes the underlying artifact-collision bug and includes the surrounding workflow changes needed for the fix to work end-to-end.",
+        "best_pr_reason": "PR 45645 is the strongest representative: it addresses the underlying artifact collision and the downstream log-collection path, making it the more complete fix.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45639",
             "right": "pull_request:45645",
             "accept": true,
-            "reason": "Same underlying bug: xdist collisions for `captured_info` artifacts in `testing_utils`. #45645 largely subsumes #45639 and adds CI log aggregation changes, so these look like variants of one fix rather than separate PRs."
+            "reason": "Same underlying bug in captured_info handling under xdist; 45645 subsumes 45639's file-path fix and adds the necessary CI-side log aggregation."
           }
         ]
       },
@@ -19238,7 +19220,7 @@
       "soft_pairs": [
         "pull_request:45975|pull_request:45977"
       ],
-      "prepared_review_unit_hash": "40b16558a5b03c92f9323b95f6a0a901f7e14161749d5b3cded4b79b00333b0e",
+      "prepared_review_unit_hash": "a297f103ea738b6621d35cb6d1f962b1f7470446b084500015ff8c3368701183",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -19260,18 +19242,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs are about Apple Silicon / GGUF / kernels-community acceleration, but they address different code paths: one adds an optional Metal dequant fast path for GGUF serving, while the other introduces a new GGUF linear layer and quantizer integration for direct inference. They are related, but not the same change and should not be merged as duplicates.",
-        "confidence": 0.93,
+        "summary": "Two open PRs touch GGUF/Apple Silicon/kernels-community, but they implement different pieces: one is an optional dequant fast-path/configuration update, the other is a new GGUF linear inference path with integration code and tests.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No duplicate PR pair here. If a representative PR is needed, #45977 is the more central/core implementation because it adds the new GGUF linear inference path and associated quantizer/modeling integration, whereas #45975 is a narrower serving-side option and help text update.",
+        "canonical_pr_reason": "PR 45977 is the stronger cluster representative: it introduces the core GGUF runtime change, adds integration points, and includes tests, whereas 45975 is a smaller optional fast-path/help/configuration update.",
         "best_issue_reason": null,
-        "best_pr_reason": "#45977 is the best standalone representative for the cluster\u2019s main technical theme: native GGUF matmul inference on Apple Silicon with kernels-community, backed by a new module and tests. #45975 is narrower and targets a different serving path.",
+        "best_pr_reason": "PR 45977 is the best standalone candidate for triage because it addresses a concrete code-path change (GGUF matmul on Apple Silicon) with implementation and test coverage.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45975",
             "right": "pull_request:45977",
             "accept": false,
-            "reason": "Related only at a broad theme level (GGUF + Apple Silicon + kernels-community). #45975 is a serving-side optional Metal fast path, while #45977 implements a new GGUF linear matmul path and quantizer changes. Different concrete problems and not mergeable as one PR."
+            "reason": "Related domain, but not the same underlying change: 45975 updates CLI/help and GGUF-related fast-path plumbing, while 45977 adds a new GgufLinear inference implementation and tests. They would not plausibly be merged into one PR as duplicate fixes."
           }
         ]
       },
diff --git a/analysis/current/manifest.json b/analysis/current/manifest.json
index ef8ba1857c0d81875cba9e210b8ac3a9b025a7fe..1ff253c75039ad3dda46ab372141dbfc8317dd29 100644
--- a/analysis/current/manifest.json
+++ b/analysis/current/manifest.json
@@ -1,8 +1,8 @@
 {
-  "analysis_id": "hybrid-model-20260515t000050z",
+  "analysis_id": "hybrid-model-20260515t060029z",
   "archived_artifacts": {
-    "hybrid": "snapshots/20260515T000050Z/analysis-runs/hybrid-model-20260515t000050z/analysis-report-hybrid.json",
-    "hybrid_reviews": "snapshots/20260515T000050Z/analysis-runs/hybrid-model-20260515t000050z/analysis-report-hybrid.llm-reviews.json"
+    "hybrid": "snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.json",
+    "hybrid_reviews": "snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.llm-reviews.json"
   },
   "artifacts": {
     "hybrid": "analysis/current/analysis-report-hybrid.json",
@@ -10,9 +10,9 @@
   },
   "channel": "canonical",
   "model": null,
-  "published_at": "2026-05-15T00:09:06Z",
+  "published_at": "2026-05-15T06:04:53Z",
   "repo": "huggingface/transformers",
   "schema_version": 1,
-  "snapshot_id": "20260515T000050Z",
+  "snapshot_id": "20260515T060029Z",
   "variant": "hybrid"
 }
diff --git a/snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.json b/snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.json
new file mode 100644
index 0000000000000000000000000000000000000000..58b9255c33d57bf53fe4b434dc1a9ba6a2a7b721
--- /dev/null
+++ b/snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.json
@@ -0,0 +1,2205 @@
+{
+  "schema_version": "1.0",
+  "repo": "huggingface/transformers",
+  "snapshot_id": "20260515T060029Z",
+  "generated_at": "2026-05-15T06:04:51Z",
+  "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/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.llm-reviews.json b/snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.llm-reviews.json
new file mode 100644
index 0000000000000000000000000000000000000000..fe13419bf490f489af924c38ba39cd1442ff3d4d
--- /dev/null
+++ b/snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.llm-reviews.json
@@ -0,0 +1,19268 @@
+{
+  "schema_version": "1.0",
+  "repo": "huggingface/transformers",
+  "snapshot_id": "20260515T060029Z",
+  "generated_at": "2026-05-15T06:04:51Z",
+  "model": "gpt-5.4-mini?service_tier=flex",
+  "reviews": [
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 1,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "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:43425",
+        "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:44805",
+        "issue:44841",
+        "issue:44863",
+        "issue:44918",
+        "issue:45081",
+        "issue:45161",
+        "issue:45237",
+        "issue:45464",
+        "issue:45507",
+        "issue:45685",
+        "issue:45812",
+        "issue:45823",
+        "issue:45859",
+        "issue:45874",
+        "issue:45901",
+        "issue:45902",
+        "issue:45920"
+      ],
+      "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:44805|issue:45823",
+        "issue:43638|issue:44805",
+        "issue:43596|issue:44805",
+        "issue:44805|issue:45161",
+        "issue:44805|issue:44918",
+        "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:45859|issue:45902",
+        "issue:43638|issue:45161",
+        "issue:34928|issue:36331",
+        "issue:43596|issue:45161",
+        "issue:43716|issue:45507",
+        "issue:39290|issue:45081",
+        "issue:45823|issue:45874",
+        "issue:43425|issue:44292",
+        "issue:43828|issue:45161",
+        "issue:34928|issue:41720",
+        "issue:44918|issue:45161",
+        "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:45237|issue:45901",
+        "issue:44841|issue:45823",
+        "issue:43039|issue:44560",
+        "issue:45812|issue:45920",
+        "issue:43866|issue:44863",
+        "issue:43854|issue:43866",
+        "issue:41093|issue:43638",
+        "issue:43792|issue:45901"
+      ],
+      "prepared_review_unit_hash": "ca841ab4584d65cd09eac411188d38fbd226e3f152249a9752dd7fc47802c51f",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 62,
+        "serialized_chars": 46363,
+        "estimated_input_tokens": 11591,
+        "estimated_eval_tokens": 23438
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly a mix of unrelated issue families, with only a few true duplicates: a ZeRO-3/Bert zero-size index bug and a broader AutoTokenizer v5 regression, plus one possible Qwen3_5Moe rotary-embedding duplicate. Most soft edges are false positives from broad semantic similarity.",
+        "confidence": 0.62,
+        "canonical_issue_reason": "issue:45920 is the broadest, most representative issue here: a silent AutoTokenizer v5 regression affecting many model families, which is the best anchor for the tokenizer-related duplicate set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45920 has the widest blast radius and clearest regression framing, making it the best overall issue candidate in this cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44322",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Different problems: a missing config attribute vs a streaming chat/completions failure."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Unrelated model/runtime failures on different hardware paths; not the same bug."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Different model paths and symptoms: vision dtype mismatch vs GPT-OSS AMD execution failure."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Both are runtime errors, but they involve different models and different device/dtype code paths."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "No shared underlying code path; one is AMD GPU loading, the other is an invalid method call in GraniteMoEHybrid."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Different model families and failure modes; not plausibly one fix."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:35545",
+            "accept": false,
+            "reason": "DeepSpeed collective mismatch and ONNX export are unrelated code paths."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "ONNX export error vs FP8 model loading failure; no common bug."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Different subsystems: export tooling vs device mapping runtime crash."
+          },
+          {
+            "left": "issue:44509",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are stale v5 docs, but they cover different removed features and should be triaged separately."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Export failure and GenerationMixin import regression are unrelated."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Mask shape mismatch and Gemma4 device mismatch do not point to the same bug."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Both are index/shape errors, but they arise in different model and execution contexts."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Same broad error class, but the model/setup differs enough that a shared fix is not clear."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Different models and failure mechanisms; not a duplicate."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Mask indexing failure vs TRL embedding unpacking failure are distinct issues."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:35545",
+            "accept": false,
+            "reason": "Processor voice_preset bug is unrelated to ModernBERT ONNX export."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Both mention DeepSpeed, but one is a collective mismatch and the other is a zero-sized index error; different bugs."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:43638",
+            "accept": true,
+            "reason": "Same ZeRO-3/Bert index-0-out-of-bounds failure pattern, with the same loading/training path and likely one fix."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Liger kernel cross-entropy routing and GPT-OSS tensor-parallel failure are different problems."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:44457",
+            "accept": false,
+            "reason": "Detokenization regression and LoRA merge/save inconsistency do not share the same underlying bug."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44457",
+            "accept": false,
+            "reason": "NVFP4 execution error vs LoRA reload mismatch are unrelated."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "FSDP activation-checkpointing size mismatch is not the same as this mask indexing bug."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:30277",
+            "accept": false,
+            "reason": "Both involve DeepSpeed, but they are different errors in different training scenarios."
+          },
+          {
+            "left": "issue:45859",
+            "right": "issue:45902",
+            "accept": true,
+            "reason": "Same Qwen3_5MoeTextRotaryEmbedding device/materialization path; both point to the same buffer initialization/offload bug."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Torch version incompatibility and Qwen NVFP4 runtime failure are separate issues."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Autocast dtype mismatch and GPT-OSS-only tensor parallel failure are unrelated."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "FSDP checkpointing mismatch and auto-device-map CUDA assert are different problems."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "TRL SFT embedding unpacking failure is not the same as GPT-OSS tensor-parallel failure."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Both are ZeRO-3 index errors, but the model contexts differ enough that a shared root cause is not established."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are documentation regressions from v5 removals, but they cover different removed entry points."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Different multimodal failures: image-preprocessor dtype mismatch vs video StopIteration."
+          },
+          {
+            "left": "issue:45812",
+            "right": "issue:45920",
+            "accept": true,
+            "reason": "Same AutoTokenizer v5 regression; 45920 is the broader multi-model report, and 45812 is a specific affected subset."
+          },
+          {
+            "left": "issue:43866",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Checkpoint corruption and load failure for a different model are not the same bug."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "Different models and different load symptoms; no clear shared fix."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Both are shape/index errors, but they involve different tensor operations and model setups."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "Whisper inference failure and table-question-answering crash are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative: it distinguishes the broad unrelated issue families, identifies the clearly related AutoTokenizer v5 regression cluster, and treats the Qwen3_5Moe pair as only a possible duplicate. The accepted soft-edge pairs are plausible and not overstated based on the available titles/body signals."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 2,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:36010",
+        "issue:39186",
+        "issue:39290",
+        "issue:40990",
+        "issue:41093",
+        "issue:41762",
+        "issue:42915",
+        "issue:43054",
+        "issue:43296",
+        "issue:43366",
+        "issue:43404",
+        "issue:43425",
+        "issue:43531",
+        "issue:43541",
+        "issue:43572",
+        "issue:43596",
+        "issue:43645",
+        "issue:43716",
+        "issue:43792",
+        "issue:43828",
+        "issue:43866",
+        "issue:43975",
+        "issue:43994",
+        "issue:44291",
+        "issue:44292",
+        "issue:44387",
+        "issue:44530",
+        "issue:44589",
+        "issue:44841",
+        "issue:44918",
+        "issue:44960",
+        "issue:45005",
+        "issue:45020",
+        "issue:45070",
+        "issue:45084",
+        "issue:45161",
+        "issue:45362",
+        "issue:45399",
+        "issue:45464",
+        "issue:45542",
+        "issue:45632",
+        "issue:45685",
+        "issue:45698",
+        "issue:45759",
+        "issue:45800",
+        "issue:45823",
+        "issue:45901",
+        "issue:45902"
+      ],
+      "soft_pairs": [
+        "issue:39290|issue:43531",
+        "issue:43425|issue:45070",
+        "issue:43596|issue:45800",
+        "issue:43975|issue:45901",
+        "issue:45362|issue:45464",
+        "issue:43792|issue:45070",
+        "issue:44292|issue:45901",
+        "issue:43404|issue:43866",
+        "issue:39290|issue:45823",
+        "issue:45464|issue:45685",
+        "issue:43541|issue:45685",
+        "issue:43541|issue:43828",
+        "issue:44841|issue:45084",
+        "issue:36010|issue:39290",
+        "issue:41093|issue:41762",
+        "issue:45800|issue:45823",
+        "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:43366|issue:45464",
+        "issue:45020|issue:45800",
+        "issue:43531|issue:45070",
+        "issue:41093|issue:43596",
+        "issue:43541|issue:44589",
+        "issue:43054|issue:43994",
+        "issue:45542|issue:45800",
+        "issue:45759|issue:45902",
+        "issue:40990|issue:44960",
+        "issue:43716|issue:44918",
+        "issue:39186|issue:44918",
+        "issue:45632|issue:45698",
+        "issue:43404|issue:45005",
+        "issue:43645|issue:45399",
+        "issue:44530|issue:45005",
+        "issue:42915|issue:45399"
+      ],
+      "prepared_review_unit_hash": "181cd044c2ce24d415ec50f27f69f21c7149553d939d945cb2b5883741f964f8",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 38,
+        "serialized_chars": 38760,
+        "estimated_input_tokens": 9690,
+        "estimated_eval_tokens": 19636
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous and not a single duplicate set. I only see two true duplicate pairs: SigLIP2 bad-output reports, and the trust_remote_code/custom-module cache-collision reports. Everything else is at best thematic similarity, not the same underlying bug.",
+        "confidence": 0.79,
+        "canonical_issue_reason": "Issue 45698 is the clearest concrete bug report in the strongest duplicate subcluster: it describes the wrong custom module being loaded after save_pretrained, which matches the cache/path collision symptom in 45632.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45698 is the best representative issue overall because it is specific, actionable, and captures the underlying custom-code loading failure more cleanly than the broader symptom report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:39290",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Both mention sliding_window-related failures, but they affect different model families and different config fields; this is similarity, not the same bug."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "One is a torch-version incompatibility report and the other is a PretrainedConfig/pydantic break; too different in root cause and code path."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Different failure modes and causes: ZeRO-3/BERT loading index error vs torch/transformers incompatibility."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "Detokenization corruption and table QA crashing are unrelated symptoms and do not point to the same code path."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Both involve Qwen3.5 and chat-like flows, but one is a transformers chat crash and the other is a streaming chat/completions failure; not the same concrete bug."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Whisper runtime failure and a pydantic/config field regression are unrelated."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "Qwen-3 NVFP4 runtime failure and table-question-answering crashes are different bugs in different paths."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "A tied lm_head bug in Mistral3 is not the same as a corrupted Ovis checkpoint."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Gemma3 missing sliding_window_pattern and Gemma4 device mismatch are different issues, despite both being model-integration regressions."
+          },
+          {
+            "left": "issue:45464",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Chat streaming failure and an MPS Int histogram kernel error are unrelated."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Grouped_mm tracing failure in Mixtral and an MPS histogram kernel limitation are different code paths."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Both are dtype-related, but one is grouped_mm tracing and the other is autocast mismatch in Phi-tiny-MoE; not the same underlying bug."
+          },
+          {
+            "left": "issue:44841",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Processor failure for Voxtral and a template-node compilation TypeError are unrelated."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:39290",
+            "accept": false,
+            "reason": "ImportError for GenerationMixin and a Gemma3 sliding_window config error are not the same issue."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:41762",
+            "accept": false,
+            "reason": "Both are IndexErrors, but one is a mask shape mismatch and the other is a zero-sized tensor during ZeRO-3 loading; different root causes."
+          },
+          {
+            "left": "issue:45800",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Torch/transformers version incompatibility and Gemma4 device-mismatch during forward are different problems."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:43366",
+            "accept": false,
+            "reason": "PaddleOCR-VL load failure in vLLM and gpt-oss GGUF support are not the same bug."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Missing pad_token_idx after 5.0 update and a pydantic model-field regression are unrelated."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "CUDA memory regression under int4 quantization is not the same as tied-weight translation-model behavior in v5."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Mixtral grouped_mm tracing and GPT-OSS TP-only failure are different model/runtime issues."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Autocast dtype mismatch in Phi-tiny-MoE and Qwen3.5 streaming API failure are different bugs."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Unexpected _is_hf_initialized argument in init_empty_weights is unrelated to the pydantic PretrainedConfig regression."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "GGUF gpt-oss support and Qwen3.5 streaming inference failures are not the same underlying issue."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Broad remote_code breakage and a specific torch/transformers incompatibility are too broad and different to treat as duplicates."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "A Qwen3-MoE sliding_window issue and a pydantic PretrainedConfig issue do not share a concrete code path."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:43596",
+            "accept": false,
+            "reason": "Both are loading-time IndexErrors, but they come from different shapes, models, and initialization paths."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Grouped_mm tracing in Mixtral and missing Float8 storage support are different runtime failures."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:43994",
+            "accept": true,
+            "reason": "Both report SigLIP2 producing bad/nonsensical outputs, pointing to the same model-quality or implementation regression."
+          },
+          {
+            "left": "issue:45542",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "TensorBoard-without-TensorFlow backend errors are unrelated to torch/transformers version incompatibility."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Composite Qwen config unwrapping and uninitialized-memory in rotary embedding are different bugs."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "High perplexity on gpt-oss and a GLM5 issue do not describe the same failure."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Gemma-3 multimodal dtype mismatch and Qwen3.5 embedding unpacking under TRL are different problems."
+          },
+          {
+            "left": "issue:39186",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "FSDP weight dimensionality and Qwen3.5/TRL embedding unpacking are unrelated."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:45698",
+            "accept": true,
+            "reason": "Both describe the same trust_remote_code/custom-code cache collision: local models with the same leaf name can resolve to the wrong loaded module."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Mistral3 lm_head tying and translation-model tied-weight issues are similar only at a high level, not the same concrete bug."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Jupyter custom model initialization and flash-attn fallback blocking are unrelated."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "PagedAttentionCache linear_attention failure and tied-weight translation-model issues are different subsystems and causes."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 failure and flash-attn fallback blocking are not the same underlying problem."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The two accepted duplicate pairs are supported by the issue titles, and the rejected soft-edge pairs are mostly framed as similarity rather than identity of bugs. No obvious overstatement of evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 3,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:36331",
+        "issue:38175",
+        "issue:39692",
+        "issue:41553",
+        "issue:41628",
+        "issue:41762",
+        "issue:42915",
+        "issue:42947",
+        "issue:43054",
+        "issue:43493",
+        "issue:43643",
+        "issue:43644",
+        "issue:43782",
+        "issue:43828",
+        "issue:43856",
+        "issue:43866",
+        "issue:43872",
+        "issue:43873",
+        "issue:43881",
+        "issue:43883",
+        "issue:43950",
+        "issue:43975",
+        "issue:44292",
+        "issue:44360",
+        "issue:44387",
+        "issue:44403",
+        "issue:44451",
+        "issue:44488",
+        "issue:44492",
+        "issue:44512",
+        "issue:44534",
+        "issue:44589",
+        "issue:44863",
+        "issue:44910",
+        "issue:44918",
+        "issue:44928",
+        "issue:44960",
+        "issue:45005",
+        "issue:45020",
+        "issue:45237",
+        "issue:45507",
+        "issue:45520",
+        "issue:45663",
+        "issue:45701",
+        "issue:45800",
+        "issue:45859",
+        "issue:45902"
+      ],
+      "soft_pairs": [
+        "issue:44360|issue:44512",
+        "issue:43873|issue:45005",
+        "issue:44863|issue:45005",
+        "issue:44492|issue:44512",
+        "issue:44918|issue:45507",
+        "issue:43975|issue:44292",
+        "issue:43950|issue:44534",
+        "issue:44292|issue:45507",
+        "issue:43644|issue:43950",
+        "issue:39692|issue:43054",
+        "issue:43828|issue:44928",
+        "issue:36331|issue:41762",
+        "issue:43866|issue:44403",
+        "issue:41628|issue:45800",
+        "issue:43950|issue:45902",
+        "issue:43872|issue:44589",
+        "issue:43054|issue:45701",
+        "issue:41553|issue:43883",
+        "issue:38175|issue:43054",
+        "issue:44910|issue:45859",
+        "issue:45663|issue:45902",
+        "issue:41553|issue:43881",
+        "issue:42947|issue:44387",
+        "issue:43054|issue:43493",
+        "issue:44403|issue:44451",
+        "issue:43643|issue:45020",
+        "issue:43975|issue:45520",
+        "issue:43975|issue:45237",
+        "issue:44403|issue:44488",
+        "issue:42947|issue:43856",
+        "issue:43782|issue:44960",
+        "issue:43881|issue:43883",
+        "issue:41762|issue:42915"
+      ],
+      "prepared_review_unit_hash": "8e89dea8719ecc725c18a5089445c5b8f28e69d3d521ae2973eda00b81899356",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 33,
+        "serialized_chars": 36583,
+        "estimated_input_tokens": 9146,
+        "estimated_eval_tokens": 18548
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a mix of unrelated issues, but it does contain a clear duplicate subcluster around the Transformers 5.x non-persistent buffer regression. A smaller SigLIP2-related subcluster also looks related, but not every SigLIP2 report is the same underlying bug.",
+        "confidence": 0.41,
+        "canonical_issue_reason": "Issue 43950 is the cleanest representative of the strongest duplicate subcluster: it describes the underlying non-persistent-buffer regression directly and matches multiple related reports (43644, 44534, 45902).",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43950 is the best issue to anchor the cluster because it captures the core bug in a concrete, reusable way rather than a single symptom.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44360",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Unrelated: one is about a DSA indexer ReLU discussion, the other is docs mentioning a removed command."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Different problems: quantization/offloading behavior versus tied-weights handling in v5."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Different failures: NemotronH checkpoint loading versus translation-model tied weights in v5."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "A typo in cache strategies is not the same as outdated docs about the removed transformers run command."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Different code paths and symptoms: Qwen3.5 embedding unpacking with TRL versus GraniteMoEHybrid invalid method calls."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Different model-specific failures: detokenization bug versus runtime error on Qwen-3-8B-NVFP4."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:44534",
+            "accept": true,
+            "reason": "Same underlying regression: both report non-persistent buffers being incorrectly populated with junk in Transformers 5.x."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "No shared concrete bug; one is a Qwen inference/runtime issue and the other is a model-method implementation problem."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43950",
+            "accept": true,
+            "reason": "Same bug: both describe the Transformers 5.x regression where non-persistent buffers are filled/corrupted on load."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43054",
+            "accept": false,
+            "reason": "Both involve SigLIP2, but one is a docs example/model-processor mismatch and quantization failure, while the other is a text-embedding quality regression."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:44928",
+            "accept": false,
+            "reason": "Related to Qwen3.5 dtype/attention behavior, but the concrete failures differ: autocast dtype mismatch versus RLHF gradient explosion."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41762",
+            "accept": false,
+            "reason": "Different training/loading errors with no clear shared code-path bug."
+          },
+          {
+            "left": "issue:43866",
+            "right": "issue:44403",
+            "accept": false,
+            "reason": "Different topics: corrupted checkpoint versus generic loading noise."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Not the same bug: import/export issue for AutoImageProcessor versus broad torch/transformers incompatibility."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:45902",
+            "accept": true,
+            "reason": "Same regression: 45902 is a specific manifestation of the non-persistent buffer materialization bug covered by 43950."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Both are compatibility/type errors, but they involve different objects and different failure modes."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "SigLIP2 embedding quality is unrelated to a generic tokenization change report."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "Different model-loading issues with no shared bug or fix path."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43054",
+            "accept": true,
+            "reason": "Likely same SigLIP2 implementation bug family: both describe bad/incorrect SigLIP2 outputs rather than distinct subsystems."
+          },
+          {
+            "left": "issue:44910",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Different problems: flash-attention crash from packed-sequence misinterpretation versus CPU-offload incompatibility in Qwen3_5MoeTextRotaryEmbedding."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Different models and failures: Gemma4 FSDP2 shared_kv_states issue versus Qwen3_5MoeTextRotaryEmbedding buffer materialization."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "Different loading failures for different models; no evidence of the same underlying bug."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Both are memory-related, but one is gradient checkpointing with LoRA and the other is int4 quantization reserved-memory/OOM."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:43493",
+            "accept": true,
+            "reason": "Likely same SigLIP2 discrepancy family: one report is poor text embeddings, the other explicitly compares HF and JAX implementations."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Generic loading noise is not the same as failure to load a specific model."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Both involve remote_code-style loading, but the concrete issues differ too much to merge as one bug."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Different bugs: detokenization versus Python 3.13 flash_attn import handling."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "No shared bug; one is a detokenization issue and the other is an AMD GPU runtime problem."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both mention loading, but they concern different models and likely different causes."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "Different performance/memory topics with no concrete shared code-path problem."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "Different model families and failure modes; not the same underlying bug."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "Different model-loading failures with distinct root causes."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Different issues: DeepSpeed ZeRO-3 loading crash versus Qwen3Moe FP8 config failure."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays mostly conservative. The non-persistent-buffer subcluster is well supported by the titles, and the SigLIP2 links are framed as likely/related rather than certain. I did not see an overstatement that would require rejection."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 4,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:24643",
+        "issue:30064",
+        "issue:36331",
+        "issue:41093",
+        "issue:41553",
+        "issue:42491",
+        "issue:42915",
+        "issue:42947",
+        "issue:43257",
+        "issue:43366",
+        "issue:43404",
+        "issue:43638",
+        "issue:43742",
+        "issue:43792",
+        "issue:43824",
+        "issue:43827",
+        "issue:43828",
+        "issue:43854",
+        "issue:43866",
+        "issue:43872",
+        "issue:43940",
+        "issue:43950",
+        "issue:44220",
+        "issue:44291",
+        "issue:44368",
+        "issue:44479",
+        "issue:44488",
+        "issue:44509",
+        "issue:44530",
+        "issue:44560",
+        "issue:44661",
+        "issue:44805",
+        "issue:44863",
+        "issue:44928",
+        "issue:44991",
+        "issue:45020",
+        "issue:45081",
+        "issue:45305",
+        "issue:45313",
+        "issue:45325",
+        "issue:45356",
+        "issue:45399",
+        "issue:45538",
+        "issue:45588",
+        "issue:45685",
+        "issue:45823",
+        "issue:45902",
+        "issue:45923"
+      ],
+      "soft_pairs": [
+        "issue:45685|issue:45823",
+        "issue:42491|issue:44368",
+        "issue:43824|issue:43872",
+        "issue:45020|issue:45823",
+        "issue:44661|issue:45020",
+        "issue:44991|issue:45399",
+        "issue:43828|issue:45823",
+        "issue:45081|issue:45399",
+        "issue:43827|issue:44509",
+        "issue:45020|issue:45356",
+        "issue:43404|issue:43950",
+        "issue:43742|issue:45399",
+        "issue:42915|issue:45020",
+        "issue:43257|issue:43866",
+        "issue:43638|issue:45823",
+        "issue:43792|issue:44220",
+        "issue:43872|issue:44291",
+        "issue:43940|issue:45313",
+        "issue:42915|issue:43872",
+        "issue:41553|issue:44991",
+        "issue:43854|issue:44863",
+        "issue:36331|issue:41093",
+        "issue:43257|issue:45305",
+        "issue:42947|issue:45923",
+        "issue:44928|issue:45588",
+        "issue:41093|issue:44805",
+        "issue:43872|issue:45538",
+        "issue:41553|issue:44488",
+        "issue:44530|issue:44863",
+        "issue:44479|issue:44560",
+        "issue:24643|issue:30064",
+        "issue:45325|issue:45902",
+        "issue:43366|issue:43828"
+      ],
+      "prepared_review_unit_hash": "0b50d24dc60dedbc7da1304521d25dbbe2dd7434a284c794748a1d046bfbc3b2",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 38066,
+        "estimated_input_tokens": 9517,
+        "estimated_eval_tokens": 19290
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly heterogeneous issue cluster. I only found a few likely true duplicates/same-bug pairs; the rest are different model, training, tokenizer, or docs regressions that should stay separate.",
+        "confidence": 0.89,
+        "canonical_issue_reason": "issue:43827 is the clearest canonical anchor: issue:44509 is the same v5 docs regression about removed pipeline tasks, just phrased differently.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43827 is the best issue to keep as the representative duplicate target because it is the earlier, cleaner report of the confirmed docs regression.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45685",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Different failures: MPS histogram on Int vs Gemma4 device-mapping mismatch."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Both involve Qwen/LoRA, but one is a HF4.x/HF5.x compatibility break and the other is a tie_word_embeddings warning."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43872",
+            "accept": false,
+            "reason": "Same keyword appears, but this is a model import error vs a bitsandbytes quantization constructor error."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "General remote_code breakage vs a specific Gemma4 forward device mismatch; not the same bug."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Tokenizer mapping helper failure vs remote_code model breakage; different code paths."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Tokenizer loading failure vs flash-attn2 fallback gating; unrelated subsystems."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Autocast dtype mismatch is not the same as a device-placement error."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Mistral tokenizer regex crash vs flash-attn fallback checks; unrelated."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44509",
+            "accept": true,
+            "reason": "Same docs regression about removed pipeline tasks in v5; 44509 is a reworded duplicate of 43827."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Remote_code regression vs Kimi-K2.5 tokenizer codec/warning regression; different model-specific bugs."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43950",
+            "accept": false,
+            "reason": "Tied lm_head weights bug vs silent non-persistent buffer corruption; different issues."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Different load-path failures: MobileLLM key error vs flash-attn fallback gating."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Both are broad compatibility complaints, but one is LoRA/Qwen3-MOE specific and the other is a remote_code regression."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "Qwen3 MOE weight-conversion failure vs Ovis2 checkpoint corruption; unrelated models and causes."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO-3 weight init error vs Gemma4 device mismatch; different problems."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Whisper load failure and fbank helper issue may be adjacent, but the titles do not indicate the same concrete bug."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:44291",
+            "accept": true,
+            "reason": "Both describe the same unexpected `_is_hf_initialized` kwarg leaking into model-loading/init code paths."
+          },
+          {
+            "left": "issue:43940",
+            "right": "issue:45313",
+            "accept": false,
+            "reason": "Both are ZeRO-3 load failures, but the model families and reported symptoms differ too much to merge."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43872",
+            "accept": false,
+            "reason": "LoRA/Qwen3-MOE compatibility issue vs bitsandbytes constructor kwarg bug; different paths."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Different tokenizer-loading failures for different models; only the symptom is generic."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "GLM-4.7-Flash unit-test load failure vs NemotronH checkpoint loading; unrelated."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41093",
+            "accept": false,
+            "reason": "Trainer API signature change vs tensor mask shape mismatch; unrelated."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45305",
+            "accept": false,
+            "reason": "DeepSpeed weight conversion bug vs DeepSpeed gradient averaging bug; different training stages."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Gradient checkpointing behavior issue vs missing support flag on a remote_code model; related theme, different bug."
+          },
+          {
+            "left": "issue:44928",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "RLHF NaN/SDPA fallback issue vs flash-attn crash on `s_aux=None`; different attention-path failures."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:44805",
+            "accept": true,
+            "reason": "Same mask/index shape-mismatch exception pattern; these look like the same bug with different sample sizes."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "bitsandbytes `_is_hf_initialized` error vs CLIPTokenizer max-length constant; unrelated."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Bad error message for Voxtral vs a different tokenizer load failure for sleng-bert; not the same issue."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "PagedAttentionCache linear_attention crash vs NemotronH checkpoint loading; unrelated."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44560",
+            "accept": true,
+            "reason": "44560 is a specific Qwen3-VL video error that fits the broader video-input regression reported in 44479."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "Training DeepSpeed weight shape error vs image-processor segmentation-map failure; unrelated."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Different rotary-embedding defects: rope-index scaling vs uninitialized memory after materialization."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "GGUF architecture support vs autocast dtype mismatch; unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays mostly conservative. The accepted duplicate pairs are plausible from the titles alone, especially 43827/44509, 43872/44291, and 41093/44805. The remaining soft-edge rejections are appropriately cautious rather than overstated."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 5,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:36010",
+        "issue:38175",
+        "issue:39692",
+        "issue:41720",
+        "issue:42491",
+        "issue:42617",
+        "issue:42915",
+        "issue:43317",
+        "issue:43366",
+        "issue:43531",
+        "issue:43550",
+        "issue:43596",
+        "issue:43632",
+        "issue:43638",
+        "issue:43644",
+        "issue:43645",
+        "issue:43646",
+        "issue:43708",
+        "issue:43720",
+        "issue:43824",
+        "issue:43950",
+        "issue:43957",
+        "issue:43994",
+        "issue:44075",
+        "issue:44155",
+        "issue:44451",
+        "issue:44488",
+        "issue:44534",
+        "issue:44560",
+        "issue:44589",
+        "issue:44661",
+        "issue:44756",
+        "issue:44805",
+        "issue:44877",
+        "issue:44910",
+        "issue:44918",
+        "issue:44960",
+        "issue:44977",
+        "issue:45030",
+        "issue:45276",
+        "issue:45313",
+        "issue:45357",
+        "issue:45488",
+        "issue:45584",
+        "issue:45663",
+        "issue:45812",
+        "issue:45823",
+        "issue:45834"
+      ],
+      "soft_pairs": [
+        "issue:45663|issue:45823",
+        "issue:44155|issue:45584",
+        "issue:43366|issue:43531",
+        "issue:43950|issue:45834",
+        "issue:43550|issue:44910",
+        "issue:43824|issue:43957",
+        "issue:43645|issue:43646",
+        "issue:43366|issue:44589",
+        "issue:42915|issue:43957",
+        "issue:44560|issue:44918",
+        "issue:42617|issue:43366",
+        "issue:36010|issue:42915",
+        "issue:44661|issue:45823",
+        "issue:38175|issue:43994",
+        "issue:43632|issue:45834",
+        "issue:44560|issue:44805",
+        "issue:44877|issue:45030",
+        "issue:43638|issue:44661",
+        "issue:43596|issue:44661",
+        "issue:44661|issue:44805",
+        "issue:43317|issue:44756",
+        "issue:43644|issue:44534",
+        "issue:38175|issue:39692",
+        "issue:44960|issue:44977",
+        "issue:42491|issue:43720",
+        "issue:45313|issue:45357",
+        "issue:44451|issue:44488",
+        "issue:41720|issue:42915",
+        "issue:42491|issue:45276",
+        "issue:42915|issue:44756",
+        "issue:43708|issue:44075",
+        "issue:45488|issue:45812"
+      ],
+      "prepared_review_unit_hash": "cfa9cb1a9a587f13152c245eaa3e74ef15733c1440600747fcbca9b81f26592f",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37263,
+        "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 mostly a grab-bag of unrelated model-specific regressions, not one tight duplicate cluster. The clearest duplicate is the Transformers 5 non-persistent-buffer bug (43644/44534), with one additional plausible duplicate around broken SigLIP2 outputs (38175/43994).",
+        "confidence": 0.64,
+        "canonical_issue_reason": "Issue 43644 is the cleanest canonical representative: it states the regression clearly and is effectively duplicated by 44534. The rest of the items are too heterogeneous to serve as one shared root issue.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43644 is the best issue to anchor the cluster because it is the most precise, broadly reusable description of the only exact duplicate pair in the set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45663",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Different Gemma4 failures: one is an FSDP2 KeyError from per-layer shared_kv_states rebuilding, the other is a device-mapping mismatch during forward."
+          },
+          {
+            "left": "issue:44155",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "AudioFlamingo3 cross-track embedding leak and Whisper empty-transcription handling are unrelated code paths."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "gpt-oss GGUF support and Qwen3-MoE sliding_window behavior are different features/issues."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "from_pretrained non-persistent buffer corruption is unrelated to the Kosmos2.5 long-OCR index error."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44910",
+            "accept": false,
+            "reason": "These are different attention/compiler bugs: SDPA+torch.compile on Bamba versus Flash Attention with Qwen3.5 position_ids packing."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Missing Qwen2_5_VL export and meta-device loading failures are distinct regressions."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:43646",
+            "accept": false,
+            "reason": "Both involve custom model init in v5, but the notebook-specific redefinition case and the general initialization failure are not clearly the same underlying bug."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "GGUF architecture support and Float8 storage lookup errors are unrelated."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Qwen3-MoE FP8 config failure and meta-device model loading are separate problems."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "A Qwen3-vl video StopIteration and a TRL SFT unpacking failure are different code paths."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:43366",
+            "accept": false,
+            "reason": "3d_parallel.py runtime failure is unrelated to gpt-oss GGUF support."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "GenerationMixin import regression and Qwen3-MoE FineGrainedFP8 failure are unrelated."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Tokenizer-mapping lookup failure and Gemma4 device mismatch are different issues."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43994",
+            "accept": true,
+            "reason": "Same SigLIP2 symptom family: both report broken outputs for google/siglip2-base-patch16-224 and point to the same model/processor regression."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "_is_hf_initialized flag breakage and Kosmos2.5 OCR indexing are unrelated."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Qwen3-vl StopIteration and a mask/tensor shape IndexError are different failures."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:45030",
+            "accept": false,
+            "reason": "Strict config loading for granite_speech and glm4v config validation are not the same bug."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "ZeRO-3 Bert initialization and add-new-model-like tokenizer-mapping handling are unrelated."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "Another ZeRO-3 Bert initialization failure; not the same as the tokenizer-mapping issue."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Tokenizer-mapping logic and mask-shape indexing are different code paths."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "gpu+cpu offload loading failure and Strix Halo mmap OOM are not the same underlying bug."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44534",
+            "accept": true,
+            "reason": "Exact duplicate regression: both describe the same Transformers 5 non-persistent-buffer junk corruption issue."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:39692",
+            "accept": false,
+            "reason": "Both are about SigLIP2, but one is bad inference outputs while the other is a docs-example model/processor mismatch and quantization failure; not clearly the same bug."
+          },
+          {
+            "left": "issue:44960",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "GLM5 is too vague and unrelated to the Qwen3.5 flash-attention generation failure."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Qwen3-MoE hf4.x/hf5.x LoRA compatibility and BitNet packed-weight loading are different problems."
+          },
+          {
+            "left": "issue:45313",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "ZeRO-3 loading for Qwen3.5 language_model and incorrect visual-encoder save_pretrained keys affect different components."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Two similar BERT loading reports, but the specific models and failure details are not enough to confirm one shared root cause."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping on A800 and Qwen3-MoE FP8 issues are unrelated."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "LoRA hf4.x/hf5.x compatibility is unrelated to Gemma4 resize_token_embeddings propagation."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Qwen3-MoE FP8 config failure and Strix Halo mmap OOM are unrelated."
+          },
+          {
+            "left": "issue:43708",
+            "right": "issue:44075",
+            "accept": false,
+            "reason": "Checkpoint max_steps math and SGD argument handling are separate trainer/optimizer bugs."
+          },
+          {
+            "left": "issue:45488",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Different tokenizer regressions: LlamaTokenizer pre-tokenizer override versus Granite token-ID corruption."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded and conservative overall. The 43644/44534 duplicate is well supported, and the SigLIP2 pair is stated as only a plausible duplicate rather than overclaimed. The remaining rejections stay close to the packet and avoid stretching across unrelated regressions."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 6,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:36010",
+        "issue:39404",
+        "issue:39692",
+        "issue:41720",
+        "issue:42491",
+        "issue:42947",
+        "issue:43257",
+        "issue:43335",
+        "issue:43381",
+        "issue:43404",
+        "issue:43425",
+        "issue:43454",
+        "issue:43644",
+        "issue:43645",
+        "issue:43854",
+        "issue:43856",
+        "issue:43873",
+        "issue:43931",
+        "issue:43957",
+        "issue:43994",
+        "issue:44155",
+        "issue:44164",
+        "issue:44291",
+        "issue:44292",
+        "issue:44387",
+        "issue:44402",
+        "issue:44488",
+        "issue:44743",
+        "issue:44756",
+        "issue:44910",
+        "issue:44912",
+        "issue:44928",
+        "issue:44960",
+        "issue:44991",
+        "issue:45216",
+        "issue:45276",
+        "issue:45292",
+        "issue:45335",
+        "issue:45362",
+        "issue:45406",
+        "issue:45588",
+        "issue:45663",
+        "issue:45684",
+        "issue:45698",
+        "issue:45704",
+        "issue:45759",
+        "issue:45823",
+        "issue:45859"
+      ],
+      "soft_pairs": [
+        "issue:43854|issue:43931",
+        "issue:39692|issue:43994",
+        "issue:43257|issue:43854",
+        "issue:44488|issue:44991",
+        "issue:42491|issue:43931",
+        "issue:43644|issue:43645",
+        "issue:43856|issue:43873",
+        "issue:45362|issue:45406",
+        "issue:43856|issue:44387",
+        "issue:44910|issue:45588",
+        "issue:42491|issue:45292",
+        "issue:45276|issue:45335",
+        "issue:44164|issue:44291",
+        "issue:43404|issue:43454",
+        "issue:36010|issue:39404",
+        "issue:41720|issue:44155",
+        "issue:44402|issue:44488",
+        "issue:39692|issue:44960",
+        "issue:43425|issue:43957",
+        "issue:42947|issue:45704",
+        "issue:45663|issue:45759",
+        "issue:43335|issue:45663",
+        "issue:43873|issue:44387",
+        "issue:45663|issue:45859",
+        "issue:44291|issue:44756",
+        "issue:44155|issue:45823",
+        "issue:43854|issue:45216",
+        "issue:44292|issue:44912",
+        "issue:44387|issue:44756",
+        "issue:44743|issue:44960",
+        "issue:43957|issue:44756",
+        "issue:45276|issue:45292",
+        "issue:45684|issue:45698",
+        "issue:43381|issue:44928",
+        "issue:44164|issue:44756",
+        "issue:42947|issue:43381"
+      ],
+      "prepared_review_unit_hash": "69ed056041d6186415d50a870e7b94d6619339214007d5a94d41d484d9eb5897",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 36,
+        "serialized_chars": 38618,
+        "estimated_input_tokens": 9655,
+        "estimated_eval_tokens": 19566
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most candidate pairs are only loosely related by subsystem or symptom. The only clear duplicate-like pair is the resize_token_embeddings/output_embeddings regression around Gemma4/T5Gemma, where the broader generic report is the better representative.",
+        "confidence": 0.83,
+        "canonical_issue_reason": "Issue 45292 is the clearest broad report of the resize_token_embeddings \u2192 output_embeddings regression; 45276 looks like a Gemma4-specific manifestation of the same defect.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "45292 best captures the underlying bug in a reusable, general form and is the strongest canonical issue for this duplicate cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43854",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Both are model-loading failures, but they affect different models and different failure modes."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Both involve SigLIP2, but one is a docs/example problem with multiple issues and the other is a runtime inference quality report; not the same concrete bug."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Different models and different root causes: Qwen3 MoE conversion vs GLM loading."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are tokenizer/load failures, but the reported models and symptoms are too different to treat as the same bug."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different model families and unrelated loading failures."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Both are Transformers 5.0 regressions, but one is buffer initialization and the other is notebook custom-model initialization."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Both mention memory/quantization, but one is Qwen3 MoE training memory use and the other is offloading behavior."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Different components and different models; both are crashes only at a very high level."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Both are memory-related, but the underlying causes and reproduction paths differ."
+          },
+          {
+            "left": "issue:44910",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Both are flash-attention crashes, but the code paths and failure signatures are unrelated."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:45292",
+            "accept": false,
+            "reason": "LoRA/Qwen3 MoE compatibility is unrelated to resize_token_embeddings behavior."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45335",
+            "accept": true,
+            "reason": "Both report resize_token_embeddings failing to update model-specific output/input embeddings; 45276 is the Gemma4-specific form of the same defect."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Different state-handling/loading regressions with different triggers and error types."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43454",
+            "accept": false,
+            "reason": "Both are lm_head tying issues, but they are on different architectures and not clearly the same bug."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:39404",
+            "accept": false,
+            "reason": "Completely different issues: an import error vs Whisper pipeline behavior."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44155",
+            "accept": false,
+            "reason": "Different model families and unrelated failure mechanisms."
+          },
+          {
+            "left": "issue:44402",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both concern tokenization/model loading, but there is no evidence they are the same underlying tokenizer bug."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "Different models and unrelated topics."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "General Torch version incompatibility is not the same as meta-device loading breakage."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Gradient checkpointing/LoRA and apex RMSNorm memory leak are unrelated."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Different model classes and different forward/configuration bugs."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "Different configuration/forward-path issues in unrelated models."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Both can lead to memory problems, but the mechanisms are different: offloading vs quantization memory overhead."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Different Qwen/Gemma subsystems and different offload compatibility issues."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Init-empty-weights argument handling is unrelated to mmap/OOM on Strix Halo."
+          },
+          {
+            "left": "issue:44155",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Different models and different runtime failures."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Different models and different save/load regressions."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Different quantization formats, models, and failure modes."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Both are OOM-related, but the causes are distinct and not mergeable into one fix."
+          },
+          {
+            "left": "issue:44743",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "Unrelated model behavior reports."
+          },
+          {
+            "left": "issue:43957",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Meta-device loading and Strix Halo mmap OOM are unrelated."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45292",
+            "accept": true,
+            "reason": "Same underlying resize_token_embeddings regression; 45292 is the broader, cleaner duplicate target."
+          },
+          {
+            "left": "issue:45684",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "One is file permission propagation, the other is wrong custom-module resolution after save/load."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:44928",
+            "accept": false,
+            "reason": "Both mention gradient/checkpointing behavior, but they are different problems with different causes."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Different bugs in unrelated code paths."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "Both involve gradient checkpointing, but the reported failures are distinct."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The canonical issue choice is grounded in the packet: 45292 is the broad resize_token_embeddings/output_embeddings report, and the Gemma4/T5Gemma issues read as specific manifestations of that same regression. The soft-edge verdicts are appropriately conservative overall, with no obvious overreach beyond the evidence in the titles and summaries."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 7,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:32090",
+        "issue:36032",
+        "issue:39404",
+        "issue:43257",
+        "issue:43299",
+        "issue:43425",
+        "issue:43452",
+        "issue:43596",
+        "issue:43611",
+        "issue:43645",
+        "issue:43646",
+        "issue:43708",
+        "issue:43716",
+        "issue:43828",
+        "issue:43856",
+        "issue:43901",
+        "issue:43906",
+        "issue:43931",
+        "issue:43957",
+        "issue:43994",
+        "issue:44079",
+        "issue:44164",
+        "issue:44265",
+        "issue:44402",
+        "issue:44451",
+        "issue:44479",
+        "issue:44509",
+        "issue:44530",
+        "issue:44589",
+        "issue:44671",
+        "issue:44683",
+        "issue:44756",
+        "issue:44805",
+        "issue:44863",
+        "issue:44877",
+        "issue:44898",
+        "issue:44912",
+        "issue:44918",
+        "issue:44991",
+        "issue:45081",
+        "issue:45084",
+        "issue:45292",
+        "issue:45335",
+        "issue:45588",
+        "issue:45663",
+        "issue:45701",
+        "issue:45759",
+        "issue:45823"
+      ],
+      "soft_pairs": [
+        "issue:44479|issue:44877",
+        "issue:43646|issue:43708",
+        "issue:45588|issue:45663",
+        "issue:45292|issue:45335",
+        "issue:43645|issue:44164",
+        "issue:44402|issue:44451",
+        "issue:43257|issue:44863",
+        "issue:43828|issue:43957",
+        "issue:43425|issue:43828",
+        "issue:44589|issue:44756",
+        "issue:44683|issue:45823",
+        "issue:39404|issue:45701",
+        "issue:43931|issue:44863",
+        "issue:43906|issue:44530",
+        "issue:43425|issue:45823",
+        "issue:43611|issue:43645",
+        "issue:43425|issue:43716",
+        "issue:44530|issue:45084",
+        "issue:43901|issue:44509",
+        "issue:39404|issue:44912",
+        "issue:44265|issue:44898",
+        "issue:44479|issue:44805",
+        "issue:43994|issue:44079",
+        "issue:43299|issue:43931",
+        "issue:43425|issue:44918",
+        "issue:44589|issue:45084",
+        "issue:39404|issue:44671",
+        "issue:44164|issue:45759",
+        "issue:36032|issue:43452",
+        "issue:43596|issue:44683",
+        "issue:44671|issue:44991",
+        "issue:32090|issue:39404",
+        "issue:43856|issue:44756",
+        "issue:44991|issue:45081"
+      ],
+      "prepared_review_unit_hash": "3b7cd63dd85484ecac9db69d604e0273178be214bdc0e2e2b1d459263bcb273a",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 38057,
+        "estimated_input_tokens": 9515,
+        "estimated_eval_tokens": 19286
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is mostly a loose thematic cluster of Transformers v5 regressions, not a true duplicate set. One pair looks like the same resize_token_embeddings bug; the rest are separate failures affecting different model families or code paths.",
+        "confidence": 0.68,
+        "canonical_issue_reason": "Issue 43611 is the broadest, most central regression report in the set: a Transformers 5 loading break around `base_model_prefix` that fits the cluster\u2019s main theme of model-initialization/loading regressions.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43611 is the best single issue to anchor the cluster because it describes a clear, general API regression with wide impact and a concrete failure mode.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44479",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Different failures: video-input regression for Qwen-family models vs strict config loading for granite_speech."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:43708",
+            "accept": false,
+            "reason": "Custom model initialization in notebooks is unrelated to Trainer resume/max_steps calculation."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "Different code paths: flash_attention `s_aux=None` crash vs Gemma4 FSDP2 shared_kv_states rebuild bug."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45335",
+            "accept": true,
+            "reason": "Same underlying bug: `resize_token_embeddings` not propagating to tied output embeddings; 45335 is a T5Gemma-specific manifestation."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:44164",
+            "accept": false,
+            "reason": "Notebook custom model init regression is not the same as save/from_pretrained handling of `extra_state`."
+          },
+          {
+            "left": "issue:44402",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Tokenizer vocab-size mismatch vs a separate tokenizer loading failure for a different model."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Both are model-loading issues, but they involve different architectures and different failure mechanisms."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Autocast dtype mismatch is not the same as meta-device loading breakage."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Torch version incompatibility is too broad and distinct from the Phi-tiny-MoE autocast dtype mismatch."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Float8 storage lookup failure and Strix Halo mmap OOM are unrelated."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Compiled flex_attention on torch>=2.9 is a different bug from Gemma4 device-map mismatch."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Whisper `return_language` regression is unrelated to tokenization changes across versions."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Qwen3-VL shape-mismatch loading failure is not the same as NemotronH checkpoint loading failure."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "An isolated reproduction of another issue is not the same as the PagedAttentionCache linear_attention crash."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Torch compatibility and Gemma4 forward device mismatch are different problems."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Both are Transformers 5 custom-model regressions, but one is about `base_model_prefix` loading and the other about notebook-defined model initialization."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43716",
+            "accept": false,
+            "reason": "Torch 2.10 incompatibility is not the same as the Mistral-3 image-preprocessor dtype mismatch."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "PagedAttentionCache group-type failure and template-node compilation failure are separate code paths."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "These are both docs-related, but they cover different removed/changed pipeline-task documentation issues."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Whisper pipeline `return_language` regression is unrelated to MXFP4 quantization fallback."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "`torch.export`/`torch_compilable_check` failure is unrelated to Perceiver non-default-resolution classification."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Video-input regression and mask-shape IndexError are different failure modes."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Siglip2 bad outputs are not the same bug as `ModelOutput` key assignment with prior `None` values."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Both mention Qwen3-VL loading, but one is a v5 break for MoE models and the other is a specific weight-shape mismatch; not enough evidence of the same bug."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Torch incompatibility is unrelated to the Qwen3.5 input-embedding unpacking issue."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Missing float8 storage and non-template-node compilation are unrelated."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Whisper pipeline regression is unrelated to CamemBERT masked-LM prediction errors."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "`extra_state` save/load handling is separate from `from_config` text_config unwrapping for composite Qwen configs."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "T5Tokenizer special-token method conflict is unrelated to gguf_file loading failures."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Deepspeed ZeRO-3 zero-sized tensor indexing is unrelated to compiled flex_attention on newer torch versions."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Tokenizer loading failure for est-roberta and Mistral regex patch backend-tokenizer access are different bugs."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded and conservative overall. The canonical issue choice is reasonable, and the summary correctly frames this as a loose Transformers v5 regression cluster rather than a true duplicate set. The one accepted soft-edge pair (45292/45335) is well supported by the issue titles and stays within the evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 8,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:34567",
+        "issue:36032",
+        "issue:40444",
+        "issue:40990",
+        "issue:42915",
+        "issue:43262",
+        "issue:43317",
+        "issue:43322",
+        "issue:43377",
+        "issue:43404",
+        "issue:43421",
+        "issue:43452",
+        "issue:43482",
+        "issue:43540",
+        "issue:43541",
+        "issue:43582",
+        "issue:43611",
+        "issue:43632",
+        "issue:43643",
+        "issue:43644",
+        "issue:43645",
+        "issue:43646",
+        "issue:43716",
+        "issue:43883",
+        "issue:43994",
+        "issue:44164",
+        "issue:44206",
+        "issue:44220",
+        "issue:44451",
+        "issue:44466",
+        "issue:44479",
+        "issue:44488",
+        "issue:44530",
+        "issue:44534",
+        "issue:44610",
+        "issue:44661",
+        "issue:44683",
+        "issue:44737",
+        "issue:44811",
+        "issue:44849",
+        "issue:44987",
+        "issue:44991",
+        "issue:45081",
+        "issue:45084",
+        "issue:45325",
+        "issue:45507",
+        "issue:45584",
+        "issue:45859"
+      ],
+      "soft_pairs": [
+        "issue:44737|issue:45859",
+        "issue:44164|issue:44479",
+        "issue:43582|issue:43632",
+        "issue:44488|issue:45081",
+        "issue:44661|issue:44683",
+        "issue:40990|issue:43421",
+        "issue:44488|issue:44987",
+        "issue:43611|issue:43644",
+        "issue:44206|issue:44220",
+        "issue:43716|issue:45084",
+        "issue:43883|issue:44488",
+        "issue:43643|issue:44530",
+        "issue:44811|issue:45584",
+        "issue:34567|issue:40444",
+        "issue:36032|issue:42915",
+        "issue:43644|issue:43646",
+        "issue:44451|issue:44987",
+        "issue:43452|issue:43482",
+        "issue:43541|issue:45084",
+        "issue:42915|issue:44530",
+        "issue:43317|issue:43482",
+        "issue:44849|issue:45325",
+        "issue:36032|issue:43645",
+        "issue:44987|issue:44991",
+        "issue:43883|issue:44991",
+        "issue:43404|issue:44466",
+        "issue:45084|issue:45507",
+        "issue:43262|issue:43377",
+        "issue:43994|issue:44610",
+        "issue:43646|issue:44534",
+        "issue:43322|issue:43540",
+        "issue:43645|issue:44534"
+      ],
+      "prepared_review_unit_hash": "b6c8734c85a5d40f89c97d8b3c6990902cf78ec89bdede0a9c0e034614bfc47a",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37334,
+        "estimated_input_tokens": 9334,
+        "estimated_eval_tokens": 18924
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly heterogeneous cluster of unrelated Transformers issues. A small subset concerns v5/custom-model initialization regressions, but the paired artifacts still describe different concrete failures, so no soft edges are accepted.",
+        "confidence": 0.71,
+        "canonical_issue_reason": "issue:43646 is the broadest, most representative issue in the cluster's main v5/custom-model-init regression theme, with a generic title that can encompass nearby initialization failures more cleanly than the more specific model- or buffer-related reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43646 is the best single issue representative because it is the most general and central to the repeated Transformers v5 initialization regression pattern in this set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44737",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Different model families and different bugs: XLNet CPU placement in a forward pass vs Qwen3.5 MoE CPU-offload incompatibility."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Serialization/extra_state handling is unrelated to a video-input regression in Qwen multimodal models."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:43632",
+            "accept": false,
+            "reason": "Apple Silicon allocator warmup TypeError and the _is_hf_initialized flag regression are distinct failures."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both are loading/tokenizer issues, but they hit different code paths and different models; not the same concrete bug."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Model-registration failure in add-new-model-like is unrelated to compiled flex_attention compatibility on torch >= 2.9."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:43421",
+            "accept": false,
+            "reason": "High perplexity on a model is not the same bug as runtime special-token post-processor updates."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Different loading failures for different models; no evidence they share the same underlying tokenizer/model bug."
+          },
+          {
+            "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; separate code paths."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Audio feature extraction failures are too vague here to conclude the same underlying bug; the titles point to different symptoms."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Mistral-3 dtype mismatch in image preprocessing is unrelated to a template compilation TypeError."
+          },
+          {
+            "left": "issue:43883",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Missing all_tied_weights_keys in Molmo is unrelated to tokenizer loading for cjvt/sleng-bert."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "AutoConfig trust_remote_code field loss and PagedAttentionCache linear_attention failure are unrelated."
+          },
+          {
+            "left": "issue:44811",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Whisper decode skip_special_tokens behavior and empty-transcription generation after align_special_tokens are related to the same area but not clearly the same concrete bug."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "TrainerState token counting and Qwen2.5-VL iterable-dataset multi-image failure are unrelated."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "T5 tokenizer special-token method conflict and Qwen3Moe FineGrainedFP8Config failure are unrelated."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43646",
+            "accept": false,
+            "reason": "Both are Transformers v5 regressions, but one concerns non-persistent buffers and the other custom model initialization; not the same fix target."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Different model-loading failures with no clear shared code path."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Both mention GGUF/loading, but one is a broad gguf_file from_pretrained breakage and the other is a specific Qwen2.5-GGUF failure; insufficient evidence they are the same bug."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Torch dynamo grouped_mm tracing error and non-template node compilation error are different failures."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "Different Qwen variants and different failure modes: FineGrainedFP8Config vs linear_attention cache group handling."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Device-map/offload dequantized-model loading issue is unrelated to a Qwen2.5-GGUF loading regression."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Qwen3.5 hidden-state output bug and Qwen2.5-VL rope-index scaling bug are distinct issues."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "Tokenizer method-name conflict is unrelated to custom model initialization in Jupyter notebooks."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Different models and likely different failure mechanisms; not the same underlying loading bug."
+          },
+          {
+            "left": "issue:43883",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Missing tied-weight attribute in Molmo is unrelated to tokenizer loading for EMBEDDIA/est-roberta."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Mistral3 lm_head tying bug and device-dependent serialization of lm_head.weight are related in broad theme only; not enough to call them the same bug."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Template compilation failure and GraniteMoEHybrid invalid method call are unrelated."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43377",
+            "accept": false,
+            "reason": "Audio sampling-rate defaulting and MIMI batching/padding-mask mismatch are separate issues."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Bad SIGLIP2 outputs and OmDet-Turbo resolution mismatch are different model/preprocessing problems."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Custom model initialization regression and non-persistent buffer junk are different concrete failures despite both being v5-related."
+          },
+          {
+            "left": "issue:43322",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Llava Next loading segfault and Qwen3OmniMoe video-input ValueError are unrelated multimodal failures."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Jupyter custom model initialization and non-persistent buffer initialization are not the same bug."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 9,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:36032",
+        "issue:40444",
+        "issue:40990",
+        "issue:41553",
+        "issue:41720",
+        "issue:42175",
+        "issue:42915",
+        "issue:43296",
+        "issue:43377",
+        "issue:43404",
+        "issue:43421",
+        "issue:43425",
+        "issue:43531",
+        "issue:43582",
+        "issue:43611",
+        "issue:43632",
+        "issue:43742",
+        "issue:43792",
+        "issue:43818",
+        "issue:43819",
+        "issue:44062",
+        "issue:44079",
+        "issue:44220",
+        "issue:44265",
+        "issue:44291",
+        "issue:44403",
+        "issue:44451",
+        "issue:44479",
+        "issue:44534",
+        "issue:44568",
+        "issue:44589",
+        "issue:44610",
+        "issue:44683",
+        "issue:44756",
+        "issue:44811",
+        "issue:44877",
+        "issue:44987",
+        "issue:44991",
+        "issue:45020",
+        "issue:45081",
+        "issue:45084",
+        "issue:45127",
+        "issue:45216",
+        "issue:45325",
+        "issue:45357",
+        "issue:45542",
+        "issue:45663"
+      ],
+      "soft_pairs": [
+        "issue:44403|issue:45081",
+        "issue:43377|issue:43819",
+        "issue:42915|issue:45084",
+        "issue:44265|issue:44610",
+        "issue:44451|issue:45081",
+        "issue:45216|issue:45357",
+        "issue:44987|issue:45081",
+        "issue:44220|issue:44683",
+        "issue:43404|issue:45127",
+        "issue:43632|issue:44756",
+        "issue:42915|issue:44291",
+        "issue:44062|issue:44987",
+        "issue:44479|issue:45325",
+        "issue:36032|issue:43742",
+        "issue:43296|issue:43531",
+        "issue:42915|issue:44589",
+        "issue:40444|issue:45325",
+        "issue:43531|issue:44877",
+        "issue:43611|issue:44534",
+        "issue:36032|issue:41720",
+        "issue:43425|issue:44403",
+        "issue:44079|issue:45663",
+        "issue:41553|issue:43582",
+        "issue:40990|issue:44811",
+        "issue:43632|issue:44534",
+        "issue:42175|issue:45542",
+        "issue:43421|issue:44568",
+        "issue:43818|issue:45357",
+        "issue:43792|issue:44589",
+        "issue:44683|issue:45020",
+        "issue:44403|issue:44991"
+      ],
+      "prepared_review_unit_hash": "0ce5cec36108bc038daf49caf19b5f64ba7ddf167eebbc6b62d1de9e726f4252",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36435,
+        "estimated_input_tokens": 9109,
+        "estimated_eval_tokens": 18474
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is still mostly heterogeneous: many issues are broad Transformers v5 regressions, but most soft-edge pairs are only loosely related by subsystem or model family rather than the same concrete bug. The two previously accepted pairs are better treated as related follow-ups/regressions, not confirmed duplicates, from the evidence shown.",
+        "confidence": 0.62,
+        "canonical_issue_reason": "issue:45020 is the best broad open umbrella for the cluster\u2019s main theme: recent Transformers versions breaking model loading/behavior, especially around remote_code and adjacent initialization paths.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45020 is the strongest representative issue because it is broad, still open, and best captures the recurring version-regression pattern across the cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44403",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both are loader-related, but this is generic load noise vs a specific Mistral regex crash; different bug paths."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Different model families and mechanisms: padding-mask batching bug vs DAC forward/from_latents mismatch."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config handling and a template-node compilation TypeError are unrelated."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Torch export/check issues are unrelated to an OmDet processor size mismatch."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Different tokenizers/models and different failure modes; not the same underlying bug."
+          },
+          {
+            "left": "issue:45216",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "They are both Qwen3.5 save_pretrained regressions, but 45357 looks like a follow-up about visual encoder keys rather than a confirmed duplicate root cause."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "One is a model-loading failure for a remote_code package; the other is a Mistral tokenizer regex crash."
+          },
+          {
+            "left": "issue:44220",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Audio feature extraction and compiled flex_attention are unrelated."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45127",
+            "accept": false,
+            "reason": "Both mention weights/tie_word_embeddings, but the concrete bugs are different: Mistral3 weight tying vs LoRA merge with extended vocab."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "A v5 initialization-flag regression is not the same as a Strix Halo mmap OOM issue."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Different loading/initialization failures with no clear shared code path."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "The AddedToken special-argument TypeError is not clearly the same as the physical-intelligence/fast load failure."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Related by Qwen VL family/version, but the titles do not clearly establish one root cause; this looks more like an adjacent video-input regression than a confirmed duplicate."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Tokenizer add_special_tokens conflict and MobileLLM key error are different loading bugs."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Different model families and different failure modes."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "FP8 config failure and missing Float8 storage are not the same underlying issue."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Both involve Qwen2.5-VL multimodal handling, but one is multi-image prompt iteration and the other is temporal position scaling; not clearly the same bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Sliding-window behavior in Qwen3-MoE is unrelated to strict config loading for granite_speech."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Base_model_prefix loading breakage and junk-filled non-persistent buffers are different v5 regressions."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Tokenizer load conflict and Qwen3 auto-device-map cuda assert are unrelated."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44403",
+            "accept": false,
+            "reason": "Torch version compatibility and generic transformer-load noise are not the same bug."
+          },
+          {
+            "left": "issue:44079",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "ModelOutput key assignment and Gemma4 FSDP2 shared_kv_states reconstruction are unrelated."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "A bad AutoTokenizer error message and an Apple Silicon caching allocator TypeError are unrelated."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Perplexity regression on GPT-OSS is unrelated to Whisper batch_decode skip_special_tokens behavior."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "These are different v5 regressions affecting different internals."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Both concern TensorFlow availability, but one is missing TF in the torch extra and the other is tensorboard-only backend detection; not the same fix."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both involve special tokens, but one is runtime post-processor refresh and the other is add_special_tokens not adding BOS/EOS for a specific tokenizer."
+          },
+          {
+            "left": "issue:43818",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Video-LLaVA temporal-attention/weight-sharing bug is unrelated to Qwen3.5 save_pretrained key serialization."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Whisper runtime failure and Float8 storage lookup error are different issues."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Compiled flex_attention on torch>=2.9 and remote_code model-loading regressions are separate problems."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Generic loading noise and a specific tokenizer load failure for EMBEDDIA/est-roberta are not the same underlying bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The summary is broadly grounded, but the soft-edge accepts are a bit too assertive for the evidence shown. In particular, issue:44479 vs issue:45325 looks related by model family/version, but the packet does not clearly establish a single root cause from titles alone, and issue:45216 vs issue:45357 likewise reads more like a continuing/follow-up regression than a confirmed duplicate. The rest of the rejects are appropriately conservative."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 10,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:33666",
+        "issue:35707",
+        "issue:40444",
+        "issue:41720",
+        "issue:42175",
+        "issue:42915",
+        "issue:43296",
+        "issue:43317",
+        "issue:43377",
+        "issue:43381",
+        "issue:43454",
+        "issue:43482",
+        "issue:43525",
+        "issue:43531",
+        "issue:43611",
+        "issue:43632",
+        "issue:43644",
+        "issue:43646",
+        "issue:43701",
+        "issue:43749",
+        "issue:43792",
+        "issue:43856",
+        "issue:43957",
+        "issue:43994",
+        "issue:44079",
+        "issue:44265",
+        "issue:44291",
+        "issue:44387",
+        "issue:44462",
+        "issue:44466",
+        "issue:44479",
+        "issue:44534",
+        "issue:44589",
+        "issue:44617",
+        "issue:44829",
+        "issue:44898",
+        "issue:45072",
+        "issue:45081",
+        "issue:45325",
+        "issue:45381",
+        "issue:45479",
+        "issue:45491",
+        "issue:45538",
+        "issue:45800",
+        "issue:45812",
+        "issue:45823",
+        "issue:45834",
+        "issue:45923"
+      ],
+      "soft_pairs": [
+        "issue:44462|issue:45812",
+        "issue:33666|issue:40444",
+        "issue:44829|issue:45479",
+        "issue:43317|issue:43856",
+        "issue:43482|issue:45538",
+        "issue:43381|issue:45923",
+        "issue:35707|issue:40444",
+        "issue:44079|issue:44534",
+        "issue:43482|issue:43525",
+        "issue:44291|issue:44589",
+        "issue:43377|issue:45491",
+        "issue:43749|issue:44079",
+        "issue:40444|issue:45381",
+        "issue:43644|issue:44079",
+        "issue:41720|issue:45081",
+        "issue:43531|issue:44291",
+        "issue:44265|issue:45072",
+        "issue:43632|issue:45072",
+        "issue:45325|issue:45381",
+        "issue:43994|issue:44898",
+        "issue:43611|issue:43646",
+        "issue:42175|issue:43296",
+        "issue:43454|issue:44466",
+        "issue:45325|issue:45834",
+        "issue:43792|issue:43957",
+        "issue:42915|issue:43482",
+        "issue:43296|issue:45800",
+        "issue:44387|issue:44617",
+        "issue:43701|issue:45325",
+        "issue:44479|issue:45823"
+      ],
+      "prepared_review_unit_hash": "033315bcf69e709c1019b06823a140019358ecb6c8cf1b01cfc5def71b2f636e",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36752,
+        "estimated_input_tokens": 9188,
+        "estimated_eval_tokens": 18632
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This is mostly a loose mix of Transformers v5 / model-specific regression reports, not one tight duplicate set. The only strong soft-edge match is the Qwen2.5-VL position-id regression pair; the rest are related by subsystem but appear to be distinct bugs.",
+        "confidence": 0.57,
+        "canonical_issue_reason": "Best umbrella representative of the cluster\u2019s recurring Transformers v5 loading/initialization regressions; broad enough to anchor triage better than the more model-specific tickets.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Most central single issue in the set: it captures the generic v5 init/loading breakage theme that several of the other issues orbit, without being overly narrow.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44462",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are AutoTokenizer regressions, but one is about ignoring tokenizer.json while the other is a Granite-specific wrong-token-IDs bug; related area, different failure mode."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "Both are Qwen VL training problems, but one is multi-GPU training and the other is IterableDataset handling with multiple images per prompt; not the same bug."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Both mention degenerate training, but the reported causes differ: flash_attention_3 behavior vs. single-label classification with num_labels=1."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "One is device_map/offload loading failure, the other is Qwen3 MoE memory inefficiency; same broad area, different issue."
+          },
+          {
+            "left": "issue:43482",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Qwen2.5-GGUF loading on transformers v5 and CLIPTokenizer model_max_length are unrelated bugs."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode is a runtime-mode bug; missing supports_gradient_checkpointing on trust_remote_code is a model config capability issue."
+          },
+          {
+            "left": "issue:35707",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "Progressive generation with inputs_embeds/past_key_values is unrelated to Qwen VL finetuning with multi-image iterable data."
+          },
+          {
+            "left": "issue:44079",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Both are Transformers v5 regressions, but one is ModelOutput key assignment and the other is non-persistent buffer initialization; different code paths."
+          },
+          {
+            "left": "issue:43482",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "GGUF loading failure and Llama4Config missing pad_token_id are separate model/load issues."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Both occur in model loading, but one is unexpected _is_hf_initialized handling and the other is Float8 storage lookup; not the same bug."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Both involve batch-vs-single discrepancies, but one is MIMI encoder padding masks and the other is Gemma3 sliding-window all-padding NaNs; different models and mechanisms."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "FSDP_CPU_RAM_EFFICIENT_LOADING and ModelOutput key assignment are unrelated."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Qwen2.5-VL multi-image finetuning failure is not the same as the video position-id regression in v5.3.0."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Non-persistent buffer junk and ModelOutput key assignment are distinct v5 regressions."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping CUDA assert and Mistral regex patch crash are unrelated."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Sliding_window behavior in Qwen3-MoE and init_empty_weights/_is_hf_initialized are different issues."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "torch.export with torch_compilable_check and dtype mismatches in bfloat16 inference do not share the same root cause."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "_is_hf_initialized breakage and bfloat16 dtype mismatch are separate Transformers v5 regressions."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45381",
+            "accept": true,
+            "reason": "Both describe Qwen2.5-VL v5.3.0 position-id generation bugs in the same code path; a single fix could plausibly cover image and video variants."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "SigLIP2 nonsensical outputs and Perceiver non-default resolution failure are different model-specific bugs."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43646",
+            "accept": false,
+            "reason": "Both are broad v5 loading/init regressions, but one is base_model_prefix handling and the other is custom model initialization; too different to merge confidently."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43296",
+            "accept": false,
+            "reason": "TensorFlow packaging/backend inclusion and PaddleOCR-VL model loading are unrelated."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "AyaVision lm_head tying and inconsistent lm_head serialization are related by weights, but they are not clearly the same defect."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Qwen2.5-VL position IDs and Kosmos2.5 OCR long-input index errors are different model bugs."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Whisper model load failure and torch.device('meta') loading regressions are separate issues."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Qwen3MoE FineGrainedFP8Config failure and Qwen2.5-GGUF loading failure are different model/config paths."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "PaddleOCR-VL loading on transformers 4.53.0/vLLM and torch 2.4.1 vs transformers 5.8.0 incompatibility are unrelated."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Int4 memory regression and Sam3Video OOM are both memory-related but not the same bug or code path."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "resume_from_checkpoint key mismatch is unrelated to Qwen2.5-VL rope/position-id scaling."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Qwen video input regression and Gemma4 PLE device mismatch are different model-specific failures."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays appropriately conservative: it frames the set as a mixed cluster with a recurring v5 loading/init theme without overclaiming a single root cause. The only accepted soft-edge pair, issue:45325 and issue:45381, is a plausible same-code-path Qwen2.5-VL position-id regression. The remaining pairs are rejected with cautious, bug-specific distinctions rather than overstating duplication."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 11,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:34634",
+        "issue:35707",
+        "issue:39692",
+        "issue:40444",
+        "issue:40990",
+        "issue:41720",
+        "issue:42947",
+        "issue:43257",
+        "issue:43262",
+        "issue:43296",
+        "issue:43317",
+        "issue:43322",
+        "issue:43377",
+        "issue:43404",
+        "issue:43454",
+        "issue:43526",
+        "issue:43550",
+        "issue:43572",
+        "issue:43606",
+        "issue:43720",
+        "issue:43746",
+        "issue:43749",
+        "issue:43819",
+        "issue:43873",
+        "issue:43881",
+        "issue:43883",
+        "issue:44008",
+        "issue:44186",
+        "issue:44265",
+        "issue:44291",
+        "issue:44466",
+        "issue:44488",
+        "issue:44534",
+        "issue:44568",
+        "issue:44610",
+        "issue:44877",
+        "issue:44898",
+        "issue:44977",
+        "issue:44991",
+        "issue:45072",
+        "issue:45357",
+        "issue:45584",
+        "issue:45701",
+        "issue:45759",
+        "issue:45823",
+        "issue:45920",
+        "issue:45923"
+      ],
+      "soft_pairs": [
+        "issue:43881|issue:44488",
+        "issue:44291|issue:44877",
+        "issue:43606|issue:44265",
+        "issue:44568|issue:45920",
+        "issue:43262|issue:43746",
+        "issue:40990|issue:45584",
+        "issue:34634|issue:35707",
+        "issue:43873|issue:45701",
+        "issue:44186|issue:44265",
+        "issue:35707|issue:41720",
+        "issue:43606|issue:43819",
+        "issue:41720|issue:45823",
+        "issue:43454|issue:43550",
+        "issue:43377|issue:44610",
+        "issue:43317|issue:43873",
+        "issue:39692|issue:40444",
+        "issue:43572|issue:45759",
+        "issue:43454|issue:45072",
+        "issue:43377|issue:44898",
+        "issue:43257|issue:45923",
+        "issue:43526|issue:43883",
+        "issue:40990|issue:42947",
+        "issue:44466|issue:45357",
+        "issue:43881|issue:44991",
+        "issue:43749|issue:44534",
+        "issue:43454|issue:43720",
+        "issue:41720|issue:43322",
+        "issue:43296|issue:44877",
+        "issue:41720|issue:44977",
+        "issue:44008|issue:45823",
+        "issue:43404|issue:45357"
+      ],
+      "prepared_review_unit_hash": "0c8521e0374cf97aa71e6872374642779e418a5a1f7c222cd174d7277dfe01a9",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36703,
+        "estimated_input_tokens": 9176,
+        "estimated_eval_tokens": 18608
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is not a true duplicate cluster: the items span unrelated failures across loading, tokenization, device mapping, quantization, audio, vision, and serialization. A few are superficially adjacent, but none look like the same underlying bug/change with enough certainty to merge.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43873 is the broadest and most central to the only recurring theme in the set (offloading/quantization/device-map behavior), but it still does not represent the full mix of unrelated issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43881",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both are model-loading failures, but they involve different models and different failure modes; not the same underlying bug."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Both touch loading/config handling in v5, but one is an init_empty_weights argument issue and the other is strict config loading; different root causes."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44265",
+            "accept": false,
+            "reason": "CPU-offload device mismatch for Bark is unrelated to torch.export failures with torch_compilable_check."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45920",
+            "accept": false,
+            "reason": "Both are tokenizer-related regressions, but one is missing special tokens and the other is wrong token IDs across models; not the same bug."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate defaulting and PEFT local checkpoint loading are unrelated."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "A perplexity regression and an empty-transcription generation failure are different issues."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "BarkProcessor voice_preset and progressive generation with inputs_embeds/past_key_values are unrelated."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Offloading/quantization behavior is unrelated to tokenizer changes across versions."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44265",
+            "accept": false,
+            "reason": "Tokenizer batching/padding crashes and torch.export check failures are different code paths."
+          },
+          {
+            "left": "issue:35707",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Progressive generation with caches and an auto device-map CUDA assert are not the same bug."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Both mention model behavior bugs, but one is offload/device mismatch and the other is DAC latent conversion mismatch."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Both involve device_map=auto, but the model-specific forward/device mismatch problems are different."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Weight tying in AyaVision is unrelated to Bamba SDPA/torch.compile failures."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Missing padding-mask support in MIMI and image-size mismatch in OmDet-Turbo are different preprocessing bugs."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "These are adjacent around offload/quantization, but one is dequantized-model loading on gpu+cpu offload and the other is generic offloading behavior; too broad to merge."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "A doc/example error plus quantization failure is unrelated to iterable-dataset multi-image finetuning failure."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Missing pad_token_idx in one config and not unwrapping text_config in another are distinct config bugs."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Weight-tying and bfloat16 dtype mismatches are unrelated."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Both are vision/multimodal preprocessing issues, but one is padding-mask consistency and the other is non-default resolution handling."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "MOE weight conversion under accelerate/deepspeed and a missing gradient_checkpointing flag are unrelated."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "Label reduction in BeitImageProcessorFast and a missing all_tied_weights_keys attribute are unrelated."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:42947",
+            "accept": false,
+            "reason": "Perplexity on a benchmark and gradient checkpointing with LoRA are different problem areas."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Both concern serialization, but one is tied lm_head.weight consistency and the other is incorrect visual encoder keys; not the same fix."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Model loading failure for one model and tokenizer loading failure for a different model are unrelated."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "FSDP CPU-efficient loading and non-persistent buffer initialization are separate v5 regressions."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Weight tying in AyaVision and BitNet packed-weight unpacking are unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43322",
+            "accept": false,
+            "reason": "Both are severe load/runtime failures, but they are model-specific and do not share a concrete code-path problem."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "PaddleOCR-VL loading in vLLM and granite_speech strict-config loading are unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Both involve Qwen models, but one is an auto device-map crash and the other is flash-attention generation behavior; not the same bug."
+          },
+          {
+            "left": "issue:44008",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Gemma3n variable-name collision in forward is unrelated to Gemma4 PLE device mismatch."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "lm_head weight tying in Mistral3 and Qwen3.5 visual key serialization are different regressions."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 12,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "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:43782",
+        "issue:43873",
+        "issue:43881",
+        "issue:44186",
+        "issue:44361",
+        "issue:44387",
+        "issue:44423",
+        "issue:44479",
+        "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:45406",
+        "issue:45632",
+        "issue:45684",
+        "issue:45701",
+        "issue:45704",
+        "issue:45812"
+      ],
+      "soft_pairs": [
+        "issue:42617|issue:42915",
+        "issue:40990|issue:45701",
+        "issue:45632|issue:45684",
+        "issue:44186|issue:45072",
+        "issue:43577|issue:43708",
+        "issue:44423|issue:45406",
+        "issue:44387|issue:45704",
+        "issue:43873|issue:44756",
+        "issue:44877|issue:44991",
+        "issue:36331|issue:42617",
+        "issue:43550|issue:43720",
+        "issue:36331|issue:42915",
+        "issue:43450|issue:45701",
+        "issue:44610|issue:44898",
+        "issue:44186|issue:44610",
+        "issue:44479|issue:44918",
+        "issue:43550|issue:43606",
+        "issue:41720|issue:42617",
+        "issue:43316|issue:44913",
+        "issue:42915|issue:43317",
+        "issue:43295|issue:45701",
+        "issue:43526|issue:43881",
+        "issue:44671|issue:44743",
+        "issue:41720|issue:45161",
+        "issue:45081|issue:45812",
+        "issue:43531|issue:43572",
+        "issue:43540|issue:43782",
+        "issue:36010|issue:42617",
+        "issue:43421|issue:45812",
+        "issue:44186|issue:44361",
+        "issue:42915|issue:43322"
+      ],
+      "prepared_review_unit_hash": "1c6c256ef97eeda4ebbaa53e9b85a3b434b91bc691d7ca2708e26b4516cc1800",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36462,
+        "estimated_input_tokens": 9116,
+        "estimated_eval_tokens": 18488
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This is a loose similarity bucket of mostly unrelated Transformers issues. The pairs share broad themes like tokenization, loading, offloading, or multimodal serving, but they do not look like the same underlying bug/change, so none should be merged as duplicates.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No true duplicate cluster is present. If a single anchor is needed, issue 43873 is the most representative cross-cutting report because it has the highest activity and broadest framing, but it is not a duplicate of the others.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43873 is the best standalone issue to keep as a representative anchor: it has strong discussion/reference activity and a broad, actionable symptom description.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42617",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Different bugs: 3d_parallel.py execution failure vs Qwen3Moe loading failure with FineGrainedFP8Config."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "One is a perplexity quality report; the other is a tokenizer-version regression. Same area, not the same bug."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:45684",
+            "accept": false,
+            "reason": "Cache-path collision for trust_remote_code is unrelated to save_pretrained permission propagation."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Tokenizer crashes on NER/padding vs dtype mismatches in bfloat16 inference are unrelated failure modes."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43708",
+            "accept": false,
+            "reason": "Blip2 dtype loading bug vs Trainer resume/max_steps calculation bug are different code paths."
+          },
+          {
+            "left": "issue:44423",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both involve serve/multimodal crashes, but the errors and likely code paths are model-specific and distinct."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Int4 quantization memory growth/OOM is unrelated to T5 using apex.FusedRMSNorm memory leak."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Quantized offloading behavior and Strix Halo mmap OOM are both memory-related but not the same underlying bug."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Strict config loading for granite_speech is unrelated to tokenizer loading for EMBEDDIA/est-roberta."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "CustomTrainer compute_loss signature breakage is unrelated to 3d_parallel.py runtime failure."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure in Bamba is a different issue from accelerate loading of packed BitNet weights."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Trainer compute_loss API change and Qwen3Moe FP8 loading failure are unrelated."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Batched video processor shape bug is unrelated to a tokenizer-version regression."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Both involve image size expectations, but OmDet-Turbo and Perceiver have different model-specific preprocessing bugs."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Tokenizer padding/truncation crash and processor output-size mismatch are different components and failures."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Video-input regression in Qwen VL models is unrelated to SFT trainer embedding unpacking."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43606",
+            "accept": false,
+            "reason": "Bamba torch.compile/SDPA failure is unrelated to Bark CPU offload device mismatch."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "Auto device mapping cudaErrorAssert and 3d_parallel.py startup failure are different problems."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Gemma3TextConfig API discrepancy is unrelated to GPTNeoX rotary_pct reload behavior."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43317",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 loading failure and dequantized model offload failure are both loading-related but not the same bug."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Processor/tokenizer regression for custom multimodal code is not the same as a general tokenization-version change report."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction bug is unrelated to GLM-4V model loading failure."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "CamemBERT masked LM prediction regression and Qwen recurrent-state reset bug are completely different model behaviors."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Auto device mapping failure on Qwen3 and TP-only failure on GPT-OSS MoE are different distributed-inference issues."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Mistral regex patch crash and Granite tokenizer wrong IDs are both tokenizer bugs, but they affect different mechanisms and models."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior and StableLm pad_token_idx config regression are unrelated."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:43782",
+            "accept": false,
+            "reason": "Video-processing ValueError in Qwen3OmniMoe and from_pretrained weight_only error in Qwen3VL are different bugs."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "GenerationMixin import failure is unrelated to 3d_parallel.py execution failure."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Special-token post-processor update bug and Granite token-ID regression are distinct tokenizer issues."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "LayoutLMv2Tokenizer NER/padding crash and MLukeTokenizer AttributeError are different tokenizer implementations and failures."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43322",
+            "accept": false,
+            "reason": "Qwen3Moe FineGrainedFP8 loading failure is unrelated to Llava Next segmentation fault."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 13,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:30064",
+        "issue:33357",
+        "issue:34567",
+        "issue:34634",
+        "issue:35707",
+        "issue:39692",
+        "issue:41720",
+        "issue:42490",
+        "issue:43329",
+        "issue:43388",
+        "issue:43421",
+        "issue:43531",
+        "issue:43550",
+        "issue:43606",
+        "issue:43653",
+        "issue:43720",
+        "issue:43746",
+        "issue:43782",
+        "issue:43873",
+        "issue:43927",
+        "issue:44062",
+        "issue:44077",
+        "issue:44368",
+        "issue:44451",
+        "issue:44492",
+        "issue:44509",
+        "issue:44514",
+        "issue:44545",
+        "issue:44589",
+        "issue:44610",
+        "issue:44625",
+        "issue:44898",
+        "issue:44987",
+        "issue:44991",
+        "issue:45042",
+        "issue:45059",
+        "issue:45072",
+        "issue:45357",
+        "issue:45381",
+        "issue:45406",
+        "issue:45464",
+        "issue:45479",
+        "issue:45701",
+        "issue:45704",
+        "issue:45759",
+        "issue:45859",
+        "issue:45923"
+      ],
+      "soft_pairs": [
+        "issue:43388|issue:45059",
+        "issue:45406|issue:45464",
+        "issue:33357|issue:35707",
+        "issue:44492|issue:44509",
+        "issue:43421|issue:45701",
+        "issue:41720|issue:43746",
+        "issue:43606|issue:43720",
+        "issue:30064|issue:34634",
+        "issue:43550|issue:43653",
+        "issue:43873|issue:45704",
+        "issue:44514|issue:44545",
+        "issue:44625|issue:45479",
+        "issue:43606|issue:43653",
+        "issue:43531|issue:45759",
+        "issue:42490|issue:43329",
+        "issue:34634|issue:41720",
+        "issue:34567|issue:35707",
+        "issue:43421|issue:43927",
+        "issue:44368|issue:45357",
+        "issue:43782|issue:45381",
+        "issue:44898|issue:45072",
+        "issue:44451|issue:44991",
+        "issue:44077|issue:44625",
+        "issue:44987|issue:45042",
+        "issue:45759|issue:45859",
+        "issue:43531|issue:44589",
+        "issue:44062|issue:44451",
+        "issue:45704|issue:45923",
+        "issue:39692|issue:44610"
+      ],
+      "prepared_review_unit_hash": "b38c460e60fb064fb2815ae4eb46bbc9d315767e9a0191c21a676116d66d14ac",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 36145,
+        "estimated_input_tokens": 9037,
+        "estimated_eval_tokens": 18330
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous issue reports from different code paths. The only clear duplicate pair is the Qwen2_5_VLProcessor batched apply_chat_template crash with padding=False (issues 44514 and 44545).",
+        "confidence": 0.96,
+        "canonical_issue_reason": "Issue 44514 is the best canonical issue because it has an exact duplicate counterpart (44545) with the same title, same failure mode, and same affected API.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44514 is the strongest representative issue in this set because it is a confirmed duplicate target; the rest of the cluster does not form a coherent duplicate group.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43388",
+            "right": "issue:45059",
+            "accept": false,
+            "reason": "Different subsystems and failure modes: metric gathering/last-batch label handling vs SAM3 text+box behavior."
+          },
+          {
+            "left": "issue:45406",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Unrelated bugs: Gemma4Processor missing _tokenizer vs streaming chat/completions failure on Qwen3.5."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "Different problems and code paths: MacOS bus error loading CLIP vs generation with inputs_embeds/past_key_values."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "A typo docs issue is not the same as stale pipeline-task documentation; both concern docs, but not the same bug."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Both relate to tokenization behavior, but one is runtime post-processor updating and the other is a version-dependent tokenization change."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Different model families and failure types: CUDA assert during device mapping vs PEFT local-checkpoint loading."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Both touch accelerate/offload/loading, but one is a CPU offload device mismatch and the other is packed-weight unpacking during loading."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:34634",
+            "accept": false,
+            "reason": "Completely different components: image processor segmentation maps vs BarkProcessor voice_preset handling."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Unrelated bugs: torch.compile+SDPA in Bamba vs tokenizer special-token registration in BigBirdTokenizer."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Different underlying issues: quantization/offloading behavior vs T5 using apex.FusedRMSNorm with a memory leak."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:44545",
+            "accept": true,
+            "reason": "Exact duplicate: same title, same API, same batched padding=False crash."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "One is config propagation of num_labels; the other is a zero-loss bug for single_label_classification across models."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Device mismatch during CPU offload is not the same as a tokenizer special-token decode bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Different Qwen-related issues: sliding_window behavior in MoE models vs from_config text_config unwrapping for multimodal configs."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Local-path trailing-slash import warning is unrelated to undefined video-token helper/merge_size in multimodal token counting."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Different failures in different models: BarkProcessor voice preset vs Qwen3 device-mapping CUDA assert."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "TrainerState token accounting bug is unrelated to progressive generation with inputs_embeds and past_key_values."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Runtime special-token/post-processor update bug is different from DiaConfig losing custom token IDs on save/load."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different Qwen3.5 issues: tie_word_embeddings warning during LoRA fine-tuning vs incorrect visual encoder keys on save_pretrained."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Different VL bugs: weight_only load failure vs incorrect video vision_position_ids."
+          },
+          {
+            "left": "issue:44898",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Different components and symptoms: Perceiver non-default-resolution image classification vs dtype mismatch in inference wrappers."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are model-loading regressions, but they affect different tokenizers/models and are not the same concrete bug."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44625",
+            "accept": false,
+            "reason": "PatchTSMixer post_init policy is unrelated to Qwen3.5 num_labels propagation."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Different root causes: loading a specific model fails vs PIL image processors requiring torchvision."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Unrelated Qwen3.5 issues: multimodal config unwrapping vs rotary embedding CPU-offload compatibility."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Sliding_window behavior in Qwen3-MoE is unrelated to missing Float8 storage registration."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "AddedToken special-argument duplication is unrelated to failing to load a specific model."
+          },
+          {
+            "left": "issue:45704",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Memory leak in T5 RMSNorm is unrelated to missing supports_gradient_checkpointing on a trust_remote_code variant."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "SigLIP2 docs/model mismatch and quantization errors are different from OmDet-Turbo processor size mismatch."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded and conservative. The packet supports exactly one clear duplicate pair: issue 44514 and 44545 share the same title and same bug description. The remaining soft pairs are plausibly non-duplicates and the reasons stay within the evidence provided."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 14,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:30064",
+        "issue:36010",
+        "issue:39692",
+        "issue:40444",
+        "issue:42886",
+        "issue:43295",
+        "issue:43335",
+        "issue:43381",
+        "issue:43421",
+        "issue:43525",
+        "issue:43531",
+        "issue:43540",
+        "issue:43550",
+        "issue:43596",
+        "issue:43618",
+        "issue:43653",
+        "issue:43697",
+        "issue:43716",
+        "issue:43720",
+        "issue:43746",
+        "issue:43756",
+        "issue:43819",
+        "issue:43844",
+        "issue:43873",
+        "issue:43994",
+        "issue:44008",
+        "issue:44186",
+        "issue:44220",
+        "issue:44661",
+        "issue:44671",
+        "issue:44843",
+        "issue:44918",
+        "issue:44936",
+        "issue:45072",
+        "issue:45081",
+        "issue:45183",
+        "issue:45356",
+        "issue:45381",
+        "issue:45479",
+        "issue:45588",
+        "issue:45663",
+        "issue:45704",
+        "issue:45759",
+        "issue:45834",
+        "issue:45859",
+        "issue:45907",
+        "issue:45923"
+      ],
+      "soft_pairs": [
+        "issue:44918|issue:44936",
+        "issue:45663|issue:45923",
+        "issue:43421|issue:43653",
+        "issue:45588|issue:45759",
+        "issue:42886|issue:44843",
+        "issue:43381|issue:43844",
+        "issue:43421|issue:44186",
+        "issue:43994|issue:44671",
+        "issue:43525|issue:43531",
+        "issue:45704|issue:45859",
+        "issue:45081|issue:45356",
+        "issue:45356|issue:45907",
+        "issue:39692|issue:43873",
+        "issue:43596|issue:44220",
+        "issue:43746|issue:43873",
+        "issue:40444|issue:43746",
+        "issue:43746|issue:45381",
+        "issue:43540|issue:44008",
+        "issue:43720|issue:43819",
+        "issue:45183|issue:45479",
+        "issue:44220|issue:44661",
+        "issue:43618|issue:43697",
+        "issue:30064|issue:36010",
+        "issue:43653|issue:43720",
+        "issue:43335|issue:43756",
+        "issue:43716|issue:44008",
+        "issue:43295|issue:45834",
+        "issue:43550|issue:45072",
+        "issue:43550|issue:43819"
+      ],
+      "prepared_review_unit_hash": "5b331124060c3d965a9dbcbc3a0bc35fc09bd008525cd780464686701497352e",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 36297,
+        "estimated_input_tokens": 9075,
+        "estimated_eval_tokens": 18406
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is mostly heterogeneous and does not form one duplicate cluster. Only one soft pair looks like the same underlying regression: the HF_HUB_OFFLINE tokenizer/cache failure in the Mistral regex patch path.",
+        "confidence": 0.71,
+        "canonical_issue_reason": "Issue 44843 is the best anchor for the only clear duplicate-like theme: it names the root cause directly (`_patch_mistral_regex` unconditionally calling `model_info()` and breaking HF_HUB_OFFLINE).",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44843 is the most concrete and root-cause-oriented issue in the set, and it best represents the accepted offline-tokenizer regression pair.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44918",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Different failures: SFT embedding unpacking vs trainer state after train/evaluate. No shared code-path bug."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Unrelated model-specific bugs: FSDP2 shared_kv_states KeyError vs missing supports_gradient_checkpointing flag."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Both tokenizer-related, but one is runtime post-processor updates and the other is BigBird special-token registration; not the same bug."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Different subsystems and symptoms: flash-attention sink handling vs Qwen config unwrapping."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:44843",
+            "accept": true,
+            "reason": "Same underlying HF_HUB_OFFLINE regression in tokenizer loading: both point to the Mistral regex patch path triggering network metadata lookup and breaking cache-only loads."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "Gradient checkpointing eval-mode behavior is unrelated to ZeRO-3 gradient growth under DeepSpeed."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44186",
+            "accept": false,
+            "reason": "Both touch tokenization, but one is post-processor updates and the other is LayoutLMv2 NER/padding crash; different bugs."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Different model families and problems: SigLIP2 pipeline output issues vs CamemBERT MLM prediction regression."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Both are model-config bugs, but pad_token_id missing on Llama4 is unrelated to sliding_window handling in Qwen3-MoE."
+          },
+          {
+            "left": "issue:45704",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "T5 apex fused RMSNorm memory leak is unrelated to Qwen3_5Moe CPU offload compatibility."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Both mention Mistral regex, but one is a backend_tokenizer AttributeError and the other is a Kimi codec/regression warning issue; not clearly the same bug."
+          },
+          {
+            "left": "issue:45356",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Kimi tokenizer regression is unrelated to `list_repo_files` needing to forward revision."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Documentation/example mismatch is unrelated to quantization/offloading behavior."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Different low-level errors: Deepspeed ZeRO-3/BertModel init vs `_torch_extract_fbank_features()`."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Local PEFT checkpoint loading for GraniteSpeech is unrelated to offloading with quantization."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Multi-image iterable dataset finetuning failure is unrelated to GraniteSpeech checkpoint loading."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Different model paths: GraniteSpeech local checkpoint loading vs Qwen2.5-VL video vision_position_ids."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44008",
+            "accept": false,
+            "reason": "Video processing ValueError in Qwen3OmniMoe is unrelated to Gemma 3n variable name collision."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "BitNet packed-weight loading and DAC.from_latents/STE mismatch are distinct bugs."
+          },
+          {
+            "left": "issue:45183",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Different validation/loss problems: input_ids/inputs_embeds XOR messaging vs single-label classification degeneracy."
+          },
+          {
+            "left": "issue:44220",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "Audio fbank extraction issue is unrelated to add-new-model-like failing on TOKENIZER_MAPPING_NAMES."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "CLIP attentions assignment regression is not the same as RTDetrV2 output changes in v5."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Void segmentation maps in Image Processor is unrelated to GenerationMixin import failure."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration and BitNet accelerate loading are unrelated."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43756",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer creation bug is unrelated to Smollm3 RoPE layer dropping."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44008",
+            "accept": false,
+            "reason": "Image preprocessor/model dtype mismatch is unrelated to Gemma 3n audio_mel_mask name collision."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Processor/tokenizer regression in v4.57.5 is unrelated to Kosmos2.5 long OCR index errors."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Different dtype/compile failures in different models; no shared concrete code-path problem."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Bamba SDPA torch.compile failure is unrelated to DAC.from_latents missing-STE mismatch."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative. The only clearly same-root-cause pair is issue:42886 and issue:44843 around the HF_HUB_OFFLINE / _patch_mistral_regex tokenizer regression, and the other soft-edge rejections do not overstate evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 15,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:30064",
+        "issue:36331",
+        "issue:39692",
+        "issue:41720",
+        "issue:43322",
+        "issue:43421",
+        "issue:43531",
+        "issue:43540",
+        "issue:43577",
+        "issue:43582",
+        "issue:43606",
+        "issue:43643",
+        "issue:43644",
+        "issue:43653",
+        "issue:43749",
+        "issue:43782",
+        "issue:43818",
+        "issue:43819",
+        "issue:43825",
+        "issue:43827",
+        "issue:43994",
+        "issue:44062",
+        "issue:44186",
+        "issue:44442",
+        "issue:44448",
+        "issue:44479",
+        "issue:44485",
+        "issue:44488",
+        "issue:44509",
+        "issue:44561",
+        "issue:44671",
+        "issue:44805",
+        "issue:44811",
+        "issue:44821",
+        "issue:44843",
+        "issue:44877",
+        "issue:44898",
+        "issue:44936",
+        "issue:44977",
+        "issue:44991",
+        "issue:45020",
+        "issue:45072",
+        "issue:45081",
+        "issue:45084",
+        "issue:45381",
+        "issue:45584",
+        "issue:45685",
+        "issue:45834"
+      ],
+      "soft_pairs": [
+        "issue:43322|issue:45685",
+        "issue:44062|issue:44488",
+        "issue:44186|issue:44898",
+        "issue:44843|issue:45081",
+        "issue:43531|issue:43643",
+        "issue:43653|issue:43819",
+        "issue:43577|issue:43818",
+        "issue:43643|issue:44877",
+        "issue:43540|issue:45381",
+        "issue:43421|issue:44811",
+        "issue:43782|issue:44977",
+        "issue:44448|issue:44671",
+        "issue:45584|issue:45834",
+        "issue:44561|issue:45020",
+        "issue:43825|issue:43994",
+        "issue:39692|issue:45834",
+        "issue:44821|issue:45084",
+        "issue:30064|issue:36331",
+        "issue:44485|issue:44509",
+        "issue:41720|issue:44991",
+        "issue:43644|issue:43749",
+        "issue:44186|issue:44442",
+        "issue:44479|issue:45834",
+        "issue:43577|issue:43582",
+        "issue:43606|issue:45072",
+        "issue:43825|issue:43827",
+        "issue:44805|issue:44936"
+      ],
+      "prepared_review_unit_hash": "1f206e81b172f9729dae36143839b80a8b3f230c8cdf697cd3dfe1b7941a3a16",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 35838,
+        "estimated_input_tokens": 8960,
+        "estimated_eval_tokens": 18176
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are mostly unrelated regression reports; I don\u2019t see a true duplicate cluster here. The only broad umbrella issue is the remote_code regression report, but it is still too general to absorb the other topics.",
+        "confidence": 0.87,
+        "canonical_issue_reason": "Issue 45020 is the broadest umbrella in the set (remote_code regressions) and is the closest thing to a central tracking issue, though the surrounding items are not true duplicates of it.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45020 is the best single issue to serve as a general tracker because it is broad, open, and aligned with multiple remote_code-related reports; the other issues are narrower and largely unrelated.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43322",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Different models and failures: Llava Next loading segfault vs MPS histogram dtype error in MoE."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Tokenizer AddedToken keyword error is unrelated to a model load failure for cjvt/sleng-bert."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Different subsystems: LayoutLMv2 tokenizer batching crash vs Perceiver image classification resize/interpolation failure."
+          },
+          {
+            "left": "issue:44843",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both involve _patch_mistral_regex, but the concrete failures differ: offline model_info call vs missing backend_tokenizer attribute."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window bug is unrelated to AutoConfig/trust_remote_code missing fields."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token decoding bug is unrelated to DAC latent/forward mismatch."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43818",
+            "accept": false,
+            "reason": "BLIP2 dtype propagation and Video-LLaVA temporal-attention/weight-sharing are separate model bugs."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Both touch config loading, but one is missing fields with trust_remote_code while the other is strict config rejection for granite_speech."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Different video-input regressions on different Qwen variants with different symptoms."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Runtime special-token/post-processor update bug is distinct from Whisper batch_decode skip_special_tokens handling."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Qwen3VL weight_only load error is unrelated to Qwen3.5 flash-attention generation behavior."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Pegasus v4/v5 output drift is unrelated to CamemBERT masked-LM prediction regression."
+          },
+          {
+            "left": "issue:45584",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Whisper empty-transcription generation after align_special_tokens is unrelated to Kosmos2.5 long-OCR index errors."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Related theme, but 44561 is a specific is_torch_fx_available removal breakage while 45020 is a broad remote_code regression report."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "One is a runtime error-message issue; the other is docs still referencing removed pipeline tasks."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "SigLIP2 nonsensical outputs with AutoModel/pipeline are not the same bug as the pipeline error-message wording."
+          },
+          {
+            "left": "issue:44821",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "AutoImageProcessor URL loading and template-node compilation errors are unrelated."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Image processor segmentation-map handling and CustomTrainer.compute_loss signature mismatch are unrelated."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "GLM-5 RoPE discussion is unrelated to v5 docs mentioning removed pipeline tasks."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Qwen3 auto-device-map CUDA assert is unrelated to tokenizer loading for EMBEDDIA/est-roberta."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Non-persistent buffer junk initialization and FSDP CPU RAM efficient loading are different issues."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "LayoutLMv2 NER/padding crash is unrelated to FastSpeech2ConformerTokenizer AutoTokenizer loading."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Qwen video-input regression and Kosmos2.5 OCR index error are different code paths."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "BLIP2 dtype propagation and Apple Silicon caching_allocator_warmup TypeError are unrelated."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "CPU offload device mismatch for bark-small is not the same as dtype mismatches in SwitchTransformers/TimmWrapperModel."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "Same broad v5 pipeline-removal theme, but one is message text and the other is documentation references."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Mask/index shape mismatch at inference is unrelated to trainer.evaluate() failing after trainer.train()."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 16,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:34634",
+        "issue:36010",
+        "issue:36331",
+        "issue:38175",
+        "issue:42175",
+        "issue:43262",
+        "issue:43329",
+        "issue:43452",
+        "issue:43526",
+        "issue:43582",
+        "issue:43638",
+        "issue:43650",
+        "issue:43653",
+        "issue:43697",
+        "issue:43720",
+        "issue:43742",
+        "issue:43746",
+        "issue:43749",
+        "issue:43782",
+        "issue:43819",
+        "issue:43854",
+        "issue:43906",
+        "issue:43927",
+        "issue:44060",
+        "issue:44291",
+        "issue:44368",
+        "issue:44554",
+        "issue:44589",
+        "issue:44610",
+        "issue:44671",
+        "issue:44811",
+        "issue:44857",
+        "issue:44912",
+        "issue:44936",
+        "issue:44977",
+        "issue:44991",
+        "issue:45020",
+        "issue:45042",
+        "issue:45072",
+        "issue:45081",
+        "issue:45084",
+        "issue:45325",
+        "issue:45406",
+        "issue:45459",
+        "issue:45538",
+        "issue:45600",
+        "issue:45701",
+        "issue:45803"
+      ],
+      "soft_pairs": [
+        "issue:44671|issue:45701",
+        "issue:34634|issue:36010",
+        "issue:43782|issue:45325",
+        "issue:44610|issue:45072",
+        "issue:43720|issue:45072",
+        "issue:42175|issue:45081",
+        "issue:45081|issue:45406",
+        "issue:43749|issue:44291",
+        "issue:43650|issue:43906",
+        "issue:43582|issue:44857",
+        "issue:44060|issue:44368",
+        "issue:43262|issue:43819",
+        "issue:43526|issue:43697",
+        "issue:43526|issue:44554",
+        "issue:45459|issue:45803",
+        "issue:34634|issue:36331",
+        "issue:45084|issue:45538",
+        "issue:43742|issue:45020",
+        "issue:43329|issue:45600",
+        "issue:43653|issue:44912",
+        "issue:43638|issue:44936",
+        "issue:43452|issue:44991",
+        "issue:43746|issue:44811",
+        "issue:38175|issue:44610",
+        "issue:44589|issue:45042",
+        "issue:43854|issue:44977",
+        "issue:43819|issue:43927"
+      ],
+      "prepared_review_unit_hash": "8566e4c2a353cb77ff130ce2319a993bc7f95384979968ba27582323f20080b1",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 36108,
+        "estimated_input_tokens": 9027,
+        "estimated_eval_tokens": 18310
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a set of unrelated Transformers issues that only weakly overlap by theme (loading, tokenization, processors, trainers). There are no strong duplicate issue/PR matches; the best representative is a broad open regression report.",
+        "confidence": 0.24,
+        "canonical_issue_reason": "issue:45020 is the broadest and most central active regression report in the set, covering breakages in recent Transformers versions with remote_code models; it is the best umbrella-style representative even though the rest are not true duplicates.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45020 is the clearest high-signal issue: it is open, well-scoped to a version regression, and broad enough to represent the cluster\u2019s general upstream-breakage theme better than the narrower model-specific reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44671",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Both mention tokenization/version effects, but one is a model-specific masked-LM regression and the other is a broad tokenization-change report; not the same concrete bug."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Different failure modes and code paths: BarkProcessor voice_preset vs GenerationMixin import error."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Both are Qwen VL-related, but one is a from_pretrained weight_only loading error and the other is a rope-index positioning bug; distinct issues."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Processor input-size mismatch and bfloat16 dtype mismatch are unrelated bug classes."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Packed-weight loading in BitNet is a different concrete problem from dtype mismatches in inference."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Packaging/backends issue vs tokenizer regex patch crash; no shared underlying bug."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both are tokenizer/processor load-time crashes, but they involve different models and different missing attributes/code paths."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Both concern loading, but one is FSDP CPU RAM-efficient loading and the other is init_empty_weights argument compatibility; not the same fix."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "Unrelated placeholder/data request vs isolated reproduction of another issue."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Apple Silicon allocator warmup TypeError and CUDA AMP image-loss crash are unrelated."
+          },
+          {
+            "left": "issue:44060",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Both mention tie-weight warnings, but they are model-specific warning reports with different symptoms and likely different causes."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Audio processor sampling-rate defaulting and DAC from_latents STE mismatch are different subsystems and bugs."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "Both involve vision processors/models, but one is a label-reduction bug and the other is an RTDetrV2 output change; not the same defect."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "No meaningful overlap beyond general multimodal/model processing."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Both are error-handling smell issues, but they affect different functions and different failure scenarios."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "BarkProcessor voice preset and CustomTrainer.compute_loss signature mismatch are unrelated."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Template compilation error and CLIPTokenizer max_length issue are unrelated."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Loading a specific model key error is not the same concrete issue as the broader remote_code regression report."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Undefined multimodal-token helper bug and auto_mappings CI breakage are different code paths."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Tokenizer special-token registration bug vs MXFP4 quantization loading/fallback issue; unrelated."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Training index error under ZeRO-3 and post-train evaluate failure are different trainer bugs."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "gguf_file support breakage and tokenizer loading failure are both load-related but not the same underlying bug."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "PEFT local checkpoint loading and Whisper batch_decode skip_special_tokens are unrelated."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Both are vision-model regressions, but one is zero-probabilities output and the other is processor/model input-size mismatch."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Float8 storage lookup failure and PIL backend torchvision dependency regression are unrelated."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Different model-specific generation/loading failures with no shared concrete code path."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "DAC forward-vs-from_latents mismatch and DiaConfig token-ID persistence 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": 1,
+      "component_count": 17,
+      "review_unit_index": 17,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:34567",
+        "issue:36010",
+        "issue:36331",
+        "issue:39692",
+        "issue:40990",
+        "issue:41720",
+        "issue:42490",
+        "issue:42915",
+        "issue:43257",
+        "issue:43262",
+        "issue:43329",
+        "issue:43450",
+        "issue:43452",
+        "issue:43454",
+        "issue:43531",
+        "issue:43540",
+        "issue:43550",
+        "issue:43577",
+        "issue:43582",
+        "issue:43606",
+        "issue:43643",
+        "issue:43645",
+        "issue:43696",
+        "issue:43697",
+        "issue:43720",
+        "issue:43742",
+        "issue:43749",
+        "issue:43756",
+        "issue:43931",
+        "issue:44291",
+        "issue:44479",
+        "issue:44568",
+        "issue:44610",
+        "issue:44756",
+        "issue:44811",
+        "issue:44849",
+        "issue:44912",
+        "issue:44987",
+        "issue:45042",
+        "issue:45356",
+        "issue:45357",
+        "issue:45381",
+        "issue:45488",
+        "issue:45538",
+        "issue:45701",
+        "issue:45759",
+        "issue:45859",
+        "issue:45907"
+      ],
+      "soft_pairs": [
+        "issue:39692|issue:44912",
+        "issue:43257|issue:43329",
+        "issue:36010|issue:43645",
+        "issue:43720|issue:44610",
+        "issue:43550|issue:44912",
+        "issue:43454|issue:43577",
+        "issue:43257|issue:43696",
+        "issue:43329|issue:44479",
+        "issue:43262|issue:44811",
+        "issue:43531|issue:45042",
+        "issue:42490|issue:43257",
+        "issue:36010|issue:36331",
+        "issue:44568|issue:45356",
+        "issue:43749|issue:44756",
+        "issue:43643|issue:45759",
+        "issue:44568|issue:45488",
+        "issue:34567|issue:40990",
+        "issue:43329|issue:43450",
+        "issue:44849|issue:45701",
+        "issue:43257|issue:43931",
+        "issue:43606|issue:45859",
+        "issue:43697|issue:45357",
+        "issue:43257|issue:43756",
+        "issue:44987|issue:45538",
+        "issue:43452|issue:43582",
+        "issue:44479|issue:45381",
+        "issue:44291|issue:45042",
+        "issue:41720|issue:43540",
+        "issue:42915|issue:43742",
+        "issue:44987|issue:45907"
+      ],
+      "prepared_review_unit_hash": "295340dcde1e22be72f849b996889019366f0c78c3039bc3e28f90fa2662f7e2",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36927,
+        "estimated_input_tokens": 9232,
+        "estimated_eval_tokens": 18720
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are mostly unrelated across different bugs and subsystems. The only plausible duplicate pair is the Qwen video-input regression around 44479/45381; everything else is too far apart to merge safely.",
+        "confidence": 0.78,
+        "canonical_issue_reason": "issue:44479 is the broadest umbrella for the Qwen video-input regression family and best represents the only near-duplicate cluster in this set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44479 is the most representative issue overall because it is broader than the other nearby Qwen video reports and can subsume the narrow duplicate-like follow-up.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:39692",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Doc example errors in SigLIP2 are unrelated to the GPT-OSS quantization/load fallback bug."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Qwen3 MoE weight conversion during loading is a different code path from the undefined video token-count helper bug."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "GenerationMixin import breakage is unrelated to custom model initialization in notebooks."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "BitNet packed-weight loading and OmDet-Turbo image-size mismatch are different failures."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "torch.compile SDPA on Bamba and MXFP4 quantization fallback are unrelated bugs."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "AyaVision weight-tying and Blip2 dtype propagation are separate model-loading issues."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43696",
+            "accept": false,
+            "reason": "Qwen3 MoE accelerate/deepspeed loading is unrelated to GPT-OSS out-of-memory."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Both are video-related, but one is an undefined-helper bug in token counting while the other is a broader Qwen video-input regression; not clearly the same defect."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate defaulting is unrelated to Whisper batch_decode skip_special_tokens handling."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior and PIL backend torchvision dependency are unrelated."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43257",
+            "accept": false,
+            "reason": "Path-trailing-slash import warning is unrelated to Qwen3 MoE weight conversion during accelerate+deepspeed loading."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "GenerationMixin import error and CustomTrainer.compute_loss signature mismatch are different trainer/import regressions."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Tokenizer special-token insertion bug and Kimi-K2.5 codec/regex regression are separate tokenizer issues."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "FSDP CPU RAM efficient loading and Strix Halo mmap OOM are different memory-path problems."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "AutoConfig missing fields with trust_remote_code is unrelated to AutoModelForCausalLM.from_config text_config unwrapping."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45488",
+            "accept": false,
+            "reason": "mDeBERTa BOS/EOS insertion and LlamaTokenizer pre-tokenizer override are different tokenizer regressions."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:40990",
+            "accept": false,
+            "reason": "Trainer token accounting and GPT-OSS perplexity are unrelated."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Undefined video token helper variables and batched video processor shape errors are not the same bug."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Qwen3.5 hidden-state output behavior and version-dependent tokenization changes are unrelated."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Qwen3 MoE loading with accelerate/deepspeed is not the same as Qwen3-VL weight-shape mismatch."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Both mention CPU offload, but the affected models and code paths differ too much to treat as duplicates."
+          },
+          {
+            "left": "issue:43697",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "RTDetrV2 output drift and Qwen3.5 save_pretrained visual-key regression are unrelated."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43756",
+            "accept": false,
+            "reason": "Qwen3 MoE loading and SmolLM3 RoPE-layer dropping are separate model/config bugs."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Physical-Intelligence loading failure and CLIPTokenizer model_max_length issue are unrelated."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "gguf_file tokenizer/model loading breakage is unrelated to Apple Silicon warmup TypeError."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45381",
+            "accept": true,
+            "reason": "Both describe the same Qwen video-input regression; 45381 is a narrower report of the broader v5.3.0 video-input bug in 44479."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "init_empty_weights TypeError and PIL backend torchvision requirement are unrelated."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Qwen3 auto device-mapping CUDA assert and Qwen3OmniMoe video ValueError are different failures."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Qwen3Moe FineGrainedFP8Config and MobileLLM key errors are unrelated loading bugs."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "physical-intelligence loading and list_repo_files revision forwarding are unrelated CI/library bugs."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The assessment is grounded in the packet and stays conservative. All soft-edge verdicts match the packet\u2019s deterministic outcomes, and the only accepted pair (44479/45381) is well-supported as the same Qwen video-input regression family."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 18,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:28282",
+        "issue:29127",
+        "issue:33357",
+        "issue:34567",
+        "issue:34689",
+        "issue:36331",
+        "issue:38175",
+        "issue:42175",
+        "issue:42915",
+        "issue:43317",
+        "issue:43493",
+        "issue:43526",
+        "issue:43540",
+        "issue:43577",
+        "issue:43582",
+        "issue:43643",
+        "issue:43742",
+        "issue:43749",
+        "issue:43756",
+        "issue:43782",
+        "issue:43824",
+        "issue:43906",
+        "issue:43927",
+        "issue:43994",
+        "issue:44062",
+        "issue:44291",
+        "issue:44479",
+        "issue:44488",
+        "issue:44560",
+        "issue:44737",
+        "issue:44821",
+        "issue:44898",
+        "issue:44912",
+        "issue:44938",
+        "issue:44991",
+        "issue:45072",
+        "issue:45081",
+        "issue:45137",
+        "issue:45325",
+        "issue:45357",
+        "issue:45375",
+        "issue:45381",
+        "issue:45405",
+        "issue:45479",
+        "issue:45538",
+        "issue:45588",
+        "issue:45823",
+        "issue:45920"
+      ],
+      "soft_pairs": [
+        "issue:33357|issue:34567",
+        "issue:42915|issue:44821",
+        "issue:42915|issue:45538",
+        "issue:43927|issue:44912",
+        "issue:43317|issue:45538",
+        "issue:42175|issue:43824",
+        "issue:43540|issue:44479",
+        "issue:43540|issue:44560",
+        "issue:43577|issue:45072",
+        "issue:44062|issue:44991",
+        "issue:42175|issue:44938",
+        "issue:38175|issue:44898",
+        "issue:42175|issue:44488",
+        "issue:45137|issue:45405",
+        "issue:43906|issue:45381",
+        "issue:44737|issue:45588",
+        "issue:28282|issue:34689",
+        "issue:43742|issue:44291",
+        "issue:42915|issue:43749",
+        "issue:43582|issue:43782",
+        "issue:43526|issue:45479",
+        "issue:43540|issue:45823",
+        "issue:43756|issue:45325",
+        "issue:28282|issue:29127",
+        "issue:43493|issue:43994",
+        "issue:34567|issue:36331",
+        "issue:44062|issue:45081",
+        "issue:44291|issue:44821",
+        "issue:45357|issue:45920",
+        "issue:43643|issue:45375"
+      ],
+      "prepared_review_unit_hash": "34fd7776a457f7274538a63f6251d11b7126695d0cbe7f60fd4aec5455b2dadb",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 37017,
+        "estimated_input_tokens": 9255,
+        "estimated_eval_tokens": 18766
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous. I see one clear duplicate pair around SigLIP2 behavior and one plausible tokenizer-loading duplicate pair; the rest look like distinct bugs in different models, error paths, or subsystems.",
+        "confidence": 0.69,
+        "canonical_issue_reason": "issue:43994 is the best anchor for the SigLIP2 correctness cluster: it is the clearest user-facing report, and it matches one of the strongest duplicate pairs in this set (with issue:43493).",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43994 is the strongest overall issue anchor because it describes the observable failure directly and is a good representative target for the SigLIP2 duplicate pair.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:33357",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "MacOS CLIP model bus error vs TrainerState token-count bookkeeping: different bugs."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Qwen3Moe FP8/config loading issue vs AutoImageProcessor URL loading: unrelated."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "FineGrainedFP8Config failure and CLIPTokenizer max-length behavior are different problems."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "DiaConfig save/load token-ID bug vs MXFP4 quantization fallback: different code paths."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "device_map/offload dequantization failure is not the same as CLIPTokenizer max_length behavior."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Packaging/backend availability issue vs model import/export name error: not the same bug."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Both involve video, but the model families and failures differ."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Qwen3OmniMoe ValueError and Qwen3-vl-embedding StopIteration look like different failures."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Both mention dtype mismatches, but they affect different model families and code paths."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44991",
+            "accept": true,
+            "reason": "Likely the same tokenizer-loading regression: 44062 shows the concrete AddedToken TypeError, while 44991 reports the broader tokenizer load failure on the same upgrade path."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "TensorFlow backend packaging issue vs Python 3.14 load failure: different causes."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "SigLIP2 zero-probabilities bug vs Perceiver resolution/interpolation failure: different model issues."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Backend installation issue vs loading a specific model repo: unrelated."
+          },
+          {
+            "left": "issue:45137",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO3 deque error is unrelated to the PEFT version bump."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Isolated reproduction of another issue vs Qwen2.5-VL video position-id regression."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "XLNet CPU placement bug and flash-attention sink handling are unrelated."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Missing PyTorch import error vs Llama 3.2 Vision loading breakage: different failures."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "MobileLLM key error vs init_empty_weights unexpected-argument error: different loading paths."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config issue vs FSDP_CPU_RAM_EFFICIENT_LOADING: unrelated."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:43782",
+            "accept": false,
+            "reason": "Apple Silicon allocator TypeError vs Qwen3VL weight_only loading error: different problems."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction bug vs sequence-classification zero-loss behavior: unrelated."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video-input ValueError and Gemma4 PLE device mismatch are different bugs."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Both involve positional/RoPE logic, but they affect different models and computations."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "Missing PyTorch import vs LayoutLMv3 error-message clarity: unrelated."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43994",
+            "accept": true,
+            "reason": "Same SigLIP2 correctness issue: 43493 describes the implementation discrepancy, and 43994 reports the user-visible bad outputs on the same model."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "TrainerState token counting and CustomTrainer compute_loss signature mismatch are unrelated training issues."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "AddedToken constructor duplication and Mistral regex patch crash are different tokenizer bugs."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "init_empty_weights loading failure is unrelated to loading AutoImageProcessor from a URL."
+          },
+          {
+            "left": "issue:45357",
+            "right": "issue:45920",
+            "accept": false,
+            "reason": "Qwen3.5 save_pretrained key regression and AutoTokenizer wrong token IDs affect different components."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "General trust_remote_code AutoConfig field loss is not the same as the Qwen3_5MoeVisionConfig strict-field drop bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The cluster summary is mostly grounded, but it overstates confidence in at least one place. In particular, the accepted 44062/44991 pair is described as the \u201csame tokenizer-loading regression\u201d and \u201cbroader manifestation,\u201d which is stronger than the packet\u2019s title-level evidence supports. Also, the canonical issue note saying 43994 matches the \u201conly strong duplicate pair in this set\u201d is inaccurate because the packet contains another accepted duplicate pair (44062/44991). Please soften those claims to plausible/likely duplicates rather than asserting the same underlying bug from titles alone."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 19,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:24643",
+        "issue:29127",
+        "issue:38175",
+        "issue:41762",
+        "issue:43295",
+        "issue:43452",
+        "issue:43531",
+        "issue:43577",
+        "issue:43643",
+        "issue:43653",
+        "issue:43688",
+        "issue:43697",
+        "issue:43720",
+        "issue:43742",
+        "issue:43749",
+        "issue:43819",
+        "issue:44361",
+        "issue:44442",
+        "issue:44464",
+        "issue:44466",
+        "issue:44484",
+        "issue:44488",
+        "issue:44514",
+        "issue:44561",
+        "issue:44568",
+        "issue:44589",
+        "issue:44625",
+        "issue:44704",
+        "issue:44743",
+        "issue:44792",
+        "issue:44811",
+        "issue:44821",
+        "issue:44843",
+        "issue:44877",
+        "issue:44898",
+        "issue:44936",
+        "issue:44964",
+        "issue:45005",
+        "issue:45042",
+        "issue:45081",
+        "issue:45216",
+        "issue:45290",
+        "issue:45356",
+        "issue:45381",
+        "issue:45405",
+        "issue:45600",
+        "issue:45823"
+      ],
+      "soft_pairs": [
+        "issue:43452|issue:44488",
+        "issue:44792|issue:44964",
+        "issue:45042|issue:45081",
+        "issue:43742|issue:44589",
+        "issue:43531|issue:44821",
+        "issue:45405|issue:45600",
+        "issue:43697|issue:45216",
+        "issue:43688|issue:43749",
+        "issue:44514|issue:45381",
+        "issue:43720|issue:44898",
+        "issue:44484|issue:44704",
+        "issue:41762|issue:44936",
+        "issue:24643|issue:29127",
+        "issue:44625|issue:44877",
+        "issue:43452|issue:43577",
+        "issue:44514|issue:45290",
+        "issue:43452|issue:44843",
+        "issue:44466|issue:45005",
+        "issue:43295|issue:45356",
+        "issue:43819|issue:44811",
+        "issue:44936|issue:45823",
+        "issue:43653|issue:44568",
+        "issue:44464|issue:44743",
+        "issue:38175|issue:43295",
+        "issue:43643|issue:44561",
+        "issue:43653|issue:44361",
+        "issue:44361|issue:44442"
+      ],
+      "prepared_review_unit_hash": "52a6634dfff0056aef956b088f035c923f7536cfc9c0f6553aa1adde7675252d",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 27,
+        "serialized_chars": 35465,
+        "estimated_input_tokens": 8867,
+        "estimated_eval_tokens": 17990
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is not a duplicate set; it\u2019s a mix of unrelated loading, tokenizer, model-shape, generation, and trainer regressions. The closest thing to a representative issue is the broad loading regression in issue 43452, but the rest are mostly model- or path-specific.",
+        "confidence": 0.9,
+        "canonical_issue_reason": "Issue 43452 is the broadest shared-loader regression in the set and is the best heuristic representative for the cluster, even though the other issues are not true duplicates.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43452 best captures the cluster\u2019s common theme of upstream loading regressions without being overly narrow to one model family.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43452",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both are loading-related, but 43452 is a gguf_file regression in tokenizer/model loading while 44488 is a specific sleng-bert load failure; different underlying bugs."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different model families and different failures: janus test instability vs Phi-4 multimodal load failure."
+          },
+          {
+            "left": "issue:45042",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Image-processor torchvision dependency bug vs Mistral regex tokenizer crash; unrelated code paths."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "MobileLLM loading key error and Float8 storage lookup error are distinct loader/runtime issues."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior and AutoImageProcessor URL loading are unrelated."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "PEFT version bump packaging issue vs removed config references in auto_mappings.py; different problems."
+          },
+          {
+            "left": "issue:43697",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "RTDetrV2 output mismatch in v5 is not the same bug as Qwen3.5 save_pretrained checkpoint corruption."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Auxiliary loss normalization and FSDP CPU RAM efficient loading are separate training/loading bugs."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both involve Qwen2.5-VL, but one crashes in chat templating with batching/padding and the other has wrong video vision_position_ids; not the same defect."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "BitNet packed-weight loading during accelerate vs Perceiver interpolation on non-default resolution are unrelated."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "Save-pretrained shard-size question and AutoProcessor kwarg forwarding are unrelated issues."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Gemma3 ZeRO-3 load failure and trainer.evaluate() after trainer.train() are different code paths and failures."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "DeepSpeed weight dimensionality error and LayoutLMv3 box validation messaging are not the same underlying bug."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Qwen3.5 num_labels propagation and strict granite_speech config loading are distinct config-handling regressions."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "gguf_file loading regression and Blip2 dtype propagation are unrelated."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both are chat-template issues, but one is batched padding behavior and the other is tool-call assistant messages with no content; different fixes."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44843",
+            "accept": false,
+            "reason": "gguf_file loading breakage and mistral-regex offline model_info behavior are unrelated loading paths."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Both mention tied weights, but 44466 is device-dependent lm_head serialization while 45005 is a broader translation-model tied-weights issue; not enough evidence they\u2019re the same bug."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Custom processor.tokenizer regression and Kimi-K2.5 codec/fix_mistral_regex regression are different tokenizer issues."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "DAC.from_latents mismatch and Whisper batch_decode skip_special_tokens handling are unrelated."
+          },
+          {
+            "left": "issue:44936",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "trainer.evaluate() state issue and Gemma4 device_map='auto' mismatch are different failures."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration and mdeberta-v3 BOS/EOS insertion are similar category-wise but clearly different tokenizers and bugs."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Chunked generation with compiled forward and recurrent state reset with cache are distinct generation-path regressions."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Siglip2 zero-probabilities and processor.tokenizer regression are unrelated."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "AutoConfig missing fields under trust_remote_code and removal of is_torch_fx_available are separate trust_remote_code compatibility issues."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token behavior and MLukeTokenizer AttributeError are unrelated tokenizer bugs."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "Both are tokenizer loading failures, but for different tokenizer classes and different failure modes; not a duplicate."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 20,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:28282",
+        "issue:30064",
+        "issue:32090",
+        "issue:33357",
+        "issue:33666",
+        "issue:34567",
+        "issue:35141",
+        "issue:36010",
+        "issue:41720",
+        "issue:43122",
+        "issue:43278",
+        "issue:43452",
+        "issue:43493",
+        "issue:43504",
+        "issue:43525",
+        "issue:43531",
+        "issue:43550",
+        "issue:43572",
+        "issue:43577",
+        "issue:43606",
+        "issue:43646",
+        "issue:43653",
+        "issue:43688",
+        "issue:43697",
+        "issue:43720",
+        "issue:43742",
+        "issue:43749",
+        "issue:43825",
+        "issue:43906",
+        "issue:43927",
+        "issue:44361",
+        "issue:44451",
+        "issue:44464",
+        "issue:44479",
+        "issue:44554",
+        "issue:44610",
+        "issue:44625",
+        "issue:44821",
+        "issue:44843",
+        "issue:44849",
+        "issue:44977",
+        "issue:44991",
+        "issue:45357",
+        "issue:45440",
+        "issue:45588",
+        "issue:45701",
+        "issue:45704"
+      ],
+      "soft_pairs": [
+        "issue:30064|issue:33666",
+        "issue:41720|issue:43906",
+        "issue:43688|issue:45440",
+        "issue:43688|issue:45704",
+        "issue:43577|issue:43697",
+        "issue:28282|issue:32090",
+        "issue:28282|issue:33357",
+        "issue:44843|issue:44991",
+        "issue:43646|issue:43749",
+        "issue:28282|issue:35141",
+        "issue:43688|issue:45701",
+        "issue:43653|issue:43927",
+        "issue:43531|issue:43742",
+        "issue:43550|issue:43927",
+        "issue:44464|issue:44977",
+        "issue:44479|issue:44625",
+        "issue:43577|issue:44554",
+        "issue:43550|issue:44361",
+        "issue:43577|issue:45588",
+        "issue:43122|issue:43906",
+        "issue:43504|issue:43550",
+        "issue:43504|issue:43720",
+        "issue:41720|issue:44464",
+        "issue:32090|issue:34567",
+        "issue:43504|issue:44610",
+        "issue:32090|issue:33666",
+        "issue:43452|issue:44451",
+        "issue:44849|issue:45357",
+        "issue:43742|issue:44821",
+        "issue:43278|issue:43688",
+        "issue:33666|issue:36010",
+        "issue:43493|issue:43825",
+        "issue:43606|issue:43927",
+        "issue:43525|issue:43572"
+      ],
+      "prepared_review_unit_hash": "c3e37aa0197d1ac495ee3f5ef25db9e97a79f75e047dd7d32d5a13f1af3958e9",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 34,
+        "serialized_chars": 37206,
+        "estimated_input_tokens": 9302,
+        "estimated_eval_tokens": 18860
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is not a duplicate set: the proposed soft edges span unrelated issues across loading, tokenization, training, generation, and model-specific regressions. No candidate pair looks like the same underlying bug/change.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "No canonical issue: the items are heterogeneous and do not form a single duplicate cluster.",
+        "canonical_pr_reason": "No canonical PR: there are no pull requests in the input.",
+        "best_issue_reason": "No global best issue: the cluster is too mixed to nominate one representative duplicate target.",
+        "best_pr_reason": "No PRs to evaluate for global best PR suitability.",
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:30064",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "Image processor segmentation-map bug vs Qwen2-VL multi-GPU training; unrelated subsystems and failure modes."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "Qwen3 auto device-mapping CUDA assert vs isolated reproduction of another issue; not the same bug."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Aux-loss normalization bug vs DeepSeekV3 implementation divergence; different code paths."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Aux-loss normalization vs T5 Apex RMSNorm memory leak; unrelated model/runtime issues."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "BLIP2 dtype propagation issue vs RTDetrV2 output mismatch; different models and symptoms."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Missing PyTorch ImportError vs Trainer NoneType broadcast error; unrelated causes."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Missing PyTorch library vs Mac bus error with CLIP; not the same problem."
+          },
+          {
+            "left": "issue:44843",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Offline-model-info patch bug vs tokenizer load failure for a specific model; different failures."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Custom model init breakage in 5.0 vs FSDP RAM-efficient loading; separate features."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "ImportError for absent PyTorch vs embedding reinit during resize; unrelated regressions."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Aux-loss normalization vs tokenization differences across versions; different layers of the stack."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token decode bug vs DiaConfig token-ID loss on save/load; different models and behavior."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window issue vs MobileLLM key error; unrelated model-specific bugs."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Bamba torch.compile/SDPA bug vs DiaConfig token-ID persistence bug; not the same code path."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Compiled-forward chunked generation mismatch vs flash-attention generation bug; different generation paths."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44625",
+            "accept": false,
+            "reason": "Video-input regression vs num_labels propagation issue; different model components."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "BLIP2 dtype mismatch vs MPS attention correctness issue; unrelated bugs."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Bamba compile/SDPA failure vs MLukeTokenizer task AttributeError; different subsystems."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "BLIP2 dtype issue vs flash_attention.py s_aux=None crash; unrelated failures."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "Tokenization changes across versions vs isolated reproduction of another issue; not a duplicate."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "BEiT legacy-field load bug vs Bamba torch.compile SDPA bug; different models and code paths."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "BEiT legacy-field load bug vs BitNet packed-weight unpacking bug; unrelated loading problems."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44464",
+            "accept": false,
+            "reason": "Qwen3 device-mapping CUDA assert vs chunked generation inconsistency; different failure modes."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "NoneType broadcast error in Trainer vs num_input_tokens_seen not updating; unrelated trainer issues."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "BEiT pretrained-load legacy field vs OmDet-Turbo processor size mismatch; different model pipelines."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "Trainer broadcast TypeError vs multi-GPU training request; not the same bug."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "gguf_file breaks AutoTokenizer/AutoModel loading vs specific model load failure; different root causes."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Qwen3.5 hidden-states bug vs incorrect visual-encoder keys on save_pretrained; separate regressions."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "MobileLLM key error vs AutoImageProcessor URL-loading failure; unrelated loading paths."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43688",
+            "accept": false,
+            "reason": "Embedding dtype drift in evaluate vs auxiliary-loss normalization; different issues."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Qwen2-VL multi-GPU training vs GenerationMixin import error; unrelated topics."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "SigLIP2 implementation discrepancy vs pipeline error-message wording; not the same bug."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "CPU offload device mismatch vs DiaConfig token-ID persistence; different models and code paths."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Llama4Config missing pad_token_id vs StableLmConfig missing pad_token_idx; similar theme, but model-specific config regressions and not one concrete bug."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 21,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:28218",
+        "issue:33357",
+        "issue:33666",
+        "issue:36331",
+        "issue:39692",
+        "issue:43257",
+        "issue:43262",
+        "issue:43278",
+        "issue:43295",
+        "issue:43299",
+        "issue:43334",
+        "issue:43493",
+        "issue:43504",
+        "issue:43540",
+        "issue:43572",
+        "issue:43606",
+        "issue:43646",
+        "issue:43653",
+        "issue:43720",
+        "issue:43756",
+        "issue:43824",
+        "issue:43844",
+        "issue:43927",
+        "issue:43994",
+        "issue:44291",
+        "issue:44361",
+        "issue:44423",
+        "issue:44464",
+        "issue:44466",
+        "issue:44488",
+        "issue:44493",
+        "issue:44568",
+        "issue:44589",
+        "issue:44625",
+        "issue:44734",
+        "issue:44843",
+        "issue:44913",
+        "issue:44964",
+        "issue:44977",
+        "issue:45083",
+        "issue:45200",
+        "issue:45229",
+        "issue:45405",
+        "issue:45520",
+        "issue:45584",
+        "issue:45759",
+        "issue:45812",
+        "issue:45823"
+      ],
+      "soft_pairs": [
+        "issue:44568|issue:45812",
+        "issue:43720|issue:43927",
+        "issue:43262|issue:43295",
+        "issue:43756|issue:44913",
+        "issue:43720|issue:44361",
+        "issue:28218|issue:33357",
+        "issue:33666|issue:36331",
+        "issue:39692|issue:43493",
+        "issue:43540|issue:44977",
+        "issue:43493|issue:43844",
+        "issue:44964|issue:45405",
+        "issue:43334|issue:43572",
+        "issue:43278|issue:43844",
+        "issue:44625|issue:45584",
+        "issue:43257|issue:43299",
+        "issue:43756|issue:45083",
+        "issue:43334|issue:45759",
+        "issue:44488|issue:44843",
+        "issue:45200|issue:45229",
+        "issue:43824|issue:44589",
+        "issue:43646|issue:44466",
+        "issue:39692|issue:43295",
+        "issue:45229|issue:45823",
+        "issue:43504|issue:43606",
+        "issue:44589|issue:45520",
+        "issue:43572|issue:44291",
+        "issue:33357|issue:33666",
+        "issue:44423|issue:44734",
+        "issue:43927|issue:44361",
+        "issue:43504|issue:43653",
+        "issue:44464|issue:45584",
+        "issue:43994|issue:44493"
+      ],
+      "prepared_review_unit_hash": "cba50e2e792c8ae1e6fe36cf5636b5eef8a04156b1fd6f1717d0b12b862726b9",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37647,
+        "estimated_input_tokens": 9412,
+        "estimated_eval_tokens": 19080
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is a loose collection of unrelated Transformers issues. The soft pairs mostly share subsystem keywords (tokenizers, multimodal configs, model loading, training) but describe different bugs and fixes, so none should be merged as duplicates.",
+        "confidence": 0.66,
+        "canonical_issue_reason": "issue:45812 is the strongest representative because it\u2019s a broad, user-facing tokenizer regression affecting many Granite models and has clear cross-model impact.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45812 is the best overall issue to anchor the cluster since it is specific, high-impact, and most likely to attract the right tokenizer-focused triage.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44568",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but one is about add_special_tokens/BOS-EOS behavior for mdeberta and the other is wrong token IDs for Granite models; different code paths and symptoms."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Packed-weight unpacking during accelerate loading and DiaConfig save/load token ID loss are unrelated bugs."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Audio chat-template sampling rate defaults and processor.tokenizer/image-handling regression are different multimodal issues."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Smollm3 RoPE-layer omission and GPTNeoX rotary_pct reload fallback are separate configuration bugs."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "BitNet packed-weight loading and MLukeTokenizer task AttributeError do not share the same underlying failure."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Tokenizer spacing after added tokens and a MacOS bus error in CLIP are clearly unrelated."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Qwen2-VL multi-GPU training and a CustomTrainer signature mismatch are different training issues."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "SigLIP2 docs/example errors and SigLIP2 HF-vs-JAX implementation discrepancies are not the same bug."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video-input processing and Qwen3.5 flash-attention generation problems are distinct code paths."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "A model implementation discrepancy and a DeepSpeed gradient anomaly are different failures."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Phi-4 multimodal load failure and a bumped unreleased PEFT minimum version are unrelated."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Both involve config fields, but Qwen3-VL pad_token_id and StableLmConfig pad_token_idx are separate model-specific regressions."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "Embedding dtype drift in eval and gradient growth under ZeRO-3 are unrelated symptoms."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Qwen3.5 num_labels propagation and Whisper empty transcription after align_special_tokens are different model families and bugs."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43299",
+            "accept": false,
+            "reason": "Both are Qwen MOE loading issues, but one is about accelerate/deepspeed weight conversion and the other is a dev-version model-loading breakage for multimodal configs."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "RoPE-layer dropping in Smollm3 and qwen3_omni_moe feature-length helper behavior are unrelated."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Pad-token-id missing in Qwen3-VL and failing to unwrap text_config in composite Qwen configs are different multimodal config problems."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44843",
+            "accept": false,
+            "reason": "A model failing to load and an unconditional hub API call breaking offline mode are separate issues."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Text-only mm_token_type_ids defaults and Gemma4 multi-GPU OOM are not the same defect."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Importing Qwen2.5-VL class and missing Float8 storage support are unrelated runtime errors."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Custom model initialization breakage and inconsistent lm_head.weight serialization are different v5 regressions."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Doc/example mistakes and a processor.tokenizer regression are not the same underlying bug."
+          },
+          {
+            "left": "issue:45229",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Gemma4 inference OOM and a device mismatch during forward are distinct multi-GPU/runtime failures."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43606",
+            "accept": false,
+            "reason": "Legacy field loading in Beit semantic segmentation and Bark CPU-offload device mismatch are unrelated."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Float8 storage lookup failure and a Python 3.13 flash_attn import-utils KeyError are different compatibility issues."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "StableLmConfig pad_token_idx regression and init_empty_weights unexpected _is_hf_initialized argument are separate API breakages."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "MacOS bus error in a CLIP community model and Qwen2-VL multi-GPU training are unrelated."
+          },
+          {
+            "left": "issue:44423",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Continuous-batching multimodal crash and KV-cache continuation indexing bug affect different serve paths."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "DiaConfig custom-token-ID loss and MLukeTokenizer task AttributeError are unrelated model/tokenizer bugs."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Beit legacy-field loading and BigBirdTokenizer special-token registration are different fixes."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Compiled-forward chunked generation inconsistency and Whisper empty-transcription failure after align_special_tokens are different generation bugs."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "SigLIP2 bad results with AutoModel/pipeline and unexpected position-id keys across models are not the same issue."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 22,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:15354",
+        "issue:28218",
+        "issue:29942",
+        "issue:30333",
+        "issue:34567",
+        "issue:37428",
+        "issue:38175",
+        "issue:42490",
+        "issue:43232",
+        "issue:43278",
+        "issue:43295",
+        "issue:43493",
+        "issue:43540",
+        "issue:43572",
+        "issue:43575",
+        "issue:43582",
+        "issue:43606",
+        "issue:43696",
+        "issue:43701",
+        "issue:43844",
+        "issue:43873",
+        "issue:43957",
+        "issue:44112",
+        "issue:44315",
+        "issue:44464",
+        "issue:44466",
+        "issue:44521",
+        "issue:44554",
+        "issue:44610",
+        "issue:44792",
+        "issue:44849",
+        "issue:44898",
+        "issue:44964",
+        "issue:44977",
+        "issue:44987",
+        "issue:45072",
+        "issue:45084",
+        "issue:45127",
+        "issue:45198",
+        "issue:45200",
+        "issue:45290",
+        "issue:45356",
+        "issue:45405",
+        "issue:45412",
+        "issue:45446",
+        "issue:45520",
+        "issue:45563",
+        "issue:45823"
+      ],
+      "soft_pairs": [
+        "issue:43606|issue:44112",
+        "issue:43575|issue:45405",
+        "issue:43278|issue:43701",
+        "issue:44792|issue:45405",
+        "issue:44521|issue:45290",
+        "issue:29942|issue:30333",
+        "issue:42490|issue:43696",
+        "issue:43844|issue:44964",
+        "issue:28218|issue:34567",
+        "issue:44554|issue:45446",
+        "issue:45084|issue:45520",
+        "issue:43582|issue:44977",
+        "issue:43540|issue:45200",
+        "issue:15354|issue:38175",
+        "issue:43232|issue:44849",
+        "issue:45072|issue:45198",
+        "issue:37428|issue:43957",
+        "issue:45200|issue:45823",
+        "issue:44315|issue:44849",
+        "issue:44554|issue:45356",
+        "issue:43295|issue:44610",
+        "issue:43493|issue:43701",
+        "issue:43873|issue:44464",
+        "issue:44898|issue:45412",
+        "issue:44466|issue:45127",
+        "issue:44792|issue:45563",
+        "issue:43572|issue:44987"
+      ],
+      "prepared_review_unit_hash": "6176ccff03352acb63903f83a924bd52aba9c06698ddee6dd706792a6a56f7ba",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 35782,
+        "estimated_input_tokens": 8946,
+        "estimated_eval_tokens": 18148
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous: the issues span unrelated tokenizer, generation, loading, device, quantization, and CI problems. All soft pairs look like superficial text similarity rather than the same underlying bug/change, so there is no dedupe-worthy canonical artifact here.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "No single canonical issue: the items cover many different models, subsystems, and failure modes, with only broad keyword overlap.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No issue is a good global representative because there is no shared defect to anchor a canonical duplicate cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43606",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Both mention device-related failures, but one is CPU offload mismatch on a specific model and the other is a stale CI device-override test; different bugs and code paths."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "OOM during tensor-parallel loading is unrelated to a packaging/version bump for MIN_PEFT_VERSION."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "BF16-to-FP32 dtype drift in evaluation is unrelated to a resume_from_checkpoint key mismatch."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "A Janus image-generation test failure has no meaningful relation to the PEFT version requirement issue."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both involve chat template processing, but one is zeroed assistant masks for multimodal inputs while the other is a crash on tool-call assistant messages; not the same bug."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:30333",
+            "accept": false,
+            "reason": "Flash Attention 2 test failures and MLflow reporting after training failure are unrelated."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43696",
+            "accept": false,
+            "reason": "Local-path import warning and CUDA OOM on GPT-oss are different loading/runtime issues."
+          },
+          {
+            "left": "issue:43844",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO-3 gradient growth and Phi-4 multimodal loading failure do not share a concrete code path."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Tokenizer spacing after added tokens is unrelated to TrainerState token-count tracking."
+          },
+          {
+            "left": "issue:44554",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "An MPS attention correctness bug is unrelated to a PyTorch version check in flex_attention imports."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Template compilation errors and a flash_attn import KeyError on Python 3.13 are different failures."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "An Apple Silicon allocator TypeError and a flash-attention generation bug are not the same underlying issue."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Video-input processing in Qwen3OmniMoe and missing mm_token_type_ids in Gemma 4 are different multimodal bugs."
+          },
+          {
+            "left": "issue:15354",
+            "right": "issue:38175",
+            "accept": false,
+            "reason": "TorchScript generator-expression support and zero probabilities in SigLIP2 are unrelated."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44849",
+            "accept": false,
+            "reason": "Generation kwargs update after sync_gpus is unrelated to the Qwen3.5 hidden-states bug."
+          },
+          {
+            "left": "issue:45072",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Bfloat16 dtype mismatches in SwitchTransformers/TimmWrapperModel are unrelated to Wav2Vec2 save_pretrained/tokenization failures."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "A missing flash attention import symbol and meta-device model loading breakage are different regression classes."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Gemma 4 text-only fine-tuning mask defaults and a device_map='auto' forward mismatch are separate issues."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44849",
+            "accept": false,
+            "reason": "Liger Kernel not applying with model_init is unrelated to the Qwen3.5 hidden-states bug."
+          },
+          {
+            "left": "issue:44554",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "MPS attention correctness and Kimi-K2.5 tokenizer codec regression do not share the same bug surface."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Processor/tokenizer regression with images is different from OmDet-Turbo's image-size mismatch."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "SigLIP2 HF-vs-JAX discrepancy is unrelated to resume_from_checkpoint key mismatch."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44464",
+            "accept": false,
+            "reason": "Quantization/offloading behavior and chunked generation with compiled forward are distinct problems."
+          },
+          {
+            "left": "issue:44898",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Perceiver non-default resolution failure is unrelated to RT-DETR memory release issues."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45127",
+            "accept": false,
+            "reason": "Serialization of tied lm_head weights and LoRA merge collapse with extended vocab are not the same concrete defect."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "A Janus image-generation test failure is unrelated to a stale warning in paged generate()."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "A missing StableLmConfig pad_token_idx after a 5.0 update is not the same as loading physical-intelligence/fast failing on transformers>=5.1.0."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 23,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:28218",
+        "issue:35141",
+        "issue:37428",
+        "issue:38175",
+        "issue:42490",
+        "issue:42915",
+        "issue:43262",
+        "issue:43278",
+        "issue:43295",
+        "issue:43334",
+        "issue:43388",
+        "issue:43425",
+        "issue:43450",
+        "issue:43493",
+        "issue:43525",
+        "issue:43526",
+        "issue:43572",
+        "issue:43575",
+        "issue:43606",
+        "issue:43653",
+        "issue:43819",
+        "issue:43824",
+        "issue:43827",
+        "issue:43844",
+        "issue:43856",
+        "issue:43873",
+        "issue:43901",
+        "issue:43957",
+        "issue:43975",
+        "issue:44112",
+        "issue:44206",
+        "issue:44442",
+        "issue:44514",
+        "issue:44521",
+        "issue:44610",
+        "issue:44857",
+        "issue:44936",
+        "issue:45083",
+        "issue:45127",
+        "issue:45198",
+        "issue:45200",
+        "issue:45245",
+        "issue:45290",
+        "issue:45325",
+        "issue:45381",
+        "issue:45412",
+        "issue:45584"
+      ],
+      "soft_pairs": [
+        "issue:37428|issue:43824",
+        "issue:45200|issue:45325",
+        "issue:43572|issue:43957",
+        "issue:43844|issue:44936",
+        "issue:43334|issue:43525",
+        "issue:43295|issue:43450",
+        "issue:43824|issue:43975",
+        "issue:43295|issue:43873",
+        "issue:43388|issue:43526",
+        "issue:38175|issue:45412",
+        "issue:42915|issue:43824",
+        "issue:44514|issue:44521",
+        "issue:38175|issue:43493",
+        "issue:44442|issue:45198",
+        "issue:28218|issue:35141",
+        "issue:43278|issue:43856",
+        "issue:42490|issue:43493",
+        "issue:43653|issue:44112",
+        "issue:43827|issue:43901",
+        "issue:43450|issue:43575",
+        "issue:44206|issue:45584",
+        "issue:45127|issue:45245",
+        "issue:43819|issue:45083",
+        "issue:43819|issue:44112",
+        "issue:43262|issue:43450",
+        "issue:44610|issue:44857",
+        "issue:37428|issue:43425",
+        "issue:45290|issue:45381",
+        "issue:43606|issue:45083"
+      ],
+      "prepared_review_unit_hash": "17576ae4d099c4a9556ce8d1e0e7903c430f386680be48fbdb3bdb9347eec6bd",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 36009,
+        "estimated_input_tokens": 9003,
+        "estimated_eval_tokens": 18262
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The set is highly heterogeneous; most soft pairs are only loosely related by broad subsystem wording (imports, processors, tokenization, chat templates) and do not look like the same underlying bug. No PRs are present. If a single issue must serve as the anchor, issue 44521 is the most representative active user-facing bug in the list.",
+        "confidence": 0.72,
+        "canonical_issue_reason": "Issue 44521 is the most central and active item here, with multiple references and discussion around apply_chat_template/multimodal masking behavior.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44521 is the best global representative because it is open, user-facing, and closest to a recurring processor/chat_template bug family in this set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:37428",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but they fail on different missing symbols in different code paths; not the same bug."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "One is about mm_token_type_ids defaults, the other about rope/position-id scaling; unrelated."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "A config attribute regression vs a broad meta-device loading breakage; different underlying failures."
+          },
+          {
+            "left": "issue:43844",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Different training/evaluation problems with no shared concrete code path."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Both mention missing pad_token_id, but they affect different configs/models and are separate bugs."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Custom processor-tokenizer regression vs batched video processor shape bug; not the same change."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43975",
+            "accept": false,
+            "reason": "Both are import/usage regressions, but for different classes and APIs; not the same issue."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Processor API regression vs quantization/offloading behavior; unrelated."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Both involve labels, but one is distributed metric gathering and the other is image processor label reduction; different code paths."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Both are memory/resource concerns, but one is model output probabilities and the other is deletion/GC memory release."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 failure is unrelated to an import error for a VL model symbol."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Both touch apply_chat_template, but one is a batched-input crash and the other is incorrect assistant_mask generation; not clearly the same bug."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Different SigLIP2 issues with different symptoms and likely different implementation defects."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Different tokenizer/model families with separate tokenization/save_pretrained failures."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Tokenizer spacing after added tokens is unrelated to embedding reinitialization during resize."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "Both mention training efficiency or dtype, but they are distinct issues in different contexts."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Import warning on local paths vs SigLIP2 implementation mismatch; different behavior."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Tokenizer special-token registration bug vs a CI device-override test failure; unrelated."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Both are docs issues, but they target different API removals and different pipeline docs."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Video processor shape bug vs tensor-parallel OOM on model loading; not the same problem."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Both are audio-related, but one is an unsupported argument in a feature extractor and the other is empty-transcription generation behavior."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "LoRA/extended-vocab collapse is unrelated to the categories limit runtime error."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "DAC latent reconstruction mismatch vs a helper-length calculation issue in qwen3_omni_moe; different bugs."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "No shared underlying code path; audio codec mismatch vs stale CI test."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate default vs video batch shape regression; unrelated."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Different detection/image-processing failures with different symptoms and likely fixes."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43425",
+            "accept": false,
+            "reason": "A missing symbol import error is not the same as generic Torch 2.10 incompatibility."
+          },
+          {
+            "left": "issue:45290",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "One is a chat-template/tool-call crash, the other is a Qwen2.5-VL position-id bug; not the same issue."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "CPU offload device mismatch is unrelated to feature-extraction output length behavior."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 24,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:29127",
+        "issue:29942",
+        "issue:32090",
+        "issue:34689",
+        "issue:35141",
+        "issue:37428",
+        "issue:38175",
+        "issue:39692",
+        "issue:40444",
+        "issue:42490",
+        "issue:43278",
+        "issue:43335",
+        "issue:43504",
+        "issue:43526",
+        "issue:43550",
+        "issue:43606",
+        "issue:43611",
+        "issue:43653",
+        "issue:43720",
+        "issue:43819",
+        "issue:43825",
+        "issue:43854",
+        "issue:43931",
+        "issue:44112",
+        "issue:44206",
+        "issue:44292",
+        "issue:44351",
+        "issue:44360",
+        "issue:44448",
+        "issue:44462",
+        "issue:44493",
+        "issue:44561",
+        "issue:44743",
+        "issue:44779",
+        "issue:44855",
+        "issue:44857",
+        "issue:44977",
+        "issue:45072",
+        "issue:45083",
+        "issue:45200",
+        "issue:45229",
+        "issue:45310",
+        "issue:45478",
+        "issue:45520",
+        "issue:45588",
+        "issue:45593",
+        "issue:45823",
+        "issue:45859"
+      ],
+      "soft_pairs": [
+        "issue:43335|issue:43550",
+        "issue:44448|issue:44743",
+        "issue:44857|issue:45593",
+        "issue:45200|issue:45588",
+        "issue:44360|issue:44561",
+        "issue:43278|issue:43825",
+        "issue:37428|issue:44351",
+        "issue:29942|issue:32090",
+        "issue:44206|issue:45823",
+        "issue:43611|issue:43931",
+        "issue:39692|issue:43526",
+        "issue:43335|issue:43653",
+        "issue:43819|issue:44493",
+        "issue:43550|issue:44112",
+        "issue:44855|issue:45083",
+        "issue:43335|issue:43504",
+        "issue:45310|issue:45478",
+        "issue:43335|issue:45083",
+        "issue:29127|issue:32090",
+        "issue:45588|issue:45859",
+        "issue:39692|issue:44493",
+        "issue:44292|issue:45520",
+        "issue:38175|issue:44493",
+        "issue:44448|issue:44779",
+        "issue:43335|issue:45072",
+        "issue:44462|issue:44561",
+        "issue:44561|issue:44779",
+        "issue:43720|issue:44112",
+        "issue:43335|issue:45593",
+        "issue:44977|issue:45229",
+        "issue:42490|issue:43854",
+        "issue:40444|issue:43526",
+        "issue:44448|issue:44977",
+        "issue:34689|issue:35141",
+        "issue:43606|issue:44857"
+      ],
+      "prepared_review_unit_hash": "bd6f8d5671147a30a964f89015d1b612adb37f6e9acf80e8056efc6b37638406",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 35,
+        "serialized_chars": 38236,
+        "estimated_input_tokens": 9559,
+        "estimated_eval_tokens": 19374
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a false-positive grouping of unrelated Transformer bug reports. The only clear duplicate pair is the two identical Qwen3.5 MoE `from_pretrained` regression issues; the rest differ by model, subsystem, or root cause.",
+        "confidence": 0.9,
+        "canonical_issue_reason": "issue:45310 is the best canonical issue because it appears to be the earlier and slightly more detailed report for the exact same Qwen3.5 MoE `from_pretrained` failure as issue:45478.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45310 is the strongest representative of the cluster\u2019s only true duplicate pair and has the clearest tie to the shared regression.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43335",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Different models and failures: SwitchTransformers sparse-layer config bug vs Bamba torch.compile/SDPA issue."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Pegasus v4/v5 output drift is unrelated to Qwen recurrent-state cache resetting."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Different detection losses with distinct failure modes; not the same bug."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Gemma4 text-only mm_token_type_ids defaulting issue is unrelated to flash_attention sink-less model attribute handling."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "DSA ReLU discussion and removal of `is_torch_fx_available` breaking trust_remote_code are unrelated."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "BF16/FP32 dtype regression is unrelated to the pipeline translation-task error message."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but for different missing symbols and likely different fixes."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Flash Attention 2 test failures are not the same as the Trainer `_gpu_broadcast_one` NoneType error."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Different models and different runtime problems; no shared code-path bug."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "General base_model_prefix loading breakage is not the same as Qwen3-VL weight-shape mismatch."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "SigLIP2 docs errors and BEiT reduce_labels bug are unrelated."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer creation is unrelated to BigBird tokenizer special-token decode failure."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "DAC latent/STE mismatch is unrelated to the position-id key warning."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Bamba SDPA compile issue is unrelated to a GraniteSpeech stale device override test failure."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Python 3.13 import indentation/script parsing bug is unrelated to qwen3_omni_moe feature-length helper behavior."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43504",
+            "accept": false,
+            "reason": "Different models and different configuration/loading bugs."
+          },
+          {
+            "left": "issue:45310",
+            "right": "issue:45478",
+            "accept": true,
+            "reason": "Exact same title and same Qwen3.5 MoE `from_pretrained` error; these look like duplicate reports of the same bug."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Unrelated model/config bug and feature-length helper behavior bug."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "LayoutLMV3 error-message clarity issue is unrelated to Trainer broadcast type failure."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Both involve MoE/flash-attention adjacent code, but they describe different failing paths and fixes."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Doc-example/model mismatch is unrelated to unexpected key warnings with position ids."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Qwen-3-8B-NVFP4 runtime failure is unrelated to the Python 3.13 flash_attn import KeyError."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "SigLIP2 zero-probability behavior is unrelated to position-id key warnings."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Different models and different regressions, despite both being v5-related output issues."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "SwitchTransformers config bug is unrelated to dtype mismatches in bfloat16 inference."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "Tokenizer.json selection issue is unrelated to removed utility function breakage."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Trust-remote-code breakage is unrelated to Deepseek tokenizer output regression."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "BitNet packed-weight loading bug is unrelated to GraniteSpeech stale override CI failure."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Different models and different failure modes; not the same underlying bug."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Qwen3.5 flash-attention generation bug and Gemma4 multi-GPU OOM are unrelated."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Local-path import warning is unrelated to GLM-4.7-Flash loading failure in tests."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Qwen2.5-VL multi-image finetuning failure is unrelated to BEiT fast processor label reduction."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Different models and distinct v5 behavior regressions; not mergeable as one bug."
+          },
+          {
+            "left": "issue:34689",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Llama 3.2 model loading regression and embedding resize reinit bug are different issues."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "CPU offload device mismatch and float16 AMP loss crash are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded and conservative. The only accepted pair is the clearly identical Qwen3.5 MoE `from_pretrained` report, and the rejection reasons for the other soft pairs stay within what the titles/body metadata support. The canonical issue choice is also reasonable, with the earlier report selected appropriately."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 25,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:10000",
+        "issue:15354",
+        "issue:28218",
+        "issue:29127",
+        "issue:29942",
+        "issue:32090",
+        "issue:33357",
+        "issue:35141",
+        "issue:37428",
+        "issue:41720",
+        "issue:42490",
+        "issue:43065",
+        "issue:43278",
+        "issue:43329",
+        "issue:43412",
+        "issue:43550",
+        "issue:43575",
+        "issue:43653",
+        "issue:43701",
+        "issue:43723",
+        "issue:43873",
+        "issue:43927",
+        "issue:43976",
+        "issue:43994",
+        "issue:44493",
+        "issue:44554",
+        "issue:44568",
+        "issue:44743",
+        "issue:44857",
+        "issue:44977",
+        "issue:44987",
+        "issue:45072",
+        "issue:45083",
+        "issue:45198",
+        "issue:45229",
+        "issue:45245",
+        "issue:45372",
+        "issue:45406",
+        "issue:45412",
+        "issue:45440",
+        "issue:45459",
+        "issue:45491",
+        "issue:45563",
+        "issue:45588",
+        "issue:45698",
+        "issue:45712",
+        "issue:45803"
+      ],
+      "soft_pairs": [
+        "issue:43723|issue:44554",
+        "issue:44857|issue:45072",
+        "issue:44977|issue:45563",
+        "issue:44568|issue:45491",
+        "issue:43927|issue:45698",
+        "issue:28218|issue:32090",
+        "issue:43701|issue:45083",
+        "issue:43550|issue:45083",
+        "issue:45198|issue:45803",
+        "issue:45372|issue:45588",
+        "issue:37428|issue:44987",
+        "issue:29942|issue:35141",
+        "issue:45459|issue:45712",
+        "issue:43873|issue:44493",
+        "issue:43412|issue:45588",
+        "issue:32090|issue:33357",
+        "issue:43873|issue:45229",
+        "issue:45245|issue:45563",
+        "issue:41720|issue:45229",
+        "issue:43278|issue:43329",
+        "issue:10000|issue:15354",
+        "issue:43976|issue:45440",
+        "issue:45372|issue:45406",
+        "issue:43065|issue:43994",
+        "issue:42490|issue:43575",
+        "issue:28218|issue:29127",
+        "issue:43065|issue:45412",
+        "issue:45372|issue:45459",
+        "issue:44857|issue:45588",
+        "issue:43653|issue:44743"
+      ],
+      "prepared_review_unit_hash": "6081a1bac8a3467e4375579576655a23f36bc4bf5845b412539d92e80fefa77c",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 30,
+        "serialized_chars": 36332,
+        "estimated_input_tokens": 9083,
+        "estimated_eval_tokens": 18422
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a grab bag of unrelated Transformers bug reports rather than a true duplicate set. None of the soft-edge pairs look like the same underlying issue or change, so I would not merge any of them.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "No single issue is a clear canonical duplicate target because the cluster is heterogeneous. If a representative anchor is needed, issue 45372 is the most broadly useful open, well-described bug report with downstream references.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45372 is the best global issue candidate here because it is open, concrete, and has inbound references, but it still does not meaningfully unify the cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43723",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "Different subsystems and failures: tokenizer loading vs MPS attention correctness."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Both involve numeric precision, but they are different model-specific bugs with different failure modes."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Flash-attention generation bug vs a stale warning about num_return_sequences."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Tokenizer special-token behavior is unrelated to Gemma3 NaN embeddings on mixed-length batches."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Both touch save/load paths, but one is config token IDs and the other is custom module selection; not the same bug."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Tokenizer spacing issue vs distributed trainer broadcast type error."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Checkpoint key mismatch is unrelated to a helper-function behavior bug."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure on Bamba is unrelated to qwen3_omni_moe feature-length logic."
+          },
+          {
+            "left": "issue:45198",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "save_pretrained/tokenization failure is not the same as a bare-except swallowing errors."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Gemma 4 processor import failure and flash_attention s_aux handling are different concrete bugs."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Shared release area, but one is a missing import symbol and the other is a model loading failure."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Flash Attention test failures are unrelated to embedding reinitialization in post_init."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Tokenizer exception masking is unrelated to leftover dummy classes leaking into dir()."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Quantization offloading and unexpected position-id keys are different issues."
+          },
+          {
+            "left": "issue:43412",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "ExecuTorch export problems are unrelated to flash_attention sink-less-model crashes."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Trainer NoneType broadcast error and MacOS bus error on CLIP are not the same bug."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Offloading/quantization behavior is unrelated to multi-GPU OOM."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Category-cardinality runtime error is unrelated to a stale generate warning."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "cudaErrorAssert during auto device mapping is not the same as plain OOM."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Embedding dtype regression is unrelated to undefined video-token helper variables."
+          },
+          {
+            "left": "issue:10000",
+            "right": "issue:15354",
+            "accept": false,
+            "reason": "Completely different models and failures: DistilBert model issue vs torch.jit.script GeneratorExp support."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Python version compatibility is unrelated to DeepSeekV3MoE divergence from remote implementation."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both mention Gemma 4, but one is processor import loading and the other is serve-time missing _tokenizer state."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Sam3PixelDecoder dummy layer and SigLIP2 nonsensical outputs are unrelated model bugs."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Import warning on trailing slash paths is unrelated to tensor-parallel OOM."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "Tokenizer adds extra space vs layoutlmv3 box-validation error messaging are not duplicates."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Dummy Conv2d in Sam3PixelDecoder is unrelated to RT-DETR memory release behavior."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "Gemma 4 import dependency issue is not the same as protobuf-related tokenizer error masking."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "AMP/CUDA loss crash and flash_attention AttributeError are different code paths."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration bug is unrelated to Qwen3 recurrent state reset."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 26,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:25251",
+        "issue:29127",
+        "issue:29942",
+        "issue:30064",
+        "issue:32090",
+        "issue:33290",
+        "issue:33357",
+        "issue:34567",
+        "issue:39692",
+        "issue:42490",
+        "issue:42915",
+        "issue:42994",
+        "issue:43066",
+        "issue:43257",
+        "issue:43278",
+        "issue:43381",
+        "issue:43421",
+        "issue:43653",
+        "issue:43756",
+        "issue:43761",
+        "issue:43824",
+        "issue:43873",
+        "issue:43994",
+        "issue:44186",
+        "issue:44206",
+        "issue:44265",
+        "issue:44279",
+        "issue:44361",
+        "issue:44442",
+        "issue:44448",
+        "issue:44451",
+        "issue:44462",
+        "issue:44466",
+        "issue:44479",
+        "issue:44610",
+        "issue:44871",
+        "issue:44977",
+        "issue:45103",
+        "issue:45198",
+        "issue:45372",
+        "issue:45412",
+        "issue:45440",
+        "issue:45491",
+        "issue:45520",
+        "issue:45588",
+        "issue:45593",
+        "issue:45823",
+        "issue:45859"
+      ],
+      "soft_pairs": [
+        "issue:44448|issue:44466",
+        "issue:45588|issue:45823",
+        "issue:43278|issue:43381",
+        "issue:44206|issue:44479",
+        "issue:43066|issue:44462",
+        "issue:43994|issue:45412",
+        "issue:44186|issue:45198",
+        "issue:25251|issue:32090",
+        "issue:44451|issue:45520",
+        "issue:44361|issue:45198",
+        "issue:42915|issue:42994",
+        "issue:44265|issue:45198",
+        "issue:44871|issue:44977",
+        "issue:43756|issue:45440",
+        "issue:29942|issue:30064",
+        "issue:43421|issue:45491",
+        "issue:43257|issue:43278",
+        "issue:33357|issue:39692",
+        "issue:44610|issue:45412",
+        "issue:29127|issue:34567",
+        "issue:42490|issue:43873",
+        "issue:32090|issue:33290",
+        "issue:45103|issue:45372",
+        "issue:45440|issue:45593",
+        "issue:43824|issue:44279",
+        "issue:43653|issue:44442",
+        "issue:43761|issue:45859"
+      ],
+      "prepared_review_unit_hash": "6e4d5bc08dda6a7feb0ea134b973387278538d22db4174365a74ddb45cdda595",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 35829,
+        "estimated_input_tokens": 8958,
+        "estimated_eval_tokens": 18172
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The items are a heterogeneous mix of unrelated bugs, regressions, tokenizer/model issues, and training/runtime problems. None of the soft-similarity pairs look like the same underlying bug or change, so all soft edges should be rejected.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No single canonical issue stands out: the set spans many unrelated subsystems and the soft pairs do not collapse into one duplicate family.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No issue is a strong global representative for this cluster, because the similarities are only superficial and the underlying problems differ.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44448",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Both mention v5 behavior changes, but one is a generation-output difference and the other is device-dependent weight serialization; different bugs and code paths."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "One is a flash-attention crash on sink-less models, the other is a Gemma4 PLE device-mismatch during forward; not the same failure."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "BF16-to-FP32 dtype drift and gradient-checkpointing-in-eval are unrelated issues."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Unsupported processor argument in LasrFeatureExtractor vs a Qwen VL video-input regression; different components and symptoms."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:44462",
+            "accept": false,
+            "reason": "Both are tokenizer-loading related, but one is a wrong decoder type in v5 and the other is ignoring tokenizer.json; not the same bug."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "SigLIP2 doc example problems and RT-DETR memory leaks are unrelated."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "LayoutLMv2 tokenizer crashes on NER/padding issues, while Wav2Vec2 has save_pretrained/tokenization failures; different code paths."
+          },
+          {
+            "left": "issue:25251",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Pipeline top_k nesting behavior and Trainer _gpu_broadcast_one TypeError are unrelated."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Model-load failure for ScandiBERT and Python 3.13 flash_attn import key error are unrelated."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "MLukeTokenizer AttributeError on tasks is not the same underlying problem as Wav2Vec2 save/tokenization failure."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:42994",
+            "accept": false,
+            "reason": "Both involve quantization, but one is a FineGrainedFP8Config failure and the other is a quantized-model save failure; different stages and causes."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "torch.export failure with torch_compilable_check is unrelated to Wav2Vec2 save_pretrained/tokenization issues."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Gemma-3 eos_token_id inconsistency and Qwen3.5 flash-attention generation regression are different bugs."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Both compare model behavior to an expected/reference implementation, but they concern different models and different defects."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "Flash Attention 2 test failures and void segmentation map processing failures are unrelated."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Runtime special-token/post-processor update logic is unrelated to Gemma3 NaN embeddings in mixed-length sliding-window batches."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43278",
+            "accept": false,
+            "reason": "Qwen3 MoE weight-conversion issues and embedding dtype changes are not the same underlying bug."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:39692",
+            "accept": false,
+            "reason": "Both involve vision models, but MacOS bus errors on CLIP and SigLIP2 doc/quantization problems are distinct."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Processor input-size mismatch and RT-DETR memory not being released are unrelated."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "LayoutLMv3 error messaging for boxes and TrainerState token-count tracking are different issues."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Local-path import warning and quantization offloading problems do not share the same code path."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "Trainer NoneType broadcast error and DeepSpeed Adafactor OOM are both training-related but clearly separate bugs."
+          },
+          {
+            "left": "issue:45103",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "Auto-docstring annotation parsing crash and Gemma4 processor import error are unrelated."
+          },
+          {
+            "left": "issue:45440",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "DeepSeekV3 parity divergence and D-FINE auxiliary-loss behavior are different model-specific problems."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Specific Qwen2.5-VL import error versus a vague dependency issue; too generic to be the same bug."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration and AutoTokenizer loading FastSpeech2ConformerTokenizer are different tokenizer issues."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "CLIPVisionModel hidden_states regression and Qwen3_5Moe rotary embedding CPU-offload incompatibility are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 27,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:29127",
+        "issue:29942",
+        "issue:33357",
+        "issue:35141",
+        "issue:36010",
+        "issue:42490",
+        "issue:42915",
+        "issue:43232",
+        "issue:43262",
+        "issue:43316",
+        "issue:43334",
+        "issue:43381",
+        "issue:43388",
+        "issue:43441",
+        "issue:43525",
+        "issue:43526",
+        "issue:43531",
+        "issue:43575",
+        "issue:43646",
+        "issue:43701",
+        "issue:43723",
+        "issue:43761",
+        "issue:43931",
+        "issue:44062",
+        "issue:44077",
+        "issue:44079",
+        "issue:44186",
+        "issue:44188",
+        "issue:44206",
+        "issue:44387",
+        "issue:44442",
+        "issue:44464",
+        "issue:44479",
+        "issue:44743",
+        "issue:44811",
+        "issue:44857",
+        "issue:44977",
+        "issue:45003",
+        "issue:45072",
+        "issue:45245",
+        "issue:45290",
+        "issue:45440",
+        "issue:45479",
+        "issue:45491",
+        "issue:45588",
+        "issue:45803",
+        "issue:45834",
+        "issue:45907"
+      ],
+      "soft_pairs": [
+        "issue:43526|issue:45834",
+        "issue:43381|issue:43701",
+        "issue:43526|issue:44186",
+        "issue:29942|issue:36010",
+        "issue:29127|issue:33357",
+        "issue:43334|issue:43531",
+        "issue:44977|issue:45245",
+        "issue:44464|issue:45491",
+        "issue:29127|issue:35141",
+        "issue:42915|issue:43931",
+        "issue:44442|issue:45072",
+        "issue:43262|issue:44206",
+        "issue:43526|issue:44857",
+        "issue:42490|issue:43381",
+        "issue:44062|issue:45907",
+        "issue:43441|issue:45588",
+        "issue:43232|issue:43388",
+        "issue:44743|issue:44811",
+        "issue:43646|issue:44077",
+        "issue:43761|issue:44079",
+        "issue:43575|issue:43701",
+        "issue:45003|issue:45440",
+        "issue:43575|issue:44387",
+        "issue:43723|issue:45479",
+        "issue:44442|issue:45803",
+        "issue:44479|issue:45290",
+        "issue:43761|issue:44188",
+        "issue:43316|issue:43525"
+      ],
+      "prepared_review_unit_hash": "472d8644d3674aa9a72d3c4179dc858701aa133f81df81ed3f29c4dca3d26760",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36170,
+        "estimated_input_tokens": 9043,
+        "estimated_eval_tokens": 18342
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is not a duplicate set; it mixes many unrelated Transformers bugs across tokenizers, vision, audio, generation, quantization, and config APIs. The soft pairs are mostly broad topical similarities only, not the same underlying defect.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "No single canonical duplicate stands out because the items are heterogeneous. If forced to pick an anchor, issue 43646 is the broadest umbrella-style regression report.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43646 is the most general, high-level regression report in the set and is the best representative anchor, though it is not a true duplicate of the others.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43526",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Both are vision/processor-related, but one is a label reduction bug and the other is an OCR length/index error; different code paths and failures."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Training-time gradient checkpointing in eval mode is unrelated to resume-from-checkpoint key mismatches."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44186",
+            "accept": false,
+            "reason": "Different tokenizer/processor bugs: label reduction vs NER padding/truncation crashes."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Flash Attention test failures and a GenerationMixin import error are unrelated."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "LayoutLMv3 box-content messaging and a MacOS CLIP bus error are clearly different defects."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Missing pad_token_id on Qwen3-VL text config is a different bug from Qwen3-MoE sliding_window behavior."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "Flash-attention generation regression and a category-count limit error are unrelated."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Both involve generation/attention, but chunked compiled generation inconsistency and Gemma3 NaN embeddings have different failure modes and likely different fixes."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Layout boxes validation/message clarity is unrelated to embedding reinitialization during post_init."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config failure and Qwen3-VL weight-shape mismatch are different loading issues."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Tokenizer loading and bfloat16 inference dtype mismatches are unrelated."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44206",
+            "accept": false,
+            "reason": "Audio chat-template sample-rate defaults and an unsupported feature-extractor argument are different audio-processing bugs."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Label reduction behavior and LwDetrImageLoss AMP/CUDA crashing are unrelated."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "Trailing-slash import warnings and eval-mode gradient checkpointing are unrelated."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "AddedToken argument duplication and list_repo_files revision forwarding are separate API issues."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Both mention flash attention, but one is a model-specific failure and the other is an s_aux=None crash in a different code path."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:43388",
+            "accept": false,
+            "reason": "Generation kwargs sync logic and metrics gathering label truncation are unrelated."
+          },
+          {
+            "left": "issue:44743",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Recurrent-state reset with cache and Whisper batch decoding skip_special_tokens are different subsystems."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "General custom model initialization breakage is not the same concrete bug as patchtsmixer's post_init allowance change."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Hidden_states disappearing in CLIPVisionModel.forward and ModelOutput key assignment when a previous key is None are different output-handling bugs."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Tensor-parallel OOM on a large model is unrelated to checkpoint key mismatches."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Unsafe sys.modules access and DeepseekV3MoE remote/native divergence are unrelated."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Both are OOM reports, but one is tensor-parallel loading and the other is increased reserved memory under int4 quantization; different causes and fixes."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "AutoTokenizer loading and single-label classification zero-loss behavior are unrelated."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Tokenizer-loading failure and bare-except masking errors in Fuyu tensor conversion are different bugs."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both involve multimodal/chat flows, but the video-input regression and tool-call chat-template crash are distinct code paths."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44188",
+            "accept": false,
+            "reason": "A CLIPVisionModel output regression is unrelated to torch.compile attention-kernel branching."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Both are config/API discrepancies, but they affect different models and different missing-attribute behaviors; not the same bug."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 28,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:33290",
+        "issue:34567",
+        "issue:34689",
+        "issue:35141",
+        "issue:36010",
+        "issue:36331",
+        "issue:39692",
+        "issue:41628",
+        "issue:42175",
+        "issue:42490",
+        "issue:42907",
+        "issue:43065",
+        "issue:43316",
+        "issue:43381",
+        "issue:43404",
+        "issue:43441",
+        "issue:43493",
+        "issue:43540",
+        "issue:43611",
+        "issue:43643",
+        "issue:43673",
+        "issue:43720",
+        "issue:43749",
+        "issue:43827",
+        "issue:43873",
+        "issue:43931",
+        "issue:44077",
+        "issue:44162",
+        "issue:44190",
+        "issue:44442",
+        "issue:44479",
+        "issue:44496",
+        "issue:44514",
+        "issue:44617",
+        "issue:44743",
+        "issue:44792",
+        "issue:44855",
+        "issue:44871",
+        "issue:44877",
+        "issue:44898",
+        "issue:44913",
+        "issue:45137",
+        "issue:45216",
+        "issue:45372",
+        "issue:45491",
+        "issue:45698",
+        "issue:45759"
+      ],
+      "soft_pairs": [
+        "issue:42907|issue:45698",
+        "issue:44442|issue:44898",
+        "issue:42490|issue:44743",
+        "issue:33290|issue:34567",
+        "issue:34689|issue:43611",
+        "issue:39692|issue:45491",
+        "issue:43931|issue:45216",
+        "issue:33290|issue:35141",
+        "issue:33290|issue:34689",
+        "issue:43381|issue:43873",
+        "issue:43493|issue:43827",
+        "issue:44077|issue:44479",
+        "issue:43316|issue:45759",
+        "issue:43720|issue:44442",
+        "issue:41628|issue:45372",
+        "issue:43643|issue:44913",
+        "issue:43749|issue:43931",
+        "issue:43873|issue:44871",
+        "issue:34689|issue:36010",
+        "issue:44190|issue:44496",
+        "issue:33290|issue:36331",
+        "issue:44855|issue:45137",
+        "issue:43316|issue:44877",
+        "issue:43540|issue:44514",
+        "issue:42175|issue:43316",
+        "issue:43673|issue:44743",
+        "issue:43065|issue:44617",
+        "issue:43404|issue:43441",
+        "issue:44162|issue:44792"
+      ],
+      "prepared_review_unit_hash": "90d10cac729bb1da9151671520641b165b5f4e0032bdf833071c2f49cb98d1e6",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 35918,
+        "estimated_input_tokens": 8980,
+        "estimated_eval_tokens": 18216
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is highly heterogeneous: most items are unrelated model-specific bugs, config/import regressions, or docs issues. I found no convincing duplicate pairs among the soft edges, so this should not be merged into a single canonical bug cluster.",
+        "confidence": 0.92,
+        "canonical_issue_reason": "If a representative issue is needed, issue 45698 is the broadest and most generic load/save regression in the set; however, the overall cluster does not look like a true duplicate group.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45698 is the best global representative because it concerns core from_pretrained/save_pretrained behavior, which is broader than the mostly model-specific or niche issues here.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42907",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Different problems: saving dequantized models vs loading the wrong custom module after save/load."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Unrelated model paths and failures: tokenizer loading vs Perceiver image classification/interpolation."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Different subsystems: local path warning vs recurrent state reset in a Qwen modular model."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Unrelated issues: DeepSpeed Adafactor OOM vs TrainerState token counter not updating."
+          },
+          {
+            "left": "issue:34689",
+            "right": "issue:43611",
+            "accept": false,
+            "reason": "Both are loading regressions, but they affect different mechanisms and code paths."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Different models and symptoms: SigLIP2 docs example errors vs Gemma3 NaN embeddings."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both involve Qwen models, but one is a shape-mismatch load error and the other is a save_pretrained regression."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "OOM in optimizer training is unrelated to embedding reinitialization during post_init."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Different bug classes: optimizer memory blowup vs model loading failure."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Eval-mode gradient checkpointing is unrelated to quantization offloading behavior."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "SigLIP2 implementation discrepancy is unrelated to docs still referencing pipeline()."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Different model families and different failures; no shared concrete code path."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both mention config/API shape, but they do not clearly describe the same underlying bug."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "BitNet packed-weight loading failure is unrelated to tokenizer loading for FastSpeech2Conformer."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols, packages, and loaders."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Both involve config values after reload, but the affected configs and failure modes differ."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "FSDP CPU RAM efficient loading is a distributed-loading issue, not the same as a Qwen weight-shape mismatch."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Quantization/offloading behavior is unrelated to eos_token_id configuration inconsistency."
+          },
+          {
+            "left": "issue:34689",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Different breakages: one is a model loading regression, the other is an import path failure for GenerationMixin."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Dataset loading in a training script is unrelated to an unrecognized model/config error."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Python 3.13 import parsing error is unrelated to DeepSpeed ZeRO3 deque failure."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Different config-loading problems; no evidence they are the same bug."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "Both are Qwen multimodal input bugs, but the failing inputs and code paths differ."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43316",
+            "accept": false,
+            "reason": "Backend dependency packaging issue is unrelated to a Gemma3TextConfig API discrepancy."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Cache-missing in chunked prefill and recurrent-state reset are different cache bugs."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "A dummy Conv2d in Sam3PixelDecoder is not the same as a Sam3Video CUDA OOM."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Both are Mistral-family issues, but one is tied weights and the other is FlashAttention compatibility."
+          },
+          {
+            "left": "issue:44162",
+            "right": "issue:44792",
+            "accept": false,
+            "reason": "ESM2 regression and janus image-generation test failure are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 29,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:16998",
+        "issue:22355",
+        "issue:25251",
+        "issue:28282",
+        "issue:32090",
+        "issue:35141",
+        "issue:41950",
+        "issue:42617",
+        "issue:43065",
+        "issue:43295",
+        "issue:43388",
+        "issue:43452",
+        "issue:43475",
+        "issue:43504",
+        "issue:43550",
+        "issue:43575",
+        "issue:43606",
+        "issue:43646",
+        "issue:43688",
+        "issue:43701",
+        "issue:43742",
+        "issue:43761",
+        "issue:43825",
+        "issue:43906",
+        "issue:43931",
+        "issue:44060",
+        "issue:44077",
+        "issue:44242",
+        "issue:44464",
+        "issue:44479",
+        "issue:44496",
+        "issue:44521",
+        "issue:44534",
+        "issue:44610",
+        "issue:44617",
+        "issue:44734",
+        "issue:44792",
+        "issue:44871",
+        "issue:44913",
+        "issue:44977",
+        "issue:45127",
+        "issue:45200",
+        "issue:45290",
+        "issue:45381",
+        "issue:45406",
+        "issue:45563",
+        "issue:45584",
+        "issue:45698"
+      ],
+      "soft_pairs": [
+        "issue:25251|issue:28282",
+        "issue:43295|issue:43825",
+        "issue:43388|issue:43688",
+        "issue:41950|issue:43475",
+        "issue:44077|issue:44534",
+        "issue:16998|issue:22355",
+        "issue:43550|issue:43701",
+        "issue:44871|issue:45584",
+        "issue:43761|issue:44479",
+        "issue:43606|issue:43701",
+        "issue:45127|issue:45200",
+        "issue:43295|issue:44871",
+        "issue:43452|issue:44496",
+        "issue:42617|issue:43575",
+        "issue:43575|issue:43742",
+        "issue:44734|issue:45406",
+        "issue:43646|issue:44913",
+        "issue:43504|issue:45698",
+        "issue:44617|issue:45584",
+        "issue:43761|issue:44242",
+        "issue:43931|issue:44060",
+        "issue:45563|issue:45584",
+        "issue:32090|issue:35141",
+        "issue:44464|issue:44871",
+        "issue:43065|issue:44610",
+        "issue:44479|issue:44734",
+        "issue:44792|issue:44977",
+        "issue:44521|issue:45381",
+        "issue:43906|issue:45290",
+        "issue:44464|issue:45563"
+      ],
+      "prepared_review_unit_hash": "63b4d9320b40dc30ba22228bb113fe4d23859ff1af3bfab868eec4d9b2bd5046",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36698,
+        "estimated_input_tokens": 9175,
+        "estimated_eval_tokens": 18606
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster does not contain any convincing duplicate pairs; the soft matches mostly share only broad subsystem vocabulary (pipelines, loading, generation, training) but describe different failures. I picked a highly active issue as the representative issue since there are no PRs and no true canonical duplicate emerges.",
+        "confidence": 0.23,
+        "canonical_issue_reason": "issue:44521 is the most active/open, with the strongest inbound reference signal and broad current relevance, so it is the best available representative despite the cluster not being a real duplicate set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44521 is the best global issue representative here because it has the most discussion/inbound reference activity among the listed issues and is still open.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:25251",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Different symptoms and code paths: pipeline top_k return-shape behavior vs AutoModel/PyTorch import failure."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Both involve v5 regressions around processors/pipeline messaging, but one is a processor API break and the other is an incorrect task-support error message."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43688",
+            "accept": false,
+            "reason": "Training metric gathering bug vs auxiliary-loss normalization bug; unrelated mechanisms and model families."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Video-classification processor lookup issue is unrelated to a SAM3 vision encoder missing attribute."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Optional post_init deprecation/allowance issue is different from non-persistent buffers being filled with junk."
+          },
+          {
+            "left": "issue:16998",
+            "right": "issue:22355",
+            "accept": false,
+            "reason": "Model_max_length question for DeBERTa-v3 is unrelated to missing transformers.onnx module import."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "torch.compile SDPA failure in Bamba is unrelated to resume_from_checkpoint key mismatches."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Gemma eos_token_id mismatch is unrelated to Whisper empty-transcription failure after align_special_tokens."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "CLIP hidden_states regression and Qwen video-input regression are both v5 issues but different models, inputs, and failure points."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "CPU offload device mismatch in bark-small is unrelated to checkpoint key mismatch."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "LoRA merge/tie_word_embeddings collapse is not the same bug as missing default mm_token_type_ids in Gemma 4 fine-tuning."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Processor/tokenizer regression and eos_token_id configuration mismatch are distinct API/config problems."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "gguf_file handling for tokenizer/model loading is not the same as unrecognized model loading due to missing model_type."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "3d_parallel runtime failure and Qwen2-57B-A14B-Instruct OOM during tensor-parallel load are different issues."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "OOM with tensor parallel loading is not the same as a key error while loading MobileLLM."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "serve KV-cache continuation indexing crash is unrelated to Gemma4Processor missing _tokenizer."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Custom model initialization breakage in Transformers 5.0 is unrelated to GPTNeoX rotary_pct not persisting on reload."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "BEiT preset legacy-field loading bug is unrelated to wrong custom module resolution after save/load."
+          },
+          {
+            "left": "issue:44617",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Sam3Video OOM and Whisper empty-transcription failure are different resource vs logic failures."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44242",
+            "accept": false,
+            "reason": "Both are model/training regressions, but hidden_states propagation and load-balancing loss inclusion are unrelated code paths."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44060",
+            "accept": false,
+            "reason": "Qwen3-VL weight-shape mismatch and Qwen3-Next tied-weights warning bug are different loading/config issues."
+          },
+          {
+            "left": "issue:45563",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Paged generate stale warning is unrelated to Whisper empty transcription crash."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Trainer gpu_broadcast NoneType error and resize-token-embeddings reinit bug affect different subsystems and behaviors."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Chunked generation inconsistency under compiled forward is unrelated to Gemma eos_token_id config mismatch."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Sam3PixelDecoder dummy Conv2d implementation issue is not the same as OmDet-Turbo processor output-size mismatch."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Qwen2.5-VL/Qwen3 video-input regression and serve KV-cache continuation indexing bug are different failures."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Janus image-generation test failure is unrelated to Qwen3.5 flash-attention generation malfunction."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Assistant mask generation for multimodal chat templates is distinct from Qwen2.5-VL video vision_position_ids being wrong."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Isolated reproduction of an external issue is unrelated to chat template crashes on tool-call assistant messages."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Compiled-forward chunked generation inconsistency is unrelated to paged generate warning text."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 30,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:35141",
+        "issue:36246",
+        "issue:39692",
+        "issue:41628",
+        "issue:41720",
+        "issue:41950",
+        "issue:42490",
+        "issue:43295",
+        "issue:43335",
+        "issue:43388",
+        "issue:43479",
+        "issue:43540",
+        "issue:43550",
+        "issue:43582",
+        "issue:43644",
+        "issue:43650",
+        "issue:43827",
+        "issue:43976",
+        "issue:43994",
+        "issue:44016",
+        "issue:44077",
+        "issue:44079",
+        "issue:44112",
+        "issue:44190",
+        "issue:44220",
+        "issue:44462",
+        "issue:44496",
+        "issue:44534",
+        "issue:44734",
+        "issue:44737",
+        "issue:44792",
+        "issue:44855",
+        "issue:44913",
+        "issue:44933",
+        "issue:44938",
+        "issue:44964",
+        "issue:45127",
+        "issue:45245",
+        "issue:45276",
+        "issue:45357",
+        "issue:45362",
+        "issue:45375",
+        "issue:45542",
+        "issue:45584",
+        "issue:45588",
+        "issue:45600",
+        "issue:45803",
+        "issue:45907"
+      ],
+      "soft_pairs": [
+        "issue:43827|issue:43994",
+        "issue:44534|issue:44913",
+        "issue:44190|issue:44964",
+        "issue:44933|issue:45362",
+        "issue:41720|issue:45245",
+        "issue:44792|issue:45600",
+        "issue:44855|issue:45588",
+        "issue:44737|issue:45803",
+        "issue:45127|issue:45357",
+        "issue:43388|issue:43582",
+        "issue:35141|issue:42490",
+        "issue:44462|issue:45907",
+        "issue:43335|issue:44855",
+        "issue:44077|issue:44079",
+        "issue:44938|issue:45542",
+        "issue:43976|issue:44855",
+        "issue:43550|issue:44190",
+        "issue:43644|issue:44913",
+        "issue:43540|issue:44734",
+        "issue:44112|issue:45600",
+        "issue:44496|issue:45803",
+        "issue:43650|issue:44220",
+        "issue:43335|issue:44190",
+        "issue:45276|issue:45584",
+        "issue:39692|issue:44190",
+        "issue:45245|issue:45584",
+        "issue:43479|issue:45375",
+        "issue:41950|issue:44016",
+        "issue:43295|issue:45276",
+        "issue:36246|issue:41628"
+      ],
+      "prepared_review_unit_hash": "696194d6cbab93cf7987e73883c3be319c417be8dc4831a484a4de086619c68c",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36921,
+        "estimated_input_tokens": 9231,
+        "estimated_eval_tokens": 18718
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous: one small token-embedding resize/output-embedding theme, plus many unrelated import, config, pipeline, and runtime bugs. No PRs are present, so the best canonical issue is the clearest representative issue in the token-embedding area.",
+        "confidence": 0.74,
+        "canonical_issue_reason": "Issue 35141 is the clearest, most general description of the resize-token-embeddings/output-embedding regression and is the best anchor among the issues here.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "35141 is the least-specific, most precise statement of the core bug, while the rest of the cluster is largely unrelated noise.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43827",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Both mention pipeline usage, but one is about removed v5 pipelines in docs and the other is a SigLIP2 model/processor correctness issue."
+          },
+          {
+            "left": "issue:44534",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "One is about non-persistent buffers being junk in v5, the other is a GPTNeoX config field not persisting on reload."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different failure modes: local dataset loading in an example script versus loading a specific multimodal model."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "Both are import/runtime problems, but they involve different symbols and code paths with no clear shared bug."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "A device-mapping CUDA assert and a category-count overflow are unrelated failures."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Both affect CI/tests, but they are about different models and different root causes."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Python 3.13 import parsing versus a flash-attention crash on missing sink parameters are unrelated."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "One is a performance/device placement bug in XLNet, the other is a broad exception-handling bug in Fuyu preprocessing."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Both are model-specific regressions, but one is about LoRA/extended vocab collapse and the other about incorrect visual encoder keys in save_pretrained."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "A metrics-gathering label truncation bug and an Apple Silicon allocator TypeError are not the same issue."
+          },
+          {
+            "left": "issue:35141",
+            "right": "issue:42490",
+            "accept": false,
+            "reason": "Token-embedding reinitialization and a trailing-slash local path import warning are unrelated."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "AutoTokenizer ignoring tokenizer.json and list_repo_files not forwarding revision are separate repository-loading bugs."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "SwitchTransformers layer construction and a DebertaV2 Python 3.13 import error are unrelated."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Optional post_init enforcement in patchtsmixer and ModelOutput key assignment are different code paths."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Python 3.14 import failure and undefined tf backend from tensorboard-only install are unrelated environment issues."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "Both are Python-version compatibility reports, but they concern different versions and distinct import failures."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "A torch.compile/SDPA model bug and a dataset loading example bug are unrelated."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Non-persistent buffer corruption and a config reload default-reset bug are not the same underlying defect."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Video input validation in Qwen3OmniMoe and KV-cache continuation indexing in serve are different failures."
+          },
+          {
+            "left": "issue:44112",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "These are separate CI breakages with different affected components and causes."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Model config recognition failure and bare-except swallowing errors are unrelated."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "The first is a nonspecific placeholder title; the second is a concrete feature-extraction bug."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Single-layer SwitchTransformers config logic and local dataset loading are unrelated."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Resize-token-embeddings/output-embedding behavior and Whisper empty-transcription generation are unrelated."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "SigLIP2 doc example issues and a local dataset loading failure are different problems."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Category-count overflow and Whisper empty-transcription generation are unrelated."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both involve multimodal config fields, but they describe different models and different serialization/defaulting bugs."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:44016",
+            "accept": false,
+            "reason": "A pipeline processor lookup bug and a notebook syntax error are clearly unrelated."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "Processor.tokenizer regression and resize_token_embeddings not updating embeddings are different bugs."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "Both are import errors, but one is a missing model-specific processor export and the other is a top-level AutoImageProcessor import issue."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 31,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:36246",
+        "issue:38175",
+        "issue:38617",
+        "issue:42175",
+        "issue:42490",
+        "issue:42907",
+        "issue:43065",
+        "issue:43295",
+        "issue:43299",
+        "issue:43381",
+        "issue:43404",
+        "issue:43441",
+        "issue:43526",
+        "issue:43531",
+        "issue:43643",
+        "issue:43673",
+        "issue:43824",
+        "issue:43854",
+        "issue:43901",
+        "issue:43976",
+        "issue:44016",
+        "issue:44062",
+        "issue:44075",
+        "issue:44112",
+        "issue:44186",
+        "issue:44220",
+        "issue:44246",
+        "issue:44279",
+        "issue:44336",
+        "issue:44368",
+        "issue:44466",
+        "issue:44496",
+        "issue:44561",
+        "issue:44683",
+        "issue:44734",
+        "issue:44913",
+        "issue:44933",
+        "issue:44964",
+        "issue:44987",
+        "issue:45003",
+        "issue:45042",
+        "issue:45216",
+        "issue:45335",
+        "issue:45446",
+        "issue:45584",
+        "issue:45593",
+        "issue:45600",
+        "issue:45632"
+      ],
+      "soft_pairs": [
+        "issue:36246|issue:38617",
+        "issue:44186|issue:44734",
+        "issue:38175|issue:43065",
+        "issue:42175|issue:44279",
+        "issue:44683|issue:45446",
+        "issue:44016|issue:44220",
+        "issue:44466|issue:44913",
+        "issue:43404|issue:45216",
+        "issue:44734|issue:45632",
+        "issue:44561|issue:45632",
+        "issue:44246|issue:45003",
+        "issue:42490|issue:45632",
+        "issue:43643|issue:44496",
+        "issue:44964|issue:45600",
+        "issue:44112|issue:44964",
+        "issue:42907|issue:43441",
+        "issue:43854|issue:44964",
+        "issue:44016|issue:44062",
+        "issue:43976|issue:44987",
+        "issue:43295|issue:45335",
+        "issue:43673|issue:44336",
+        "issue:43531|issue:45600",
+        "issue:43824|issue:44933",
+        "issue:44496|issue:45042",
+        "issue:43526|issue:43901",
+        "issue:43299|issue:45216",
+        "issue:44075|issue:44368",
+        "issue:45335|issue:45584",
+        "issue:43976|issue:44279",
+        "issue:43381|issue:45593"
+      ],
+      "prepared_review_unit_hash": "9347925bcd14e02f3fc2002dc678a532b3a33d2a3db900099977a75e12095b8d",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36718,
+        "estimated_input_tokens": 9180,
+        "estimated_eval_tokens": 18616
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are mostly unrelated Transformers issues. The soft pairs share broad keywords or subsystems, but they do not describe the same concrete bug/change, so none should be merged as duplicates. There is no meaningful canonical issue/PR for this set.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:36246",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Both are import errors, but for different missing symbols in different modules; not the same bug."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Different code paths: tokenizer batching/NER crash vs serve KV-cache continuation indexing."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43065",
+            "accept": false,
+            "reason": "Unrelated models and failures; one is zero probabilities, the other is a dummy Conv2d in a decoder."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "General dependency/install complaints, but no shared concrete failure mode."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Both mention flex_attention/versioning, but they point to different import/runtime issues and fixes."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "A notebook syntax error and an audio feature extraction issue are unrelated."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Both concern serialization/reload behavior, but different fields and different model classes."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both involve tied weights/save_pretrained, but they are different regressions on different models."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:45632",
+            "accept": false,
+            "reason": "Serve KV-cache indexing and local trust_remote_code cache collisions are unrelated."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:45632",
+            "accept": false,
+            "reason": "Removal of an API vs cache-path collision are different trust_remote_code failures."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Import slowness and unsafe sys.modules access are related at most indirectly, not the same bug."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:45632",
+            "accept": false,
+            "reason": "Both are path-related, but one is a trailing-slash warning and the other is a cache collision."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Different config-loading failures: missing fields with trust_remote_code vs unrecognized model_type."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Model loading failure vs stale auto_mappings CI breakage; different causes."
+          },
+          {
+            "left": "issue:44112",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "CI test flake in GraniteSpeech is unrelated to Phi-4 multimodal loading."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Same family name, but save-dequantized behavior and FlashAttention loading are distinct bugs."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different multimodal model loading problems with different failure modes."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44062",
+            "accept": false,
+            "reason": "Notebook syntax error vs tokenizer AddedToken argument collision; unrelated."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "A Python version support issue is not the same as a specific model-loading regression."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "Processor/tokenizer regression and resize_token_embeddings mismatch are different code paths."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "Generation cache bug and ANSI-code logging issue are unrelated."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior and stale config mappings are unrelated."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Both are import-related, but they involve different missing exports from different modules."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Model config discovery and image processor torchvision dependency are unrelated."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "A BEiT processor bug and a docs/behavior mismatch for text classification are not duplicates."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Different regressions affecting different Qwen model families and different save/load behavior."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Optimizer argument handling and tie_word_embeddings warnings are unrelated."
+          },
+          {
+            "left": "issue:45335",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Embedding resize behavior and Whisper empty-transcription generation are different bugs."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Python/package compatibility and a generic dependency issue do not match the same failure."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Gradient checkpointing eval-mode limitation and D-FINE auxiliary-loss behavior are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 32,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:33290",
+        "issue:38175",
+        "issue:39692",
+        "issue:41950",
+        "issue:42994",
+        "issue:43065",
+        "issue:43295",
+        "issue:43299",
+        "issue:43317",
+        "issue:43441",
+        "issue:43450",
+        "issue:43479",
+        "issue:43525",
+        "issue:43577",
+        "issue:43650",
+        "issue:43854",
+        "issue:43901",
+        "issue:43976",
+        "issue:44016",
+        "issue:44242",
+        "issue:44279",
+        "issue:44291",
+        "issue:44297",
+        "issue:44351",
+        "issue:44462",
+        "issue:44496",
+        "issue:44521",
+        "issue:44617",
+        "issue:44683",
+        "issue:44779",
+        "issue:44898",
+        "issue:45020",
+        "issue:45081",
+        "issue:45200",
+        "issue:45216",
+        "issue:45278",
+        "issue:45325",
+        "issue:45357",
+        "issue:45405",
+        "issue:45542",
+        "issue:45600",
+        "issue:45656",
+        "issue:45712",
+        "issue:45759",
+        "issue:45812",
+        "issue:45823",
+        "issue:45834",
+        "issue:45907"
+      ],
+      "soft_pairs": [
+        "issue:43299|issue:45357",
+        "issue:43901|issue:44496",
+        "issue:44496|issue:45759",
+        "issue:44351|issue:45278",
+        "issue:44521|issue:45325",
+        "issue:41950|issue:45020",
+        "issue:43976|issue:45405",
+        "issue:44242|issue:45200",
+        "issue:43479|issue:43525",
+        "issue:44779|issue:45812",
+        "issue:41950|issue:45823",
+        "issue:38175|issue:44617",
+        "issue:43577|issue:45216",
+        "issue:44683|issue:45542",
+        "issue:43299|issue:43854",
+        "issue:39692|issue:43901",
+        "issue:44279|issue:45405",
+        "issue:41950|issue:43650",
+        "issue:43295|issue:44779",
+        "issue:44462|issue:45081",
+        "issue:44297|issue:44462",
+        "issue:45081|issue:45907",
+        "issue:44016|issue:44291",
+        "issue:43441|issue:45834",
+        "issue:45600|issue:45712",
+        "issue:43450|issue:44279",
+        "issue:33290|issue:45656",
+        "issue:43065|issue:44898",
+        "issue:42994|issue:43317"
+      ],
+      "prepared_review_unit_hash": "4a97838951d538e5edc969679a45c3532e518ec50a824bf92e66e69cf743fcff",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36352,
+        "estimated_input_tokens": 9088,
+        "estimated_eval_tokens": 18432
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a grab bag of distinct Transformers regressions and docs issues, with only superficial similarity around loading/tokenizer/model-compatibility failures. I do not see any pair that is clearly the same underlying bug or change.",
+        "confidence": 0.82,
+        "canonical_issue_reason": "issue:45020 is the broadest open regression report in the set and the best umbrella-style representative for the overall 'recent Transformers versions break model loading' theme.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45020 is the strongest global representative: open, detailed, and broad enough to anchor the cluster even though the other items are not true duplicates.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43299",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different Qwen3VL loading failure vs Qwen3.5 save_pretrained checkpoint-key bug; separate code paths."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Docs mismatch for text classification pipeline vs unrecognized-model/config loading failure."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both involve config/model resolution, but one is unrecognized model loading and the other is AutoModelForCausalLM not unwrapping text_config; not the same bug."
+          },
+          {
+            "left": "issue:44351",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Importing HybridCache missing vs broad post-upgrade import errors; too different and not a single code-path issue."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "assistant_masks generation vs temporal position-id scaling in Qwen2.5-VL; unrelated multimodal bugs."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Video pipeline processor lookup vs remote_code model breakage; different subsystem and failure mode."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Python-version compatibility regression vs PEFT minimum-version bump; no shared bug."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "MoE load-balancing loss gating vs Gemma4 mm_token_type_ids defaulting; separate training semantics."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Phi4 multimodal config defaults vs Llama4Config missing pad_token_id; different model families and errors."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Deepseek tokenizer regression vs Granite tokenizer regression; same broad area, but not the same concrete tokenizer bug."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Video-classification pipeline processor lookup vs Gemma4 PLE device mismatch; unrelated."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Unexpected zero probabilities in SigLIP2 vs Sam3Video CUDA OOM; entirely different failures."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "BLIP2 dtype propagation bug vs Qwen3.5 save_pretrained regression; different code paths."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "torch >= 2.9 flex_attention failure vs tensorboard/TensorFlow backend import issue; unrelated."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Qwen3VL MoE loading regression vs GLM-4.7-Flash unit-test loading issue; different models and causes."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "SigLIP2 documentation example errors vs text-classification docs mismatch; separate documentation bugs."
+          },
+          {
+            "left": "issue:44279",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Generic dependency issue vs PEFT version pin bump; not the same underlying problem."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:43650",
+            "accept": false,
+            "reason": "Video processor lookup bug vs unrelated 'ADD THE DATA' report."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "processor.tokenizer/images-to-tokenizer regression vs Deepseek tokenizer wrong results; both tokenizer-related but not the same concrete defect."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "AutoTokenizer ignoring tokenizer.json vs Mistral regex patch crash; different failure modes."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:44462",
+            "accept": false,
+            "reason": "tokenizer_class mismatch on save vs AutoTokenizer ignoring tokenizer.json on load; related area but distinct bug."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Mistral regex crash vs list_repo_files revision forwarding; unrelated."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Notebook syntax error vs init_empty_weights unexpected argument TypeError; separate issues."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Ministral FlashAttention failure vs Kosmos2.5 OCR index error; different models and code paths."
+          },
+          {
+            "left": "issue:45600",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "auto_mappings stale references vs leftover dummy classes leaking into dir(transformers); both cleanup-related but not the same bug."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Batched video processor shape bug vs generic dependency issue; no overlap."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "DeepSpeed adafactor OOM vs optimizer.step being called twice under DeepSpeed; not the same underlying problem."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Dummy Conv2d in Sam3PixelDecoder vs Perceiver non-default-resolution failure; different components and symptoms."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43317",
+            "accept": false,
+            "reason": "Quantized model saving failure vs dequantized model offload loading failure; separate lifecycle stages and bugs."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 33,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:41628",
+        "issue:42175",
+        "issue:42491",
+        "issue:43122",
+        "issue:43262",
+        "issue:43299",
+        "issue:43441",
+        "issue:43450",
+        "issue:43525",
+        "issue:43643",
+        "issue:43650",
+        "issue:43653",
+        "issue:43673",
+        "issue:43723",
+        "issue:43761",
+        "issue:43784",
+        "issue:43824",
+        "issue:43881",
+        "issue:43901",
+        "issue:43976",
+        "issue:43994",
+        "issue:44016",
+        "issue:44162",
+        "issue:44188",
+        "issue:44261",
+        "issue:44263",
+        "issue:44297",
+        "issue:44336",
+        "issue:44361",
+        "issue:44466",
+        "issue:44521",
+        "issue:44589",
+        "issue:44617",
+        "issue:44623",
+        "issue:44671",
+        "issue:44683",
+        "issue:44749",
+        "issue:44779",
+        "issue:44987",
+        "issue:45276",
+        "issue:45290",
+        "issue:45362",
+        "issue:45397",
+        "issue:45542",
+        "issue:45561",
+        "issue:45563",
+        "issue:45593",
+        "issue:45823"
+      ],
+      "soft_pairs": [
+        "issue:44466|issue:45276",
+        "issue:44261|issue:44263",
+        "issue:43650|issue:45290",
+        "issue:44623|issue:45397",
+        "issue:43441|issue:44683",
+        "issue:44623|issue:44749",
+        "issue:43653|issue:45593",
+        "issue:43262|issue:44521",
+        "issue:43723|issue:44162",
+        "issue:44016|issue:45823",
+        "issue:43673|issue:43761",
+        "issue:42175|issue:43976",
+        "issue:44361|issue:45593",
+        "issue:44671|issue:44779",
+        "issue:43784|issue:45542",
+        "issue:43901|issue:43994",
+        "issue:44987|issue:45542",
+        "issue:43122|issue:43650",
+        "issue:43824|issue:43976",
+        "issue:44297|issue:45397",
+        "issue:43525|issue:43881",
+        "issue:44188|issue:44336",
+        "issue:43525|issue:43643",
+        "issue:43450|issue:44617",
+        "issue:44589|issue:45362",
+        "issue:41628|issue:43976",
+        "issue:45561|issue:45563",
+        "issue:42491|issue:43299"
+      ],
+      "prepared_review_unit_hash": "e0bcd1c29a12ccea85da4a2944b392759e46589ba59e0ea6c2134bc0e46768e7",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 35947,
+        "estimated_input_tokens": 8987,
+        "estimated_eval_tokens": 18230
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The items are mostly separate regressions across tokenizer, chat-template, model-loading, and backend/version compatibility areas. I found no soft pair that clearly represents the same concrete bug, so I kept all soft edges rejected. For the canonical issue, I picked the most broadly representative, high-activity regression report in the set.",
+        "confidence": 0.34,
+        "canonical_issue_reason": "issue:44521 has the strongest evidence of being a central user-facing regression in this cluster: it is active, has multiple inbound references, and describes a concrete multimodal chat-template failure rather than a narrow model-specific edge case.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44521 is the best representative issue because it is concrete, broadly relevant, and has the highest apparent external impact/activity among the listed issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44466",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "Both involve v5 model weight handling, but one is serialization of tied lm_head weights and the other is resize_token_embeddings not updating Gemma4 embedding/output layers; different bugs and code paths."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "These are unrelated GLM/MoE issues: one is a missing rms_norm_eps in MLA q_a_layernorm, the other is torch.split return handling in GlmMoeDsaIndexer."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "'ADD THE DATA' is not a technical duplicate, and the assistant-message/tool-call crash is a specific chat-template parsing bug."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "One is processor.save_pretrained missing files; the other is gemma-4 zero3 from_pretrained. Different failure points and symptoms."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Both mention attention kernels, but one is Ministral-3 with FlashAttention and the other is compiled flex_attention on torch>=2.9; not the same concrete bug."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "Missing processor files and slow filtering after upgrade are unrelated issues."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration and a stale generate warning are unrelated."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Both touch apply_chat_template, but one is audio sampling-rate defaulting and the other is multimodal assistant mask generation; different code paths."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "Tokenizer loading in v5 and ESM2 model breakage are different regressions."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Notebook syntax error and Gemma4 device-mismatch forward failure are unrelated."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:43761",
+            "accept": false,
+            "reason": "Generation cache missing during chunked_prefill and CLIPVisionModel hidden_states handling are different model/runtime bugs."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "Backend install packaging and Python-version compatibility are distinct environment issues."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "MLukeTokenizer AttributeError is unrelated to the stale num_return_sequences warning."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "CamemBERT masked-LM prediction regression and Deepseek tokenizer incorrect results are separate model/tokenizer bugs."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Importing sentence-transformers fails with undefined nn, while the other issue is about TensorFlow being absent when only tensorboard is installed."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Docs mismatch for return_all_scores is not the same as SigLIP2 producing nonsensical outputs."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Loading physical-intelligence/fast on v5.1.0 and undefined tf backend error are different failures."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:43650",
+            "accept": false,
+            "reason": "Tokenizer output changes across versions are unrelated to the placeholder 'ADD THE DATA' issue."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "A missing Qwen2.5-VL export and a Python version support problem are not duplicates."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "Tokenizer class serialization mismatch and gemma-4 zero3 loading are different serialization/loading bugs."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "Missing pad_token_id on Llama4Config and glm-4v-9b loading failure are model-specific but not the same bug."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "torch.compile attention-kernel divergence and ANSI codes in loading_report are unrelated."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "Llama4Config missing pad_token_id and trust_remote_code returning missing fields are separate config-loading issues."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Video batched-shape handling and Sam3Video CUDA OOM are different problems."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "Float8 storage lookup failure and Qwen3.5-35B chat crash are unrelated runtime issues."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "AutoImageProcessor import failure and Python-version compatibility are unrelated."
+          },
+          {
+            "left": "issue:45561",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "pytest-xdist race on captured_info.txt and a stale generation warning are unrelated."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43299",
+            "accept": false,
+            "reason": "Both involve Qwen3/MoE on v5, but one is a LoRA model incompatibility and the other is a Qwen3VL MoE loading breakage; not clearly the same concrete bug."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 34,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:36246",
+        "issue:41950",
+        "issue:42757",
+        "issue:42886",
+        "issue:43065",
+        "issue:43232",
+        "issue:43257",
+        "issue:43262",
+        "issue:43352",
+        "issue:43408",
+        "issue:43526",
+        "issue:43531",
+        "issue:43596",
+        "issue:43673",
+        "issue:43901",
+        "issue:44162",
+        "issue:44220",
+        "issue:44242",
+        "issue:44246",
+        "issue:44295",
+        "issue:44336",
+        "issue:44351",
+        "issue:44373",
+        "issue:44485",
+        "issue:44496",
+        "issue:44561",
+        "issue:44749",
+        "issue:44779",
+        "issue:44857",
+        "issue:44871",
+        "issue:44877",
+        "issue:44987",
+        "issue:44995",
+        "issue:45003",
+        "issue:45137",
+        "issue:45200",
+        "issue:45229",
+        "issue:45250",
+        "issue:45292",
+        "issue:45362",
+        "issue:45375",
+        "issue:45397",
+        "issue:45459",
+        "issue:45561",
+        "issue:45584",
+        "issue:45593",
+        "issue:45701",
+        "issue:45759"
+      ],
+      "soft_pairs": [
+        "issue:43901|issue:44246",
+        "issue:44336|issue:44373",
+        "issue:44242|issue:45593",
+        "issue:44295|issue:45397",
+        "issue:41950|issue:44220",
+        "issue:44246|issue:44496",
+        "issue:43531|issue:45362",
+        "issue:36246|issue:42757",
+        "issue:36246|issue:44351",
+        "issue:43352|issue:45250",
+        "issue:43526|issue:45003",
+        "issue:44336|issue:44485",
+        "issue:45003|issue:45561",
+        "issue:43673|issue:44561",
+        "issue:44246|issue:44871",
+        "issue:44242|issue:44857",
+        "issue:44162|issue:45200",
+        "issue:44295|issue:44749",
+        "issue:36246|issue:43596",
+        "issue:44877|issue:45375",
+        "issue:44779|issue:45292",
+        "issue:45292|issue:45584",
+        "issue:43901|issue:45003",
+        "issue:45292|issue:45701",
+        "issue:43232|issue:44336",
+        "issue:44987|issue:45459",
+        "issue:42886|issue:45229",
+        "issue:43065|issue:43408",
+        "issue:44496|issue:45003",
+        "issue:45375|issue:45759",
+        "issue:43262|issue:44246",
+        "issue:43257|issue:45137",
+        "issue:44995|issue:45200"
+      ],
+      "prepared_review_unit_hash": "9f497729fa03af61a97d57eeda2a6d32c2c4b5b196b6960869e0ca77e6af43a7",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 37300,
+        "estimated_input_tokens": 9325,
+        "estimated_eval_tokens": 18906
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is very heterogeneous: most pairs are only loosely related by broad themes like import failures, config loading, generation, or loss computation, but they do not look like the same underlying bug. I would not merge any of the soft edges as duplicates.",
+        "confidence": 0.82,
+        "canonical_issue_reason": "issue:44246 is the most central node in the cluster by soft-link connectivity, so it is the best single anchor if one issue must represent the cluster, even though the cluster itself is not a true duplicate set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44246 has the strongest cross-links and substantial discussion activity, making it the best representative issue for triage.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43901",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "Docs wording about return_all_scores and an intermittent import slowdown are unrelated problems."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "ANSI output in loading_report and a wrong docstring for position_ids are not the same bug."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Both concern loss terms, but they affect different models and different code paths."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "A position_ids buffer access error is distinct from a Gemma-4 ZeRO3 from_pretrained loading failure."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Video-classification processor lookup and _torch_extract_fbank_features() are unrelated audio/video issues."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Import performance and unrecognized model config loading are different failure modes."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "Both mention Qwen, but one is a sliding_window bug and the other is a chat/runtime crash; not the same issue."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "These are different missing-import errors for different symbols."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but they concern different missing exports and different root causes."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:45250",
+            "accept": false,
+            "reason": "A specific model not supporting Flash Attention 2.0 is not the same as a generic FA2 issue title."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "A BeitImageProcessorFast label reduction bug is unrelated to unsafe sys.modules access."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "ANSI logging artifacts and GLM-5 RoPE implementation are unrelated."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "sys.modules access and a pytest-xdist file race are separate concurrency/import problems."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "Chunked-prefill cache handling and removal of is_torch_fx_available affect different code paths."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Import slowdown and Gemma eos_token_id inconsistency are unrelated."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Missing load-balancing loss and an AMP crash in LwDetrImageLoss are different bugs."
+          },
+          {
+            "left": "issue:44162",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "ESM2 breakage and Gemma 4 token-type defaults are unrelated model-specific issues."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "A buffer read error and a tokenizer slowdown after upgrade are not the same underlying bug."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:43596",
+            "accept": false,
+            "reason": "An import-name failure and a DeepSpeed zero3 IndexError are unrelated."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both involve config strictness, but they are different models and likely different missing fields."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:45292",
+            "accept": false,
+            "reason": "Tokenizer correctness regression and resize_token_embeddings not syncing output embeddings are different issues."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Embedding resize behavior and Whisper empty-transcription generation are unrelated."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "A docs mismatch and unsafe sys.modules access do not share the same bug."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Tokenization changes across versions and embedding resize/output mismatch are not the same defect."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "Generation kwargs handling and ANSI codes in loading_report are unrelated."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "A model-loading regression and tokenizer exception masking are different tokenizer/loading problems."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Offline cache loading and multi-GPU CUDA OOM are unrelated."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:43408",
+            "accept": false,
+            "reason": "Dummy Conv2d in Sam3PixelDecoder and sam3_video vs sam3_tracker config mismatch are not the same bug."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Unrecognized model config and unsafe sys.modules access are different loading failures."
+          },
+          {
+            "left": "issue:45375",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Missing a strict config field and failing to unwrap text_config are different composite-model config bugs."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate defaults and import slowdown are unrelated."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Qwen3 MoE weight conversion and DeepSpeed deque underflow are different distributed-training failures."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "A stale indexer cache crash and a missing mm_token_type_ids default are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 35,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:28282",
+        "issue:30990",
+        "issue:41628",
+        "issue:42371",
+        "issue:42994",
+        "issue:43232",
+        "issue:43475",
+        "issue:43482",
+        "issue:43525",
+        "issue:43540",
+        "issue:43643",
+        "issue:43646",
+        "issue:43653",
+        "issue:43673",
+        "issue:43931",
+        "issue:44038",
+        "issue:44060",
+        "issue:44077",
+        "issue:44162",
+        "issue:44297",
+        "issue:44361",
+        "issue:44368",
+        "issue:44466",
+        "issue:44623",
+        "issue:44661",
+        "issue:44734",
+        "issue:44749",
+        "issue:44869",
+        "issue:44933",
+        "issue:44977",
+        "issue:44995",
+        "issue:45081",
+        "issue:45092",
+        "issue:45103",
+        "issue:45137",
+        "issue:45230",
+        "issue:45278",
+        "issue:45310",
+        "issue:45335",
+        "issue:45356",
+        "issue:45381",
+        "issue:45397",
+        "issue:45459",
+        "issue:45563",
+        "issue:45588",
+        "issue:45600",
+        "issue:45656"
+      ],
+      "soft_pairs": [
+        "issue:44060|issue:44077",
+        "issue:43475|issue:43540",
+        "issue:43653|issue:44162",
+        "issue:44466|issue:45335",
+        "issue:45397|issue:45656",
+        "issue:44368|issue:45563",
+        "issue:44995|issue:45588",
+        "issue:44368|issue:44977",
+        "issue:43673|issue:44734",
+        "issue:45137|issue:45600",
+        "issue:43931|issue:44368",
+        "issue:44933|issue:45459",
+        "issue:44869|issue:45356",
+        "issue:43232|issue:43673",
+        "issue:44297|issue:44623",
+        "issue:43673|issue:45278",
+        "issue:43646|issue:44162",
+        "issue:44361|issue:44749",
+        "issue:44623|issue:45656",
+        "issue:44749|issue:45656",
+        "issue:44749|issue:45310",
+        "issue:42994|issue:43482",
+        "issue:28282|issue:30990",
+        "issue:45230|issue:45397",
+        "issue:42371|issue:45092",
+        "issue:43475|issue:43643",
+        "issue:44038|issue:44623",
+        "issue:43475|issue:45381",
+        "issue:41628|issue:44933",
+        "issue:45081|issue:45103",
+        "issue:43475|issue:44661",
+        "issue:42994|issue:43525"
+      ],
+      "prepared_review_unit_hash": "91ced0ba32dab54dee0754efe7f9f22fbdf26360ed0c5cf94c178b48cb2352ad",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36715,
+        "estimated_input_tokens": 9179,
+        "estimated_eval_tokens": 18614
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are mostly unrelated issue reports, with a loose common thread of Transformers v5 regressions and import/model-loading breakages. None of the soft-edge pairs look like the exact same underlying bug or change, so I would not merge any of them.",
+        "confidence": 0.38,
+        "canonical_issue_reason": "issue:45278 is the broadest and most cluster-representative report: it covers widespread import breakages after upgrading from 4.57.0 to 5.5.0, which matches the main theme better than the model-specific or symptom-specific issues.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the best global issue candidate because it is broad, current, and describes a user-facing regression class that can subsume several of the other v5-related reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44060",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "Both mention warnings, but they concern different models and different code paths; no evidence of the same bug."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Both are video/model issues, but they affect different architectures with distinct failures and are not plausibly the same fix."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "Different models and different failure modes; too broad a similarity to treat as one underlying bug."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "Both involve embeddings/tied weights, but one is serialization and the other is resize propagation; not the same concrete defect."
+          },
+          {
+            "left": "issue:45397",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "These are unrelated: one is a from_pretrained/ZeRO3 loading problem, the other is a deepspeed optimizer-step issue."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Only the presence of a warning is shared; the model, API, and bug mechanism are different."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Different subsystems and symptoms: stale cache in one model versus flash-attention crash in another."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Both involve Qwen3.5 and generation/fine-tuning, but the reported failures are different and not the same code-path problem."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Both are cache-related generation issues, but one is missing cache in chunked prefill and the other is wrong indexing in serve responses; not one bug."
+          },
+          {
+            "left": "issue:45137",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Completely different topics: DeepSpeed ZeRO3 deque error versus removed auto-mapping config references."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Different models and unrelated behaviors; no shared underlying defect is evident."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "Both touch imports, but one is a missing public import and the other is tokenizer error handling; not the same issue."
+          },
+          {
+            "left": "issue:44869",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Tokenizer/decoding regressions, but for different models and different failure conditions."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "Both relate to generation internals, but one is sync_gpus kwarg handling and the other is cache availability; too different to merge."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Both are save_pretrained bugs, but one is tokenizer metadata and the other is processor file output; not the same concrete defect."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "One is a generation cache bug, the other is a broad import-regression report; they do not describe the same problem."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "Custom model initialization breakage versus a model-family workflow regression; too generic to be the same bug."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "An AttributeError in a tokenizer versus a slowdown after upgrade; no common underlying code-path."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Save-pretrained file emission and optimizer stepping are unrelated."
+          },
+          {
+            "left": "issue:44749",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Performance regression versus Qwen3.5-Moe from_pretrained error; different failures."
+          },
+          {
+            "left": "issue:44749",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Slow filtering after upgrade and duplicate optimizer stepping are unrelated."
+          },
+          {
+            "left": "issue:44749",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "No shared underlying bug: one is a runtime slowdown, the other is a model-loading failure."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Both involve model formats and loading/saving, but quantized model saving and GGUF loading under v5 are distinct issues."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:30990",
+            "accept": false,
+            "reason": "A missing PyTorch dependency error and a sentence-transformers loading hang are different environment/load problems."
+          },
+          {
+            "left": "issue:45230",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "One is an unspecified bug report and the other is a specific ZeRO3/from_pretrained failure; not enough to connect them."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "TF32 API guidance and remote-code/meta-init incompatibility are unrelated."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "SAM3 video output fields and AutoConfig missing fields with trust_remote_code=True are different bug classes."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe v5 breakage and processor.save_pretrained missing files are not the same problem."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both concern vision/video models, but they affect different models and different internal fields/logic."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Both are import-related, but they refer to different missing symbols and likely different export paths."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Tokenizer regex patching and auto-docstring annotation handling are unrelated."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "A SAM3 vision encoder attribute error and add-new-model-like tokenizer mapping failure are not the same defect."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "No shared underlying code-path; one is a model-specific v5 bug and the other is serialization output missing files."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 36,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:30990",
+        "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:43479",
+        "issue:43504",
+        "issue:43526",
+        "issue:43673",
+        "issue:43704",
+        "issue:43784",
+        "issue:43824",
+        "issue:43873",
+        "issue:43976",
+        "issue:44038",
+        "issue:44079",
+        "issue:44164",
+        "issue:44188",
+        "issue:44222",
+        "issue:44242",
+        "issue:44351",
+        "issue:44373",
+        "issue:44496",
+        "issue:44568",
+        "issue:44623",
+        "issue:44792",
+        "issue:44869",
+        "issue:44933",
+        "issue:44945",
+        "issue:44964",
+        "issue:45137",
+        "issue:45200",
+        "issue:45310",
+        "issue:45375",
+        "issue:45478",
+        "issue:45561",
+        "issue:45712",
+        "issue:45753",
+        "issue:45812"
+      ],
+      "soft_pairs": [
+        "issue:45375|issue:45478",
+        "issue:43232|issue:44188",
+        "issue:42673|issue:43704",
+        "issue:45310|issue:45375",
+        "issue:43873|issue:44945",
+        "issue:41628|issue:44351",
+        "issue:42831|issue:44222",
+        "issue:44945|issue:45753",
+        "issue:44222|issue:45310",
+        "issue:42994|issue:44164",
+        "issue:36010|issue:41628",
+        "issue:43673|issue:44188",
+        "issue:42175|issue:45712",
+        "issue:42222|issue:45200",
+        "issue:43232|issue:44792",
+        "issue:44038|issue:45310",
+        "issue:43377|issue:44242",
+        "issue:43526|issue:45561",
+        "issue:43504|issue:45561",
+        "issue:44496|issue:45561",
+        "issue:44964|issue:45137",
+        "issue:42175|issue:44933",
+        "issue:42898|issue:44568",
+        "issue:30990|issue:35141",
+        "issue:43479|issue:44869",
+        "issue:43976|issue:44933",
+        "issue:43479|issue:44079",
+        "issue:44188|issue:44373",
+        "issue:43784|issue:43824",
+        "issue:44623|issue:45310",
+        "issue:42898|issue:45812",
+        "issue:43278|issue:45137"
+      ],
+      "prepared_review_unit_hash": "5588d8a3cea2150ffa6ae5878845ffefc83403c50e8e48eb316d0e182466d40e",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36428,
+        "estimated_input_tokens": 9107,
+        "estimated_eval_tokens": 18470
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly a set of unrelated transformer bug reports, with one clear duplicate pair around the Qwen3ForCausalLM VRAM leak. I would canonicalize that pair; the rest should stay separate.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "issue:42673 is the stronger canonical: it is the earlier report, has much more discussion/references, and matches issue:43704 exactly on the same Qwen3ForCausalLM multi-dataloader-thread VRAM leak.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:42673 is the best representative issue because it is the clearest true duplicate target and the most complete/active report among the accepted pair.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45375",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Same model family, but these describe different failures: missing config field being dropped vs a from_pretrained error. Not the same underlying bug."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44188",
+            "accept": false,
+            "reason": "Both are generation-related, but one is about cache updates after sync_gpus and the other is about attention-kernel divergence under torch.compile. Different code paths."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:43704",
+            "accept": true,
+            "reason": "Exact same title and symptom: Qwen3ForCausalLM VRAM leak when used in multiple dataloader threads. Clear duplicate."
+          },
+          {
+            "left": "issue:45310",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both involve Qwen3.5/Moe configs, but one is a from_pretrained error while the other is a specific missing field silently dropped by strict config handling. Not the same bug."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Offloading/quantization behavior and pipeline-parallel output correctness are different problems."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and likely different missing exports. Too broad to treat as one bug."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "FineGrainedFP8 accuracy issues and FP8 save_pretrained failures are distinct symptoms and code paths."
+          },
+          {
+            "left": "issue:44945",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Pipeline-parallel output corruption and a Qwen3_5 infinite-loop-on-specific-image report do not look like the same defect."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "One is about saving FP8 MoE checkpoints; the other is a from_pretrained load error. Related area, different bug."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:44164",
+            "accept": false,
+            "reason": "Both involve saving/loading, but quantized model saving and extra_state handling are not the same concrete failure."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "Different missing imports (GenerationMixin vs AutoImageProcessor) and different breakages."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44188",
+            "accept": false,
+            "reason": "Chunked-prefill cache behavior and torch.compile branching in attention kernels are separate generation issues."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Packaging backend inclusion and leftover dummy class cleanup are unrelated."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "VitPose model breakage and Gemma 4 mm_token_type_ids behavior are unrelated model-specific bugs."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44792",
+            "accept": false,
+            "reason": "Generation kwargs handling after sync_gpus is unrelated to the janus test failure."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Both touch Qwen3.5/Qwen3-VL-Moe, but they appear to be different version/model load failures rather than the same concrete bug."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44242",
+            "accept": false,
+            "reason": "Padding-mask batching bug in MIMI encoder and missing load-balancing loss when output_router_logits=False are different issues."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction and xdist race on captured_info.txt are unrelated."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Beit legacy-field loading failure and the pytest-xdist race are unrelated."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Model auto-loading failure and a test-utility race condition are unrelated."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Phi-4 multimodal loading and DeepSpeed ZeRO3 deque underflow are unrelated."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Package-install backend composition and missing image_utils import are unrelated."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both are v5 tokenizer behavior regressions, but one is cleanup spacing and the other is add_special_tokens/BOS-EOS behavior. Different bugs."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Sentence Transformers loading hang and embedding reinitialization after resize are unrelated."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44869",
+            "accept": false,
+            "reason": "Phi4Multimodal config defaults and Whisper timestamp decode crash are unrelated."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Python version compatibility and a missing image_utils import are unrelated."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Both mention None-handling, but one is multimodal config initialization and the other is ModelOutput key assignment. Not the same defect."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "Attention-kernel divergence and a wrong docstring are unrelated."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import failures with latest transformers, but for different packages/symbols and different root causes."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Processor save_pretrained missing files and Qwen3.5 from_pretrained error are different save/load problems."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Tokenizer cleanup behavior and Granite token ID regression are different tokenizer bugs."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Embedding dtype drift between train/eval and ZeRO3 deque underflow are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: 42673 and 43704 have identical titles, and 42673 is the earlier/more active report. The soft-edge verdicts are conservative and do not overclaim duplicate status for clearly different issues."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 37,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:22355",
+        "issue:28282",
+        "issue:30333",
+        "issue:33290",
+        "issue:34689",
+        "issue:35141",
+        "issue:36331",
+        "issue:41628",
+        "issue:42757",
+        "issue:42831",
+        "issue:42898",
+        "issue:43116",
+        "issue:43232",
+        "issue:43479",
+        "issue:43644",
+        "issue:43723",
+        "issue:43874",
+        "issue:44060",
+        "issue:44190",
+        "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:44568",
+        "issue:44623",
+        "issue:44734",
+        "issue:44779",
+        "issue:44933",
+        "issue:44977",
+        "issue:45092",
+        "issue:45103",
+        "issue:45125",
+        "issue:45310",
+        "issue:45356",
+        "issue:45405",
+        "issue:45406",
+        "issue:45656",
+        "issue:45712",
+        "issue:45834"
+      ],
+      "soft_pairs": [
+        "issue:44734|issue:45103",
+        "issue:30333|issue:33290",
+        "issue:43723|issue:44568",
+        "issue:42898|issue:45356",
+        "issue:43874|issue:45092",
+        "issue:44060|issue:45125",
+        "issue:42898|issue:44779",
+        "issue:45405|issue:45712",
+        "issue:42898|issue:44448",
+        "issue:44295|issue:45656",
+        "issue:44933|issue:45712",
+        "issue:44222|issue:44297",
+        "issue:44297|issue:45656",
+        "issue:43116|issue:44190",
+        "issue:43479|issue:43644",
+        "issue:44222|issue:44623",
+        "issue:44206|issue:45103",
+        "issue:42831|issue:44623",
+        "issue:41628|issue:42757",
+        "issue:22355|issue:28282",
+        "issue:30333|issue:36331",
+        "issue:44297|issue:45310",
+        "issue:30333|issue:34689",
+        "issue:44222|issue:44484",
+        "issue:43232|issue:45834",
+        "issue:43723|issue:44977",
+        "issue:28282|issue:30333",
+        "issue:44263|issue:45092",
+        "issue:30333|issue:35141",
+        "issue:44483|issue:45406",
+        "issue:44263|issue:44360",
+        "issue:44373|issue:44485",
+        "issue:44295|issue:44623",
+        "issue:44514|issue:45103",
+        "issue:44261|issue:45092"
+      ],
+      "prepared_review_unit_hash": "69471f0b9929c6b273ae978d4f365705981e2922d2501e19c3c4aa266b04b124",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 35,
+        "serialized_chars": 38112,
+        "estimated_input_tokens": 9528,
+        "estimated_eval_tokens": 19312
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This is a heterogeneous cluster of Transformer issues with only loose topical similarity. None of the soft pairs look like the same underlying bug/change strongly enough to merge, so all soft edges should be rejected.",
+        "confidence": 0.84,
+        "canonical_issue_reason": "Issue 42898 is the clearest high-level regression report in the set: it describes a broad v5 tokenizer behavior change with a well-scoped symptom and likely captures the most representative regression theme among the mostly unrelated items.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "42898 is the best representative issue because it is explicit, general, and regression-focused, whereas most other issues are narrow model- or component-specific failures.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44734",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Unrelated crashes: serving KV-cache indexing vs. auto-docstring annotation handling."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "Different problems: MLflow status reporting vs. deepspeed/Adafactor OOM."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but one is a generic AutoTokenizer load failure and the other is a specific BOS/EOS behavior bug."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Different tokenizer regressions with different models and different root causes."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Image patch counting API bug vs. remote-code checkpoint/meta-init incompatibility."
+          },
+          {
+            "left": "issue:44060",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "Different Qwen3 issues: incorrect tie warning vs. missing tensor-parallel plan."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both involve tokenizers in v5, but the model-specific symptoms and likely code paths differ."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Unrelated: dependency version pinning vs. dummy class cleanup in import surfaces."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44448",
+            "accept": false,
+            "reason": "Shared v4/v5 regression theme, but the reported behavior and affected model differ."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Buffer access after registration is unrelated to a double optimizer-step bug."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Import-from-image_utils failure is not the same as leftover dummy classes leaking into dir()."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "FP8 save_pretrained behavior and tokenizer_config metadata mismatch are different bugs."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Tokenizer save/load metadata bug vs. Qwen3.5 MoE from_pretrained failure."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Different example-script failures with different input pipelines and error modes."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:43644",
+            "accept": false,
+            "reason": "Config default initialization bug is unrelated to non-persistent buffer corruption."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Both mention save_pretrained, but one is FP8/MoE and the other is processor file serialization."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Feature-extractor argument regression is unrelated to auto-docstring annotation crashes."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "FP8 accuracy issue is not the same as missing processor save_pretrained files."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Two separate import errors from different packages and causes."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Missing transformers.onnx module is unrelated to the absent PyTorch dependency error."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "MLflow reporting vs. trainer compute_loss signature mismatch are different regressions."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Tokenizer config mismatch is not the same as the Qwen3.5 MoE from_pretrained regression."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Training-job reporting bug and Llama model-loading regression are unrelated."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Save_pretrained shard-size question is unrelated to the FP8 MoE save bug."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Generation kwargs handling and OCR long-input indexing are different code paths."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but the affected models and symptoms are different."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:30333",
+            "accept": false,
+            "reason": "ImportError about PyTorch is unrelated to MLflow reporting after failed training jobs."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "DSA indexer split semantics are unrelated to multimodal checkpoint/meta-init compatibility."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Training-job reporting bug is unrelated to embedding resize reinitialization."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Chat-completions rejection and missing Gemma4Processor attribute are different serve failures."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Same subsystem, but different bugs: split return handling vs. missing ReLU."
+          },
+          {
+            "left": "issue:44373",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Wrong docstring for position_ids is not the same as the GLM-5 RoPE implementation issue."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Reading a registered buffer and processor serialization are unrelated."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Batched chat-template crash and auto-docstring annotation crash have no shared code path."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "MLA config epsilon precision bug is unrelated to the DSA indexer ReLU issue."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 38,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:22355",
+        "issue:28218",
+        "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:43976",
+        "issue:44038",
+        "issue:44060",
+        "issue:44117",
+        "issue:44263",
+        "issue:44295",
+        "issue:44297",
+        "issue:44373",
+        "issue:44442",
+        "issue:44488",
+        "issue:44514",
+        "issue:44521",
+        "issue:44568",
+        "issue:44704",
+        "issue:44743",
+        "issue:44871",
+        "issue:44913",
+        "issue:44945",
+        "issue:44991",
+        "issue:45092",
+        "issue:45278",
+        "issue:45290",
+        "issue:45310",
+        "issue:45357",
+        "issue:45593",
+        "issue:45753"
+      ],
+      "soft_pairs": [
+        "issue:45092|issue:45357",
+        "issue:43723|issue:44871",
+        "issue:30990|issue:38175",
+        "issue:43723|issue:44991",
+        "issue:43723|issue:44442",
+        "issue:41628|issue:43824",
+        "issue:43479|issue:44913",
+        "issue:42831|issue:45310",
+        "issue:44263|issue:44373",
+        "issue:39401|issue:44568",
+        "issue:42831|issue:44038",
+        "issue:42757|issue:45278",
+        "issue:42886|issue:42994",
+        "issue:43232|issue:45593",
+        "issue:43756|issue:44060",
+        "issue:43576|issue:43784",
+        "issue:42831|issue:45753",
+        "issue:42886|issue:43066",
+        "issue:43122|issue:44521",
+        "issue:42371|issue:44263",
+        "issue:22355|issue:28218",
+        "issue:44117|issue:44704",
+        "issue:30990|issue:33357",
+        "issue:43232|issue:44743",
+        "issue:44295|issue:44297",
+        "issue:44704|issue:45290",
+        "issue:42907|issue:44945",
+        "issue:43122|issue:44514",
+        "issue:41628|issue:45278",
+        "issue:43976|issue:44945",
+        "issue:39401|issue:44488"
+      ],
+      "prepared_review_unit_hash": "6f701ecddad96af777ba9487b398f64bd2cb5f021cecfbafd88f7ff35bf50652",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 36563,
+        "estimated_input_tokens": 9141,
+        "estimated_eval_tokens": 18538
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are largely heterogeneous regressions across tokenizer loading, model-specific multimodal behavior, config reloads, and unrelated import/runtime errors. I do not see any soft pair that clearly describes the same concrete bug, so all candidate duplicate edges should be rejected.",
+        "confidence": 0.78,
+        "canonical_issue_reason": "Issue 43723 is the broadest representative of the tokenizer-loading regression family in this set (AutoTokenizer.from_pretrained in v5). It is the best single issue to anchor the cluster, though the surrounding items still look mostly unrelated rather than true duplicates.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43723 is the most general and user-facing issue here, with the clearest umbrella scope over tokenizer-loading failures. It is the best representative issue, even if the cluster does not collapse into a single duplicate group.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45092",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different model-specific bugs: meta-initialization incompatibility vs incorrect visual encoder keys on save_pretrained."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "One is a tokenizer loading failure in v5; the other is an eos_token_id config mismatch for Gemma-3."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:38175",
+            "accept": false,
+            "reason": "Different symptoms and models: Sentence Transformers loading stall vs SigLIP2 zero-probability outputs."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both involve tokenizer loading, but the concrete failures and affected models differ; not clearly the same bug."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "Broad AutoTokenizer loading regression vs a FastSpeech2ConformerTokenizer-specific load failure."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but for different missing symbols/classes and different code paths."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Unrelated config bugs: default multimodal sub-config initialization vs GPTNeoX rotary_pct reload behavior."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Different problems: FineGrainedFP8 accuracy degradation vs Qwen3.5-Moe from_pretrained error."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "A runtime/indexer issue is not the same as a docstring correction."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Different tokenizer bugs: wrong offset_mapping vs missing BOS/EOS insertion with add_special_tokens."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44038",
+            "accept": false,
+            "reason": "Different model families and failure modes; one is FP8 accuracy, the other is a Qwen3-VL-Moe loading bug."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "One is a specific is_offline_mode import failure; the other is a broader collection of import errors after upgrade."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:42994",
+            "accept": false,
+            "reason": "Offline tokenizer cache loading and quantized model saving are unrelated subsystems."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Generation kwargs handling after sync_gpus is unrelated to D-FINE auxiliary-loss behavior."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44060",
+            "accept": false,
+            "reason": "Different model internals: RoPE-layer dropping vs a tied-weights warning in Qwen3-Next."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:43784",
+            "accept": false,
+            "reason": "transformers env CLI breakage is not the same bug as sentence-transformers failing with nn undefined."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Accuracy regression in FineGrainedFP8 is unrelated to a Qwen3_5 image-specific infinite loop."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:43066",
+            "accept": false,
+            "reason": "Offline cache loading and wrong tokenizer decoder type in v5 are different issues."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Both touch tokenization/chat templating, but one is a version-to-version tokenization change and the other is a multimodal assistant_mask bug."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "TF32 settings guidance is unrelated to a torch.split return-value bug."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:28218",
+            "accept": false,
+            "reason": "Missing transformers.onnx module and an extra-space tokenizer behavior are different bugs."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "TOKENIZER_MAPPING_NAMES returning None is a different loader issue than AutoProcessor forwarding kwargs to cached_file."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "A loading hang in Sentence Transformers is not the same as a Mac bus error with a CLIP model."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Generation sync_gpus state handling and Qwen3_5 recurrent-state reset are different code paths."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "Reading a buffer after registration and saving the wrong tokenizer_class are unrelated."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "AutoProcessor cached_file kwarg forwarding is unrelated to chat-template crashes on tool-call assistant messages."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Quantized-model saving failure and pipeline-parallelism output corruption are different bugs."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "Tokenization differences across versions are not the same as a batched apply_chat_template crash with padding=False."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "A specific AutoImageProcessor import failure is not the same as a broad multi-import regression report."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Python version compatibility is unrelated to pipeline-parallelism output errors."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Wrong offset_mapping in Qwen3 tokenizer is unrelated to loading a specific model/tokenizer checkpoint."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 39,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:22355",
+        "issue:29127",
+        "issue:30990",
+        "issue:36296",
+        "issue:38617",
+        "issue:39401",
+        "issue:42371",
+        "issue:42673",
+        "issue:42757",
+        "issue:42907",
+        "issue:43066",
+        "issue:43116",
+        "issue:43122",
+        "issue:43278",
+        "issue:43408",
+        "issue:43452",
+        "issue:43475",
+        "issue:43493",
+        "issue:43784",
+        "issue:43824",
+        "issue:43874",
+        "issue:43937",
+        "issue:43992",
+        "issue:44038",
+        "issue:44222",
+        "issue:44261",
+        "issue:44295",
+        "issue:44297",
+        "issue:44351",
+        "issue:44360",
+        "issue:44373",
+        "issue:44485",
+        "issue:44521",
+        "issue:44625",
+        "issue:44704",
+        "issue:44779",
+        "issue:44829",
+        "issue:44945",
+        "issue:44964",
+        "issue:44977",
+        "issue:45072",
+        "issue:45200",
+        "issue:45216",
+        "issue:45278",
+        "issue:45446",
+        "issue:45478",
+        "issue:45600",
+        "issue:45753"
+      ],
+      "soft_pairs": [
+        "issue:43784|issue:45600",
+        "issue:43408|issue:43475",
+        "issue:42907|issue:43937",
+        "issue:44261|issue:44373",
+        "issue:44360|issue:44485",
+        "issue:44625|issue:44829",
+        "issue:39401|issue:44521",
+        "issue:43116|issue:45200",
+        "issue:29127|issue:30990",
+        "issue:44038|issue:44297",
+        "issue:42907|issue:44964",
+        "issue:43122|issue:44779",
+        "issue:44038|issue:44295",
+        "issue:36296|issue:43116",
+        "issue:42907|issue:43452",
+        "issue:43824|issue:44351",
+        "issue:44038|issue:44222",
+        "issue:45216|issue:45478",
+        "issue:42757|issue:44351",
+        "issue:43278|issue:44945",
+        "issue:43493|issue:45753",
+        "issue:42907|issue:44222",
+        "issue:43122|issue:44977",
+        "issue:43992|issue:44704",
+        "issue:43784|issue:45446",
+        "issue:43066|issue:44779",
+        "issue:38617|issue:45278",
+        "issue:42673|issue:45072",
+        "issue:43937|issue:44485",
+        "issue:22355|issue:29127",
+        "issue:42371|issue:43874"
+      ],
+      "prepared_review_unit_hash": "bc29e2105b6271bb31659554d6fcbdded8ac98a1a459c05409eae5b6c0f82235",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 36602,
+        "estimated_input_tokens": 9151,
+        "estimated_eval_tokens": 18558
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly topical rather than duplicate-heavy: many items share broad themes like import regressions, tokenizer issues, or model-specific bugs, but they do not look like the same underlying defect. No soft edge is strong enough to accept.",
+        "confidence": 0.46,
+        "canonical_issue_reason": "issue:45278 is the best umbrella representative because it captures the broad import-regression theme after the 4.57.x -> 5.5.0 upgrade, which is the closest thing to a shared cluster topic here.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the most globally suitable issue because it is broad, central, and representative of the recurring import-failure reports in this set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43784",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Both are import-related, but one is a NameError in sentence-transformers and the other is a stale auto_mappings reference in transformers CI; different code paths and bugs."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Same SAM 3 area, but one is a model-type warning and the other is a missing attribute on vision output; not the same defect."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43937",
+            "accept": false,
+            "reason": "Dequantized save failure for Ministrals/Devstrals vs invalid GenerationConfig for GLM-5; unrelated model paths."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "One is a numerical/config precision issue in MLA q_a_layernorm, the other is a docstring mismatch for position_ids; not duplicates."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "DSA indexer ReLU complaint and GLM-5 RoPE implementation are unrelated subsystems and symptoms."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Qwen3.5 num_labels propagation and flash_attention_3 degenerate training are different bugs with different code paths."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Both concern tokenization/multimodal behavior, but one is wrong offset_mapping for Qwen3 and the other is all-zero assistant masks in apply_chat_template; not the same issue."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Example script multi-label output bug vs Gemma 4 mm_token_type_ids defaulting; different components and failure modes."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:30990",
+            "accept": false,
+            "reason": "LayoutLMv3 error-message clarity and Sentence Transformers loading hang are unrelated."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe compatibility bug vs tokenizer.save_pretrained class-name mismatch; not the same concrete problem."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Dequantized save failure and Phi-4-multimodal load failure are distinct model-management bugs."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both are tokenizer regressions in v5, but one is a generic version-to-version tokenization change and the other is Deepseek-specific incorrect output; not clearly the same bug."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44295",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe error and position_ids buffer-read error are unrelated."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43116",
+            "accept": false,
+            "reason": "Tensor parallel training bug and multi-label classification example bug are different training/example paths."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "Saving dequantized Ministrals/Devstrals and gguf_file breaking AutoTokenizer/AutoModel loading are separate loading/saving issues."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Cannot import Qwen2_5_VLForConditionalGeneration and cannot import HybridCache are both import errors, but they stem from different symbols and fixes."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe compatibility and FP8 save_pretrained moe are different model serialization/runtime bugs."
+          },
+          {
+            "left": "issue:45216",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both involve Qwen3.5 and Transformers 5.4+, but one is save_pretrained checkpoint correctness and the other is from_pretrained loading failure; different code paths."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Missing is_offline_mode import from huggingface_hub versus missing HybridCache export from transformers; both imports fail, but not the same missing symbol or fix."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Embedding dtype drift between train/eval and incorrect LLM output under pipeline parallelism are unrelated issues."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "SigLIP2 implementation discrepancy and Qwen3_5 infinite loop on a specific image are different model behaviors."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Dequantized save failure and FP8 moe save_pretrained bug are separate serialization problems."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Generic tokenizer regression across versions vs Qwen3.5 flash-attention generation bug; not the same underlying defect."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "UMT5Encoder missing embed_tokens.weight and AutoProcessor not passing kwargs to cached_file are unrelated loader issues."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Sentence-transformers import NameError and flex_attention PyTorch version check are entirely different failures."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Tokenizer decoder type migration issue in Transformers v5 and Deepseek tokenization regression are different tokenizer bugs."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "A specific missing import in configuration_utils is not the same as the broad multi-import regression report, even if both are import-related."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Qwen3ForCausalLM VRAM leak in dataloader threads and dtype mismatches in bfloat16 inference are different runtime bugs."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Invalid GenerationConfig for GLM-5 and GLM-5 RoPE implementation are separate model configuration vs positional encoding issues."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "Transformers.onnx import failure and LayoutLMv3 error-message clarity do not describe the same bug."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:43874",
+            "accept": false,
+            "reason": "TF32 behavior API guidance and GLM46V image patch-count AttributeError are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 40,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:33453",
+        "issue:38617",
+        "issue:39401",
+        "issue:41628",
+        "issue:42222",
+        "issue:42371",
+        "issue:42617",
+        "issue:42673",
+        "issue:43064",
+        "issue:43116",
+        "issue:43408",
+        "issue:43421",
+        "issue:43441",
+        "issue:43450",
+        "issue:43493",
+        "issue:43582",
+        "issue:43673",
+        "issue:43742",
+        "issue:43825",
+        "issue:43827",
+        "issue:43906",
+        "issue:43937",
+        "issue:43976",
+        "issue:43986",
+        "issue:44038",
+        "issue:44075",
+        "issue:44117",
+        "issue:44261",
+        "issue:44291",
+        "issue:44410",
+        "issue:44484",
+        "issue:44779",
+        "issue:44829",
+        "issue:44936",
+        "issue:44977",
+        "issue:45071",
+        "issue:45081",
+        "issue:45083",
+        "issue:45216",
+        "issue:45310",
+        "issue:45341",
+        "issue:45357",
+        "issue:45405",
+        "issue:45446",
+        "issue:45478",
+        "issue:45561",
+        "issue:45563",
+        "issue:45588"
+      ],
+      "soft_pairs": [
+        "issue:43064|issue:43673",
+        "issue:43976|issue:45446",
+        "issue:43116|issue:43493",
+        "issue:42222|issue:43441",
+        "issue:44075|issue:44936",
+        "issue:43116|issue:43825",
+        "issue:45357|issue:45478",
+        "issue:44038|issue:45478",
+        "issue:45083|issue:45446",
+        "issue:43937|issue:45563",
+        "issue:44117|issue:45588",
+        "issue:33453|issue:39401",
+        "issue:39401|issue:44779",
+        "issue:43827|issue:45446",
+        "issue:43937|issue:44977",
+        "issue:42673|issue:43906",
+        "issue:44291|issue:45071",
+        "issue:43450|issue:43986",
+        "issue:45405|issue:45446",
+        "issue:45341|issue:45561",
+        "issue:44117|issue:45081",
+        "issue:38617|issue:41628",
+        "issue:39401|issue:43421",
+        "issue:44410|issue:45310",
+        "issue:44484|issue:45310",
+        "issue:44829|issue:45216",
+        "issue:42371|issue:44261",
+        "issue:42673|issue:43582",
+        "issue:43408|issue:43742",
+        "issue:42617|issue:45341"
+      ],
+      "prepared_review_unit_hash": "deb290f9e8456531e99e9add2da481f0d1505dd510f9d09959d1ceb9e1d9454a",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36560,
+        "estimated_input_tokens": 9140,
+        "estimated_eval_tokens": 18536
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is centered on Qwen3.5 loading/saving regressions, but it is not a single duplicate pair. The only strong overlap is the exact-title pair issue:45310 and issue:45478; issue:45357 is a separate save_pretrained visual-encoder-key regression and should not be merged with the from_pretrained failure.",
+        "confidence": 0.79,
+        "canonical_issue_reason": "issue:45478 is a good canonical representative because it is the latest report with the exact Qwen3.5 MoE from_pretrained failure title; however, issue:45310 is an equally important duplicate-like report with the same title, so 45478 should be treated as representative rather than uniquely authoritative.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45478 best anchors the cluster because it most directly names the concrete failing code path and is the clearest representative of the repeated Qwen3.5 MoE from_pretrained reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45357",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Different regressions: 45357 is about incorrect save_pretrained visual encoder keys, while 45478 is a from_pretrained loading error. Same model family, but not the same bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The writeup overstates the evidence in the packet. The strongest counterexample is issue:45310, which has the exact same title as issue:45478, so 45478 is not the sole or clearly latest duplicate-like report. Also, the accepted pair issue:45357 \u2194 issue:45478 is not well grounded: 45357 is about incorrect `save_pretrained` visual encoder keys, not the same `from_pretrained` failure described in 45478. Overall the cluster may still center on Qwen3.5 loading/saving regressions, but the summary is too strong in calling it a single duplicate pair and in treating the non-matching Qwen3.5 issue as the same bug."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 41,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:38617",
+        "issue:41628",
+        "issue:42175",
+        "issue:42222",
+        "issue:42371",
+        "issue:42490",
+        "issue:42673",
+        "issue:42831",
+        "issue:43012",
+        "issue:43066",
+        "issue:43097",
+        "issue:43122",
+        "issue:43335",
+        "issue:43408",
+        "issue:43441",
+        "issue:43452",
+        "issue:43577",
+        "issue:43673",
+        "issue:43874",
+        "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:44933",
+        "issue:44987",
+        "issue:45003",
+        "issue:45071",
+        "issue:45278",
+        "issue:45341",
+        "issue:45459",
+        "issue:45464",
+        "issue:45478",
+        "issue:45542",
+        "issue:45656"
+      ],
+      "soft_pairs": [
+        "issue:44933|issue:45278",
+        "issue:42222|issue:45542",
+        "issue:42490|issue:42673",
+        "issue:43937|issue:44496",
+        "issue:42175|issue:42222",
+        "issue:44483|issue:45464",
+        "issue:38617|issue:44220",
+        "issue:42371|issue:42831",
+        "issue:43066|issue:43122",
+        "issue:44075|issue:45656",
+        "issue:43673|issue:45341",
+        "issue:43012|issue:43577",
+        "issue:44484|issue:45478",
+        "issue:43874|issue:44261",
+        "issue:44410|issue:45478",
+        "issue:43097|issue:44493",
+        "issue:44263|issue:44485",
+        "issue:43012|issue:43408",
+        "issue:41628|issue:45071",
+        "issue:42222|issue:45459",
+        "issue:43441|issue:44829",
+        "issue:45071|issue:45542",
+        "issue:43931|issue:44410",
+        "issue:43335|issue:45071",
+        "issue:43931|issue:44038",
+        "issue:45003|issue:45341",
+        "issue:43097|issue:43452",
+        "issue:43986|issue:45341",
+        "issue:44117|issue:44987",
+        "issue:43986|issue:44484",
+        "issue:43408|issue:44368",
+        "issue:43874|issue:44263",
+        "issue:43976|issue:45071"
+      ],
+      "prepared_review_unit_hash": "14226c2a3e5a77ef3753d358bd9c1559f12c6bc6c7cd0a53dbfc8ee2956e842f",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 37337,
+        "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 a loose mix of unrelated Transformers regressions. I would keep the broad import-regression report as the canonical issue, but I do not see any soft pair that is strong enough to merge as the same underlying bug/change.",
+        "confidence": 0.39,
+        "canonical_issue_reason": "issue:45278 is the broadest umbrella report here (many import errors after upgrading) and is the best place to anchor related import-regression duplicates.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45464 is the most detailed and actionable standalone report in the set, with a large reproduction write-up and active discussion.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44933",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "One is a specific missing import from image_utils; the other is a broad post-upgrade import-error report. Same symptom class, not the same concrete bug."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "VitPose model breakage and TensorFlow backend installation behavior are unrelated failures."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:42673",
+            "accept": false,
+            "reason": "Local-path import warning and Qwen3ForCausalLM VRAM leakage are different code paths and symptoms."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "GLM-5 generation config validation and an unrelated model-type loading error are not the same bug."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:42222",
+            "accept": false,
+            "reason": "Package backend installation coverage and broken VitPose models are unrelated."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Both concern chat/completions, but one is a generic API rejection and the other is a streaming-specific Qwen3.5-0.8B failure; not clearly the same code-path problem."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "ImportError for layer_type_validation is unrelated to the audio feature-extraction issue."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:42831",
+            "accept": false,
+            "reason": "TF32 behavior warnings and FineGrainedFP8 accuracy problems are different precision-related issues."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:43122",
+            "accept": false,
+            "reason": "Tokenizer decoder-type regression and changed tokenization output are related to v5 tokenizer behavior, but not clearly the same concrete defect."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "SGD argument handling and double optimizer-step behavior under deepspeed are different optimizer bugs."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Generation cache missing in chunked_prefill and a testing_utils bug are unrelated."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "A PyTorch warning on compiling bf16 models is not the same as Blip2 dtype fields staying float32 on load."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "max_shard_size behavior and Qwen3.5-MoE from_pretrained errors are unrelated."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:44261",
+            "accept": false,
+            "reason": "Missing get_number_of_image_patches in GLM46V and MLA rms_norm_eps precision issues are different multimodal bugs."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Missing projections in qwen3next layer 0 and Qwen3.5-MoE loading failures are different model implementations."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Removed tie_embeddings_and_encoder_decoder and unexpected position-id keys are not the same regression."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "torch.split return-value handling in GlmMoeDsaIndexer is unrelated to GLM-5 RoPE implementation."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43408",
+            "accept": false,
+            "reason": "bf16 compile warnings and sam3_tracker/sam3_video model-type mismatch are unrelated."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Cannot import AutoImageProcessor and PretrainedConfig type-checking regressions are different API breakages."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "VitPose model breakage and protobuf-tokenizer error masking are unrelated."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Both mention flash_attention_3, but one is an inference/load failure and the other is a training degradation issue; not the same code-path."
+          },
+          {
+            "left": "issue:45071",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "PretrainedConfig type checking and missing TensorFlow backend errors are unrelated."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44410",
+            "accept": false,
+            "reason": "Weight-shape mismatch for Qwen3-VL and missing projections in qwen3next are different model-loading failures."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer construction and PretrainedConfig type checking are different regressions."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44038",
+            "accept": false,
+            "reason": "Qwen3-VL weight mismatch and Qwen3.5/Qwen3 transformers v5 bug reports are unrelated."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "sys.modules access in modeling_utils and a testing_utils bug are unrelated."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "Removed tie_embeddings_and_encoder_decoder and gguf_file/from_pretrained loading failures are different issues."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Video-model loading without torchvision and a testing_utils bug are unrelated."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Tokenizer mapping None handling and a specific physical-intelligence/fast loading failure may be related in theme, but not enough evidence they are the same underlying bug."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Video processor import failure and max_shard_size behavior are unrelated."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "sam3 model-type warning and tie_word_embeddings warning are different model-configuration problems."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "GLM46V image patch counting and GlmMoeDsaIndexer torch.split behavior are unrelated."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Python version compatibility in 5.1.0 and PretrainedConfig type checking in 5.4.0 are different release issues."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 42,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:38617",
+        "issue:42757",
+        "issue:42886",
+        "issue:42907",
+        "issue:43012",
+        "issue:43064",
+        "issue:43066",
+        "issue:43097",
+        "issue:43208",
+        "issue:43352",
+        "issue:43381",
+        "issue:43502",
+        "issue:43526",
+        "issue:43531",
+        "issue:43550",
+        "issue:43824",
+        "issue:43825",
+        "issue:43873",
+        "issue:43901",
+        "issue:43906",
+        "issue:43976",
+        "issue:43986",
+        "issue:43992",
+        "issue:44038",
+        "issue:44062",
+        "issue:44117",
+        "issue:44261",
+        "issue:44315",
+        "issue:44351",
+        "issue:44368",
+        "issue:44380",
+        "issue:44410",
+        "issue:44483",
+        "issue:44484",
+        "issue:44485",
+        "issue:44514",
+        "issue:44655",
+        "issue:44704",
+        "issue:44829",
+        "issue:45103",
+        "issue:45245",
+        "issue:45341",
+        "issue:45372",
+        "issue:45375",
+        "issue:45479",
+        "issue:45588",
+        "issue:45753"
+      ],
+      "soft_pairs": [
+        "issue:43986|issue:45372",
+        "issue:43531|issue:45341",
+        "issue:44829|issue:45245",
+        "issue:38617|issue:42757",
+        "issue:44261|issue:44485",
+        "issue:43064|issue:44315",
+        "issue:38617|issue:44351",
+        "issue:44410|issue:45375",
+        "issue:44062|issue:44117",
+        "issue:43097|issue:45588",
+        "issue:44038|issue:44410",
+        "issue:44704|issue:45103",
+        "issue:43502|issue:43992",
+        "issue:43986|issue:44117",
+        "issue:43097|issue:43873",
+        "issue:43824|issue:43986",
+        "issue:43066|issue:43906",
+        "issue:43097|issue:43901",
+        "issue:43825|issue:44655",
+        "issue:43992|issue:44484",
+        "issue:43208|issue:43550",
+        "issue:43012|issue:43550",
+        "issue:43012|issue:44368",
+        "issue:42886|issue:43381",
+        "issue:43352|issue:44380",
+        "issue:42907|issue:45753",
+        "issue:42907|issue:44655",
+        "issue:44380|issue:45479",
+        "issue:43352|issue:43976",
+        "issue:43064|issue:44380",
+        "issue:43208|issue:43526",
+        "issue:44483|issue:44514"
+      ],
+      "prepared_review_unit_hash": "5e65e64cbc5263addea6f2ff54360451a4872be60ca8fcd49eaf5bf2ce488885",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36713,
+        "estimated_input_tokens": 9179,
+        "estimated_eval_tokens": 18614
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is very heterogeneous: most items are distinct Transformers bugs, regressions, or API/deprecation reports. None of the soft pairs look like true duplicates of the same underlying issue or change.",
+        "confidence": 0.26,
+        "canonical_issue_reason": "issue:43097 is the strongest cluster anchor: it is a concrete v5-breaking behavior change, has inbound references, and is broad enough to sit near several other v5 compatibility reports even though they are not duplicates.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43097 is the best representative issue in this set because it is well-scoped, release-relevant, and likely the most central of the v5 compatibility regressions here.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43986",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "Both are import/crash reports during processor loading, but they involve different missing symbols, different optional dependencies, and different code paths."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding-window behavior and a small testing_utils bug are unrelated."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "torch.compile/FlashAttention training degeneration is unrelated to the category-cardinality RuntimeError."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Different import errors from different modules and symbols; not the same bug."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Missing rms_norm_eps precision issue and GLM-5 RoPE implementation are different model-specific problems."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44315",
+            "accept": false,
+            "reason": "FSDP2/PEFT optimizer-state corruption and Liger-kernel model_init behavior are different training paths."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Cannot import layer_type_validation and cannot import HybridCache are separate API surface regressions."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both mention Qwen-family vision models, but they concern different configs and different missing fields."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44117",
+            "accept": false,
+            "reason": "AddedToken keyword collision and TOKENIZER_MAPPING_NAMES returning None are unrelated tokenizer bugs."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Config removal warning and flash_attention s_aux=None crash are different code paths and symptoms."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44410",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe model compatibility and qwen3next missing projections are distinct model implementation issues."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "AutoProcessor kwargs passthrough and auto_docstring annotation parsing are unrelated."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:43992",
+            "accept": false,
+            "reason": "local_files_only still making API requests is unrelated to missing embed_tokens.weight on UMT5Encoder load."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44117",
+            "accept": false,
+            "reason": "Missing torchvision on video model load and tokenizer mapping None are different failures."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Removed tie_embeddings_and_encoder_decoder and offloading with quantization are separate release/feature problems."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43986",
+            "accept": false,
+            "reason": "Different missing-import failures for different model families and dependencies."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "Wrong tokenizer decoder type in v5 and a reproduction of another issue are not the same concrete bug."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Missing configuration API and pipeline docs mentioning return_all_scores are unrelated."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Incorrect translation-task messaging and inability to save Pipeline objects are different pipeline issues."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Missing pretrained weight handling and max_shard_size default behavior are unrelated save/load topics."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "xLSTM training bugs and Bamba SDPA compile failure are unrelated model-specific regressions."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "A PyTorch bfloat16 warning and a compile/SDPA failure are not the same underlying issue."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "bfloat16 precision warning and tie_word_embeddings warning are different warnings with different causes."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "Offline cache loading failure and gradient checkpointing in eval mode are unrelated behaviors."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Nemotron FA2 support error and GPT2 attention scaling ignored under SDPA/FlashAttention are different attention/backends bugs."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Trainer/FSDP optimizer-state corruption and GPT2 attention scaling are separate subsystems."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "xLSTM training bugs and BeitImageProcessorFast reduce_labels behavior are unrelated."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "Chat/completions request handling and batched apply_chat_template padding behavior are different APIs and failures."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 43,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:36010",
+        "issue:36683",
+        "issue:38617",
+        "issue:42490",
+        "issue:42757",
+        "issue:42994",
+        "issue:43010",
+        "issue:43012",
+        "issue:43064",
+        "issue:43066",
+        "issue:43299",
+        "issue:43352",
+        "issue:43475",
+        "issue:43502",
+        "issue:43525",
+        "issue:43582",
+        "issue:43611",
+        "issue:43618",
+        "issue:43756",
+        "issue:43761",
+        "issue:43824",
+        "issue:43867",
+        "issue:43901",
+        "issue:43931",
+        "issue:43937",
+        "issue:43976",
+        "issue:44038",
+        "issue:44077",
+        "issue:44188",
+        "issue:44222",
+        "issue:44230",
+        "issue:44242",
+        "issue:44263",
+        "issue:44380",
+        "issue:44483",
+        "issue:44485",
+        "issue:44492",
+        "issue:44655",
+        "issue:44683",
+        "issue:44938",
+        "issue:45216",
+        "issue:45278",
+        "issue:45290",
+        "issue:45381",
+        "issue:45440",
+        "issue:45478",
+        "issue:45593",
+        "issue:45753"
+      ],
+      "soft_pairs": [
+        "issue:36683|issue:43582",
+        "issue:44222|issue:45478",
+        "issue:42490|issue:43012",
+        "issue:43012|issue:44683",
+        "issue:43618|issue:43761",
+        "issue:42757|issue:43824",
+        "issue:43618|issue:44380",
+        "issue:38617|issue:43502",
+        "issue:44483|issue:45290",
+        "issue:44038|issue:44230",
+        "issue:43299|issue:43867",
+        "issue:44483|issue:45381",
+        "issue:43756|issue:44485",
+        "issue:43901|issue:44655",
+        "issue:36683|issue:43475",
+        "issue:43299|issue:45753",
+        "issue:42994|issue:44655",
+        "issue:43352|issue:43937",
+        "issue:44077|issue:44380",
+        "issue:43867|issue:45216",
+        "issue:36010|issue:42757",
+        "issue:43066|issue:45381",
+        "issue:36683|issue:44077",
+        "issue:43611|issue:44655",
+        "issue:43976|issue:44938",
+        "issue:36683|issue:43525",
+        "issue:43352|issue:45593",
+        "issue:44263|issue:45440",
+        "issue:43618|issue:45216",
+        "issue:36683|issue:44242",
+        "issue:44188|issue:44380",
+        "issue:43064|issue:44492",
+        "issue:43352|issue:43931",
+        "issue:43010|issue:45278"
+      ],
+      "prepared_review_unit_hash": "b54861e2774782127a222d5bf6cc3926024c7e7208094f047ac7aa69429a7938",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 37412,
+        "estimated_input_tokens": 9353,
+        "estimated_eval_tokens": 18962
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly a set of unrelated Transformers regression reports; the soft pairs are mostly superficial similarity matches (imports, loading, MoE, CLIP, or save_pretrained) rather than the same underlying bug. One broad import-regression issue stands out as the best representative.",
+        "confidence": 0.42,
+        "canonical_issue_reason": "issue:45278 is the broadest umbrella report in the set: multiple import errors after upgrading from 4.57.0 to 5.5.0. It best represents the shared import/regression theme without being overly specific.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the most representative issue for the cluster because it describes a broad, user-visible regression class rather than a single model-specific edge case.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:36683",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "Different failures: Gemma3Config missing vocab_size vs AppleSilicon TypeError in caching_allocator_warmup. Same broad severity, not the same bug."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both involve MoE/save/load workflows, but one is FP8 save_pretrained failure and the other is a from_pretrained loading error for Qwen3.5 MoE. Different code paths."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43012",
+            "accept": false,
+            "reason": "One is a local model path trailing-slash warning; the other is a bfloat16 compile precision warning. No shared underlying defect."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Both mention compile-related issues, but one is a warning and the other is a hard failure in compiled flex_attention on newer torch. Not the same bug."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:43761",
+            "accept": false,
+            "reason": "Both are CLIP v5 regressions, but they affect different outputs: attentions missing vs hidden_states=None. Separate return-value bugs."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but for different missing symbols from different modules. This looks like separate packaging/API breakages."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "CLIP attentions regression and GPT2 attention scaling behavior are unrelated model/backend issues."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:43502",
+            "accept": false,
+            "reason": "ImportError for layer_type_validation is unrelated to local_files_only still making API requests."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "One is chat/completions request validation, the other is apply_chat_template crashing on tool-call assistant messages. Different layers of the stack."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "Both mention Qwen3 VL MoE/fp8, but one is a loading regression and the other is fp8 inference support. Not the same concrete bug."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43867",
+            "accept": false,
+            "reason": "Both are model-loading problems, but one is a Qwen3VL MoE load breakage and the other is a sorted state_dict loading error. Different triggers and fixes."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Chat completions request acceptance and qwen2.5-vl video vision_position_ids are unrelated issues."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Both mention RoPE, but one is Smollm3 dropping layers and the other is GLM-5 RoPE implementation discussion. Different model implementations."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Docs mismatch for return_all_scores is unrelated to save_pretrained failing on Pipeline objects."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Different missing-attribute bugs in different models/configs: vocab_size vs fpn_position_embeddings."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Qwen3VL MoE loading breakage and Qwen3_5 infinite loop on a specific image are distinct failures."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Both involve saving, but one is quantized model saving and the other is Pipeline.save_pretrained. Not the same issue."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:43937",
+            "accept": false,
+            "reason": "Flash Attention support error for NemotronHForCausalLM is unrelated to GLM-5 GenerationConfig validation."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Patchtsmixer post_init validation and GPT2 attention scaling under SDPA/FlashAttention are unrelated model/config bugs."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "One is a load failure with sorted state_dict keys; the other is incorrect checkpoint saving for Qwen3.5. Different directions of the save/load path."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are import errors from Transformers-related APIs, but they concern different missing symbols and likely separate dependency/version regressions."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Wrong tokenizer decoder type and wrong vision_position_ids are different v5 regressions in different components."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:44242",
+            "accept": false,
+            "reason": "Gemma3Config missing vocab_size and load balancing loss omission in MoE training are unrelated bugs."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "torch.compile attention-kernel divergence and GPT2 attention scaling ignorement are separate attention/backend issues."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "FSDP2 optimizer state corruption and a typo in cache strategies are unrelated."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Nemotron Flash Attention 2 support and Qwen3-VL weight-shape mismatch are not the same model-loading problem."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Cache update no_grad decoration and broad import errors after upgrade are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 44,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:28282",
+        "issue:33453",
+        "issue:36296",
+        "issue:36683",
+        "issue:41669",
+        "issue:42490",
+        "issue:42548",
+        "issue:42890",
+        "issue:43010",
+        "issue:43064",
+        "issue:43065",
+        "issue:43502",
+        "issue:43575",
+        "issue:43618",
+        "issue:43723",
+        "issue:43867",
+        "issue:43976",
+        "issue:43992",
+        "issue:44062",
+        "issue:44075",
+        "issue:44077",
+        "issue:44112",
+        "issue:44222",
+        "issue:44246",
+        "issue:44351",
+        "issue:44355",
+        "issue:44360",
+        "issue:44387",
+        "issue:44484",
+        "issue:44485",
+        "issue:44556",
+        "issue:44855",
+        "issue:44857",
+        "issue:44938",
+        "issue:44964",
+        "issue:44977",
+        "issue:45071",
+        "issue:45125",
+        "issue:45310",
+        "issue:45341",
+        "issue:45357",
+        "issue:45375",
+        "issue:45381",
+        "issue:45440",
+        "issue:45478",
+        "issue:45561",
+        "issue:45600",
+        "issue:45850"
+      ],
+      "soft_pairs": [
+        "issue:43992|issue:45310",
+        "issue:43867|issue:45375",
+        "issue:44355|issue:44855",
+        "issue:41669|issue:44246",
+        "issue:36683|issue:43064",
+        "issue:42890|issue:45561",
+        "issue:44938|issue:45071",
+        "issue:43867|issue:45357",
+        "issue:43010|issue:45850",
+        "issue:43992|issue:45478",
+        "issue:43010|issue:43976",
+        "issue:44938|issue:45381",
+        "issue:43618|issue:44077",
+        "issue:42490|issue:44075",
+        "issue:36296|issue:43575",
+        "issue:44360|issue:45440",
+        "issue:28282|issue:33453",
+        "issue:43867|issue:44484",
+        "issue:43618|issue:44857",
+        "issue:43618|issue:44977",
+        "issue:42890|issue:44964",
+        "issue:42890|issue:44112",
+        "issue:42890|issue:45600",
+        "issue:43502|issue:44062",
+        "issue:43723|issue:44556",
+        "issue:42890|issue:45341",
+        "issue:44222|issue:45125",
+        "issue:44387|issue:44556",
+        "issue:42548|issue:44351",
+        "issue:43010|issue:43065",
+        "issue:43502|issue:44485"
+      ],
+      "prepared_review_unit_hash": "fb171a1b41d9c3c3eb88869d9dce9469e1dd09359effb1873a12f95f1499fce5",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 36423,
+        "estimated_input_tokens": 9106,
+        "estimated_eval_tokens": 18468
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These are mostly unrelated issues that only share broad themes like loading, imports, tensor-parallelism, or test flakiness. I found no soft pair that looks like the same concrete bug/change, so all soft edges are rejected. The most representative issue for the overall set is the v5 tokenizer/loading regression.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "issue:43723 is the broadest, most representative loading-regression report here: it targets AutoTokenizer.from_pretrained behavior in v5 and overlaps conceptually with several other load-failure reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43723 is the best umbrella issue for this set because it is a clear, user-facing regression around model/tokenizer loading in v5 rather than a narrow model-specific or test-only failure.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43992",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Different failures and model families: missing embed_tokens.weight in UMT5Encoder vs a Qwen3.5 MoE from_pretrained error."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "One is a sorted-state_dict loading error; the other is a missing config field in Qwen3_5MoeVisionConfig. Same broad area, not the same bug."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "Both are import-related, but one is about compiled Python files and the other is a Python 3.13 torch.jit.script parsing/IndentationError issue."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "Both mention import slowness, but one is caused by import * fanout and the other is a general intermittent import-time slowdown; too broad to merge."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43064",
+            "accept": false,
+            "reason": "Unrelated problems: missing Gemma3Config.vocab_size vs wrong optimizer state/params under v5 + FSDP2 + PEFT + cpu_ram_efficient_loading."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Both are test/CI issues, but one is missing set_seed coverage and the other is an xdist race on captured_info.txt."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Different regressions: Python 3.14 import/load failure vs PretrainedConfig type-checking breakage in v5.4.0."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Loading a sorted state_dict and saving wrong visual encoder keys are different code paths and failure modes."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "One is a specific no_grad decoration request for cache/layer updates; the other is a vague complaint about many models breaking."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both involve Qwen3.5/Qwen-style loading, but they are different model variants and different from_pretrained failures."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "No_grad on cache updates is unrelated to a Python-version compatibility problem in Transformers 5.1.0."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Python 3.14 loading failure and Qwen2.5-VL video position_id corruption are unrelated bugs."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "CLIPOutput attentions regression and patchtsmixer post_init policy are unrelated model-specific changes."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:44075",
+            "accept": false,
+            "reason": "Local-path trailing-slash import warning is unrelated to SGD optimizer arguments not being used."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Both mention tensor parallelism, but one is a training bug and the other is OOM when loading a specific large model with tp."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "DSA indexer ReLU discussion and DeepseekV3MoE remote-divergence are unrelated."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "PyTorch-not-found ImportError and tokenizer-loading regression are different root causes."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "State_dict sorting load error has nothing to do with the 50GB default max_shard_size question."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "CLIPOutput attentions regression and LwDetrImageLoss AMP/CUDA crash are unrelated."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "CLIPOutput attentions assignment and Qwen3.5 flash-attention generation failure are different bugs."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Test seeding flakiness is unrelated to a Phi-4 multimodal model loading failure."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Both are CI/test-related, but one is set_seed coverage and the other is a stale device override failure in GraniteSpeech."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Different test-suite problems: xdist seed/captured_info race vs auto_mappings referencing removed configs."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44062",
+            "accept": false,
+            "reason": "local_files_only still making API requests is unrelated to AddedToken receiving duplicate special kwargs."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Both concern loading regressions after v5, but one is tokenizer loading and the other is checkpoint reload after upgrade; not the same concrete bug."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "A generic testing_utils bug does not match the specific set_seed-related SAM HQ integration failure."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "FP8 save_pretrained for MoE and missing _tp_plan for tensor parallelism are different functionality paths."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Int4 quantization memory regression and checkpoint reload failure after upgrading are both v5-related, but not the same issue."
+          },
+          {
+            "left": "issue:42548",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are top-level import errors, but they involve different symbols and likely different export paths."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43065",
+            "accept": false,
+            "reason": "Cache update no_grad decoration and Sam3PixelDecoder dummy Conv2d are unrelated."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "local_files_only network access bug and GLM-5 RoPE implementation discussion are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 45,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:30333",
+        "issue:30990",
+        "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:43673",
+        "issue:43824",
+        "issue:43827",
+        "issue:43874",
+        "issue:43901",
+        "issue:44222",
+        "issue:44230",
+        "issue:44261",
+        "issue:44273",
+        "issue:44315",
+        "issue:44355",
+        "issue:44360",
+        "issue:44485",
+        "issue:44492",
+        "issue:44556",
+        "issue:44623",
+        "issue:44655",
+        "issue:44749",
+        "issue:44908",
+        "issue:44938",
+        "issue:44961",
+        "issue:45003",
+        "issue:45071",
+        "issue:45125",
+        "issue:45230",
+        "issue:45362",
+        "issue:45375",
+        "issue:45468",
+        "issue:45478",
+        "issue:45542",
+        "issue:45850"
+      ],
+      "soft_pairs": [
+        "issue:44355|issue:44749",
+        "issue:45125|issue:45478",
+        "issue:43010|issue:43232",
+        "issue:30990|issue:31795",
+        "issue:45125|issue:45375",
+        "issue:30333|issue:33453",
+        "issue:43299|issue:44556",
+        "issue:43502|issue:44222",
+        "issue:44315|issue:44360",
+        "issue:42831|issue:43874",
+        "issue:43502|issue:44961",
+        "issue:44355|issue:45071",
+        "issue:30333|issue:36296",
+        "issue:44355|issue:44655",
+        "issue:43827|issue:44961",
+        "issue:44222|issue:44230",
+        "issue:33453|issue:34689",
+        "issue:42757|issue:43576",
+        "issue:33453|issue:36296",
+        "issue:43576|issue:45071",
+        "issue:43576|issue:43824",
+        "issue:31795|issue:38617",
+        "issue:41628|issue:42548",
+        "issue:44961|issue:45468",
+        "issue:38617|issue:43576",
+        "issue:43576|issue:44938",
+        "issue:43673|issue:44492",
+        "issue:44908|issue:44961",
+        "issue:44485|issue:44961",
+        "issue:36296|issue:43901",
+        "issue:41628|issue:44355",
+        "issue:36296|issue:43526",
+        "issue:45362|issue:45850",
+        "issue:44261|issue:44360",
+        "issue:44623|issue:45230",
+        "issue:44273|issue:45542",
+        "issue:44273|issue:45003"
+      ],
+      "prepared_review_unit_hash": "843c4fe4b6249675e3a5d333c47fbfbf024bef61affc44bca82cb98fbc104f22",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 37,
+        "serialized_chars": 38005,
+        "estimated_input_tokens": 9502,
+        "estimated_eval_tokens": 19260
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a mix of distinct Transformers regressions, import/load failures, and docs-only issues rather than true duplicates. No soft edge is strong enough to merge with confidence from titles alone.",
+        "confidence": 0.36,
+        "canonical_issue_reason": "Issue 45850 is the broadest umbrella-style report in the set, covering many v5 breakages; if one issue must represent the cluster, it is the most general fit, though it is not very specific.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43299 is the strongest concrete example of a v5 loading regression and is more actionable than the generic umbrella reports, so it is the best representative issue overall.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44355",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "Different symptom classes: compiled-file execution errors vs a performance regression while filtering data."
+          },
+          {
+            "left": "issue:45125",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Same model family, but one is tensor-parallel metadata and the other is a from_pretrained load failure; related, not clearly the same bug."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43232",
+            "accept": false,
+            "reason": "Both involve generation/cache internals, but the reported failures are on different code paths and look separately fixable."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:31795",
+            "accept": false,
+            "reason": "A stuck-loading runtime bug and a documentation clarity issue are not duplicates."
+          },
+          {
+            "left": "issue:45125",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Different missing fields/metadata in the same model family; too little evidence that they are one underlying defect."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Training-job reporting in MLflow is unrelated to tokenizer-loading regressions."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Both are version-regression loading failures, but they concern different model/checkpoint situations and are not the same concrete bug."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Offline-download behavior and FP8 save_pretrained are unrelated."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Liger-kernel application and a DSA indexer activation-function issue are unrelated."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:43874",
+            "accept": false,
+            "reason": "An FP8 accuracy problem and a multimodal-token count AttributeError are different bugs."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "No meaningful overlap between local_files_only network requests and an unrelated placeholder title."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Compiled-file execution errors are unrelated to PretrainedConfig type checking."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "MLflow failure reporting and tensor-parallel training are different subsystems and failure modes."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Running compiled Python files and saving Pipeline objects are distinct problems."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "Docs referencing pipeline() and a meaningless single-word issue title are not duplicates."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "FP8 save_pretrained and FP8 inference support are related only at a broad feature level, not the same bug."
+          },
+          {
+            "left": "issue:33453",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Different loading regressions affecting different model classes and code paths."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:43576",
+            "accept": false,
+            "reason": "An import error from huggingface_hub and a broken env command are not the same defect."
+          },
+          {
+            "left": "issue:33453",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "Tokenizer loading and tensor-parallel training are unrelated."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "CLI breakage and config type-checking are different issues."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "A broken CLI command is not the same as a missing Qwen2.5-VL export."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Documentation confusion and an import error are not duplicates."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:42548",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and likely different missing exports."
+          },
+          {
+            "left": "issue:44961",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "No clear shared underlying bug; one title is non-informative."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:43576",
+            "accept": false,
+            "reason": "ImportError for a specific symbol and a broken env command are not the same code-path problem."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "CLI breakage and Python 3.14 load failure are too broad to treat as one duplicate."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "Generation cache missing and a cache-strategy typo are not the same bug."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "Scheduler kwargs handling is unrelated to the other title."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "RoPE implementation discussion and the other title are unrelated."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Tensor-parallel training and docs mentioning return_all_scores are unrelated."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44355",
+            "accept": false,
+            "reason": "An import-name regression and compiled-file execution errors do not match."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Tensor-parallel training and BeitImageProcessorFast label reduction are different bugs."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "A specific Qwen3.5 chat crash is not the same concrete issue as a generic complaint about many broken models."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Missing rms_norm_eps precision handling and a missing ReLU in DSA indexer are unrelated."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "Missing processor.save_pretrained files is too specific to merge with a generic bug report."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Lazy loading behavior and a TensorFlow backend error from tensorboard-only installs are unrelated."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Lazy loading failure and unsafe sys.modules access are different mechanisms."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 46,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:31795",
+        "issue:33453",
+        "issue:34689",
+        "issue:37428",
+        "issue:41669",
+        "issue:42371",
+        "issue:42890",
+        "issue:43352",
+        "issue:43408",
+        "issue:43519",
+        "issue:43525",
+        "issue:43572",
+        "issue:43576",
+        "issue:43673",
+        "issue:43704",
+        "issue:43827",
+        "issue:43874",
+        "issue:43976",
+        "issue:44038",
+        "issue:44077",
+        "issue:44246",
+        "issue:44273",
+        "issue:44295",
+        "issue:44297",
+        "issue:44315",
+        "issue:44351",
+        "issue:44360",
+        "issue:44393",
+        "issue:44485",
+        "issue:44492",
+        "issue:44556",
+        "issue:44655",
+        "issue:44704",
+        "issue:44829",
+        "issue:44861",
+        "issue:44908",
+        "issue:44945",
+        "issue:45071",
+        "issue:45092",
+        "issue:45125",
+        "issue:45230",
+        "issue:45278",
+        "issue:45310",
+        "issue:45341",
+        "issue:45468",
+        "issue:45478",
+        "issue:45753",
+        "issue:45850"
+      ],
+      "soft_pairs": [
+        "issue:44393|issue:45753",
+        "issue:44393|issue:45071",
+        "issue:44038|issue:45125",
+        "issue:43576|issue:44556",
+        "issue:43525|issue:45850",
+        "issue:41669|issue:43673",
+        "issue:44556|issue:45092",
+        "issue:44492|issue:45850",
+        "issue:43572|issue:45850",
+        "issue:42371|issue:44492",
+        "issue:44704|issue:44908",
+        "issue:42890|issue:44908",
+        "issue:45278|issue:45850",
+        "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:41669|issue:44351",
+        "issue:41669|issue:45341",
+        "issue:44492|issue:44945",
+        "issue:43874|issue:45850",
+        "issue:44393|issue:45478",
+        "issue:44077|issue:44861",
+        "issue:43827|issue:44908",
+        "issue:31795|issue:33453",
+        "issue:44315|issue:44829",
+        "issue:41669|issue:44908",
+        "issue:43576|issue:44273",
+        "issue:44655|issue:44861",
+        "issue:44360|issue:45468",
+        "issue:45230|issue:45310",
+        "issue:44273|issue:44908",
+        "issue:43519|issue:44485",
+        "issue:43519|issue:44908",
+        "issue:31795|issue:34689",
+        "issue:41669|issue:43704",
+        "issue:31795|issue:37428",
+        "issue:44315|issue:45092",
+        "issue:44485|issue:45468",
+        "issue:43704|issue:44485",
+        "issue:42371|issue:43704"
+      ],
+      "prepared_review_unit_hash": "98ff1b978ef90c3908b75cededd85ba4571d25410243827f679875f62290f311",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 45,
+        "serialized_chars": 40585,
+        "estimated_input_tokens": 10147,
+        "estimated_eval_tokens": 20550
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a grab-bag of mostly unrelated Transformers v5 regressions, model-specific loading/runtime bugs, docs issues, and performance complaints. None of the soft-edge pairs look like the same underlying fix; most only share a broad subsystem or upgrade context.",
+        "confidence": 0.76,
+        "canonical_issue_reason": "issue:45278 is the best umbrella-style issue: it\u2019s open, broad, and explicitly captures the post-4.57/5.x import-breakage pattern affecting many models, which fits the cluster better than the narrower model-specific reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the strongest global issue candidate because it summarizes the broadest recurring failure mode in this set (upgrade-related import/loading regressions) without being tied to one model or one code path.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44393",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Same model family, but different symptoms and likely different bugs: 2D bbox hallucination/error vs infinite loop on a specific image."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Qwen3-VL output bug and PretrainedConfig type-checking regression are unrelated."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "Both mention Qwen3.5/MoE, but one is a from_pretrained error and the other is missing _tp_plan for tensor parallelism; not the same concrete fix."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Different failures: broken `transformers env` command vs checkpoint reload incompatibility after upgrade."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "A specific `Llama4Config.pad_token_id` attribute error is not the same as a generic \u201cmany models break\u201d report."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "Import-star performance problem and missing GenerationMixin cache during chunked prefill are separate issues."
+          },
+          {
+            "left": "issue:44556",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Both are upgrade/loading-related, but one is a general reload failure and the other is a specific remote-code/meta-init incompatibility."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "A typo in cache strategy text is not the same underlying bug as a broad v5 regression report."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "Missing `pad_token_idx` in StableLmConfig is model-specific and distinct from the generic breakage umbrella."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "TF32 API/settings guidance and a cache-strategy typo are unrelated."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "AutoProcessor kwargs forwarding and inverse_sqrt scheduler kwargs are different code paths and different bugs."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Test flakiness from missing `set_seed()` is unrelated to scheduler kwarg handling."
+          },
+          {
+            "left": "issue:45278",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "Both are broad v5 complaint-style issues, but they don\u2019t point to the same concrete defect or fix; one is about import errors, the other about many models breaking generally."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44273",
+            "accept": false,
+            "reason": "Import-time slowness and lazy-loading malfunction are related only at a high level, not the same bug."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:44315",
+            "accept": false,
+            "reason": "Model-type warning and Liger Kernel application with `model_init` are unrelated."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44393",
+            "accept": false,
+            "reason": "Python version compatibility and Qwen3-VL bbox output are unrelated."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:44393",
+            "accept": false,
+            "reason": "Flash Attention 2.0 support error for Nemotron and Qwen3-VL output bug are unrelated."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "Tokenizer save metadata mismatch and a generic bug report are not the same underlying issue."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Broken `transformers env` command and scheduler kwarg handling are unrelated."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "`position_ids` buffer access error is a concrete model/runtime bug, not the same as a placeholder bug report."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Scheduler kwargs handling and PretrainedConfig type-checking are unrelated."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Import-star performance and missing `HybridCache` import are different problems."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Import-star performance and a small testing_utils bug do not share the same fix."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Cache-strategy typo and incorrect pipeline-parallel LLM output are unrelated."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "GLM46V missing method is a model-specific implementation gap, not the same as a broad breakage report."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Different Qwen3.5/VL failures with different symptoms and likely different code paths."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Optional `post_init` policy and `_get_tied_weight_keys` list crash are unrelated."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Docs still referencing `pipeline()` and scheduler kwarg handling are unrelated."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Documentation confusion about `input_ids`/`past_key_values` is not the same as a tokenizer loading regression."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Liger Kernel not applied with `model_init` and degenerate training under flash_attention_3 are separate issues."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Import-star performance and inverse_sqrt scheduler kwargs are unrelated."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44273",
+            "accept": false,
+            "reason": "Broken env command and lazy-loading malfunction are different bugs."
+          },
+          {
+            "left": "issue:44655",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Pipeline save/load failure and tied-weights key crash are unrelated."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "DSA indexer missing ReLU and Gemma-4 audio positional encoding are unrelated."
+          },
+          {
+            "left": "issue:45230",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Generic bug report and Qwen3.5 MoE from_pretrained error are not the same issue."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Lazy loading bug and scheduler kwarg handling are unrelated."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Qwen3VL timestamp calculation and GLM-5 RoPE implementation are unrelated model-specific bugs."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Timestamp calculation bug and scheduler kwargs bug are unrelated."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Docs confusion around forward args and Llama 3.2 Vision model loading failure are unrelated."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:43704",
+            "accept": false,
+            "reason": "Import-star performance and VRAM leak in dataloader threads are different problems."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:37428",
+            "accept": false,
+            "reason": "Forward-doc wording and flash-attention import error are unrelated."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Liger Kernel/model_init behavior and remote-code meta-init incompatibility are different fixes."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "GLM-5 RoPE implementation and Gemma-4 audio positional encoding are unrelated."
+          },
+          {
+            "left": "issue:43704",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "VRAM leak across threads and RoPE implementation are unrelated."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:43704",
+            "accept": false,
+            "reason": "TF32 API guidance and VRAM leak in dataloader threads are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 17,
+      "review_unit_index": 47,
+      "review_unit_count": 47,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-400",
+      "nodes": [
+        "issue:30333",
+        "issue:31515",
+        "issue:31795",
+        "issue:33453",
+        "issue:36296",
+        "issue:38617",
+        "issue:42913",
+        "issue:43066",
+        "issue:43298",
+        "issue:43502",
+        "issue:43519",
+        "issue:43525",
+        "issue:43572",
+        "issue:43606",
+        "issue:43618",
+        "issue:43704",
+        "issue:44230",
+        "issue:44261",
+        "issue:44263",
+        "issue:44492",
+        "issue:44623",
+        "issue:44861",
+        "issue:44908",
+        "issue:44998",
+        "issue:45290",
+        "issue:45310",
+        "issue:45341",
+        "issue:45468",
+        "issue:45850"
+      ],
+      "soft_pairs": [
+        "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:42913|issue:43066",
+        "issue:44908|issue:44998",
+        "issue:44861|issue:45290",
+        "issue:44492|issue:44998",
+        "issue:44998|issue:45850",
+        "issue:43618|issue:44998",
+        "issue:44998|issue:45341",
+        "issue:43525|issue:44861",
+        "issue:43572|issue:44861",
+        "issue:31515|issue:38617",
+        "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": "424dc500e96f887e45fe2a581e971db5217a8398fcf0083cb93d1ec134ed6ba9",
+      "component_budget": {
+        "node_count": 400,
+        "item_count": 400,
+        "soft_pair_count": 1499,
+        "serialized_chars": 772949,
+        "estimated_input_tokens": 193238,
+        "estimated_eval_tokens": 386732
+      },
+      "budget": {
+        "node_count": 29,
+        "item_count": 29,
+        "soft_pair_count": 23,
+        "serialized_chars": 23090,
+        "estimated_input_tokens": 5773,
+        "estimated_eval_tokens": 11802
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This looks like a mostly noisy, title-similarity cluster: the paired issues share broad Transformer/modeling vocabulary, but the underlying bugs are different enough that none should be merged as duplicates.",
+        "confidence": 0.41,
+        "canonical_issue_reason": "issue:43502 is the clearest concrete bug report in the set and a reasonable representative of the loading/runtime-failure group; the others are either narrower model-specific defects, documentation, or unrelated noise.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43502 is the strongest standalone issue for triage because it states a reproducible user-facing failure plainly and is less ambiguous than the more model-specific or off-topic candidates.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43502",
+            "right": "issue:43519",
+            "accept": false,
+            "reason": "One is network access despite local_files_only=True; the other is a Qwen3VL timestamp-calculation bug. Different code paths and symptoms."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Timestamp math in a processor is unrelated to missing files from processor.save_pretrained()."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:43704",
+            "accept": false,
+            "reason": "Processor timestamp logic and VRAM leakage in multi-threaded dataloader use are distinct bugs."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "torch.split return handling in GLM MoE is unrelated to Gemma-4 audio positional encoding."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "CPU offload device-mismatch in suno/bark-small is a different failure than Gemma-4 audio rel-positional encoding."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "Missing rms_norm_eps in MLA q_a_layernorm is a different model-specific precision issue, not the Gemma-4 audio bug."
+          },
+          {
+            "left": "issue:42913",
+            "right": "issue:43066",
+            "accept": false,
+            "reason": "Both mention tokenizer v5 behavior, but one is a broad behavior change and the other is a decoder-type mismatch; not enough evidence of the same bug."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Scheduler kwargs handling is unrelated to the off-topic 'Unemployment' issue."
+          },
+          {
+            "left": "issue:44861",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "_get_tied_weight_keys AttributeError and chat-template crashes with tool calls are different APIs and failure modes."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "A cache-strategy typo is unrelated to the off-topic unemployment issue."
+          },
+          {
+            "left": "issue:44998",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "The off-topic unemployment report is not the same as the broad 'many models break after 5.0' complaint."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "CLIPOutput attentions assignment is a concrete model-output bug, not related to the off-topic unemployment report."
+          },
+          {
+            "left": "issue:44998",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "An off-topic issue and a testing_utils bug are unrelated."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Missing pad_token_id in Llama4Config and an AttributeError in tied-weight key extraction are different config/runtime paths."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "StableLmConfig missing pad_token_idx is unrelated to _get_tied_weight_keys crashing."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Slow checkpoint loading from from_pretrained and an ImportError for layer_type_validation are unrelated."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "FP8 support for Qwen3-VL/Qwen3.5 MoE and missing processor.save_pretrained files are different problems."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:31515",
+            "accept": false,
+            "reason": "MLflow failure-reporting on training-job failure is unrelated to slow checkpoint loading."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "Checkpoint loading slowness and tensor-parallel training behavior are different issues."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:31795",
+            "accept": false,
+            "reason": "Loading performance and confusing forward() documentation are not the same bug."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "FP8 inference support and a from_pretrained error for Qwen3.5 MoE are related by model family but not the same concrete failure."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "from_pretrained checkpoint slowness and tokenizer-loading regression are different code paths."
+          },
+          {
+            "left": "issue:43298",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "Continuous batching not supporting audio models is unrelated to Qwen3-VL/Qwen3.5 FP8 inference support."
+          }
+        ]
+      },
+      "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": 17,
+      "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": "858dc1e43cba70bec3314b07910c2e3be01217118fdec36f4b4aa5b6c473e2cf",
+      "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 three issues are not true duplicates: one is about Tapas tokenization mismatch, one is an auth/error-message problem when downloading a private model, and one is a general model-loading failure for bert-base-uncased. The cluster is loosely related by NLP/model loading, but the underlying problems differ.",
+        "confidence": 0.89,
+        "canonical_issue_reason": "Issue #16618 is the best representative only because it is the broadest and most active of the three, but the cluster is heterogeneous rather than a real duplicate set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#16618 has the most general model-loading framing and the most discussion, so it is the least misleading single anchor for this noisy cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:13244",
+            "right": "issue:16618",
+            "accept": false,
+            "reason": "Tokenization parity for Tapas is a different bug from a model-loading failure for bert-base-uncased."
+          },
+          {
+            "left": "issue:15980",
+            "right": "issue:16618",
+            "accept": false,
+            "reason": "Private-model download auth/error-message behavior is not the same underlying problem as a generic inability to load bert-base-uncased."
+          }
+        ]
+      },
+      "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": 17,
+      "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": "ee5967d31c785b6051de8857ffb7e15fe7c90333094ab8cc8b77869e3601789c",
+      "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 related in that both are community contribution requests about expanding architecture support, but they target different features: Flash Attention 2 vs GGUF. They should not be deduplicated.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #26350 is the more established discussion thread by age and activity, but it is not a duplicate of #33260 because the requested support is for a different mechanism and code path.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:26350",
+            "right": "issue:33260",
+            "accept": false,
+            "reason": "Both are feature requests for broader architecture support, but they concern different underlying changes: Flash Attention 2 support versus GGUF support. They are not the same bug or mergeable PR-equivalent change."
+          }
+        ]
+      },
+      "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": 17,
+      "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-27088-5",
+      "nodes": [
+        "issue:27088",
+        "issue:28936",
+        "issue:43834",
+        "issue:44365",
+        "issue:44370"
+      ],
+      "soft_pairs": [
+        "issue:27088|issue:44370",
+        "issue:27088|issue:44365",
+        "issue:27088|issue:43834",
+        "issue:27088|issue:28936",
+        "issue:28936|issue:43834",
+        "issue:28936|issue:44365",
+        "issue:28936|issue:44370",
+        "issue:43834|issue:44370",
+        "issue:43834|issue:44365",
+        "issue:44365|issue:44370"
+      ],
+      "prepared_review_unit_hash": "9aaf826beffaac7266a69254b9b26d749f875a7150656a3edae3d9516bef46aa",
+      "component_budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 10,
+        "serialized_chars": 6441,
+        "estimated_input_tokens": 1611,
+        "estimated_eval_tokens": 3478
+      },
+      "budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 10,
+        "serialized_chars": 6441,
+        "estimated_input_tokens": 1611,
+        "estimated_eval_tokens": 3478
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster is mostly parallel docs-translation requests, not true duplicates: Turkish and Spanish are distinct, and the three generic placeholder issues are nearly identical but still lack evidence they refer to the same concrete translation task. I would keep them separate, with the oldest generic placeholder issue as the best representative only for that narrow subgroup.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "Issue 43834 is the oldest and most generic of the near-identical placeholder-template issues, so it is the cleanest representative if one must pick a canonical anchor; however, it does not subsume the language-specific Turkish/Spanish requests.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43834 is the best single representative of the generic placeholder subgroup because it is the earliest and most broadly worded, but it should not be treated as a duplicate of the actual Turkish or Spanish translation requests.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:27088",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Different concrete translation requests: Turkish vs a generic placeholder issue; same i18n theme is not enough to merge."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Different concrete translation requests: Turkish vs a generic placeholder issue; no evidence they are the same task."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:43834",
+            "accept": false,
+            "reason": "Turkish translation request and a generic placeholder issue are not the same underlying change."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:28936",
+            "accept": false,
+            "reason": "They target different languages, Turkish and Spanish, so they are distinct requests."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:43834",
+            "accept": false,
+            "reason": "Spanish translation request vs generic placeholder issue; not the same concrete work item."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Spanish translation request and a generic placeholder issue are distinct tasks."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Spanish translation request and a generic placeholder issue are distinct tasks."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "The titles are nearly identical, but they still look like separate placeholder issues rather than a clearly identical concrete translation task."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Near-identical placeholder issues, but no explicit target or other evidence that they are the same underlying change."
+          },
+          {
+            "left": "issue:44365",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Almost identical template issues, but the available data does not prove they are the same concrete request."
+          }
+        ]
+      },
+      "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": 17,
+      "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-36817-15",
+      "nodes": [
+        "issue:36817",
+        "issue:36879",
+        "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:36817|issue:36879",
+        "issue:36879|issue:41211",
+        "issue:36879|issue:43909",
+        "issue:36879|issue:44121",
+        "issue:36879|issue:43837",
+        "issue:36879|issue:43671",
+        "issue:37015|issue:43837",
+        "issue:37015|issue:44121",
+        "issue:37015|issue:43909",
+        "issue:36879|issue:37015",
+        "issue:37015|issue:42503",
+        "issue:41115|issue:43837",
+        "issue:41115|issue:44121",
+        "issue:41115|issue:42503",
+        "issue:41115|issue:45027",
+        "issue:41211|issue:43671",
+        "issue:42503|issue:43837",
+        "issue:42503|issue:43671",
+        "issue:42503|issue:42738",
+        "issue:42738|issue:45027",
+        "issue:43671|issue:43837",
+        "issue:43671|issue:45027",
+        "issue:43671|issue:45177",
+        "issue:43671|issue:44121",
+        "issue:43671|issue:43909",
+        "issue:43837|issue:43909",
+        "issue:43837|issue:45322",
+        "issue:43837|issue:44121",
+        "issue:43909|issue:44205",
+        "issue:44121|issue:45322",
+        "issue:43837|issue:45027",
+        "issue:37015|issue:45027",
+        "issue:43837|issue:45177"
+      ],
+      "prepared_review_unit_hash": "3c31d82c4886f731a4fbe5feb7861bca4b95cbc96436a636f7c4a3372ef77812",
+      "component_budget": {
+        "node_count": 15,
+        "item_count": 15,
+        "soft_pair_count": 35,
+        "serialized_chars": 20711,
+        "estimated_input_tokens": 5178,
+        "estimated_eval_tokens": 10612
+      },
+      "budget": {
+        "node_count": 15,
+        "item_count": 15,
+        "soft_pair_count": 35,
+        "serialized_chars": 20711,
+        "estimated_input_tokens": 5178,
+        "estimated_eval_tokens": 10612
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This is a loose similarity cluster of unrelated model-request issues; the overlap is mostly generic \u201cadd/support model\u201d wording, not the same underlying bug or change.",
+        "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": "Different model requests (EuroBert vs DEIMv2); only the generic add-model template overlaps."
+          },
+          {
+            "left": "issue:36817",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "EuroBert and Qwen3-TTS are unrelated model requests, not one underlying change."
+          },
+          {
+            "left": "issue:36817",
+            "right": "issue:36879",
+            "accept": false,
+            "reason": "Different requested models (EuroBert vs RF-DETR); same broad \u201cadd model\u201d phrasing only."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:41211",
+            "accept": false,
+            "reason": "RF-DETR and DEIMv2 are different models with no shared concrete fix."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "RF-DETR and LFM2.5 Audio are distinct model additions in different domains."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "RF-DETR vs weight-sparse transformer request; no same code-path problem."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "RF-DETR and Qwen3-ASR are unrelated model requests."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "RF-DETR and Qwen3-TTS are different models; similarity is only superficial."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "NeoBERT and Qwen3-ASR are unrelated requests."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "NeoBERT and circuit-sparsity/circuitgpt are different model additions."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "NeoBERT and LFM2.5 Audio target different model families and modalities."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:37015",
+            "accept": false,
+            "reason": "RF-DETR and NeoBERT are different requested models; not the same change."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:42503",
+            "accept": false,
+            "reason": "NeoBERT and ModernVBERT are separate model requests, not duplicates."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "MiniCPM3 architecture request vs Qwen3-ASR request; different models and tasks."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "MiniCPM3 and circuit-sparsity/circuitgpt are unrelated additions."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:42503",
+            "accept": false,
+            "reason": "MiniCPM3 and ModernVBERT are distinct model requests."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "MiniCPM3 and Voxtral TTS are different model families and modalities."
+          },
+          {
+            "left": "issue:41211",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "DEIMv2 and Qwen3-TTS are unrelated model requests."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "ModernVBERT and Qwen3-ASR are different models, not the same bug/change."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "ModernVBERT and Qwen3-TTS are distinct requests."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:42738",
+            "accept": false,
+            "reason": "ModernVBERT vs BERT-like models with RoPE is not the same concrete request."
+          },
+          {
+            "left": "issue:42738",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "BERT-like RoPE models and Voxtral TTS are unrelated additions."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "Qwen3-TTS and Qwen3-ASR are same family but different tasks; not one mergeable PR."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "Qwen3-TTS and Voxtral-4B-TTS-2603 are different TTS model requests."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:45177",
+            "accept": false,
+            "reason": "Qwen3-TTS and DeepSeek-OCR2 are unrelated model requests."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "Qwen3-TTS and circuit-sparsity/circuitgpt are different model changes."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "Qwen3-TTS and LFM2.5 Audio are separate audio-model requests, not duplicates."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "Qwen3-ASR and LFM2.5 Audio are different models and implementation paths."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45322",
+            "accept": false,
+            "reason": "Qwen3-ASR and EUPE are unrelated model requests."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "Qwen3-ASR and circuit-sparsity/circuitgpt are different changes."
+          },
+          {
+            "left": "issue:43909",
+            "right": "issue:44205",
+            "accept": false,
+            "reason": "LFM2.5 Audio and SAM3-LiteText are unrelated audio vs vision/text requests."
+          },
+          {
+            "left": "issue:44121",
+            "right": "issue:45322",
+            "accept": false,
+            "reason": "Weight-sparse transformer request and EUPE are different model additions."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "Qwen3-ASR and Voxtral TTS are different speech-model requests."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "NeoBERT and Voxtral TTS are unrelated model requests."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45177",
+            "accept": false,
+            "reason": "Qwen3-ASR and DeepSeek-OCR2 are different domains and 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": 6,
+      "component_count": 17,
+      "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": "69eaeebe2f75ca112a039daa8b0fda47db97aa12346879a97bf5280f607039b0",
+      "component_budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 8,
+        "serialized_chars": 6293,
+        "estimated_input_tokens": 1574,
+        "estimated_eval_tokens": 3404
+      },
+      "budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 8,
+        "serialized_chars": 6293,
+        "estimated_input_tokens": 1574,
+        "estimated_eval_tokens": 3404
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These issues are not duplicates of one another; they span unrelated models, configs, tokenizer warnings, and typo/propagation bugs. The cluster is mostly soft-similarity noise rather than one underlying defect.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:37934",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Different components and bug types: a model normalization design question vs. a config type-hint issue."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43240",
+            "accept": false,
+            "reason": "Mask2former ignore_value handling is unrelated to missing kwargs in a loss-call path."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:44031",
+            "accept": false,
+            "reason": "Unrelated subsystems: segmentation loss config vs. tokenizer regex warning regression."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43813",
+            "accept": false,
+            "reason": "No shared code path; this is a mask2former parameter usage bug, not a typo in quantization naming."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Different models and different failure modes; not the same underlying bug."
+          },
+          {
+            "left": "issue:43240",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Loss-function kwargs forwarding and a config annotation mistake are not the same issue."
+          },
+          {
+            "left": "issue:43668",
+            "right": "issue:44031",
+            "accept": false,
+            "reason": "A type-hint bug in ModernBERTConfig is unrelated to tokenizer regex warnings."
+          },
+          {
+            "left": "issue:43668",
+            "right": "issue:43813",
+            "accept": false,
+            "reason": "These concern unrelated areas: config typing vs. a typo in a quantization attribute name."
+          }
+        ]
+      },
+      "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": 17,
+      "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": "3fee830e0daf28fa54ee2f361c024149a0bad964c3953740a03d18b45f63089b",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1479,
+        "estimated_input_tokens": 370,
+        "estimated_eval_tokens": 996
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1479,
+        "estimated_input_tokens": 370,
+        "estimated_eval_tokens": 996
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues are not duplicates: one is a generic request for \"Set Block Decoding,\" while the other asks for \"DeepSeek v3.2 support.\" They describe different kinds of work and different likely code paths.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #43037 is the more concrete and actionable request; it names a specific model support gap, whereas #41084 is a broader/less-defined decoding feature request.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41084",
+            "right": "issue:43037",
+            "accept": false,
+            "reason": "Different underlying requests: block decoding support vs. DeepSeek v3.2 model support. They do not appear to be the same bug or change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 8,
+      "component_count": 17,
+      "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": "cb0d8d3692a4529927193bdad6982cc11ca006d45d84fa2eedd173fc5b9d63de",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1584,
+        "estimated_input_tokens": 396,
+        "estimated_eval_tokens": 1048
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1584,
+        "estimated_input_tokens": 396,
+        "estimated_eval_tokens": 1048
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues are both about Trainer customization, but they describe different problems and likely different fixes: one is about `remove_unused_column` behavior, the other about FLOPs estimation for custom trainers.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "Issue #43598 is the broader and more representative Trainer customizability request; #43600 is a separate FLOPs-estimation edge case.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#43598 is the better cluster anchor because it frames a general Trainer customization limitation, while #43600 is narrower and unrelated in concrete fix path.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43598",
+            "right": "issue:43600",
+            "accept": false,
+            "reason": "Both mention custom trainers, but one concerns unused-column handling and the other concerns FLOPs estimation; they are not the same underlying bug or change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 9,
+      "component_count": 17,
+      "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": "8679537367b512bdfd81e55c2afc525f8cc7d177e4e6c9f60f697629777df31a",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1612,
+        "estimated_input_tokens": 403,
+        "estimated_eval_tokens": 1062
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1612,
+        "estimated_input_tokens": 403,
+        "estimated_eval_tokens": 1062
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are related at a very high level (both ask for new AutoModel support for Qwen variants), but they are not the same underlying request: one is for Qwen3-Omni registration in base/conditional-generation auto classes, the other is for Qwen3.5 sequence-classification support.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "Issue 43684 is the broader and more generic registration request, so it is the better representative issue for this cluster topic.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43684 best represents the cluster because it targets core AutoModel registration, whereas 44405 is a narrower, task-specific classifier-head request for a different Qwen family.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43684",
+            "right": "issue:44405",
+            "accept": false,
+            "reason": "Different model families and different auto-task surfaces (conditional generation vs sequence classification); they are not the same concrete bug or change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 10,
+      "component_count": 17,
+      "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": "bcf2c67d1c1393c59985e3bb355de6de4eb39ac97584a5bdf1a29a54de5512c7",
+      "component_budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4264,
+        "estimated_input_tokens": 1066,
+        "estimated_eval_tokens": 2388
+      },
+      "budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4264,
+        "estimated_input_tokens": 1066,
+        "estimated_eval_tokens": 2388
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "All four are extremely low-signal closed issues with generic or likely spammy titles and no body detail. If one issue must stand in as canonical, the earliest repeated-title issue is the best anchor, but there is not enough evidence to confidently merge any pair.",
+        "confidence": 0.63,
+        "canonical_issue_reason": "Issue #43845 is the earliest of the three identical 'huggingface ' reports, so it is the most reasonable canonical anchor for that repeated cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#43845 is the best overall issue representative because it is the first repeated-title entry and has the strongest superficial similarity to #43846 and #43859.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43835",
+            "right": "issue:43845",
+            "accept": false,
+            "reason": "Different titles ('Fett' vs 'huggingface') and no body or target evidence linking them to the same underlying issue."
+          },
+          {
+            "left": "issue:43835",
+            "right": "issue:43846",
+            "accept": false,
+            "reason": "Different titles and no descriptive evidence; cannot treat as the same bug or change."
+          },
+          {
+            "left": "issue:43835",
+            "right": "issue:43859",
+            "accept": false,
+            "reason": "Different titles and no supporting context, so the shared cluster signal is too weak to accept."
+          },
+          {
+            "left": "issue:43845",
+            "right": "issue:43846",
+            "accept": false,
+            "reason": "Exact same title, but there is no issue description or concrete bug detail to prove they report the same underlying problem."
+          },
+          {
+            "left": "issue:43845",
+            "right": "issue:43859",
+            "accept": false,
+            "reason": "Same generic title, but the available evidence is insufficient to confirm identical underlying content."
+          },
+          {
+            "left": "issue:43846",
+            "right": "issue:43859",
+            "accept": false,
+            "reason": "Same generic title, but no body, targets, or references provide enough proof that these are true duplicates."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 11,
+      "component_count": 17,
+      "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": "917e05cd0f2a64692dddcb8f1ab6881e1bf003f1a687b70907bc634d3b286635",
+      "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 issues look like unrelated spam/placeholder entries plus one vague documentation request; there is no concrete evidence they describe the same underlying bug or change.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "Issue 44418 is the most substantive of the set, with a non-spam title suggesting an actual documentation request, while the others are generic spam/placeholder posts.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 44418 is the best representative issue in the cluster because it is at least semantically meaningful; the others are clearly low-signal spam or off-topic titles.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44371",
+            "right": "issue:44572",
+            "accept": false,
+            "reason": "Both are generic '<spam>' placeholders; no concrete bug, change, or shared detail to justify duplicate linkage."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "Both are generic '<spam>' placeholders with empty bodies; identical spam text is not enough to establish the same underlying issue."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44418",
+            "accept": false,
+            "reason": "One is spam and the other is a vague documentation suggestion; they do not describe the same problem."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "The titles are unrelated spam vs. a site-style prompt; no shared bug or actionable change is evident."
+          },
+          {
+            "left": "issue:44418",
+            "right": "issue:44572",
+            "accept": false,
+            "reason": "A documentation enhancement request does not match a spam placeholder issue."
+          },
+          {
+            "left": "issue:44418",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "A documentation enhancement request does not match a spam placeholder issue."
+          },
+          {
+            "left": "issue:44572",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "Although the titles match exactly, both are empty spam placeholders and do not provide enough evidence of the same underlying bug/change."
+          },
+          {
+            "left": "issue:44572",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "Spam placeholder vs. unrelated branded prompt; no concrete overlap in issue content."
+          },
+          {
+            "left": "issue:44573",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "Spam placeholder vs. unrelated branded prompt; no concrete overlap in issue content."
+          }
+        ]
+      },
+      "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": 17,
+      "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": "c3722fed8b89e350fc33eb1a0439d550bff7d2a63bbad35ec94f601354544e66",
+      "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": "These two closed issues are highly similar and likely duplicates: they have the same title, very close body lengths, and a soft-similarity signal, but the available evidence does not prove their bodies are identical. #45145 remains the best canonical issue because it was opened earlier and already has an inbound reference.",
+        "confidence": 0.89,
+        "canonical_issue_reason": "#45145 is earlier and already has an inbound reference, so it is the better canonical issue for this likely-duplicate pair.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#45145 best represents the underlying report because it predates #45151 and appears to be the original submission.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45145",
+            "right": "issue:45151",
+            "accept": true,
+            "reason": "Likely duplicate: same title, very close body size, and soft-similarity support the same underlying report. The evidence is strong, but not enough to claim an exact duplicate."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The conclusion is a bit overstated for the evidence available. The packet supports that these issues are highly similar (same title, close body lengths, soft_similarity), but it does not establish that they are exact duplicates or that the bodies/content are the same. The canonical-choice reasoning is fine, but the duplicate verdict should stay conservative and say \"likely duplicate\" rather than \"exact duplicate.\""
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 13,
+      "component_count": 17,
+      "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": 17,
+      "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": "fb6c3949747a04aef77cd2a107568ce6e21d1796518b38d191ab8cac74cc3cd9",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1669,
+        "estimated_input_tokens": 418,
+        "estimated_eval_tokens": 1092
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1669,
+        "estimated_input_tokens": 418,
+        "estimated_eval_tokens": 1092
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues both involve 4-bit inference with BitsAndBytes, but they describe different failure modes: one is a Gemma4 audio-module uint8/torch.finfo TypeError, the other is a BitsAndBytesConfig lm_head exclusion/assertion regression. They are related by subsystem, not the same underlying bug.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "Issue #45674 is the better canonical issue because it is more broadly framed around a specific BitsAndBytesConfig regression, has more discussion activity and inbound references, and describes a clearer configuration-driven failure path.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #45674 is the best representative for the cluster since it appears more central and actionable, while #45672 is a model/module-specific manifestation that does not match the same defect.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45672",
+            "right": "issue:45674",
+            "accept": false,
+            "reason": "Both concern 4-bit inference, but the concrete bugs differ: Gemma4 audio modules hit a torch.finfo TypeError on uint8 weights, while the other issue is about llm_int8_skip_modules removing the default lm_head exclusion and triggering an AssertionError. Too dissimilar to be duplicates."
+          }
+        ]
+      },
+      "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": 17,
+      "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": "7f3591a82ecb08fbd76066debc5adbc149a144f00c90e8c5a6e43369076ff2de",
+      "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": "These two issues do not look like duplicates; they are separate, unrelated low-content reports with different personal-name titles and dates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45762",
+            "right": "issue:45763",
+            "accept": false,
+            "reason": "Only weak textual similarity from the shared surname/date pattern; there is no evidence they describe the same underlying bug or request."
+          }
+        ]
+      },
+      "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": 17,
+      "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": "629f3f0d7104d0f39058a9afd252c277b126c748583f3b5979bd5e34c7bfb8f8",
+      "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": "Two closed issues with empty bodies and unrelated personal-name/date titles; there is no evidence they describe the same underlying bug or request.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "No real canonical issue emerges here: the titles point to different names/dates, and there is no description, linkage, or shared context to unify them.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Neither issue is a strong canonical choice; if forced to pick one as a placeholder, #45764 is only marginally earlier, not substantively better.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45764",
+            "right": "issue:45765",
+            "accept": false,
+            "reason": "Reject: the only similarity is superficial string overlap in surname/date formatting; the titles refer to different individuals and provide no shared bug or task context."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 17,
+      "component_count": 17,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45869-5",
+      "nodes": [
+        "issue:45869",
+        "issue:45870",
+        "issue:45871",
+        "issue:45872",
+        "issue:45873"
+      ],
+      "soft_pairs": [
+        "issue:45869|issue:45870",
+        "issue:45869|issue:45871",
+        "issue:45870|issue:45871",
+        "issue:45870|issue:45872",
+        "issue:45870|issue:45873",
+        "issue:45871|issue:45872",
+        "issue:45871|issue:45873",
+        "issue:45872|issue:45873"
+      ],
+      "prepared_review_unit_hash": "4297ba6904697ca57d12cd03910a9c18285e68ad0551c60af44d8864e1cca0a1",
+      "component_budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 8,
+        "serialized_chars": 5723,
+        "estimated_input_tokens": 1431,
+        "estimated_eval_tokens": 3118
+      },
+      "budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 8,
+        "serialized_chars": 5723,
+        "estimated_input_tokens": 1431,
+        "estimated_eval_tokens": 3118
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster splits into two very likely duplicate issue pairs: the two \"Example: Financial time series prediction with DataForge\" issues and the two \"DataForge: Financial time series + scraping service\" issues. The dataset promo issue is only loosely related and should not be treated as a duplicate of either pair. There is no single issue that cleanly represents the whole cluster.",
+        "confidence": 0.95,
+        "canonical_issue_reason": "No single issue cleanly represents the whole cluster because it contains two separate likely-duplicate pairs plus a loosely related promo-style post. If a representative is needed, issue 45870 is a reasonable canonical for the \"Example: Financial time series prediction with DataForge\" pair because 45871 matches it in title and body size, but that only covers one subcluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45870 is the best overall issue representative because it sits in a clear likely-duplicate pair and is more self-contained than the dataset promo issue or the broader \"scraping service\" request. It is still only a representative for one subcluster, not the whole cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45869",
+            "right": "issue:45870",
+            "accept": false,
+            "reason": "Different subjects: a dataset promotion versus an example/tutorial request. They are only loosely related by finance/data context."
+          },
+          {
+            "left": "issue:45869",
+            "right": "issue:45871",
+            "accept": false,
+            "reason": "Different subjects: a dataset promotion versus an example/tutorial request. Similar wording is not enough to make them the same issue."
+          },
+          {
+            "left": "issue:45870",
+            "right": "issue:45871",
+            "accept": true,
+            "reason": "Same title and same body length, with only soft-similarity evidence; this is very likely the same underlying example request duplicated."
+          },
+          {
+            "left": "issue:45870",
+            "right": "issue:45872",
+            "accept": false,
+            "reason": "Both mention financial time series and DataForge, but one is an example request and the other adds a scraping service. Those are distinct asks."
+          },
+          {
+            "left": "issue:45870",
+            "right": "issue:45873",
+            "accept": false,
+            "reason": "Related theme, but the concrete request differs from the example issue. Not the same underlying issue."
+          },
+          {
+            "left": "issue:45871",
+            "right": "issue:45872",
+            "accept": false,
+            "reason": "Related DataForge/financial content, but different concrete requests. Not the same bug or change."
+          },
+          {
+            "left": "issue:45871",
+            "right": "issue:45873",
+            "accept": false,
+            "reason": "Related DataForge/financial content, but different concrete requests. Not the same bug or change."
+          },
+          {
+            "left": "issue:45872",
+            "right": "issue:45873",
+            "accept": true,
+            "reason": "Same title and same body length, with only soft-similarity evidence; this is very likely the same underlying request duplicated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The cluster interpretation is mostly grounded, but a few phrases are a bit too strong for the packet. The two same-title pairs are likely duplicates, but the evidence shown is only identical titles/body lengths plus soft-similarity, so avoid calling them an \"exact\" or \"perfect\" duplicate pair as if that were proven. Also, \"dataset promo issue is unrelated\" is slightly overstated; the safer wording is that it is only loosely related and not a duplicate."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 1,
+      "component_count": 27,
+      "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-39895-15",
+      "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",
+        "pull_request:45919",
+        "pull_request:45922"
+      ],
+      "soft_pairs": [
+        "pull_request:39895|pull_request:43424",
+        "pull_request:41855|pull_request:43823",
+        "pull_request:41855|pull_request:43424",
+        "pull_request:42668|pull_request:43823",
+        "pull_request:43251|pull_request:43823",
+        "pull_request:43251|pull_request:44827",
+        "pull_request:43424|pull_request:45919",
+        "pull_request:43424|pull_request:44827",
+        "pull_request:43424|pull_request:43823",
+        "pull_request:43823|pull_request:45655",
+        "pull_request:43823|pull_request:45922",
+        "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:44827|pull_request:45922",
+        "pull_request:43251|pull_request:43254"
+      ],
+      "prepared_review_unit_hash": "96ea4833ea1cb2ce921e5550a0c00cd4c21fe491dcdb3813e0833889bd189ec9",
+      "component_budget": {
+        "node_count": 15,
+        "item_count": 15,
+        "soft_pair_count": 22,
+        "serialized_chars": 42217,
+        "estimated_input_tokens": 10555,
+        "estimated_eval_tokens": 21366
+      },
+      "budget": {
+        "node_count": 15,
+        "item_count": 15,
+        "soft_pair_count": 22,
+        "serialized_chars": 42217,
+        "estimated_input_tokens": 10555,
+        "estimated_eval_tokens": 21366
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The cluster is mostly unrelated PRs that only share broad Transformer-maintenance themes. The only strong duplicate pair is the two `fixed_cross_entropy` PRs, which address the same loss-helper change for the same underlying `cross_entropy` kwargs support.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "`pull_request:43251` is the better canonical PR because it directly targets issue 43240, clearly states the bug fix, and is the more complete implementation of the shared `fixed_cross_entropy` change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "`pull_request:43251` is the best representative of the duplicate pair: same file, same bug, same functional goal, and it has the explicit tracking issue link. The other PR (`43254`) is a near-duplicate variant of the same fix.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:39895",
+            "right": "pull_request:43424",
+            "accept": false,
+            "reason": "Unrelated PRs: one adds VideoPrism, the other adds an Executorch dynamic-shape test. No shared code-path or concrete bug fix."
+          },
+          {
+            "left": "pull_request:41855",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Different model/task areas: Mistral tokenizer helper methods vs adding MobileLLM. Shared tokenization context is too broad to be the same change."
+          },
+          {
+            "left": "pull_request:41855",
+            "right": "pull_request:43424",
+            "accept": false,
+            "reason": "Tokenizer utility additions and an Executorch export test are unrelated changes."
+          },
+          {
+            "left": "pull_request:42668",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Both touch processor/model infrastructure, but they target different models and different problems. Not the same underlying bug or change."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Loss-helper kwargs support and adding MobileLLM are unrelated; no shared code-path."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "Both are small fixes, but one is cross-entropy kwargs handling and the other is Mistral4 test cleanup. Different code paths and goals."
+          },
+          {
+            "left": "pull_request:43424",
+            "right": "pull_request:45919",
+            "accept": false,
+            "reason": "Executorch dynamic-shape export test vs adding Sapiens2 model. Not a duplicate change."
+          },
+          {
+            "left": "pull_request:43424",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "Different concerns: dynamic-shape export testing versus Mistral4 test fixes. No shared underlying bug."
+          },
+          {
+            "left": "pull_request:43424",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Executorch export test and MobileLLM model addition are unrelated."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:45655",
+            "accept": false,
+            "reason": "New model addition vs config resolution order fix. Too broad and unrelated."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:45922",
+            "accept": false,
+            "reason": "MobileLLM model addition and vision-model memory-leak fix are unrelated changes."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:43995",
+            "accept": false,
+            "reason": "Adding MobileLLM and refactoring Falcon output collection are different model-specific changes."
+          },
+          {
+            "left": "pull_request:43995",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "Falcon output-interface refactor and a model-creation skill are unrelated."
+          },
+          {
+            "left": "pull_request:43995",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "Falcon model refactor and Mistral4 test fixes do not address the same bug or code path."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45170",
+            "accept": false,
+            "reason": "Mistral4 test fixes and CLIP-like `layrnorm`\u2192`layernorm` renaming are unrelated."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45795",
+            "accept": false,
+            "reason": "Mistral4 test fixes and docs for audio/video processors are not the same change."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45655",
+            "accept": false,
+            "reason": "Mistral4 test fixes and config resolution ordering are unrelated maintenance changes."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "Mistral4 test fixes and a model-creation skill PR are unrelated."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45922",
+            "accept": false,
+            "reason": "Mistral4 test fixes and vision-model memory-leak fixes are different issues in different subsystems."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:43254",
+            "accept": true,
+            "reason": "Same underlying fix in the same file: both update `fixed_cross_entropy` to support extra `nn.functional.cross_entropy` kwargs such as `weight` and `label_smoothing`. They could plausibly be merged into one PR."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The duplicate judgment is well grounded in the packet. The only accepted pair, `pull_request:43251` and `pull_request:43254`, clearly targets the same `fixed_cross_entropy` change in the same file and the same issue (`43240`), so mergeability as one bugfix is plausible. The other soft pairs are conservatively rejected and the summary does not overstate the evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 2,
+      "component_count": 27,
+      "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": "75cb1bd96934f1d1dddaa97af0355bbf0daf8bdcf072e0e20ba3f047fc1fa9de",
+      "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": "The two PRs overlap heavily on VibeVoice model scaffolding, but they are not duplicates: 40546 implements the base VibeVoice model, while 42781 extends it with realtime/acoustic-tokenizer work and extra generation changes.",
+        "confidence": 0.88,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 40546 is the core standalone VibeVoice implementation and has the stronger review history; 42781 is a later realtime extension rather than the same change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 40546 is the best representative of the cluster because it covers the primary VibeVoice feature without the additional realtime follow-on scope.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:40546",
+            "right": "pull_request:42781",
+            "accept": false,
+            "reason": "They share the VibeVoice subsystem, but 42781 adds realtime/acoustic-tokenizer functionality and extra generation changes, so it is an extension rather than the same concrete 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": 3,
+      "component_count": 27,
+      "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": "e4b37e2b939feb7dd563df270c79ee68a22e6817d18608b68d582897629bc888",
+      "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": "False positive cluster: one PR adds TimesFM covariate forecasting support, while the other adds DINOv3 image-classification support. They touch different models, code paths, and user-facing features.",
+        "confidence": 0.99,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No canonical PR here: these are unrelated feature PRs for different model families and should not be deduplicated into one cluster.",
+        "best_issue_reason": null,
+        "best_pr_reason": "No global best PR for the cluster; each PR is independently valid but they solve different problems and belong in separate clusters.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:40755",
+            "right": "pull_request:41224",
+            "accept": false,
+            "reason": "Different model families and changesets: TimesFM covariate forecasting vs DINOv3 image classification. Not the same underlying bug or mergeable 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": 4,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-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": 27,
+      "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": "2ec9a732cb7a092f672c42e2ea003f88ca1d890bacfc8aa0b90aa6746e991cdf",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 10408,
+        "estimated_input_tokens": 2602,
+        "estimated_eval_tokens": 5460
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 10408,
+        "estimated_input_tokens": 2602,
+        "estimated_eval_tokens": 5460
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Three pull requests add different model integrations (DeepSeek OCR, Isaac, Kimi2-6). They overlap in boilerplate registration/docs, but they do not fix the same underlying code-path problem, so they should not be merged as duplicates.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45630 is the strongest representative of the cluster: it has the broadest integration footprint and the highest review activity, but it is still a distinct Kimi2-6 model addition rather than the same change as the others.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45630 is the best single representative because it is the most developed and actively reviewed of the three, making it the most useful anchor if this cluster needs one canonical PR.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:41797",
+            "right": "pull_request:45186",
+            "accept": false,
+            "reason": "Both are new model additions, but they target different models and different implementation stacks (DeepSeek OCR vs Isaac). Shared auto-registration/docs files are boilerplate and not evidence of the same fix."
+          },
+          {
+            "left": "pull_request:45186",
+            "right": "pull_request:45630",
+            "accept": false,
+            "reason": "These are separate model integrations for Isaac and Kimi2-6. They overlap in model registry plumbing, but they do not address the same concrete bug or change and would not plausibly 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": 6,
+      "component_count": 27,
+      "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": "cdb9eaaf4d1c58320dcd8eb575f4b0c13ad8cb483009f504d858f3c7297b7a7c",
+      "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": "Two open PRs both update GitHub Actions, but they target different dependency families and change sets: one is a broad Node 24 compatibility sweep for workflow actions, while the other upgrades Docker/build-related actions to newer major versions. They overlap in a few workflow files, but they do not fix the same concrete problem.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 42975 is the better cluster representative because it is the broader, explicitly scoped workflow-actions compatibility upgrade and touches many of the same CI workflows.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 42975 is the best single representative of this cluster since it captures the general GitHub Actions upgrade theme more comprehensively than the narrower Docker-action bump in 42976.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:42975",
+            "right": "pull_request:42976",
+            "accept": false,
+            "reason": "Do not merge: these are separate maintenance changes. 42975 upgrades multiple GitHub Actions for Node 24 compatibility, while 42976 primarily bumps docker/build-push-action and a few other actions to latest versions. Shared workflow filenames alone are not enough to treat them as 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": 7,
+      "component_count": 27,
+      "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-42978-7",
+      "nodes": [
+        "pull_request:42978",
+        "pull_request:43448",
+        "pull_request:43451",
+        "pull_request:43973",
+        "pull_request:44601",
+        "pull_request:45490",
+        "pull_request:45838"
+      ],
+      "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",
+        "pull_request:43973|pull_request:45838",
+        "pull_request:44601|pull_request:45838",
+        "pull_request:45490|pull_request:45838"
+      ],
+      "prepared_review_unit_hash": "f23405e4678881cb0d7623cf445a7666a13ad9bd6356a79ca848daaa7b7ecaf2",
+      "component_budget": {
+        "node_count": 7,
+        "item_count": 7,
+        "soft_pair_count": 9,
+        "serialized_chars": 23541,
+        "estimated_input_tokens": 5886,
+        "estimated_eval_tokens": 12028
+      },
+      "budget": {
+        "node_count": 7,
+        "item_count": 7,
+        "soft_pair_count": 9,
+        "serialized_chars": 23541,
+        "estimated_input_tokens": 5886,
+        "estimated_eval_tokens": 12028
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster is mostly a set of unrelated PRs that happen to share Transformers boilerplate (docs TOC, auto registries, model package wiring). They cover different model families and one separate distributed-loading feature, so the soft similarities are false positives rather than duplicates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45838 is the largest self-contained model-support addition in the set, but it is not actually a duplicate of the others; the cluster does not have a true canonical PR.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:43448",
+            "accept": false,
+            "reason": "ViT NEPA vs Molmo are different model-addition PRs; overlap is only shared docs/auto wiring."
+          },
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:43451",
+            "accept": false,
+            "reason": "ViT NEPA vs Molmo2 are separate model implementations, not the same underlying change."
+          },
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "ViT NEPA and CTSM are distinct model-support PRs with unrelated codepaths."
+          },
+          {
+            "left": "pull_request:43448",
+            "right": "pull_request:43451",
+            "accept": false,
+            "reason": "Molmo and Molmo2 are related in name only; they add different model families and different support code."
+          },
+          {
+            "left": "pull_request:43448",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "Molmo and CTSM are unrelated model additions; shared registry/docs edits are boilerplate."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "Molmo2 and CTSM are different models with different modeling/configuration changes."
+          },
+          {
+            "left": "pull_request:43973",
+            "right": "pull_request:45838",
+            "accept": false,
+            "reason": "Lfm2 Audio and PP-OCRv6 text recognition are different modalities and different feature/model stacks."
+          },
+          {
+            "left": "pull_request:44601",
+            "right": "pull_request:45838",
+            "accept": false,
+            "reason": "Pipeline-parallel loading support is a distributed-systems change, not the same bug/change as PP-OCRv6 model support."
+          },
+          {
+            "left": "pull_request:45490",
+            "right": "pull_request:45838",
+            "accept": false,
+            "reason": "CTSM and PP-OCRv6 are separate model additions; only generic docs/auto files overlap."
+          }
+        ]
+      },
+      "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": 27,
+      "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-43291-4",
+      "nodes": [
+        "pull_request:43291",
+        "pull_request:43665",
+        "pull_request:44815",
+        "pull_request:45981"
+      ],
+      "soft_pairs": [
+        "pull_request:43291|pull_request:43665",
+        "pull_request:43291|pull_request:44815",
+        "pull_request:43291|pull_request:45981",
+        "pull_request:43665|pull_request:44815",
+        "pull_request:43665|pull_request:45981",
+        "pull_request:44815|pull_request:45981"
+      ],
+      "prepared_review_unit_hash": "bf9a4f5f20e7a18001cb024aa147875a5af1f87dbd7f4889f1876000048ba476",
+      "component_budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 9635,
+        "estimated_input_tokens": 2409,
+        "estimated_eval_tokens": 5074
+      },
+      "budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 9635,
+        "estimated_input_tokens": 2409,
+        "estimated_eval_tokens": 5074
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "All four PRs are only loosely similar at the title level; they do not appear to fix the same underlying bug or change. The only substantive code fix is the FP8 dequant/loading PR, while the others are unrelated test-only changes or Whisper tokenizer work.",
+        "confidence": 0.9,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "pull_request:44815 is the strongest representative because it contains the only concrete production code fix, centered on FP8 dequant/loading behavior, with related tests.",
+        "best_issue_reason": null,
+        "best_pr_reason": "pull_request:44815 is the best global representative of the cluster\u2019s substantive work; the other PRs are either unrelated model tests or Whisper-specific test/tokenization changes.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:43665",
+            "accept": false,
+            "reason": "Different targets and symptoms: Whisper tokenizer decoding vs CLIP/ViT test edits. Same broad \"tests\" theme only."
+          },
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:44815",
+            "accept": false,
+            "reason": "Unrelated domains and code paths: Whisper decode logic vs FP8 dequant/loading internals."
+          },
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:45981",
+            "accept": false,
+            "reason": "Whisper tokenizer changes are unrelated to a ColQwen2 test device fix."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:44815",
+            "accept": false,
+            "reason": "43665 is test-only and spans CLIP/ViT; 44815 is a production FP8 fix. They do not address the same bug."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:45981",
+            "accept": false,
+            "reason": "Both are test-related, but they touch different model test suites and different failure modes."
+          },
+          {
+            "left": "pull_request:44815",
+            "right": "pull_request:45981",
+            "accept": false,
+            "reason": "FP8 loading/dequantization code changes are unrelated to a ColQwen2 test tensor-device adjustment."
+          }
+        ]
+      },
+      "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": 27,
+      "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": "93b9476c93395f3daf0e7392de20c081bf47dc1e4ef2205b5b3fa34a5685e685",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4562,
+        "estimated_input_tokens": 1141,
+        "estimated_eval_tokens": 2538
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4562,
+        "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 PRs are not duplicates: #43488 is a small repo-bot/check-formatting style change around setup/versioning and a Vit test, while #43532 is a broader workflow/test-harness update affecting CI automation and several files. They do not look like the same underlying change.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #43532 is the more representative standalone change in this cluster because it touches the actual CI/test automation path and spans multiple related files, whereas #43488 is a narrow experimental/check PR.",
+        "best_issue_reason": null,
+        "best_pr_reason": "#43532 is the best PR candidate to represent the cluster because it is more substantive and coherent as a single code-path change; #43488 is too small and unrelated in scope to serve as the canonical representative.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43488",
+            "right": "pull_request:43532",
+            "accept": false,
+            "reason": "Different underlying purposes and code paths: #43488 is about setup/versioning and a Vit test, while #43532 modifies CI workflow logic, testing utilities, and notification paths. Shared Vit test file is not enough to make them 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": 10,
+      "component_count": 27,
+      "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": "4b13690079e2dd099370f78c27d5bb8ab72845d88103288b6d99d6a23b458f2e",
+      "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 separate feature PRs with similar scaffolding but different pipeline targets: promptable concept segmentation vs promptable visual segmentation. Shared docs/registry plumbing is not enough to treat them as the same underlying change.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No canonical PR: the two PRs implement different pipeline features and touch different model integrations, so neither is a duplicate of the other.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 43613 is the broader, more established feature change, but it still represents a different pipeline than 43612 and should not be merged as a duplicate representative.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43612",
+            "right": "pull_request:43613",
+            "accept": false,
+            "reason": "Reject: these add different pipelines and model integrations (concept segmentation vs visual segmentation), so they are not the same concrete change and would not be merged into one PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 11,
+      "component_count": 27,
+      "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": "dddd36c9880d5e7be004f176771629ef9ac40f1da379074e6fa717c9b37a0f16",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4291,
+        "estimated_input_tokens": 1073,
+        "estimated_eval_tokens": 2402
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4291,
+        "estimated_input_tokens": 1073,
+        "estimated_eval_tokens": 2402
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are two separate `trainer.py` pull requests with similar extensibility themes but different code paths: one changes gradient-accumulation loss scaling, the other extracts signature-column logic for overrideability. They should not be treated as duplicates.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 43651 is the more concrete functional change: it alters `training_step` behavior around gradient accumulation loss scaling and adds a dedicated property to control that logic. PR 43663 is a separate refactor/API hook for signature-column filtering.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 43651 is the best representative of the cluster because it targets a specific runtime behavior fix, while PR 43663 is an unrelated trainer extensibility change despite touching the same file.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43651",
+            "right": "pull_request:43663",
+            "accept": false,
+            "reason": "Same file and broad trainer extensibility area, but different underlying changes: GA loss scaling vs. model signature/column filtering. They do not fix the same bug and would not plausibly 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": 12,
+      "component_count": 27,
+      "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": "71930c34305857e475f8b58f5d295e317c9c9027cf8b0a764d7ed78dfc8d96fd",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 5992,
+        "estimated_input_tokens": 1498,
+        "estimated_eval_tokens": 3252
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 5992,
+        "estimated_input_tokens": 1498,
+        "estimated_eval_tokens": 3252
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These PRs cluster around the same CLI `serve.py` runtime `TypeAdapter` annotation NameError. PR 43842 is the cleanest, most targeted fix; the other two are broader or alternate implementations of the same underlying issue.",
+        "confidence": 0.79,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 43842 is the best canonical PR: it is narrowly scoped to the reported `TypeAdapter` NameError, has an explicit issue link, and makes the smallest localized change to `serve.py`.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 43842 most directly fixes the concrete runtime failure with minimal surface area, making it the strongest representative of this bug cluster.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43656",
+            "right": "pull_request:43842",
+            "accept": true,
+            "reason": "Both address the same `serve.py` runtime `TypeAdapter` NameError caused by evaluated annotations; they are alternate fixes for the same code-path problem."
+          },
+          {
+            "left": "pull_request:43836",
+            "right": "pull_request:43842",
+            "accept": true,
+            "reason": "Both are fixing the same `serve.py` annotation-evaluation NameError around `TypeAdapter`; despite different implementation style, they target the same underlying bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: all three PRs touch `src/transformers/cli/serve.py`, two explicitly target the same `TypeAdapter` NameError issue, and PR 43842 is the smallest localized fix with an explicit issue link. The pairwise soft-edge acceptances are conservative enough given the titles, file overlap, and shared TypeAdapter/annotation-evaluation context."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 13,
+      "component_count": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-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": 27,
+      "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: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:44071|pull_request:44085",
+        "pull_request:44068|pull_request:44086",
+        "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: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:44056|pull_request:44101",
+        "pull_request:44073|pull_request:44161",
+        "pull_request:44059|pull_request:44068",
+        "pull_request:44018|pull_request:44059",
+        "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:44018|pull_request:44161",
+        "pull_request:43997|pull_request:44071",
+        "pull_request:44010|pull_request:44071",
+        "pull_request:44018|pull_request:44129",
+        "pull_request:44010|pull_request:44073",
+        "pull_request:43997|pull_request:44074",
+        "pull_request:43997|pull_request:44129",
+        "pull_request:44071|pull_request:44129",
+        "pull_request:44010|pull_request:44018",
+        "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"
+      ],
+      "prepared_review_unit_hash": "bd7d6a06897f5bf88e0cd97b9af0a031b28b5fdbf55607e20ed3f38fbad1a23a",
+      "component_budget": {
+        "node_count": 28,
+        "item_count": 28,
+        "soft_pair_count": 78,
+        "serialized_chars": 84896,
+        "estimated_input_tokens": 21224,
+        "estimated_eval_tokens": 42704
+      },
+      "budget": {
+        "node_count": 27,
+        "item_count": 27,
+        "soft_pair_count": 72,
+        "serialized_chars": 75095,
+        "estimated_input_tokens": 18774,
+        "estimated_eval_tokens": 37804
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a wave of standardized output-tracing refactors across many different model families. Most pairings are not duplicates because they touch different code paths/models; only the GPT-Neo and GPT-J/CodeGen variants clearly overlap as the same underlying refactor.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "Issue #43979 is the shared tracking issue explicitly referenced by the output-tracing refactor PRs, so it is the natural canonical issue for the cluster.",
+        "canonical_pr_reason": "PR #44722 is the strongest representative of the duplicate GPT-J/CodeGen branch: it targets the same output-tracing refactor, shares the key files, and appears to be the most complete version of that concrete change.",
+        "best_issue_reason": "Issue #43979 best represents the cluster because it is the common umbrella for the standardized output-tracing work across models.",
+        "best_pr_reason": "PR #44722 is the best single PR representative because it covers the core GPT-J/CodeGen output-tracing change and most directly reflects the shared refactor direction in the cluster.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "Different model families (CVT/FNet vs GPT-J); shared tracking theme is too broad to be a duplicate."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "Different model families and different concrete code paths; only the output-tracing theme overlaps."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44085",
+            "accept": true,
+            "reason": "Both are GPT-J output-tracing refactors on the same file and same issue; one is just a broader variant of the same underlying change."
+          },
+          {
+            "left": "pull_request:44007",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "RegNet/ResNet-style refactors are separate concrete model implementations, not the same bug or change."
+          },
+          {
+            "left": "pull_request:44072",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "EfficientNet and GPT-J/CodeGen are unrelated code paths despite sharing the standardized-tracing theme."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "GPT-J/CodeGen vs EfficientNet is a broad architectural similarity, not the same underlying change."
+          },
+          {
+            "left": "pull_request:44013",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "MobileNetV2 and DeBERTa are unrelated implementations; this is only a thematic similarity."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-J/CodeGen and MGP-STR affect different model code paths and are not the same fix."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-J/CodeGen and MPT are different model-specific refactors, not one duplicate change."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44068",
+            "accept": true,
+            "reason": "These are two GPT-Neo output-tracing refactors on the same model/file set and same underlying change."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "GPT-J/CodeGen and GPT-Neo are different model families; same refactor pattern is not enough."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44066",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J/CodeGen are separate model implementations, so this is not a duplicate fix."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "ResNet and GPT-J/CodeGen are different concrete model paths; only the refactor theme overlaps."
+          },
+          {
+            "left": "pull_request:44086",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "MGP-STR and GPT-J/CodeGen are unrelated model-specific changes."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "ResNet and GPT-J are different code paths; shared issue linkage is not enough."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44019",
+            "accept": false,
+            "reason": "GPT-Neo and ResNet are different model families and changes."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "ResNet and MPT are unrelated concrete implementations."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44116",
+            "accept": false,
+            "reason": "GPT-Neo and Flaubert are different model-specific output-tracing refactors."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J/CodeGen are different code paths despite similar refactor language."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "MPT and GPT-J/CodeGen are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44116",
+            "accept": false,
+            "reason": "GPT-Neo and Flaubert are different models and separate changes."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo and MGP-STR are different concrete implementations."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J are different model families; not the same underlying bug."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "MPT and MGP-STR are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT2 and MGP-STR are different model-specific refactors."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-J and MGP-STR are unrelated concrete changes."
+          },
+          {
+            "left": "pull_request:44098",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "ViLT and XLM/Flaubert are distinct model implementations; same theme only."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "GPT2 and GPT-J/CodeGen are different code paths; theme overlap is too broad."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-Neo and MPT are separate model-specific refactors."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "SegFormer and GPT-J/CodeGen are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "SegFormer and LongT5 are different models and different concrete changes."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "VisualBert and Flaubert are different model code paths."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "MPNet and LongT5 are unrelated concrete refactors."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44017",
+            "accept": false,
+            "reason": "RegNet and SegFormer are different model families; only the output-tracing theme overlaps."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44056",
+            "accept": false,
+            "reason": "RegNet and MPNet are different code paths, not the same bug or change."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "MPT and ImageGPT are different model-specific changes."
+          },
+          {
+            "left": "pull_request:44074",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "TextNet and Flaubert are unrelated implementations."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44059",
+            "accept": false,
+            "reason": "GPT-Neo and GPT2 are separate model families."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "SegFormer and MPT are different concrete changes."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "SegFormer and TextNet are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "GPT2 and GPT-Neo are different model families and code paths."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT2 and MPT are unrelated concrete changes."
+          },
+          {
+            "left": "pull_request:44101",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "Flaubert/XLM and LongT5 are different model implementations."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "RegNet and TextNet are unrelated model-specific refactors."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "RegNet and GPT-Neo are different code paths."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "SqueezeBert and DeBERTa are different model families and refactors."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "DeBERTa and TextNet are unrelated concrete changes."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "GPT-Neo and SpeechT5 are different model code paths."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "SqueezeBert and VisualBert are different implementations."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "SqueezeBert and GPT-Neo are different model families."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "CVT/FNet and LongT5 are unrelated concrete changes."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44017",
+            "accept": false,
+            "reason": "CVT/FNet and SegFormer are different code paths."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "SqueezeBert and LongT5 are unrelated model-specific changes."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "DeBERTa and SpeechT5 are different model implementations."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44154",
+            "accept": false,
+            "reason": "DeBERTa and ViTS are unrelated concrete refactors."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet, and the duplicate/near-duplicate calls stay conservative. The selected canonical issue/PR are supported by shared issue linkage and overlapping GPT-J/CodeGen files, and the false edges are mostly justified by distinct model families or code paths."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 14,
+      "component_count": 27,
+      "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:44024",
+        "pull_request:44056",
+        "pull_request:44073",
+        "pull_request:44074",
+        "pull_request:44076",
+        "pull_request:44129"
+      ],
+      "soft_pairs": [
+        "pull_request:44024|pull_request:44076",
+        "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": "57e47d7c258906794f23d431761dea9f72023560b4f726b2dd09d2dca4d80dfd",
+      "component_budget": {
+        "node_count": 28,
+        "item_count": 28,
+        "soft_pair_count": 78,
+        "serialized_chars": 84896,
+        "estimated_input_tokens": 21224,
+        "estimated_eval_tokens": 42704
+      },
+      "budget": {
+        "node_count": 8,
+        "item_count": 8,
+        "soft_pair_count": 6,
+        "serialized_chars": 17122,
+        "estimated_input_tokens": 4281,
+        "estimated_eval_tokens": 8818
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These are all model-specific PRs modernizing output tracing to use capture_outputs/can_return_tuple patterns. They share a theme, but each changes a different model implementation, so they are not duplicates of the same underlying fix/change.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 44129 is the best representative of the cluster because it has an explicit issue target and shows a concrete, self-contained output-tracing refactor for one model; the others are similar but are separate per-model variants.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 44129 is the strongest standalone PR in the set: it is clearly scoped, tied to a target issue, and implements the standardized output-capture migration in a concrete code path.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44024",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "Same refactor theme, but different models and code paths (FocalNet vs ImageGPT); not the same underlying change."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "Both modernize output tracing, but they touch different model implementations (VisualBert vs TextNet) and are not mergeable as one PR."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "Related output-capture refactors, but MPNet and VisualBert are separate code paths with distinct implementation details."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "Same broad standardization effort, but MPNet and TextNet are different model-specific changes."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44010",
+            "accept": false,
+            "reason": "Both are output-tracing migrations, but RegNet and SqueezeBert are unrelated model implementations and not the same fix."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "Different architectures and files (SqueezeBert vs SpeechT5); they share a modernization pattern but not the same concrete bug/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": 15,
+      "component_count": 27,
+      "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": "e0fc149cba7cdefdaf54a1a3d2e515a6b8cb5e2a8bf1bbc1c7ac5233c6d053aa",
+      "component_budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 40,
+        "serialized_chars": 40178,
+        "estimated_input_tokens": 10045,
+        "estimated_eval_tokens": 20346
+      },
+      "budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 40,
+        "serialized_chars": 40178,
+        "estimated_input_tokens": 10045,
+        "estimated_eval_tokens": 20346
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These PRs are all variants of the same broad refactor theme ('output tracing'), but they target different model implementations and forward paths. I do not see any pair that looks like the same concrete bug/change, so none should be merged as duplicates.",
+        "confidence": 0.87,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 44025 is the cleanest representative of the cluster: it applies the tracing decorator pattern to two closely related depth-anything model files and stays narrowly scoped.",
+        "best_issue_reason": null,
+        "best_pr_reason": "44025 is the best overall representative because it is self-contained, model-specific, and shows the core output-tracing refactor with minimal extra scope.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Different model families and code paths (timm_backbone vs upernet); only the refactor pattern is shared."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44000",
+            "accept": false,
+            "reason": "Different architectures and forward implementations; same tracing cleanup theme only."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Separate model code paths; not the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "timm_backbone and mamba/falcon_mamba are unrelated concrete code paths despite similar refactor wording."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Different model families; no evidence of one shared fix."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "MobileNetV1 and UnivNet are unrelated implementations; the similarity is only at the refactor level."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Different model code paths and behavior; not the same change."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "MobileNetV1 vs DepthAnything are distinct models; same output-tracing pattern is too broad to duplicate."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44026",
+            "accept": false,
+            "reason": "Different encoder-decoder targets and code paths; only the wrapper refactor theme matches."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "MobileNetV1 and SpeechEncoderDecoder are separate model implementations with no shared concrete bug."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Vision-text dual encoder and UnivNet are unrelated code paths; shared wording is not enough."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Different model families and distinct forward logic."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "The PRs touch different architectures and fix different model-specific tracing paths."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "Different models and output pathways; not the same bug/change."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "UnivNet and UperNet are unrelated; only the refactor pattern is common."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "Different model families and implementation details; not mergeable as one fix."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "Different architectures; same tracing cleanup pattern is too generic."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "UperNet and RWKV are separate code paths with no concrete shared bug."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "Different model subsystems and different forward behavior."
+          },
+          {
+            "left": "pull_request:44003",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Falcon/Mamba vs SuperPoint are unrelated model paths; no same underlying change."
+          },
+          {
+            "left": "pull_request:44003",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Mamba-family refactor and CodeGen refactor are different concrete code paths."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "TimmBackbone and CodeGen are unrelated implementations; same broad refactor theme only."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "UnivNet and CodeGen touch different model logic and cannot be treated as one duplicate fix."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "UperNet and CodeGen are unrelated model paths."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44026",
+            "accept": false,
+            "reason": "DepthAnything and VisionEncoderDecoder are different model wrappers; same decorator pattern is not enough."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "DepthAnything and SpeechEncoderDecoder are distinct code paths and model families."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "DepthAnything vs SuperPoint: different architectures and different output handling details."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "DepthAnything and RWKV are unrelated concrete changes."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "DepthAnything and DPR are separate model implementations; no shared underlying bug."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "VisionEncoderDecoder and RWKV are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "VisionEncoderDecoder and SpeechEncoderDecoder are different wrapper implementations, not the same bug."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Different model families and forward paths; shared refactor style only."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "VisionEncoderDecoder and DPR are unrelated concrete changes."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "SpeechEncoderDecoder and SuperPoint are distinct model code paths."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Different model families and output semantics; not the same fix."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "SpeechEncoderDecoder and DPR are unrelated implementations."
+          },
+          {
+            "left": "pull_request:44028",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "SuperPoint and RWKV are unrelated model paths."
+          },
+          {
+            "left": "pull_request:44028",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "SuperPoint and DPR do not share a concrete bug/change."
+          },
+          {
+            "left": "pull_request:44029",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "RWKV and DPR are unrelated model implementations; same broad refactor wording is insufficient."
+          }
+        ]
+      },
+      "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": 27,
+      "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": "4b9104eea730c11d022d34a781a6ceafe7f5a03da7af449298ede5be27a17716",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 10911,
+        "estimated_input_tokens": 2728,
+        "estimated_eval_tokens": 5712
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 10911,
+        "estimated_input_tokens": 2728,
+        "estimated_eval_tokens": 5712
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are three unrelated model-addition PRs (PenguinVL, DeepSeek-OCR-2, Voxtral TTS). They share only generic Transformers plumbing and docs registration, not the same underlying bug or change.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No true canonical PR for deduping here; the cluster is mixed. If one must be singled out, DeepSeek-OCR-2 is the most complete standalone implementation and not draft.",
+        "best_issue_reason": null,
+        "best_pr_reason": "DeepSeek-OCR-2 is the strongest overall candidate: non-draft, substantial model + conversion + test coverage, and appears more merge-ready than the draft Voxtral TTS PR.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44662",
+            "right": "pull_request:45401",
+            "accept": false,
+            "reason": "Different models and different additions (PenguinVL vs Voxtral TTS). Shared auto/docs files are generic scaffolding, not the same code-path fix."
+          },
+          {
+            "left": "pull_request:44662",
+            "right": "pull_request:45075",
+            "accept": false,
+            "reason": "Different model implementations (PenguinVL vs DeepSeek-OCR-2) with separate architecture, processing, and test code; only common Transformers registration plumbing overlaps."
+          },
+          {
+            "left": "pull_request:45075",
+            "right": "pull_request:45401",
+            "accept": false,
+            "reason": "Different model families and unrelated functionality (DeepSeek-OCR-2 vs Voxtral TTS). Similar filenames are broad integration points, not evidence of one mergeable change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 17,
+      "component_count": 27,
+      "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": "353f85c8de0b7494a87319632d23f91a56181bd3fd07362d0755208ebfc4018f",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4407,
+        "estimated_input_tokens": 1102,
+        "estimated_eval_tokens": 2460
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4407,
+        "estimated_input_tokens": 1102,
+        "estimated_eval_tokens": 2460
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are related around fusion, but they target different mechanisms: one introduces a generic Module Fusion API, while the other implements n-to-1 kernel fusion through KernelConfig and integration changes. They are not the same underlying change and should not be merged as duplicates.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45363 is the more concrete end-user feature: it wires kernel fusion into model loading/integration and includes an example. PR 44979 is a separate API layer for module chaining/fusion, so 45363 is the better representative if one PR must be selected.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44979",
+            "right": "pull_request:45363",
+            "accept": false,
+            "reason": "Both involve fusion, but they address different code paths and abstractions. 44979 adds a generic module-fusion API in a new module, while 45363 adds KernelConfig-driven kernel fusion and integration hooks. They could coexist, but they are 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": "pull_request",
+      "component_index": 18,
+      "component_count": 27,
+      "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": "2bb7317ae619e4e6dbe73d4f9ebccf463c69e4a2add67e84e2a1117d56b78e00",
+      "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 PRs target the same Mistral regex patch bug in the same file; 45317 is the stronger candidate because it includes the same fix plus tests and a clearer scope.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45317 is the best canonical PR: it fixes the same AttributeError as 45086, has the same code change, and adds test coverage for the behavior.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45317 is the best overall PR because it is the same fix as 45086 but more complete, with accompanying tests and clearer intent.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45086",
+            "right": "pull_request:45317",
+            "accept": true,
+            "reason": "Same underlying AttributeError fix in _patch_mistral_regex on the same code path; 45317 is effectively the more complete version of 45086 and they could plausibly be merged as one PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 19,
+      "component_count": 27,
+      "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": "0fc5e678f311b76f1b7f3e3403137bd96f580f7492baaa76dc099ada7873cecb",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6066,
+        "estimated_input_tokens": 1517,
+        "estimated_eval_tokens": 3290
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6066,
+        "estimated_input_tokens": 1517,
+        "estimated_eval_tokens": 3290
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs touch NemotronH internals, but they are different changes: #45115 is a refactor/conversion-mapping inheritance cleanup, while #45569 is a broader model split into NemotronHDense/NemotronHSparse with docs and tests. They are related but not the same underlying change.",
+        "confidence": 0.91,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45569 is the broader, more representative NemotronH change: it introduces the dense/sparse model split, updates auto-mapping/docs, and includes tests. It better captures the main architectural change in this cluster.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45569 is the best single representative because it is the more complete and user-visible NemotronH update, whereas #45115 is a narrower refactor focused on conversion mapping and inheritance details.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45115",
+            "right": "pull_request:45569",
+            "accept": false,
+            "reason": "They share NemotronH files, but the concrete changes differ: #45115 refactors inheritance/conversion mappings, while #45569 splits NemotronH into dense/sparse variants with docs and tests. Not the same code-path fix and not plausibly one merged 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": 27,
+      "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": "5069797d27ef4cf3ab18fdd602a819f9ce97aeda6762560bc7483f40b572ed1b",
+      "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 open PRs for heterogeneous model support: one broad end-to-end implementation and one smaller config-only foundation. They are related, but not the same concrete change.",
+        "confidence": 0.91,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45332 is the broader umbrella change: it includes heterogeneous config handling plus model/cache/masking and GPT-OSS integration, so it best represents the overall feature effort.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45332 is the best representative PR because it covers the full heterogeneous-model support stack and subsumes the narrower config-only work in #45333.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45332",
+            "right": "pull_request:45333",
+            "accept": false,
+            "reason": "Related feature slices, but not the same concrete change. #45333 is config-only per-layer configuration support, while #45332 extends into modeling/cache/masking and model integration; they could ship independently and are not safe to treat 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": 21,
+      "component_count": 27,
+      "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": "44f2fb48ed7693da5d93ccff5fc9274216dc7986da13657706bfbe31082d895f",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5514,
+        "estimated_input_tokens": 1379,
+        "estimated_eval_tokens": 3014
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5514,
+        "estimated_input_tokens": 1379,
+        "estimated_eval_tokens": 3014
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are related typing-focused PRs, but they are not duplicates: #45415 is a broad repo-wide type-checking cleanup and config change, while #45425 adds specific typing protocols/aliases and updates a few concrete call sites. They could be merged independently.",
+        "confidence": 0.89,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "#45425 is the more focused and concrete change: it adds missing typing abstractions and updates the affected implementation sites, rather than making a broad type-checking sweep.",
+        "best_issue_reason": null,
+        "best_pr_reason": "#45425 is the better representative PR for this cluster because it targets a specific typing gap and is less sprawling than the large type-checking enablement in #45415.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45415",
+            "right": "pull_request:45425",
+            "accept": false,
+            "reason": "Shared typing-related files are incidental; #45415 is a broad type-checking/config cleanup, while #45425 introduces specific typing protocols and aliases for modeling_utils/peft-related code. They do not look like the same underlying change and should not be merged as duplicates."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 22,
+      "component_count": 27,
+      "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-45487-6",
+      "nodes": [
+        "pull_request:45487",
+        "pull_request:45817",
+        "pull_request:45843",
+        "pull_request:45858",
+        "pull_request:45890",
+        "pull_request:45926"
+      ],
+      "soft_pairs": [
+        "pull_request:45487|pull_request:45843",
+        "pull_request:45487|pull_request:45926",
+        "pull_request:45817|pull_request:45843",
+        "pull_request:45858|pull_request:45890",
+        "pull_request:45858|pull_request:45926",
+        "pull_request:45890|pull_request:45926"
+      ],
+      "prepared_review_unit_hash": "df0d4c7d66cd44a67337ae5153d24069ca9e38a7a7ea64bac87e0b938c7e8512",
+      "component_budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 6,
+        "serialized_chars": 14717,
+        "estimated_input_tokens": 3680,
+        "estimated_eval_tokens": 7616
+      },
+      "budget": {
+        "node_count": 6,
+        "item_count": 6,
+        "soft_pair_count": 6,
+        "serialized_chars": 14717,
+        "estimated_input_tokens": 3680,
+        "estimated_eval_tokens": 7616
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "No true duplicates here: the cluster only shares a broad theme of model-parallel or XPU-related changes, but the PRs target different models and different code paths. The most representative runtime fix is the AltCLIP/ChineseCLIP device-placement PR.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45487 is the clearest substantive bug fix in model code, addressing device placement for token_type_ids in AltCLIP/ChineseCLIP rather than only adjusting tests or touching unrelated model internals.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45487 is the best representative of the cluster because it fixes an actual model-parallel runtime issue and includes focused tests; the other PRs are either different models, unrelated refactors, or test-only XPU expectation updates.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45487",
+            "right": "pull_request:45843",
+            "accept": false,
+            "reason": "Different model families and different fixes; shared 'model parallel' theme is too broad to treat as the same change."
+          },
+          {
+            "left": "pull_request:45487",
+            "right": "pull_request:45926",
+            "accept": false,
+            "reason": "AltCLIP/ChineseCLIP runtime device-placement fix vs HyperCLOVAX XPU test expectations; not the same underlying bug."
+          },
+          {
+            "left": "pull_request:45817",
+            "right": "pull_request:45843",
+            "accept": false,
+            "reason": "Gemma4 model-parallel helper/device-placement fix is unrelated to the DeIMv2 changes shown here."
+          },
+          {
+            "left": "pull_request:45858",
+            "right": "pull_request:45890",
+            "accept": false,
+            "reason": "Both are test-only XPU expectation updates, but they target different models and different logits, so they are not the same change."
+          },
+          {
+            "left": "pull_request:45858",
+            "right": "pull_request:45926",
+            "accept": false,
+            "reason": "Same broad XPU-testing motif, but different model/test files and different expected outputs; not a duplicate bug fix."
+          },
+          {
+            "left": "pull_request:45890",
+            "right": "pull_request:45926",
+            "accept": false,
+            "reason": "These are separate model-specific test expectation updates, not one shared code-path 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": 27,
+      "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": "a421a3ba3b9725aabbe735eb169bbcfdb1084c88ff77b9b8b7b43c514da36698",
+      "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": "Both PRs are EP-related loading fixes, but they address different failure modes: one serializes distributed_config and adds has_ep for accelerate+ZeRO-3 loading, while the other fixes expert parameter handling so rank-0 broadcast does not overwrite weights under FSDP2.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45662 is the better canonical representative: it targets a specific runtime corruption bug in distributed loading and spans the concrete loading/MoE code path.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45662 is the strongest standalone fix in the cluster because it addresses a concrete expert-weight overwrite bug with a clear code-path change.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45548",
+            "right": "pull_request:45662",
+            "accept": false,
+            "reason": "Not the same underlying bug. 45548 is about serializing distributed_config and detecting EP for loading via accelerate launch; 45662 is about expert parameters being overwritten by rank-0 broadcast during FSDP2 loading. Shared EP context only."
+          }
+        ]
+      },
+      "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": 27,
+      "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": "b0784b4571d29a4173e94337cb1dd5e5f6ad14e8110ed976f8ae7ee4f81d5c7a",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4590,
+        "estimated_input_tokens": 1148,
+        "estimated_eval_tokens": 2552
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4590,
+        "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 target the same xdist/captured_info logging collision path in `testing_utils.py`; PR 45645 is the broader superset because it includes the worker-safe log file fix plus CI workflow changes to read multiple log files.",
+        "confidence": 0.91,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45645 is the better canonical PR because it covers the same core xdist-safe captured_info fix as 45639 and also updates CI to preserve and aggregate the per-worker logs.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45645 is the strongest representative: it addresses the underlying artifact collision and the downstream log-collection path, making it the more complete fix.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45639",
+            "right": "pull_request:45645",
+            "accept": true,
+            "reason": "Same underlying bug in captured_info handling under xdist; 45645 subsumes 45639's file-path fix and adds the necessary CI-side log aggregation."
+          }
+        ]
+      },
+      "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": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-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": 27,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45737-2",
+      "nodes": [
+        "pull_request:45737",
+        "pull_request:45738"
+      ],
+      "soft_pairs": [
+        "pull_request:45737|pull_request:45738"
+      ],
+      "prepared_review_unit_hash": "cc2ade73a49ec6a8883a7a24c433988645588bcd049e8aad2bce56b1b24daa1c",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4131,
+        "estimated_input_tokens": 1033,
+        "estimated_eval_tokens": 2322
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4131,
+        "estimated_input_tokens": 1033,
+        "estimated_eval_tokens": 2322
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs address the same MusicgenMelody cache initialization bug in `forward()`: replacing `EncoderDecoderCache(DynamicCache, DynamicCache)` with a single `DynamicCache`. PR #45737 is the more complete duplicate because it bundles the fix with a regression test, while #45738 is the fix-only variant.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45737 is the best canonical PR because it contains the same code fix as #45738 plus a regression test, making it the most complete representation of the bug fix.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45737 is the strongest overall candidate: it fixes the concrete cache bug and adds a targeted regression test, whereas #45738 only changes the implementation.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45737",
+            "right": "pull_request:45738",
+            "accept": true,
+            "reason": "Same underlying code-path bug in `musicgen_melody` cache setup; the diffs match on the core fix and the test-only addition in #45737 could plausibly live in the same PR as #45738."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The assessment is grounded in the packet: both PRs make the same `musicgen_melody` cache initialization change, and #45737 only adds a regression test on top of that fix. The soft-edge accept is conservative enough because the diffs are mergeable into one PR for the same bugfix."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 27,
+      "component_count": 27,
+      "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-45975-2",
+      "nodes": [
+        "pull_request:45975",
+        "pull_request:45977"
+      ],
+      "soft_pairs": [
+        "pull_request:45975|pull_request:45977"
+      ],
+      "prepared_review_unit_hash": "a297f103ea738b6621d35cb6d1f962b1f7470446b084500015ff8c3368701183",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4523,
+        "estimated_input_tokens": 1131,
+        "estimated_eval_tokens": 2518
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4523,
+        "estimated_input_tokens": 1131,
+        "estimated_eval_tokens": 2518
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open PRs touch GGUF/Apple Silicon/kernels-community, but they implement different pieces: one is an optional dequant fast-path/configuration update, the other is a new GGUF linear inference path with integration code and tests.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45977 is the stronger cluster representative: it introduces the core GGUF runtime change, adds integration points, and includes tests, whereas 45975 is a smaller optional fast-path/help/configuration update.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45977 is the best standalone candidate for triage because it addresses a concrete code-path change (GGUF matmul on Apple Silicon) with implementation and test coverage.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45975",
+            "right": "pull_request:45977",
+            "accept": false,
+            "reason": "Related domain, but not the same underlying change: 45975 updates CLI/help and GGUF-related fast-path plumbing, while 45977 adds a new GgufLinear inference implementation and tests. They would not plausibly be merged into one PR as duplicate fixes."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    }
+  ]
+}
diff --git a/snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/manifest.json b/snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..1f5244e1dbf92ac8570437c0c32f58bd217ef3a8
--- /dev/null
+++ b/snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/manifest.json
@@ -0,0 +1,14 @@
+{
+  "analysis_id": "hybrid-model-20260515t060029z",
+  "artifacts": {
+    "hybrid": "snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.json",
+    "hybrid_reviews": "snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.llm-reviews.json"
+  },
+  "channel": "canonical",
+  "model": null,
+  "published_at": "2026-05-15T06:04:53Z",
+  "repo": "huggingface/transformers",
+  "schema_version": 1,
+  "snapshot_id": "20260515T060029Z",
+  "variant": "hybrid"
+}
diff --git a/snapshots/20260515T060029Z/manifest.json b/snapshots/20260515T060029Z/manifest.json
index 38025d60ebfbc94ca4aecad38f89837eb5f8a2ea..8af6ac594c38585e63022e1f8241d6142c3d03a5 100644
--- a/snapshots/20260515T060029Z/manifest.json
+++ b/snapshots/20260515T060029Z/manifest.json
@@ -32,6 +32,24 @@
     "timeline_events": 193
   },
   "extracted_at": "2026-05-15T06:00:29Z",
+  "published_analysis": {
+    "canonical_analysis_id": "hybrid-model-20260515t060029z",
+    "runs": {
+      "hybrid-model-20260515t060029z": {
+        "analysis_id": "hybrid-model-20260515t060029z",
+        "artifacts": {
+          "hybrid": "snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.json",
+          "hybrid_reviews": "snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/analysis-report-hybrid.llm-reviews.json"
+        },
+        "channel": "canonical",
+        "manifest_path": "snapshots/20260515T060029Z/analysis-runs/hybrid-model-20260515t060029z/manifest.json",
+        "model": null,
+        "published_at": "2026-05-15T06:04:53Z",
+        "variant": "hybrid"
+      }
+    },
+    "schema_version": 1
+  },
   "repo": "huggingface/transformers",
   "snapshot_id": "20260515T060029Z",
   "watermark": {