diff --git a/analysis/current/analysis-report-hybrid.json b/analysis/current/analysis-report-hybrid.json
index c05df08201a6f6b404e713c74463ae443a11691b..d2fb682d4bb171ec8dc77d70076b8c649e305fe8 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": "20260613T120022Z",
-  "generated_at": "2026-06-13T12:08:49Z",
+  "snapshot_id": "20260613T180123Z",
+  "generated_at": "2026-06-13T18:06:03Z",
   "evidence_quality": "full",
   "llm_enrichment": true,
   "meta_bugs": [
diff --git a/analysis/current/analysis-report-hybrid.llm-reviews.json b/analysis/current/analysis-report-hybrid.llm-reviews.json
index c3224c36052a1876f4049d177a644a5d91a18600..696c79d648ef284791b8f200224faf0a5361864a 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": "20260613T120022Z",
-  "generated_at": "2026-06-13T12:08:49Z",
+  "snapshot_id": "20260613T180123Z",
+  "generated_at": "2026-06-13T18:06:03Z",
   "model": "gpt-5.4-mini?service_tier=flex",
   "reviews": [
     {
@@ -13,8 +13,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -493,8 +493,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -868,8 +868,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -1190,8 +1190,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -1526,8 +1526,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -1833,8 +1833,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -2161,8 +2161,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -2504,8 +2504,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -2825,8 +2825,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -3136,8 +3136,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -3472,8 +3472,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -3798,8 +3798,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -4120,8 +4120,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -4445,8 +4445,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -4742,8 +4742,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -5043,8 +5043,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -5379,8 +5379,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -5656,8 +5656,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -5940,8 +5940,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -6245,8 +6245,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -6564,8 +6564,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -6865,8 +6865,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -7148,8 +7148,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -7466,8 +7466,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -7792,8 +7792,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -8083,8 +8083,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -8395,8 +8395,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -8779,7 +8779,7 @@
         "issue:44360|issue:44561",
         "issue:45290|issue:45381"
       ],
-      "prepared_review_unit_hash": "421ae77831b4183c98dabc2ce3f4af5f2651695f646caee4e06c17a15a3a2913",
+      "prepared_review_unit_hash": "e9590743cde7ab1ff71dd5871741553b1bf6ca4a4810f653538f0bb4b5a345d0",
       "component_budget": {
         "node_count": 467,
         "item_count": 467,
@@ -8801,192 +8801,198 @@
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "This cluster is mostly a set of unrelated bug reports spanning different models, tokenizer paths, training/inference flows, and documentation/runtime issues. The only clear duplicate is the repeated Qwen3.5 Moe from_pretrained regression.",
-        "confidence": 0.97,
-        "canonical_issue_reason": "issue:45310 is the better canonical issue: it reports the Qwen3.5 Moe from_pretrained failure first, and issue:45478 is the same bug under the same title with a later report.",
+        "summary": "The cluster is mostly noise: the items span unrelated bugs across tokenizer, training, model-loading, and CI failures. The only clear duplicate pair is the repeated Qwen3.5 Moe from_pretrained error (issues 45310 and 45478).",
+        "confidence": 0.84,
+        "canonical_issue_reason": "Issue 45310 is the earlier and clearer report of the exact same Qwen3.5 Moe from_pretrained regression seen again in 45478.",
         "canonical_pr_reason": null,
-        "best_issue_reason": "issue:45310 is the strongest representative because it is an exact duplicate target and a concrete runtime regression report rather than a broader or more auxiliary complaint.",
+        "best_issue_reason": "Issue 45310 is the best representative because it is a concrete, self-contained bug report and has a near-identical duplicate in 45478.",
         "best_pr_reason": null,
         "soft_edge_verdicts": [
           {
             "left": "issue:45290",
             "right": "issue:46326",
             "accept": false,
-            "reason": "Both are chat-template/tokenizer problems, but one is about assistant tool-call formatting and the other about continue_final_message behavior in mistral_common; different code paths."
+            "reason": "Both involve chat/tokenizer behavior, but they describe different failure modes and likely different code paths."
           },
           {
             "left": "issue:44448",
             "right": "issue:44743",
             "accept": false,
-            "reason": "These affect different models and bugs: Pegasus output drift vs Qwen3.5 recurrent state reset with cache."
+            "reason": "Unrelated regressions in different models and subsystems; not the same bug."
           },
           {
             "left": "issue:46431",
             "right": "issue:46566",
             "accept": false,
-            "reason": "One is a missing chat template error on a processor; the other is a device-map/meta-weight loading issue in DiffusionGemma."
+            "reason": "A missing chat template error and a device-map/meta-weight issue are distinct problems."
           },
           {
             "left": "issue:44855",
             "right": "issue:45083",
             "accept": false,
-            "reason": "Python 3.13 import parsing failure in DebertaV2Model is unrelated to qwen3_omni_moe feature-length helper behavior."
+            "reason": "One is a Python import/parsing failure; the other is a feature-length helper behavior bug in a different model."
           },
           {
             "left": "issue:43335",
             "right": "issue:43504",
             "accept": false,
-            "reason": "SwitchTransformers sparse-layer creation and Beit pretrained preset loading are separate configuration/loading bugs."
+            "reason": "Different models and different root causes; not mergeable as one fix."
           },
           {
             "left": "issue:37428",
             "right": "issue:44351",
             "accept": false,
-            "reason": "Both are import errors, but they involve different missing symbols in different modules and are not the same underlying failure."
+            "reason": "Both are import errors, but they concern different missing symbols and unrelated fixes."
           },
           {
             "left": "issue:43278",
             "right": "issue:43825",
             "accept": false,
-            "reason": "Embedding dtype drift during train/eval is unrelated to a pipeline error-message regression about translation tasks."
+            "reason": "Different symptom domains: dtype drift vs an incorrect pipeline error message."
           },
           {
             "left": "issue:29127",
             "right": "issue:32090",
             "accept": false,
-            "reason": "LayoutLMv3 error-message clarity and Trainer GPU broadcast NoneType failure are unrelated issues."
+            "reason": "LayoutLMv3 error-message clarity and Trainer GPU broadcast TypeError are unrelated."
           },
           {
             "left": "issue:43335",
             "right": "issue:45083",
             "accept": false,
-            "reason": "These are different model-specific bugs in different helper/config code paths."
+            "reason": "No shared underlying bug beyond both touching model internals."
           },
           {
             "left": "issue:45200",
             "right": "issue:45588",
             "accept": false,
-            "reason": "Gemma mm_token_type_ids defaulting and flash_attention s_aux=None handling are distinct runtime problems."
+            "reason": "Gemma text fine-tuning defaults and flash-attention auxiliary-state handling are unrelated."
           },
           {
             "left": "issue:43720",
             "right": "issue:44112",
             "accept": false,
-            "reason": "BitNet packed-weight loading and a stale GraniteSpeech CI test do not describe the same bug."
+            "reason": "Different CI/runtime failures in different code paths and models."
           },
           {
             "left": "issue:45127",
             "right": "issue:45245",
             "accept": false,
-            "reason": "LoRA merge collapse with extended vocabulary and the categories > 2^24 error are unrelated."
+            "reason": "LoRA/embedding collapse and category-count limits are unrelated defects."
           },
           {
             "left": "issue:45588",
             "right": "issue:45859",
             "accept": false,
-            "reason": "Both involve offload/device behavior, but one crashes on s_aux=None in flash attention and the other is a rotary embedding CPU-offload incompatibility."
+            "reason": "Both mention offload/attention-related failures, but they affect different modules and models."
           },
           {
             "left": "issue:43819",
             "right": "issue:44493",
             "accept": false,
-            "reason": "DAC.from_latents mismatch and unexpected position-id keys affect different subsystems and failure modes."
+            "reason": "A DAC forward/from_latents mismatch is unrelated to position-id key warnings."
           },
           {
             "left": "issue:43335",
             "right": "issue:44112",
             "accept": false,
-            "reason": "SwitchTransformers config bug and GraniteSpeech CI failure are unrelated."
+            "reason": "Sparse-layer config bug and GraniteSpeech CI failure do not share the same root cause."
           },
           {
             "left": "issue:39692",
             "right": "issue:43526",
             "accept": false,
-            "reason": "Documentation example errors for SigLIP2 are unrelated to BeitImageProcessorFast reduce_labels returning the wrong shape."
+            "reason": "Documentation-example issues and an image-processor label bug are not the same underlying change."
           },
           {
             "left": "issue:43335",
             "right": "issue:45593",
             "accept": false,
-            "reason": "Sparse-layer configuration in SwitchTransformers is unrelated to D-FINE omitting auxiliary losses when denoising is off."
+            "reason": "SwitchTransformers config construction and D-FINE auxiliary-loss behavior are separate issues."
           },
           {
             "left": "issue:44206",
             "right": "issue:45823",
             "accept": false,
-            "reason": "LasrFeatureExtractor center-arg crash and Gemma4 PLE device mismatch are different model/path failures."
+            "reason": "Different models, different device/dtype symptoms, and different code paths."
           },
           {
             "left": "issue:26413",
             "right": "issue:31501",
             "accept": false,
-            "reason": "Checkpoint resume epoch_iterator failure and FineWeb training not starting are too broad and do not point to the same root cause."
+            "reason": "Checkpoint resume failure and FineWeb startup failure are too vague and not clearly the same bug."
           },
           {
             "left": "issue:34689",
             "right": "issue:35141",
             "accept": false,
-            "reason": "Llama 3.2 Vision model loading breakage and post_init reinitializing resized embeddings are separate model-loading vs weight-initialization bugs."
+            "reason": "Model-loading regression and embedding reinitialization on resize are different issues."
           },
           {
             "left": "issue:45310",
             "right": "issue:45478",
             "accept": true,
-            "reason": "Exact same Qwen3.5 Moe from_pretrained regression, same title and same underlying failure."
+            "reason": "Exact same reported failure and same Qwen3.5 Moe from_pretrained bug; clear duplicate."
           },
           {
             "left": "issue:39692",
             "right": "issue:44493",
             "accept": false,
-            "reason": "SigLIP2 docs issues and position-id key warnings are unrelated."
+            "reason": "A docs example problem and a position-id warning regression are unrelated."
           },
           {
             "left": "issue:43606",
             "right": "issue:44857",
             "accept": false,
-            "reason": "CPU offload device mismatch for bark-small is unrelated to float16 AMP loss crashing in LwDetrImageLoss."
+            "reason": "Both are device/dtype mismatches, but they involve different models and different failure mechanisms."
           },
           {
             "left": "issue:42490",
             "right": "issue:43854",
             "accept": false,
-            "reason": "Local path trailing-slash warning and GLM-4.7-Flash test loading failure are not the same bug."
+            "reason": "Local-path import warning and GLM model-loading failure are not the same bug."
           },
           {
             "left": "issue:45459",
             "right": "issue:46234",
             "accept": false,
-            "reason": "Tokenizer error masking without protobuf and a missing import-utils availability helper are unrelated."
+            "reason": "Tokenizer protobuf fallback behavior and a missing utility symbol are unrelated."
           },
           {
             "left": "issue:44448",
             "right": "issue:44779",
             "accept": false,
-            "reason": "Pegasus v4/v5 output differences and Deepseek tokenizer regressions are separate tokenizer/model issues."
+            "reason": "Both are v4/v5 behavior regressions, but they affect different tokenizers/models and likely different fixes."
           },
           {
             "left": "issue:44857",
             "right": "issue:45072",
             "accept": false,
-            "reason": "These are different dtype-related failures in different components and models."
+            "reason": "Both are precision-related, but the failing code paths and model components are different."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Tokenizer spacing behavior and Trainer broadcast TypeError are unrelated."
           },
           {
             "left": "issue:44360",
             "right": "issue:44561",
             "accept": false,
-            "reason": "DSA indexer activation design and trust_remote_code breakage from removing is_torch_fx_available are unrelated."
+            "reason": "A DSA indexer architectural comment and a v5 API removal breakage are not the same issue."
           },
           {
             "left": "issue:45290",
             "right": "issue:45381",
             "accept": false,
-            "reason": "Chat-template tool-call handling and qwen2.5-vl video vision_position_ids are different bugs."
+            "reason": "Chat-template tool-call handling and Qwen2.5-VL video position IDs are distinct bugs."
           }
         ]
       },
       "evaluator_result": {
         "accept": true,
-        "feedback": "Grounded overall. The duplicate call on issue:45310/45478 is supported by the identical titles and timestamps, and the other soft-edge rejections stay conservative given only the packet evidence. No clear overstatement or mergeability issue detected."
+        "feedback": "The duplicate call on 45310/45478 is well grounded by the identical title and earlier timestamp, and the summary stays conservative about the rest of the cluster. The soft-edge rejections are appropriately cautious and do not overclaim shared root causes."
       },
       "evaluator_used": true,
       "retried": false,
@@ -9002,8 +9008,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -9307,8 +9313,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -9632,8 +9638,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -9950,8 +9956,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -10233,8 +10239,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -10516,8 +10522,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -10807,8 +10813,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -11115,8 +11121,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -11441,8 +11447,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -11732,8 +11738,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -12037,8 +12043,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -12342,8 +12348,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -12653,8 +12659,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -12968,8 +12974,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -13279,8 +13285,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -13597,8 +13603,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -13876,8 +13882,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -14188,8 +14194,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -14486,8 +14492,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -14805,8 +14811,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -15106,8 +15112,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -15431,8 +15437,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -15757,8 +15763,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -16100,8 +16106,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -16436,8 +16442,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -16763,8 +16769,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -17136,8 +17142,8 @@
       "review_unit_count": 55,
       "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-10000-467",
       "nodes": [
@@ -17522,8 +17528,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-10519-3",
       "nodes": [
@@ -17600,8 +17606,8 @@
       "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": [
@@ -17671,8 +17677,8 @@
       "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": [
@@ -17734,8 +17740,8 @@
       "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-27088-8",
       "nodes": [
@@ -17967,8 +17973,8 @@
       "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-36817-16",
       "nodes": [
@@ -18289,8 +18295,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-38740-2",
       "nodes": [
@@ -18352,8 +18358,8 @@
       "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": [
@@ -18415,8 +18421,8 @@
       "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": [
@@ -18478,8 +18484,8 @@
       "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": [
@@ -18541,8 +18547,8 @@
       "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": [
@@ -18644,8 +18650,8 @@
       "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": [
@@ -18766,8 +18772,8 @@
       "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": [
@@ -18886,8 +18892,8 @@
       "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": [
@@ -18949,8 +18955,8 @@
       "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": [
@@ -19012,8 +19018,8 @@
       "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": [
@@ -19075,8 +19081,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-45810-2",
       "nodes": [
@@ -19138,8 +19144,8 @@
       "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": [
@@ -19256,8 +19262,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-46439-2",
       "nodes": [
@@ -19319,8 +19325,8 @@
       "review_unit_count": 1,
       "status": "reviewed",
       "reason": null,
-      "source": "llm",
-      "cache_hit": false,
+      "source": "cache",
+      "cache_hit": true,
       "model": "gpt-5.4-mini?service_tier=flex",
       "cluster_id": "cluster-46551-2",
       "nodes": [
@@ -19407,8 +19413,8 @@
       ],
       "soft_pairs": [
         "pull_request:39895|pull_request:43424",
-        "pull_request:41855|pull_request:43823",
         "pull_request:41855|pull_request:43424",
+        "pull_request:41855|pull_request:43823",
         "pull_request:42668|pull_request:43823",
         "pull_request:43251|pull_request:43823",
         "pull_request:43251|pull_request:44827",
@@ -19426,7 +19432,7 @@
         "pull_request:43823|pull_request:45213",
         "pull_request:43251|pull_request:43254"
       ],
-      "prepared_review_unit_hash": "dd735f129471cb5603499fa060d9b97a355899bab093c95ba6b075c305cb4e5b",
+      "prepared_review_unit_hash": "b00fe57e94a2c0594d257c899e348af4097121c7273a9ce43b9375424fd19076",
       "component_budget": {
         "node_count": 14,
         "item_count": 14,
@@ -19448,138 +19454,132 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "This cluster is mostly heterogeneous: the PRs cover unrelated model adds, docs, refactors, and test updates. The only genuine duplicate-like pair is the two `fixed_cross_entropy` PRs, which both implement the same loss-kwargs fix on the same code path.",
-        "confidence": 0.96,
+        "summary": "This cluster is mostly unrelated pull requests. The only strong duplicate is the pair of loss-function PRs that both add `weight`/`label_smoothing` support to `fixed_cross_entropy`; the rest are separate model additions, docs, or refactors that should stay distinct.",
+        "confidence": 0.92,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 43251 is the best canonical representative of the duplicate pair because it has the explicit issue target, broader description of the `cross_entropy` kwargs fix, and the same underlying change as 43254.",
+        "canonical_pr_reason": "PR #43251 is the better canonical representative for the `fixed_cross_entropy` kwargs fix: it explicitly targets issue #43240 and carries the more complete discussion/history, while #43254 is a near-identical smaller variant of the same code-path change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 43251 is the best overall PR representative because it is the clearest, most complete version of the only near-duplicate change in the set; the rest are unrelated and should not be merged into a single canonical PR.",
+        "best_pr_reason": "PR #43251 is the best single PR in this cluster because it is the clearest, issue-linked implementation of the shared loss-function bug fix and can subsume the duplicate variant (#43254).",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:39895",
             "right": "pull_request:43424",
             "accept": false,
-            "reason": "Both are test/model-export related, but they target different features (`Videoprism` model add vs ExecuTorch dynamic-shape export tests) and do not fix the same code path."
+            "reason": "Both are unrelated feature/test PRs; they do not fix the same code path or change."
           },
           {
             "left": "pull_request:41855",
-            "right": "pull_request:43823",
+            "right": "pull_request:43424",
             "accept": false,
-            "reason": "Tokenizer missing-methods cleanup for Mistral is unrelated to adding the MobileLLM model."
+            "reason": "Different tokenizer utility vs executorch export test; same broad library, not the same bug/change."
           },
           {
             "left": "pull_request:41855",
-            "right": "pull_request:43424",
+            "right": "pull_request:43823",
             "accept": false,
-            "reason": "Mistral tokenizer API additions and ExecuTorch dynamic-shape export tests are different subsystems and different changes."
+            "reason": "Mistral tokenizer methods and MobileLLM model addition are unrelated."
           },
           {
             "left": "pull_request:42668",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Processor-from-pretrained robustness work is unrelated to adding MobileLLM."
+            "reason": "Processor robustness changes are not the same underlying work as adding a new model."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "Loss-function kwargs handling and adding MobileLLM are unrelated."
+            "reason": "Loss-function kwargs support and MobileLLM model addition are unrelated."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "`fixed_cross_entropy` kwargs support and Mistral4 test fixes are different problems in different areas."
+            "reason": "A cross-entropy helper fix and Mistral4 test/model work are different code paths."
           },
           {
             "left": "pull_request:43424",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "ExecuTorch dynamic-shape export tests do not overlap with Mistral4 test fixes as the same underlying bug/change."
+            "reason": "Executorch dynamic-shape export testing and Mistral4 fixes are separate concerns."
           },
           {
             "left": "pull_request:43424",
             "right": "pull_request:43823",
             "accept": false,
-            "reason": "ExecuTorch exportability tests and MobileLLM model addition are unrelated."
+            "reason": "Dynamic-shape export tests do not overlap with model addition work."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:45655",
             "accept": false,
-            "reason": "Adding MobileLLM and fixing `cls.config` resolution are separate changes with no shared bug."
+            "reason": "New MobileLLM model code and config-resolution ordering fix are unrelated."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:43995",
             "accept": false,
-            "reason": "MobileLLM model addition and Falcon output-interface refactor are unrelated."
+            "reason": "Model addition and Falcon output-collection refactor are not the same change."
           },
           {
             "left": "pull_request:43995",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "Falcon model refactoring and the model-creation skill are not the same code-path change."
+            "reason": "Falcon refactoring is unrelated to the model-creation skill PR."
           },
           {
             "left": "pull_request:43995",
             "right": "pull_request:44827",
             "accept": false,
-            "reason": "Falcon output collection refactor and Mistral4 test fixes are unrelated."
+            "reason": "Falcon output-interface refactor and Mistral4 fixes are different model-specific changes."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45170",
             "accept": false,
-            "reason": "Mistral4 test/model updates and the `layrnorm`->`layernorm` rename fix are different issues."
+            "reason": "Mistral4 test fixes and layrnorm/layernorm renaming are different bugs."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:46267",
             "accept": false,
-            "reason": "Mistral4 test fixes and chat parsing updates do not address the same underlying bug or change."
+            "reason": "Mistral4 test fixes and chat parsing updates do not share the same concrete code path."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:46048",
             "accept": false,
-            "reason": "Mistral4 test fixes and the TDT loss kernel addition are unrelated."
+            "reason": "Mistral4 test fixes and TDT loss kernel integration are unrelated."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45655",
             "accept": false,
-            "reason": "Mistral4 test work and `cls.config` resolution order are different refactors/fixes."
+            "reason": "Mistral4 work and config-resolution fix are unrelated."
           },
           {
             "left": "pull_request:44827",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "Mistral4 test fixes and the model-creation skill are not the same concrete change."
-          },
-          {
-            "left": "pull_request:43823",
-            "right": "pull_request:45213",
-            "accept": false,
-            "reason": "Adding MobileLLM and the model-creation skill are unrelated; one is model code, the other is tooling/docs."
+            "reason": "They only overlap on a broad auto/modeling area; the actual changes are unrelated."
           },
           {
             "left": "pull_request:43823",
             "right": "pull_request:45213",
             "accept": false,
-            "reason": "Adding MobileLLM and the model-creation skill are unrelated; one is model code, the other is tooling/docs."
+            "reason": "New model implementation and model-creation skill scaffolding are not the same artifact or bug."
           },
           {
             "left": "pull_request:43251",
             "right": "pull_request:43254",
             "accept": true,
-            "reason": "Both PRs change `fixed_cross_entropy` to pass additional kwargs to `nn.functional.cross_entropy`; they fix the same concrete loss-path issue and could plausibly be merged into one PR."
+            "reason": "Both PRs implement the same `fixed_cross_entropy` kwargs support in the same file for the same issue (#43240); one is a near-duplicate variant of the other."
           }
         ]
       },
       "evaluator_result": {
         "accept": true,
-        "feedback": "The classification is grounded in the packet. The cluster is heterogeneous, and the only pair with direct shared evidence is 43251/43254: both touch `src/transformers/loss/loss_utils.py`, share issue target 43240, and implement the same `fixed_cross_entropy`/`cross_entropy` kwargs fix. The other soft-edge rejections are conservative and consistent with the diffs. One minor wording caveat: the 43251/43254 rationale says \"additional kwargs,\" but the concrete change is specifically about `weight` and `label_smoothing` rather than arbitrary passthrough kwargs."
+        "feedback": "The cluster summary is grounded in the packet and stays conservative: the only accepted duplicate pair, PR #43251/#43254, is a clear same-file same-issue `fixed_cross_entropy` kwargs fix and looks mergeable into a single PR. The other verdicts are appropriately negative and supported by distinct titles/files; I don\u2019t see any overstatement that would require rejection."
       },
       "evaluator_used": true,
       "retried": false,
@@ -19606,12 +19606,12 @@
       "soft_pairs": [
         "pull_request:40546|pull_request:42781"
       ],
-      "prepared_review_unit_hash": "5f1d3ddcf1f9ce7f63d6f111a69aea3184290ec0f66e985e898a88fc50378dd8",
+      "prepared_review_unit_hash": "91c9c06ec8e714322654fca35759a57d59906006a232854bc83b7f7d2387b357",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 7225,
+        "serialized_chars": 7226,
         "estimated_input_tokens": 1807,
         "estimated_eval_tokens": 3870
       },
@@ -19619,7 +19619,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 7225,
+        "serialized_chars": 7226,
         "estimated_input_tokens": 1807,
         "estimated_eval_tokens": 3870
       },
@@ -19628,18 +19628,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related VibeVoice pull requests, but they target different deliverables: PR 40546 is the base VibeVoice model integration, while PR 42781 adds VibeVoice Realtime plus acoustic-tokenizer work. They share filenames because the second builds on the same model family, but they are not the same change set.",
-        "confidence": 0.93,
+        "summary": "The two PRs are related to VibeVoice, but they are not duplicates: #40546 adds the core VibeVoice integration, while #42781 adds a separate realtime/acoustic-tokenizer extension on top of that base work.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 40546 is the better canonical representative for the cluster because it introduces the core VibeVoice implementation and broad model integration; PR 42781 looks like a later extension focused on realtime/acoustic-tokenizer support rather than the same change.",
+        "canonical_pr_reason": "PR #40546 is the foundational VibeVoice implementation and the broader representative of this cluster; #42781 is a later, narrower realtime extension rather than the same change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 40546 is the strongest single representative of the VibeVoice work overall: it covers the main model, auto-mapping, processing, docs, and pipeline integration. PR 42781 is narrower and more feature-specific.",
+        "best_pr_reason": "#40546 is the best cluster representative because it introduces the main model integration and core plumbing, whereas #42781 focuses on a distinct realtime variant and tokenizer stack.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:40546",
             "right": "pull_request:42781",
             "accept": false,
-            "reason": "Reject: both are VibeVoice-related, but 40546 implements the base model while 42781 adds realtime and acoustic-tokenizer functionality. Different scope and code paths, so they are not the same underlying PR."
+            "reason": "Reject: these are adjacent VibeVoice features, not the same underlying fix or change. #40546 implements the base model; #42781 adds realtime/acoustic-tokenizer support and additional generation/runtime changes. Shared filenames come from the same model family, but the code paths and merged scope differ."
           }
         ]
       },
@@ -19669,7 +19669,7 @@
       "soft_pairs": [
         "pull_request:40755|pull_request:41224"
       ],
-      "prepared_review_unit_hash": "506e2650f54a079fd52a703feac87f394e9baa6580e57f3a7c922f8966111173",
+      "prepared_review_unit_hash": "a127d89a8356226894a044e9dcf6cf8e43584bd4103201d8fbcbb2c70d3cc048",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -19691,18 +19691,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are two unrelated pull requests: one adds TimesFM covariate forecasting support, the other adds DINOv3 image classification support. They do not fix the same code-path problem and should not be clustered as duplicates.",
+        "summary": "The two PRs are unrelated feature additions in different model families: TimesFM covariate forecasting vs. DINOv3 image classification support.",
         "confidence": 0.99,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No canonical PR: the two PRs address different models, different features, and different files, so neither can represent the other as a duplicate.",
+        "canonical_pr_reason": "pull_request:41224 is the more focused representative: it targets one concrete model capability (DINOv3 image classification) with a narrow implementation scope.",
         "best_issue_reason": null,
-        "best_pr_reason": "Neither PR is a better global representative for the cluster because the cluster is not a true duplicate set; the PRs are unrelated feature additions.",
+        "best_pr_reason": "pull_request:41224 is the cleanest standalone PR in this cluster because it is narrowly scoped and directly tied to a single code path, unlike the broader TimesFM covariate work.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:40755",
             "right": "pull_request:41224",
             "accept": false,
-            "reason": "TimesFM covariate forecasting and DINOv3 image classification are separate features in different model families with no shared concrete bug or change path."
+            "reason": "Different underlying changes and code paths: TimesFM forecasting with covariates vs. DINOv3 image classification support. They do not appear mergeable into one PR."
           }
         ]
       },
@@ -19797,7 +19797,7 @@
         "pull_request:41797|pull_request:45186",
         "pull_request:45186|pull_request:45630"
       ],
-      "prepared_review_unit_hash": "992de9fac3c0b23cbecb2fafd0876be70e8a83b09800e0666c292f87a623ee30",
+      "prepared_review_unit_hash": "b8d030bf284ad626a8b8ef032485fbc5c8f33e0c93323d5ff8d38f9cc525f9ef",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
@@ -19819,24 +19819,24 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "All three items are separate open PRs adding different models (DeepSeek OCR, Isaac, and Kimi2-6). They share the usual Transformers model-integration scaffolding, but they do not describe the same underlying code change or bug fix.",
-        "confidence": 0.98,
+        "summary": "These are three separate model-addition PRs that share repository plumbing and docs scaffolding, but they target different models and codepaths. The overlap is superficial, not a duplicate change set.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No single canonical PR: each pull request adds a different model family with its own files, tests, and registration updates.",
+        "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "#45630 is the most substantial and reviewed of the three, so it is the best representative PR in the cluster, but it is not a duplicate of the others.",
+        "best_pr_reason": null,
         "soft_edge_verdicts": [
           {
             "left": "pull_request:41797",
             "right": "pull_request:45186",
             "accept": false,
-            "reason": "Different model additions: DeepSeek OCR vs Isaac. Shared auto-registration/docs scaffolding is generic and not enough to treat them as the same change."
+            "reason": "Different models (DeepSeek OCR vs Isaac) with distinct configs, model code, processors, and tests; shared auto-registration/docs files are template overlap only."
           },
           {
             "left": "pull_request:45186",
             "right": "pull_request:45630",
             "accept": false,
-            "reason": "Different model additions: Isaac vs Kimi2-6. They touch similar integration files, but they implement separate model code paths and test suites."
+            "reason": "Different models (Isaac vs Kimi2-6) and different implementation details; they share broad model-addition plumbing but not the same underlying bug or change."
           }
         ]
       },
@@ -19866,7 +19866,7 @@
       "soft_pairs": [
         "pull_request:42975|pull_request:42976"
       ],
-      "prepared_review_unit_hash": "f4e067aa07579d47bdeace3634b87d93e47d8dee55812d9487d7122db49f4eb4",
+      "prepared_review_unit_hash": "d4c6892c3cc063345aa024f3fc96e819d9441121a294d9fca2410b673630ae57",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -19888,18 +19888,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related workflow-maintenance PRs but not duplicates: one upgrades assorted GitHub Actions for Node 24 compatibility, while the other updates Docker build actions and other latest action versions. They overlap in files but fix different dependency/version bumps.",
-        "confidence": 0.92,
+        "summary": "Two open workflow-maintenance PRs that both bump GitHub Actions dependencies, but they target different action families and different concrete compatibility concerns.",
+        "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #42975 is the better cluster anchor because it is the broader, more explicit workflow-action upgrade set and its title/bodies clearly state the compatibility goal.",
+        "canonical_pr_reason": "PR #42975 is the better representative only because it is broader and more clearly centered on the Node 24-compatible Actions refresh across many workflows; PR #42976 is a narrower Docker action version bump.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #42975 is the best representative because it covers the larger cross-workflow Actions migration and is the more semantically central maintenance change in the cluster.",
+        "best_pr_reason": "PR #42975 is the strongest cluster anchor: it updates multiple core workflow actions (checkout/cache/upload-artifact) across many files. PR #42976 is still substantial, but it specifically bumps docker/build-push-action and related workflow versions, so it is a separate change stream.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:42975",
             "right": "pull_request:42976",
             "accept": false,
-            "reason": "Both are GitHub Actions version bumps, but they target different actions and different change sets (Node 24 compatibility vs. docker/build-push-action upgrades). They do not look like the same concrete bug or mergeable single PR."
+            "reason": "Do not merge: both are workflow dependency upgrades, but one is about general GitHub Actions Node 24 compatibility while the other is about docker/build-push-action and related version bumps. They touch overlapping files, not the same underlying bug or change."
           }
         ]
       },
@@ -19938,12 +19938,12 @@
         "pull_request:43451|pull_request:45490",
         "pull_request:45490|pull_request:46562"
       ],
-      "prepared_review_unit_hash": "ebce8a3edd5b75848a6239a6bc9ab2f44f89eeac2a0970a9ed60b8aeda6027f9",
+      "prepared_review_unit_hash": "c727ec421c4e6cc014d54dab4377afae0489ec6f45071d399c117ecd765eea43",
       "component_budget": {
         "node_count": 5,
         "item_count": 5,
         "soft_pair_count": 7,
-        "serialized_chars": 17574,
+        "serialized_chars": 17573,
         "estimated_input_tokens": 4394,
         "estimated_eval_tokens": 9044
       },
@@ -19951,7 +19951,7 @@
         "node_count": 5,
         "item_count": 5,
         "soft_pair_count": 7,
-        "serialized_chars": 17574,
+        "serialized_chars": 17573,
         "estimated_input_tokens": 4394,
         "estimated_eval_tokens": 9044
       },
@@ -19960,54 +19960,54 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are separate model-addition PRs for different architectures (ViT NEPA, Molmo, Molmo2, CTSM, Step3p7). They share generic docs/auto-registry scaffolding, but not the same underlying bug or change.",
-        "confidence": 0.98,
+        "summary": "All items are distinct model-addition PRs with only superficial overlap in docs/auto-registration boilerplate. They do not describe the same underlying change or fix, so none should be merged as duplicates.",
+        "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 43451 (Molmo2) is the most active and substantial of the group, but it is still a distinct model-addition change and not a duplicate of the others.",
+        "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "PR 43451 is the strongest representative PR in this cluster because it has the highest review activity and a broad, complete model integration, while still being clearly separate from the other model families.",
+        "best_pr_reason": null,
         "soft_edge_verdicts": [
           {
             "left": "pull_request:42978",
             "right": "pull_request:43448",
             "accept": false,
-            "reason": "ViT NEPA vs Molmo are different model integrations with different files, configs, and code paths; shared auto/docs edits are just boilerplate."
+            "reason": "Different models (ViT NEPA vs Molmo) with separate model, processor, and test files; shared docs/auto registration is boilerplate, not the same code-path change."
           },
           {
             "left": "pull_request:42978",
             "right": "pull_request:43451",
             "accept": false,
-            "reason": "ViT NEPA and Molmo2 add different models and processors; they do not fix or implement the same concrete change."
+            "reason": "ViT NEPA and Molmo2 are unrelated model additions; the overlap is only in generic transformers registration/docs, not a single mergeable fix."
           },
           {
             "left": "pull_request:42978",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "ViT NEPA and CTSM target unrelated model families; overlap is only in common registry/docs updates."
+            "reason": "CTSM and ViT NEPA are different model implementations and tests; no shared bug or identical change to merge together."
           },
           {
             "left": "pull_request:43448",
             "right": "pull_request:43451",
             "accept": false,
-            "reason": "Molmo and Molmo2 are related by name but are still distinct model implementations with different files and behavior, so they should not be merged as duplicates."
+            "reason": "Molmo and Molmo2 are related by naming but are distinct model families with different files and functionality; not the same concrete change."
           },
           {
             "left": "pull_request:43448",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "Molmo and CTSM are unrelated model additions; the shared subsystem edits are too broad to indicate the same change."
+            "reason": "Molmo and CTSM are unrelated model-addition PRs; shared auto-map/docs edits are too broad to indicate duplication."
           },
           {
             "left": "pull_request:43451",
             "right": "pull_request:45490",
             "accept": false,
-            "reason": "Molmo2 and CTSM add different model stacks and do not address the same code-path problem."
+            "reason": "Molmo2 and CTSM touch different model codepaths and introduce different architectures; no evidence of a single underlying fix."
           },
           {
             "left": "pull_request:45490",
             "right": "pull_request:46562",
             "accept": false,
-            "reason": "CTSM and Step3p7 are separate model-addition PRs for different architectures; they are not plausibly mergeable into one PR beyond generic registry/documentation updates."
+            "reason": "CTSM and Step3p7 are separate model additions with different configs, processors, and tests; similar repo plumbing does not make them duplicates."
           }
         ]
       },
@@ -20056,12 +20056,12 @@
         "pull_request:46380|pull_request:46539",
         "pull_request:46380|pull_request:46579"
       ],
-      "prepared_review_unit_hash": "ecc012c2454201fd32df94964991a6b782576b841cbfd85785c232f8e2ac202b",
+      "prepared_review_unit_hash": "589b9f7ab521f675f5156cac7accf6ca484225e9c95acdc59ea032aca2de374b",
       "component_budget": {
         "node_count": 8,
         "item_count": 8,
         "soft_pair_count": 14,
-        "serialized_chars": 23153,
+        "serialized_chars": 23156,
         "estimated_input_tokens": 5789,
         "estimated_eval_tokens": 11834
       },
@@ -20069,7 +20069,7 @@
         "node_count": 8,
         "item_count": 8,
         "soft_pair_count": 14,
-        "serialized_chars": 23153,
+        "serialized_chars": 23156,
         "estimated_input_tokens": 5789,
         "estimated_eval_tokens": 11834
       },
@@ -20078,96 +20078,96 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "This cluster is not a true duplicate set: it mixes unrelated PRs for whisper decoding, tie-weights compatibility, fp8 dequantization, doctest cleanup, generation validation, CI memory limits, and typo fixes.",
-        "confidence": 0.91,
+        "summary": "These PRs are mostly unrelated fixes with generic titles or superficial overlap in tests/docs. I don\u2019t see a shared concrete bug or change set that would justify treating any pair as duplicates.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No single canonical PR fits as a duplicate target; the PRs address different code paths and changes. If forced to name one, 46579 is the broadest maintenance-style PR, but it is not a duplicate of the others.",
+        "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "46579 is the broadest standalone maintenance PR in the set, but it still does not represent the others' concrete bugs/changes.",
+        "best_pr_reason": null,
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43291",
             "right": "pull_request:43665",
             "accept": false,
-            "reason": "Whisper tokenization/tests vs unrelated CLIP/ViT test edits; no shared concrete bug."
+            "reason": "Both touch test-related files, but one is a Whisper tokenizer fix and the other is unrelated CLIP/ViT test edits; different code paths and no shared bug."
           },
           {
             "left": "pull_request:43291",
             "right": "pull_request:44815",
             "accept": false,
-            "reason": "Whisper decode fix and fp8 dequant fix are different subsystems and failures."
+            "reason": "Whisper decoding/tests are unrelated to FP8 dequant/loading logic; overlap is only broad 'fix' similarity."
           },
           {
             "left": "pull_request:43291",
             "right": "pull_request:46380",
             "accept": false,
-            "reason": "Whisper decoding issue vs generation watermarking/validation change; not the same bug."
+            "reason": "Whisper tokenization/timestamp handling and generation watermarking validation are separate issues in different subsystems."
           },
           {
             "left": "pull_request:43498",
             "right": "pull_request:46579",
             "accept": false,
-            "reason": "Backward-compat alias for tie_weights vs a broad typo cleanup; unrelated changes."
+            "reason": "One adds a backward-compat alias for tie_weights; the other is a broad typo cleanup across many files. No same underlying bug."
           },
           {
             "left": "pull_request:43665",
             "right": "pull_request:46539",
             "accept": false,
-            "reason": "Inserted dummy test assertions vs CI memory-cap changes; no common underlying fix."
+            "reason": "Model test assertion edits are unrelated to the Phimoe integration test/CI memory workaround."
           },
           {
             "left": "pull_request:43665",
             "right": "pull_request:44815",
             "accept": false,
-            "reason": "Unrelated test edits vs fp8 dequantization logic."
+            "reason": "Test failures in CLIP/ViT are unrelated to the dequant fix in core loading and FP8 integration."
           },
           {
             "left": "pull_request:43665",
             "right": "pull_request:46380",
             "accept": false,
-            "reason": "Dummy test edits do not match the generation-config fix in 46380."
+            "reason": "These touch different areas: vision model tests vs generation config/Whisper changes; no common concrete fix."
           },
           {
             "left": "pull_request:43665",
             "right": "pull_request:46579",
             "accept": false,
-            "reason": "Test noise vs code typo cleanup; not the same concrete change."
+            "reason": "Generic test-file edits do not match a code-typo sweep across many source files."
           },
           {
             "left": "pull_request:44815",
             "right": "pull_request:46539",
             "accept": false,
-            "reason": "fp8 dequant logic vs CI resource-limit adjustments; different problems."
+            "reason": "FP8/dequant changes are unrelated to Phimoe integration testing and CI environment memory limits."
           },
           {
             "left": "pull_request:44815",
             "right": "pull_request:46380",
             "accept": false,
-            "reason": "Different runtime paths: fp8 conversion/loading versus generation config validation."
+            "reason": "Both involve generation-related files, but one is a dequant/loading fix and the other is a validation/Whisper integration fix; not the same bug."
           },
           {
             "left": "pull_request:44815",
             "right": "pull_request:46579",
             "accept": false,
-            "reason": "fp8 kernel/loading fix is unrelated to typo cleanup across many files."
+            "reason": "Shared file names are incidental; FP8 dequant fixes and typo cleanups address different changes."
           },
           {
             "left": "pull_request:45114",
             "right": "pull_request:46579",
             "accept": false,
-            "reason": "Doctest/documentation edits are not the same as a general code typo sweep."
+            "reason": "Documentation doctest cleanup is not the same as code typo fixes across implementation files."
           },
           {
             "left": "pull_request:46380",
             "right": "pull_request:46539",
             "accept": false,
-            "reason": "Generation config validation fix vs Phimoe CI memory workaround; unrelated."
+            "reason": "Generation/Whisper validation fixes are unrelated to the Phimoe integration test CI memory cap change."
           },
           {
             "left": "pull_request:46380",
             "right": "pull_request:46579",
             "accept": false,
-            "reason": "Generation-watermarking validation bug is unrelated to typo fixes."
+            "reason": "One fixes generation config validation and Whisper behavior; the other is a broad typo cleanup, so they are not mergeable as one PR."
           }
         ]
       },
@@ -20197,12 +20197,12 @@
       "soft_pairs": [
         "pull_request:43488|pull_request:43532"
       ],
-      "prepared_review_unit_hash": "7d0f26d9df6b661b5e30220c94315bb04108204b4db8d8be1a03371e4b942902",
+      "prepared_review_unit_hash": "dbc629f0b0bad3a827a6db245d9d9e9633ed2684b8b13700ff09dd0a4179e514",
       "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
       },
@@ -20210,7 +20210,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4563,
+        "serialized_chars": 4562,
         "estimated_input_tokens": 1141,
         "estimated_eval_tokens": 2538
       },
@@ -20219,18 +20219,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are not duplicates: PR #43488 is a small bot/config check tweak, while PR #43532 is a workflow/testing utilities change. The shared Vit test filename is incidental, not evidence of the same underlying change.",
-        "confidence": 0.97,
+        "summary": "Two open '[don't merge]' pull requests with only superficial overlap on a Vit test file; they target different code paths and are not duplicates.",
+        "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": null,
+        "canonical_pr_reason": "No canonical PR: 43488 is a packaging/test-dependency/version-bump check, while 43532 is a broader workflow/testing-utility change; they do not represent the same underlying fix.",
         "best_issue_reason": null,
-        "best_pr_reason": "Neither PR is a good canonical duplicate target for the cluster because they fix different things and would not plausibly be merged into one PR.",
+        "best_pr_reason": "If one must be chosen as the more representative cluster member, 43532 has the richer, broader change set, but it still should not be treated as a duplicate of 43488.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43488",
             "right": "pull_request:43532",
             "accept": false,
-            "reason": "Only a superficial overlap on tests/models/vit/test_modeling_vit.py; the diffs change unrelated code paths (setup/versioning vs CI/workflow/testing infrastructure)."
+            "reason": "Reject: these are different synthetic maintenance changes. 43488 edits setup/dependency/version files and a Vit test, while 43532 changes GitHub workflows plus testing/notification utilities. Shared test-file touch is incidental, not the same bug or fix."
           }
         ]
       },
@@ -20260,12 +20260,12 @@
       "soft_pairs": [
         "pull_request:43612|pull_request:43613"
       ],
-      "prepared_review_unit_hash": "ded3c2de5cc785b64df84b66e41b7b4fe9502320dca976872a8f609107b14d6c",
+      "prepared_review_unit_hash": "4a9f2f6b9bcccf3ebeff2e0c5b832acb050c6a2fc1c77cff4954d0390f79af34",
       "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
       },
@@ -20273,7 +20273,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 6083,
+        "serialized_chars": 6082,
         "estimated_input_tokens": 1521,
         "estimated_eval_tokens": 3298
       },
@@ -20282,18 +20282,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related pipeline-addition PRs, but they implement different features: promptable concept segmentation vs promptable visual segmentation. They share documentation and registration scaffolding, yet target different code paths and model integrations, so they should not be deduplicated.",
-        "confidence": 0.96,
+        "summary": "These are two related but distinct feature PRs: one adds a Promptable Concept Segmentation pipeline, the other adds a Promptable Visual Segmentation pipeline. They share some scaffolding, but they target different tasks, model integrations, and test surfaces, so they should not be treated as duplicates.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No canonical PR for this cluster: the two pull requests add different pipelines and are not the same underlying change.",
+        "canonical_pr_reason": "No single canonical PR: the two PRs implement different pipelines with different underlying functionality and model coverage rather than the same fix/change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #43613 is the broader, more integrated pipeline addition, but it still should not subsume #43612 because the underlying feature is different.",
+        "best_pr_reason": "Neither PR is a global best representative for a duplicate cluster, because they are separate feature additions instead of alternative implementations of the same code-path problem.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43612",
             "right": "pull_request:43613",
             "accept": false,
-            "reason": "Reject: despite similar scaffolding and overlapping files, these PRs add different concrete pipelines with different task docs, model hooks, and tests, so they are not the same change."
+            "reason": "Different features: concept segmentation vs visual segmentation. The diffs show different pipeline modules, docs pages, and model/test integrations (SAM3-centric versus SAM/SAM2/EdgeTam), so they are not the same underlying change and would not plausibly be merged into one PR."
           }
         ]
       },
@@ -20323,7 +20323,7 @@
       "soft_pairs": [
         "pull_request:43651|pull_request:43663"
       ],
-      "prepared_review_unit_hash": "00a47e41afe4f34db0a09f92068679dac901e54a65e2bb4b7f407ba73fb5f0c2",
+      "prepared_review_unit_hash": "6f659977cc638e23fc61e39d66304fbfb0e843a989462bc90fb4d5435d14f327",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -20345,18 +20345,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These two PRs touch the same file but implement unrelated Trainer extension points: one changes gradient-accumulation loss scaling, the other refactors signature-column filtering for custom trainers. They are not duplicates and should not be merged together.",
-        "confidence": 0.96,
+        "summary": "Two open PRs in `trainer.py` add separate trainer extension hooks: one for gradient-accumulation loss scaling and one for overriding signature-based column filtering. They touch the same file but solve different problems.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": null,
+        "canonical_pr_reason": "PR #43651 is the more representative change because it introduces a concrete new trainer behavior around loss scaling in the training step, which is a clearer standalone capability than the helper refactor in #43663.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 43651 is the more concrete bugfix-oriented change, with a narrow, well-defined behavior change in training_step. PR 43663 is an API refactor for overrideability rather than the same underlying problem.",
+        "best_pr_reason": "PR #43651 is the best standalone PR in this cluster: it addresses a specific training-path behavior change and is easy to understand as a distinct user-facing fix/feature.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43651",
             "right": "pull_request:43663",
             "accept": false,
-            "reason": "Same file, but different code paths and goals: 43651 fixes gradient accumulation loss scaling; 43663 adds an overridable helper for model input column filtering. They do not appear to be the same bug or change."
+            "reason": "They are different changes in the same file: #43651 adds `_loss_is_scaled_for_ga` to control gradient accumulation loss normalization, while #43663 adds `_get_signature_columns` to customize column filtering. They do not appear to fix the same underlying bug or be mergeable into one PR."
           }
         ]
       },
@@ -20386,7 +20386,7 @@
       "soft_pairs": [
         "pull_request:43656|pull_request:43842"
       ],
-      "prepared_review_unit_hash": "cc9a35f1e7b16ef9d635d4bb1ce79f887ae021a50e9d642374435cee7ae14799",
+      "prepared_review_unit_hash": "d3a697ec81c71df5b72e767cb08408e0f8e36f426d83c0b44167b7b179390f17",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -20408,18 +20408,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Two open PRs touch the same CLI file and both address TypeAdapter-related runtime issues, but they target different failure modes and use different fixes. They are related, not duplicates.",
-        "confidence": 0.79,
+        "summary": "Two open PRs touch the same CLI file and the same TypeAdapter area, but they address different NameError scenarios: one is a broader annotation/runtime typing fix, the other is specifically for missing pydantic. They are related, not duplicates.",
+        "confidence": 0.89,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 43842 is the more focused, explicit fix: it directly addresses the NameError when pydantic is absent by guarding the TypeAdapter import under TYPE_CHECKING and stringizing the annotation.",
+        "canonical_pr_reason": "pull_request:43656 is the better canonical representative because it is the broader CLI typing fix with an explicit issue link; pull_request:43842 is a separate, narrower pydantic-optional import/annotation guard.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 43842 is the best representative of the cluster because it has a clear explicit issue target and a narrowly scoped, concrete runtime fix.",
+        "best_pr_reason": "pull_request:43656 is the strongest single representative of the cluster: it fixes a concrete CLI NameError path and makes the typing/runtime behavior coherent. pull_request:43842 targets a different failure mode and should remain separate.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43656",
             "right": "pull_request:43842",
             "accept": false,
-            "reason": "Both touch src/transformers/cli/serve.py and mention TypeAdapter, but 43656 is a broader annotations/type-hint cleanup while 43842 fixes the specific missing-pydantic NameError path. They do not look like the same underlying bug or a single mergeable change."
+            "reason": "Reject: both PRs touch src/transformers/cli/serve.py and TypeAdapter, but they fix different bugs. 43656 addresses a broader annotation/runtime NameError in the CLI, while 43842 only guards the TypeAdapter type import when pydantic is absent. They are not the same underlying change and would not naturally merge into one PR."
           }
         ]
       },
@@ -20556,8 +20556,8 @@
         "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:44018|pull_request:44116",
+        "pull_request:44068|pull_request:44722",
         "pull_request:44071|pull_request:44722",
         "pull_request:44018|pull_request:44085",
         "pull_request:44018|pull_request:44086",
@@ -20566,8 +20566,8 @@
         "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:44059|pull_request:44722",
         "pull_request:44017|pull_request:44161",
         "pull_request:44017|pull_request:44722",
         "pull_request:44017|pull_request:44073",
@@ -20577,8 +20577,8 @@
         "pull_request:44017|pull_request:44071",
         "pull_request:44073|pull_request:44101",
         "pull_request:44073|pull_request:44161",
-        "pull_request:44059|pull_request:44071",
         "pull_request:44056|pull_request:44101",
+        "pull_request:44059|pull_request:44071",
         "pull_request:44018|pull_request:44059",
         "pull_request:44059|pull_request:44068",
         "pull_request:44101|pull_request:44161",
@@ -20611,112 +20611,118 @@
         "pull_request:44044|pull_request:44129",
         "pull_request:44044|pull_request:44154"
       ],
-      "prepared_review_unit_hash": "0b605753579ac199c989903040bbf62d1b108b9df01dcf61d52ae2942910bff8",
+      "prepared_review_unit_hash": "a47f17aec74d9ec27f72c99c6785b5090cd477b1195e63656bcbc1fde62b7e3c",
       "component_budget": {
         "node_count": 28,
         "item_count": 28,
         "soft_pair_count": 78,
-        "serialized_chars": 84890,
-        "estimated_input_tokens": 21223,
-        "estimated_eval_tokens": 42702
+        "serialized_chars": 84896,
+        "estimated_input_tokens": 21224,
+        "estimated_eval_tokens": 42704
       },
       "budget": {
         "node_count": 27,
         "item_count": 27,
         "soft_pair_count": 72,
-        "serialized_chars": 75086,
-        "estimated_input_tokens": 18772,
-        "estimated_eval_tokens": 37800
+        "serialized_chars": 75093,
+        "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 family of standardized output-tracing refactors across many different model implementations, not one duplicate stack. The only clear duplicates are within the same model/code path: GPT-Neo, and the GPT-J/CodeGen variants; the rest are related by theme only.",
-        "confidence": 0.88,
+        "summary": "This cluster is mostly a set of model-specific output-tracing decorator refactors under the same tracking theme (#43979), not true duplicates across architectures. The only strong duplicate candidates are the GPT-J/CodeGen variants and the GPT-Neo variants; the rest are related but distinct model migrations.",
+        "confidence": 0.77,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No issue artifact is present. PR 44722 is the best canonical representative because it is the clearest consolidated GPT-J/CodeGen standardized output-tracing refactor and matches the accepted overlap in that code path.",
+        "canonical_pr_reason": "PR #44007 is the best representative of the cluster: it explicitly targets #43979, has a coherent standardized output-tracing migration, and spans multiple vision model files rather than a narrow one-off tweak.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 44722 is the strongest representative of the cluster\u2019s duplicate core: it covers the GPT-J/CodeGen tracing rewrite in the same files and reads like the later, more complete version of that change.",
+        "best_pr_reason": "PR #44007 is the strongest single representative of the broader output-tracing refactor work: explicit tracking issue, broad cross-model implementation, and a clear decorator-based migration path.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43996",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "Same broad output-tracing theme, but different model implementations and code paths; not a safe duplicate."
+            "reason": "Different model families and different code paths; sharing #43979 is not enough to make them duplicates."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "Different model families and different forward paths; only the refactor theme overlaps."
+            "reason": "Both are output-tracing refactors, but for different architectures (CVT/FNet vs DeBERTa) and not the same underlying change."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44085",
             "accept": true,
-            "reason": "Both touch GPT-J output tracing in the same file and are reshaping the same forward/return behavior; they look like overlapping variants of the same change."
+            "reason": "Both are GPT-J output-tracing refactors in the same model file and appear to be iterative versions of the same change."
           },
           {
             "left": "pull_request:44007",
             "right": "pull_request:44072",
             "accept": false,
-            "reason": "Both are standardized output-tracing refactors, but they affect different vision models and separate code paths."
+            "reason": "Different model families and implementation details; they are separate output-tracing migrations, not one fix."
           },
           {
             "left": "pull_request:44072",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "Different models and different implementation paths; same refactor pattern only at a very high level."
+            "reason": "EfficientNet vs GPT-J/CodeGen are unrelated code paths despite both following the same refactor pattern."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44072",
             "accept": false,
-            "reason": "Different model code paths; sharing the same tracking issue/theme is not enough to call them duplicates."
+            "reason": "Different models and different files; same initiative, but not the same underlying bug/change."
           },
           {
             "left": "pull_request:44013",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "MobileNetV2 vs DeBERTaV2 are unrelated implementations; these are not the same bug/change."
+            "reason": "MobileNetV2 and DeBERTa are unrelated model implementations, so these are not duplicates."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "Different models and different tracing logic; same refactor umbrella only."
+            "reason": "GPT-J/CodeGen output tracing is unrelated to MGP-STR output tracing; different concrete code paths."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-J/CodeGen vs MPT are distinct code paths; not the same underlying change."
+            "reason": "GPT-J/CodeGen and MPT are different architectures; this is the same refactor theme, not the same change."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44068",
             "accept": true,
-            "reason": "Both are GPT-Neo output-tracing refactors in the same model file/tests and appear to target the same concrete implementation change."
+            "reason": "Both are GPT-Neo output-tracing refactors touching the same model file and look like alternate/updated versions of the same work."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44068",
             "accept": false,
-            "reason": "GPT-J/CodeGen and GPT-Neo are different model implementations, so these are not duplicates."
+            "reason": "GPT-J/CodeGen vs GPT-Neo are different models and code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44019",
+            "accept": false,
+            "reason": "GPT-Neo and ResNet are unrelated; the similar wording is only the shared standardization pattern."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44722",
             "accept": false,
-            "reason": "ResNet vs GPT-J/CodeGen: same general refactor style, but not the same code path."
+            "reason": "ResNet output tracing and GPT-J/CodeGen output tracing are separate model-specific migrations."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "Different model families and different forward implementations."
+            "reason": "ResNet vs GPT-J are different concrete implementations, so they are not duplicates."
           },
           {
             "left": "pull_request:44086",
@@ -20728,259 +20734,475 @@
             "left": "pull_request:44018",
             "right": "pull_request:44019",
             "accept": false,
-            "reason": "Different models; only the standardized output-tracing motif overlaps."
+            "reason": "GPT-Neo and ResNet are different architectures; no shared underlying bug/change."
           },
           {
             "left": "pull_request:44019",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "ResNet and MPT are distinct implementations, so these are not the same change."
+            "reason": "ResNet and MPT are separate model refactors, not the same fix."
           },
           {
             "left": "pull_request:44068",
             "right": "pull_request:44116",
             "accept": false,
-            "reason": "GPT-Neo and Flaubert are unrelated code paths."
+            "reason": "GPT-Neo and Flaubert/XLM output tracing are different model-specific changes."
           },
           {
-            "left": "pull_request:44068",
-            "right": "pull_request:44722",
+            "left": "pull_request:44018",
+            "right": "pull_request:44116",
             "accept": false,
-            "reason": "Different models and different concrete refactors; same umbrella issue only."
+            "reason": "GPT-Neo and Flaubert are unrelated code paths."
           },
           {
-            "left": "pull_request:44018",
-            "right": "pull_request:44116",
+            "left": "pull_request:44068",
+            "right": "pull_request:44722",
             "accept": false,
-            "reason": "GPT-Neo vs Flaubert are different implementations."
+            "reason": "GPT-Neo is unrelated to GPT-J/CodeGen despite the shared refactor theme."
           },
           {
             "left": "pull_request:44071",
-            "right": "pull_request:44722",
+            "right": "pull_request:44086",
             "accept": false,
-            "reason": "MPT and GPT-J/CodeGen are separate model-specific changes."
+            "reason": "MPT and MGP-STR are different architectures and separate changes."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44085",
             "accept": false,
-            "reason": "GPT-Neo vs GPT-J are different model code paths."
+            "reason": "GPT-Neo and GPT-J are different model implementations."
           },
           {
             "left": "pull_request:44018",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-Neo and MGP-STR are different implementations."
+            "reason": "GPT-Neo vs MGP-STR are not the same code-path problem."
           },
           {
             "left": "pull_request:44068",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-Neo and MGP-STR are unrelated model paths."
+            "reason": "GPT-Neo and MGP-STR are unrelated model migrations."
           },
           {
-            "left": "pull_request:44059",
-            "right": "pull_request:44086",
+            "left": "pull_request:44068",
+            "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-2 vs MGP-STR are different code paths; not duplicates."
+            "reason": "GPT-Neo and MPT are different models with separate output-tracing implementations."
           },
           {
-            "left": "pull_request:44085",
+            "left": "pull_request:44059",
             "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-J and MGP-STR are unrelated implementations."
+            "reason": "GPT-2 and MGP-STR are unrelated code paths."
           },
           {
-            "left": "pull_request:44098",
-            "right": "pull_request:44101",
+            "left": "pull_request:44085",
+            "right": "pull_request:44086",
             "accept": false,
-            "reason": "Different model families and different files; only the refactor theme overlaps."
+            "reason": "GPT-J output tracing and MGP-STR output tracing are different model-specific changes."
           },
           {
             "left": "pull_request:44059",
-            "right": "pull_request:44722",
-            "accept": false,
-            "reason": "GPT-2 and GPT-J/CodeGen are separate model implementations."
-          },
-          {
-            "left": "pull_request:44018",
             "right": "pull_request:44071",
             "accept": false,
-            "reason": "GPT-Neo and MPT are different code paths."
+            "reason": "GPT-2 and MPT are different architectures and not duplicates."
           },
           {
-            "left": "pull_request:44017",
-            "right": "pull_request:44161",
+            "left": "pull_request:44018",
+            "right": "pull_request:44059",
             "accept": false,
-            "reason": "SegFormer and LongT5 are unrelated implementations."
+            "reason": "GPT-Neo and GPT-2 are distinct model refactors, even if both standardize output handling."
           },
           {
-            "left": "pull_request:44017",
-            "right": "pull_request:44722",
+            "left": "pull_request:43997",
+            "right": "pull_request:44017",
             "accept": false,
-            "reason": "Different model code paths; shared output-tracing refactor theme only."
+            "reason": "RegNet and SegFormer are different model families; same issue target does not make them duplicates."
           },
           {
-            "left": "pull_request:44017",
-            "right": "pull_request:44073",
+            "left": "pull_request:44071",
+            "right": "pull_request:44076",
             "accept": false,
-            "reason": "SegFormer and VisualBert are different model implementations."
+            "reason": "MPT and ImageGPT are separate model code paths."
           },
           {
-            "left": "pull_request:44068",
-            "right": "pull_request:44071",
+            "left": "pull_request:44074",
+            "right": "pull_request:44101",
             "accept": false,
-            "reason": "GPT-Neo and MPT are separate code paths."
+            "reason": "TextNet and Flaubert/XLM are unrelated implementations."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44101",
             "accept": false,
-            "reason": "MPNet and Flaubert/XLM are unrelated model-specific changes."
+            "reason": "MPNet and Flaubert/XLM are different model-specific migrations."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44017",
-            "accept": false,
-            "reason": "RegNet and SegFormer are different vision model implementations."
-          },
-          {
-            "left": "pull_request:43997",
-            "right": "pull_request:44056",
-            "accept": false,
-            "reason": "RegNet and MPNet are separate code paths."
-          },
-          {
-            "left": "pull_request:44071",
-            "right": "pull_request:44076",
+            "left": "pull_request:44056",
+            "right": "pull_request:44161",
             "accept": false,
-            "reason": "MPT and ImageGPT are different models; same refactor umbrella only."
+            "reason": "MPNet and LongT5 are separate architectures and not the same underlying change."
           },
           {
-            "left": "pull_request:44074",
-            "right": "pull_request:44101",
+            "left": "pull_request:44059",
+            "right": "pull_request:44722",
             "accept": false,
-            "reason": "TextNet and Flaubert/XLM are unrelated code paths."
+            "reason": "GPT-2 and GPT-J/CodeGen are different models and code paths."
           },
           {
-            "left": "pull_request:44056",
+            "left": "pull_request:44017",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "MPNet and LongT5 are different model implementations."
+            "reason": "SegFormer and LongT5 are unrelated model implementations."
           },
           {
             "left": "pull_request:43997",
             "right": "pull_request:44018",
             "accept": false,
-            "reason": "RegNet and GPT-Neo are unrelated model code paths."
+            "reason": "RegNet and GPT-Neo are different model families."
           },
           {
             "left": "pull_request:43997",
-            "right": "pull_request:44161",
+            "right": "pull_request:44056",
             "accept": false,
-            "reason": "RegNet and LongT5 are different implementations."
+            "reason": "RegNet and MPNet are separate output-tracing migrations."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "SqueezeBERT and MPT are unrelated code paths."
           },
           {
             "left": "pull_request:44074",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "TextNet and LongT5 are separate model-specific changes."
+            "reason": "TextNet and LongT5 are distinct models with separate changes."
           },
           {
-            "left": "pull_request:43997",
-            "right": "pull_request:44071",
+            "left": "pull_request:44056",
+            "right": "pull_request:44068",
             "accept": false,
-            "reason": "RegNet and MPT are different code paths."
+            "reason": "MPNet and GPT-Neo are different architectures."
           },
           {
             "left": "pull_request:44010",
-            "right": "pull_request:44071",
+            "right": "pull_request:44018",
             "accept": false,
-            "reason": "SqueezeBert and MPT are unrelated implementations."
+            "reason": "SqueezeBERT and GPT-Neo are unrelated model migrations."
           },
           {
-            "left": "pull_request:44010",
-            "right": "pull_request:44073",
+            "left": "pull_request:44071",
+            "right": "pull_request:44129",
             "accept": false,
-            "reason": "SqueezeBert and VisualBert are different code paths."
+            "reason": "MPT and SpeechT5 are different code paths."
           },
           {
-            "left": "pull_request:44018",
-            "right": "pull_request:44129",
+            "left": "pull_request:44071",
+            "right": "pull_request:44086",
             "accept": false,
-            "reason": "GPT-Neo and SpeechT5 are distinct model implementations."
+            "reason": "MPT and MGP-STR are different models; not the same bug/change."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44161",
             "accept": false,
-            "reason": "CVT/FNet and LongT5 are unrelated model paths."
+            "reason": "CVT/FNet and LongT5 are unrelated model implementations."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44017",
             "accept": false,
-            "reason": "Different model families and different concrete implementations."
+            "reason": "CVT/FNet and SegFormer are different model-specific refactors."
           },
           {
             "left": "pull_request:44010",
             "right": "pull_request:44044",
             "accept": false,
-            "reason": "SqueezeBert and DeBERTaV2 are separate code paths."
+            "reason": "SqueezeBERT and DeBERTa are unrelated code paths."
           },
           {
             "left": "pull_request:44044",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "DeBERTaV2 and TextNet are different model implementations."
+            "reason": "DeBERTa and TextNet are separate model migrations."
           },
           {
             "left": "pull_request:44066",
             "right": "pull_request:44722",
             "accept": true,
-            "reason": "Same GPT-J/CodeGen output-tracing refactor in the same files; 44722 looks like a later consolidated version of the same change."
+            "reason": "Both target the same GPT-J/CodeGen output-tracing refactor and 44722 looks like a later, broader revision of the same change."
           },
           {
             "left": "pull_request:44085",
             "right": "pull_request:44722",
             "accept": true,
-            "reason": "Both are GPT-J tracing refactors over the same code path, and they could plausibly be merged into one PR."
+            "reason": "Both modify GPT-J output tracing in the same file; 44722 is a natural superset/revision of the same work."
           },
           {
             "left": "pull_request:44085",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "GPT-J and SpeechT5 are unrelated model code paths."
+            "reason": "GPT-J and SpeechT5 are unrelated models."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44007",
             "accept": false,
-            "reason": "Different model implementations; only the output-tracing refactor style matches."
+            "reason": "CVT/FNet and ResNet/RegNet/RT-DETR are different model migrations, not one duplicate change."
           },
           {
             "left": "pull_request:43996",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "CVT/FNet and SpeechT5 are unrelated code paths."
+            "reason": "CVT/FNet and SpeechT5 are unrelated output-tracing changes."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "DeBERTa and SpeechT5 are different architectures and code paths."
           },
           {
             "left": "pull_request:44044",
+            "right": "pull_request:44154",
+            "accept": false,
+            "reason": "DeBERTa and VITS are separate model-specific changes."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "GPT-J/CodeGen and EfficientNet are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:44072",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "EfficientNet and GPT-J/CodeGen are different code paths."
+          },
+          {
+            "left": "pull_request:44013",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "MobileNetV2 and DeBERTa are unrelated models."
+          },
+          {
+            "left": "pull_request:44098",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "ViLT and Flaubert/XLM are separate model-specific migrations."
+          },
+          {
+            "left": "pull_request:44098",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "ViLT and VisualBERT are related visually, but they are different implementations and not the same concrete change."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "MPNet and Flaubert/XLM are different models and not duplicates."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-2 and MGP-STR are unrelated output-tracing changes."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "RegNet and TextNet are different model implementations."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "RegNet and SpeechT5 are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "GPT-Neo and LongT5 are distinct architectures."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "RegNet and GPT-Neo are different models."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "MPT and SpeechT5 are not the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "GPT-2 and GPT-Neo are different model families."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "VisualBERT and Flaubert/XLM are separate model migrations."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "VisualBERT and LongT5 are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "MPNet and Flaubert/XLM are different model-specific refactors."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-2 and MPT are unrelated models."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44059",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-2 are different architectures."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44066",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J/CodeGen are different code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J are different model implementations."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo and MGP-STR are unrelated changes."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo and MGP-STR are separate model migrations."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-Neo and MPT are different architectures."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "GPT-2 and TextNet are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "ResNet and Flaubert/XLM are different model families."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-2 and MGP-STR are distinct model-specific refactors."
+          },
+          {
+            "left": "pull_request:44098",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "ViLT and Flaubert/XLM are not the same underlying change."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "RegNet and GPT-Neo are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "MPT and ImageGPT are different code paths."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "SqueezeBERT and GPT-Neo are unrelated."
+          },
+          {
+            "left": "pull_request:44071",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "DeBERTaV2 and SpeechT5 are different model implementations."
+            "reason": "MPT and SpeechT5 are separate model migrations."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "SegFormer and VisualBERT are different model-specific changes."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44059",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-2 are different output-tracing implementations."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "SqueezeBERT and VisualBERT are unrelated model families."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "RegNet and TextNet are distinct model code paths."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "SqueezeBERT and DeBERTa are unrelated."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "DeBERTa and TextNet are different architectures."
           },
           {
             "left": "pull_request:44044",
             "right": "pull_request:44154",
             "accept": false,
-            "reason": "DeBERTaV2 and ViT/S are separate code paths and not the same bug/change."
+            "reason": "DeBERTa and VITS are separate model-specific migrations."
           }
         ]
       },
       "evaluator_result": {
         "accept": true,
-        "feedback": "The summary is grounded in the packet and stays conservative on cross-model pairs. The accepted soft-edge pairs are same-model/code-path refactors that look mergeable into a single bugfix PR, while the rejected pairs correctly avoid treating shared output-tracing style as duplication."
+        "feedback": "The summary is grounded in the packet: it correctly treats most PRs as architecture-specific output-tracing refactors under #43979, and it only elevates the GPT-J/CodeGen and GPT-Neo pairs as likely duplicates/revisions. The rejection calls are conservative, and the accepted pairs are plausibly mergeable revisions of the same model-specific bugfix."
       },
       "evaluator_used": true,
       "retried": false,
@@ -21018,70 +21240,70 @@
         "pull_request:43997|pull_request:44010",
         "pull_request:44010|pull_request:44129"
       ],
-      "prepared_review_unit_hash": "fa152c006355a7a3c3b6d07fba825a1f5ecb45a8ed3b18c49edbf8f4cfabdf25",
+      "prepared_review_unit_hash": "03f80aa1483d98ee382c6f68bd4070f4cc10996d29505197da02bee326ccf972",
       "component_budget": {
         "node_count": 28,
         "item_count": 28,
         "soft_pair_count": 78,
-        "serialized_chars": 84890,
-        "estimated_input_tokens": 21223,
-        "estimated_eval_tokens": 42702
+        "serialized_chars": 84896,
+        "estimated_input_tokens": 21224,
+        "estimated_eval_tokens": 42704
       },
       "budget": {
         "node_count": 8,
         "item_count": 8,
         "soft_pair_count": 6,
-        "serialized_chars": 17125,
-        "estimated_input_tokens": 4282,
-        "estimated_eval_tokens": 8820
+        "serialized_chars": 17124,
+        "estimated_input_tokens": 4281,
+        "estimated_eval_tokens": 8818
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": true,
       "analyst_result": {
-        "summary": "These PRs all follow the same broad theme\u2014migrating model forward passes to standardized output capture/tracing\u2014but they are model-specific refactors, not the same underlying bug or change. None of the soft-edge pairs should be merged as duplicates.",
-        "confidence": 0.94,
+        "summary": "These PRs share a broad theme of migrating model forward passes to standardized output tracing, but they are separate model-specific refactors rather than duplicates of one underlying bug or change.",
+        "confidence": 0.93,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "44129 is the best standalone anchor because it has an explicit issue target and a concrete, model-specific output-tracing refactor; the others are related migration examples but each touches a different model and code path.",
+        "canonical_pr_reason": "PR #44129 is the most well-scoped and explicit representative of the theme because it has a linked issue and a concrete SpeechT5 output-tracing refactor, but it is still model-specific rather than a duplicate of the others.",
         "best_issue_reason": null,
-        "best_pr_reason": "44129 is the strongest representative of the cluster theme, but it is still only one model-specific implementation rather than a duplicate of the others.",
+        "best_pr_reason": "PR #44129 is the strongest single representative of the cluster\u2019s topic, with an explicit issue target and a focused code change; however, the cluster does not contain a single shared PR-worthy change across all items.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44024",
             "right": "pull_request:44076",
             "accept": false,
-            "reason": "Same migration pattern, but different models (FocalNet vs ImageGPT) and different code paths; too broad to be a duplicate."
+            "reason": "Both are model-specific output-tracing migrations, but FocalNet and ImageGPT are different code paths and architectures, so they are not the same concrete change."
           },
           {
             "left": "pull_request:44073",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "Both are standardized output-tracing refactors, but they target different architectures (VisualBert vs TextNet) and are independent changes."
+            "reason": "VisualBert and TextNet are separate models; despite similar decorator refactors, these are distinct per-model implementations, not one mergeable PR."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44073",
             "accept": false,
-            "reason": "Shared output-capture refactor theme only; MPNet and VisualBert are separate model implementations with distinct forward logic."
+            "reason": "MPNet and VisualBert touch different model stacks and output-capture wiring, so they are related only by pattern, not by the same underlying fix."
           },
           {
             "left": "pull_request:44056",
             "right": "pull_request:44074",
             "accept": false,
-            "reason": "Different models and different refactor details; the overlap is only the general standardized outputs migration."
+            "reason": "These are separate model migrations to standardized output capture, not the same code-path problem or a single PR that could plausibly be merged together."
           },
           {
             "left": "pull_request:43997",
             "right": "pull_request:44010",
             "accept": false,
-            "reason": "RegNet and SqueezeBert are unrelated model files; both adopt decorator-based output tracing, but not the same underlying fix."
+            "reason": "RegNet and SqueezeBert are unrelated model implementations; the shared decorator migration is too broad to treat as a duplicate."
           },
           {
             "left": "pull_request:44010",
             "right": "pull_request:44129",
             "accept": false,
-            "reason": "Both involve output-capture refactoring, but SqueezeBert and SpeechT5 are distinct code paths and 44129 is tied to a specific issue while 44010 is a separate migration."
+            "reason": "Although both refactor output collection, SqueezeBert and SpeechT5 are distinct model-specific changes and should not be merged as one PR."
           }
         ]
       },
@@ -21161,274 +21383,274 @@
         "pull_request:44028|pull_request:44030",
         "pull_request:44029|pull_request:44030"
       ],
-      "prepared_review_unit_hash": "afbe5ab49a6b7a3e6d846b11bd32fb77162a5beb18e03952e4d7b3ba5c8dc829",
+      "prepared_review_unit_hash": "6bcf3a2530400554388b5ade249c8422d76e6a577d1c6136b0f777ec9ac1fa39",
       "component_budget": {
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 40,
-        "serialized_chars": 40181,
-        "estimated_input_tokens": 10046,
-        "estimated_eval_tokens": 20348
+        "serialized_chars": 40172,
+        "estimated_input_tokens": 10043,
+        "estimated_eval_tokens": 20342
       },
       "budget": {
         "node_count": 13,
         "item_count": 13,
         "soft_pair_count": 40,
-        "serialized_chars": 40181,
-        "estimated_input_tokens": 10046,
-        "estimated_eval_tokens": 20348
+        "serialized_chars": 40172,
+        "estimated_input_tokens": 10043,
+        "estimated_eval_tokens": 20342
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These PRs share a common output-tracing refactor theme, but they target different model families and different concrete code paths. None of the soft-edge pairs look like true duplicates or mergeable into one PR.",
-        "confidence": 0.96,
+        "summary": "These are separate model-specific output-tracing refactors, not duplicate fixes; they share the same tooling pattern but not the same code path or mergeable change.",
+        "confidence": 0.95,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "44003 is the strongest representative of the cluster: it spans both mamba and falcon_mamba and adds the output-capture plumbing, rather than a small one-off wrapper change.",
+        "canonical_pr_reason": "PR 44025 is the best representative of the cluster: it applies the output-tracing refactor across two closely related depth_anything variants and clearly shows the new can_return_tuple pattern.",
         "best_issue_reason": null,
-        "best_pr_reason": "44003 is the broadest and most infrastructure-heavy PR in the set, so it is the best single representative for triage.",
+        "best_pr_reason": "PR 44025 is the strongest single stand-in for the group because it is a clean, substantive example of the refactor family without being limited to a single tiny code tweak.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:43998",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different model families (timm_backbone vs upernet); same refactor style only."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44000",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different architectures and code paths; both are output-tracing cleanups, not the same change."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Model-specific refactors in unrelated models; not one underlying bug or fix."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "timm_backbone and mamba are unrelated code paths; shared theme is too broad."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different models and implementations; cannot plausibly merge as one PR."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different model families; same output-capturing pattern only."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "MobileNetV1 and UperNet touch different forward paths and model outputs."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different models and API surfaces; not the same concrete fix."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44026",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "MobileNetV1 refactor is unrelated to vision_encoder_decoder's wrapper logic."
           },
           {
             "left": "pull_request:43999",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different multimodal wrappers and code paths; not duplicates."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44001",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Both are refactors, but for different model implementations and forward signatures."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different model subsystems; same broad cleanup theme only."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "vision_text_dual_encoder and falcon_mamba do not share the same underlying change."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different model families and forward paths; not mergeable into one PR."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44002",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "UnivNet and UperNet are unrelated model code paths."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Same refactor pattern, but different model-specific behavior."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44025",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different models and output contracts; not the same fix."
           },
           {
             "left": "pull_request:44002",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different architectures and output plumbing; too broad to treat as duplicate."
           },
           {
             "left": "pull_request:44002",
             "right": "pull_request:44003",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "UperNet vs mamba are distinct code paths with no shared concrete bug."
           },
           {
             "left": "pull_request:44003",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different model families and implementations; only the refactor motif matches."
           },
           {
             "left": "pull_request:44003",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "mamba and codegen changes are unrelated forward-path edits."
           },
           {
             "left": "pull_request:43998",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different models and distinct output handling; not the same underlying issue."
           },
           {
             "left": "pull_request:44001",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "UnivNet and CodeGen are separate code paths; same cleanup theme only."
           },
           {
             "left": "pull_request:44002",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "UperNet and CodeGen are unrelated; cannot be merged as one bug fix."
           },
           {
             "left": "pull_request:44000",
             "right": "pull_request:44004",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different model types and APIs; shared title is not enough for duplication."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44026",
             "accept": false,
-            "reason": "Analogous output-tracing refactors in different wrapper models; not the same concrete change."
+            "reason": "Depth Anything and Vision-Encoder-Decoder are different model wrappers and fixes."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Analogous output-tracing refactors in different wrapper models; not the same concrete change."
+            "reason": "Different encoder-decoder stacks; same refactor pattern only."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Depth Anything and SuperPoint are unrelated model paths."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different model architectures; not the same concrete change."
           },
           {
             "left": "pull_request:44025",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Depth Anything and DPR do different things and touch different forward logic."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Vision encoder-decoder and RWKV are unrelated implementations."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44027",
             "accept": false,
-            "reason": "Different wrapper models; similar refactor, but not the same code-path problem."
+            "reason": "Vision and speech encoder-decoder wrappers are separate code paths."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different models and different output-tracing behavior; not duplicates."
           },
           {
             "left": "pull_request:44026",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Distinct model families; same broad cleanup is insufficient."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44028",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Speech encoder-decoder and SuperPoint are unrelated model code paths."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different architectures and forward signatures; not the same fix."
           },
           {
             "left": "pull_request:44027",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Speech encoder-decoder vs DPR are separate changes with no shared code path."
           },
           {
             "left": "pull_request:44028",
             "right": "pull_request:44029",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "SuperPoint and RWKV are unrelated; only the refactor theme overlaps."
           },
           {
             "left": "pull_request:44028",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "Different model families and behavior; not one underlying bug."
           },
           {
             "left": "pull_request:44029",
             "right": "pull_request:44030",
             "accept": false,
-            "reason": "Different model families and code paths; only the refactor pattern is shared."
+            "reason": "RWKV and DPR are distinct implementations; same output-tracing label only."
           }
         ]
       },
@@ -21458,7 +21680,7 @@
       "soft_pairs": [
         "pull_request:44662|pull_request:45401"
       ],
-      "prepared_review_unit_hash": "c4e0c40fe62698cabbf4d84e34c9c34b897e315aa9c3b93c72d161604b61e80c",
+      "prepared_review_unit_hash": "6b32b9128918ddf7d9cb847d10f46607950449b9c83aa54de04f35b6fc2b9417",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -21480,18 +21702,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are two unrelated model-addition PRs: one adds PenguinVL, the other adds Voxtral TTS. They share generic Transformers boilerplate files, but they do not fix the same code-path problem and should not be merged together.",
-        "confidence": 0.98,
+        "summary": "These are two separate model-integration PRs: one adds PenguinVL, the other adds Voxtral TTS. They share framework registration and docs plumbing, but they target different models and code paths, so they are not duplicates.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "Pick PR #45401 as the canonical PR: it has an explicit issue target, is narrowly scoped to Voxtral TTS support, and has a clear standalone implementation path. PR #44662 is a separate PenguinVL addition.",
+        "canonical_pr_reason": "PR 45401 is the more canonical representative of its own change because it has an explicit issue target (#45027) and a tightly scoped Voxtral TTS implementation.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45401 is the better representative because it is explicitly tied to issue 45027 and has a focused feature addition. PR #44662 is equally valid as a separate PR, but it addresses a different model and is not a duplicate.",
+        "best_pr_reason": "If one PR must be picked as the better representative, 45401 is more clearly anchored and narrowly scoped; 44662 is a separate PenguinVL feature, not the same change.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44662",
             "right": "pull_request:45401",
             "accept": false,
-            "reason": "Different models and different concrete features: PenguinVL implementation vs Voxtral TTS support. Shared auto/docs files are generic scaffolding, not evidence of the same underlying bug or change."
+            "reason": "Different model additions with different implementation targets (PenguinVL vs Voxtral TTS). Shared auto-registration/docs files are boilerplate and do not indicate the same underlying bug or feature."
           }
         ]
       },
@@ -21521,12 +21743,12 @@
       "soft_pairs": [
         "pull_request:44979|pull_request:45363"
       ],
-      "prepared_review_unit_hash": "196d653c55eab60ae70a4e90157d6500ab5ed7e88c055295e1ce554cd7e96705",
+      "prepared_review_unit_hash": "a511a5642e3e0221792a75ac56dcffb5d3668018a395d428bbd33304dbc83ae1",
       "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
       },
@@ -21534,7 +21756,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4408,
+        "serialized_chars": 4407,
         "estimated_input_tokens": 1102,
         "estimated_eval_tokens": 2460
       },
@@ -21543,18 +21765,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These two PRs are related in theme, but they are not duplicates: #44979 introduces a general Module Fusion API, while #45363 builds a separate KernelConfig-based n-to-1 fusion feature on top of integration/model-utils changes. They address different layers of the stack and could not reasonably be merged as one PR without losing separability.",
-        "confidence": 0.91,
+        "summary": "The two PRs are related around module/kernel fusion, but they are not the same underlying change: #44979 introduces a general Module Fusion API, while #45363 adds a specific KernelConfig-based n-to-1 kernel fusion integration and example. This is a relationship/dependency, not a duplicate.",
+        "confidence": 0.88,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "#45363 is the better representative of the cluster because it is the more concrete end-user feature and spans the actual kernel-fusion integration path, whereas #44979 is lower-level API scaffolding.",
+        "canonical_pr_reason": "#44979 is the foundational API PR: it introduces the generic module fusion abstraction itself, while #45363 appears to build on that idea with a specific kernel-fusion integration.",
         "best_issue_reason": null,
-        "best_pr_reason": "#45363 is the most substantial and user-visible PR in the set; #44979 is foundational but narrower and primarily enables later work.",
+        "best_pr_reason": "#45363 is the more concrete end-to-end user-facing feature PR, with integration points and an example, so it is the stronger representative of the applied fusion workflow.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:44979",
             "right": "pull_request:45363",
             "accept": false,
-            "reason": "#44979 adds a generic module-fusion API in a new module, while #45363 implements KernelConfig-driven n-to-1 fusion across integrations/model loading. They share a fusion theme but not the same concrete change or code path, and they would not naturally be merged into one PR."
+            "reason": "Different changes and codepaths: one adds a generic fusion API module, the other wires KernelConfig into model loading/integrations for a specific n-to-1 fusion use case. They are related, but not the same bug or mergeable duplicate PR."
           }
         ]
       },
@@ -21584,12 +21806,12 @@
       "soft_pairs": [
         "pull_request:45086|pull_request:45317"
       ],
-      "prepared_review_unit_hash": "9906365a5fa6eebd2769ad8ae63c796727d870f861b7da3168df6f3ec4ec0305",
+      "prepared_review_unit_hash": "adb29ae74101b09d97da167fc060cc09ce7c23b18e00e9445b0f1ea1a5b10d2b",
       "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
       },
@@ -21597,7 +21819,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4413,
+        "serialized_chars": 4414,
         "estimated_input_tokens": 1104,
         "estimated_eval_tokens": 2464
       },
@@ -21606,18 +21828,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs fix the same AttributeError in `_patch_mistral_regex` by switching from `backend_tokenizer.pre_tokenizer` to `tokenizer.pre_tokenizer`; 45317 is the more complete PR because it includes a test and slightly broader review-ready context.",
-        "confidence": 0.97,
+        "summary": "Both PRs implement the same Mistral regex fix in the same code path; 45317 is the more complete version because it includes tests.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45317 is the better canonical PR because it covers the same fix as 45086 but also adds test coverage and has the more explicit, final title for the bug.",
+        "canonical_pr_reason": "PR 45317 is the better canonical representative: it fixes the same AttributeError in `_patch_mistral_regex` and also adds test coverage, making it the more complete change.",
         "best_issue_reason": null,
-        "best_pr_reason": "45317 is the strongest representative of the change set: same code-path fix, same target issue, and added regression test coverage.",
+        "best_pr_reason": "PR 45317 is the best overall artifact because it contains the same fix as 45086 plus validation in tests, so it is more suitable as the cluster representative.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45086",
             "right": "pull_request:45317",
             "accept": true,
-            "reason": "Same concrete bug fix in the same function and same file; the diff shows the same patch, and 45317 simply adds tests, so these are duplicates for triage."
+            "reason": "Same concrete bug fix in `src/transformers/tokenization_utils_tokenizers.py` for `_patch_mistral_regex`; the second PR is the same patch with added tests, so they are merge-equivalent."
           }
         ]
       },
@@ -21651,7 +21873,7 @@
         "pull_request:45569|pull_request:46565",
         "pull_request:46332|pull_request:46565"
       ],
-      "prepared_review_unit_hash": "da2e2ecb2d37c1fcf65b5187e73cfed2c3598eaa020468a3ff13f86bbe4f846a",
+      "prepared_review_unit_hash": "f5074be12dac8269bc9d8afdfa31b3875aa37baff9d3bfbced568153e124b56c",
       "component_budget": {
         "node_count": 4,
         "item_count": 4,
@@ -21673,30 +21895,30 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These PRs split into two unrelated streams: NemotronH refactoring/model split work, and Nemotron ASR streaming work for two different model generations. They share some infrastructure files, but they do not look like duplicates of the same concrete change.",
-        "confidence": 0.95,
+        "summary": "This is a mixed cluster: two NemotronH architecture/model-split PRs and two ASR streaming PRs. The overlaps are mostly broad subsystem similarity (auto mapping, docs, model scaffolding), not the same underlying change.",
+        "confidence": 0.86,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #46332 is the clearest representative of the ASR-streaming work: it introduces the Nemotron ASR stack, supporting losses, auto mappings, and the new model package. The other PRs are either a separate NemotronH refactor (#45569/#45115) or a follow-up variant for Nemotron 3.5 ASR (#46565).",
+        "canonical_pr_reason": "PR 46565 is the newest and broadest ASR-streaming implementation in the set, so it is the most reasonable single representative, though the cluster is not a true duplicate group.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #46332 is the best global representative because it is the foundational ASR streaming addition and the closest thing to a central change in this cluster. #46565 is related but targets a different model generation, and #45115/#45569 are a separate NemotronH architecture refactor.",
+        "best_pr_reason": "46565 is the strongest standalone representative because it has the clearest end-to-end feature scope and the most direct overlap with one other PR in the cluster.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45115",
             "right": "pull_request:45569",
             "accept": false,
-            "reason": "Both are NemotronH-related, but #45115 is a conversion-mapping/inheritance refactor while #45569 is a much broader model split into nemotron_h_dense/sparse plus docs/auto updates. Same family, not the same concrete change."
+            "reason": "Both concern NemotronH, but one is a checkpoint/key-mapping refactor and the other is a larger model-split/backward-compat implementation; related area, not the same concrete change."
           },
           {
             "left": "pull_request:45569",
             "right": "pull_request:46565",
             "accept": false,
-            "reason": "They touch some shared auto/docs files, but the actual work is different: NemotronH model refactor vs Nemotron ASR streaming. Different model families and code paths."
+            "reason": "They touch different model families entirely: NemotronH vs Nemotron ASR. Shared auto/docs files are too broad to treat as the same fix."
           },
           {
             "left": "pull_request:46332",
             "right": "pull_request:46565",
             "accept": false,
-            "reason": "Both are ASR streaming PRs, but one adds Nemotron ASR and the other adds Nemotron 3.5 ASR. That is a related feature area, not the same underlying change, and they would not plausibly collapse into one duplicate PR."
+            "reason": "Both are ASR streaming PRs, but they target different model versions (Nemotron ASR vs Nemotron 3.5 ASR) with separate model files and support code, so they are not the same underlying change."
           }
         ]
       },
@@ -21726,12 +21948,12 @@
       "soft_pairs": [
         "pull_request:45332|pull_request:45333"
       ],
-      "prepared_review_unit_hash": "98e578bc09c5e7b651d6a3a24799a4054cdf8e0e22e104e0c135bb971a9513a8",
+      "prepared_review_unit_hash": "3dd5c3885e559939a4f8932a99d2070ba877476223cb5b80a587256e4d5ad801",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5465,
+        "serialized_chars": 5466,
         "estimated_input_tokens": 1367,
         "estimated_eval_tokens": 2990
       },
@@ -21739,7 +21961,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5465,
+        "serialized_chars": 5466,
         "estimated_input_tokens": 1367,
         "estimated_eval_tokens": 2990
       },
@@ -21748,18 +21970,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related feature PRs for the same heterogeneous-model initiative, but they are not duplicates: #45333 adds per-layer heterogeneous config support, while #45332 extends that into broader heterogeneous modeling/runtime support.",
-        "confidence": 0.84,
+        "summary": "These are related feature PRs for the same heterogeneous-model initiative, but they are not duplicates: one adds per-layer config support, while the other extends that work into model/cache/masking integration and broader heterogeneous modeling behavior.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #45332 is the broader umbrella change: it includes heterogeneous model/runtime support plus the cache/masking/modeling updates, whereas #45333 is limited to per-layer config plumbing.",
+        "canonical_pr_reason": "PR 45332 is the broader umbrella change and includes the downstream modeling/cache integration built on top of heterogeneous configuration support.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45332 is the best representative of the cluster because it subsumes the larger end-to-end heterogeneous-model feature and touches the main runtime paths.",
+        "best_pr_reason": "PR 45332 is the best representative for the cluster because it spans the full heterogeneous-model feature set rather than just the config layer.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45332",
             "right": "pull_request:45333",
             "accept": false,
-            "reason": "Related but not the same concrete change. #45333 focuses on config-layer support; #45332 adds broader modeling/cache/masking behavior and would not be a clean duplicate merge target for the same fix."
+            "reason": "Related feature work, but not the same concrete change: 45333 is focused on introducing per-layer configuration primitives, while 45332 builds on that to add heterogeneous modeling and cache/masking integration. They could be sequenced, not merged as one PR."
           }
         ]
       },
@@ -21789,7 +22011,7 @@
       "soft_pairs": [
         "pull_request:45415|pull_request:45425"
       ],
-      "prepared_review_unit_hash": "222274c9f369488930e7a8f014ab5657066393f09b7d6eed23778381c98fdfec",
+      "prepared_review_unit_hash": "acf5393e3c86341c3874e329fbd5bdb37600cd232763b738006c09b12526f8ab",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -21811,18 +22033,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related typing-focused PRs in the same codebase, but they are not the same fix/change. One is a broad repo-wide type-checking sweep, while the other is a narrower typing refinement around modeling_utils and related protocols.",
-        "confidence": 0.89,
+        "summary": "These are related typing-focused PRs, but they are not the same underlying change: one is a broad repo-wide type-checking sweep, the other adds typing helpers and fixes in core internals.",
+        "confidence": 0.9,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45415 is the better canonical PR because it is the broader umbrella change: a repository-wide typing pass with many file edits and mypy/pyproject adjustments, whereas PR 45425 is a narrower follow-up typing cleanup.",
+        "canonical_pr_reason": "No canonical PR: 45415 is a wide type-checking cleanup across many files, while 45425 introduces typing protocol/alias helpers and a smaller set of modeling changes; they solve different problems.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45415 is the best representative of the cluster because it covers the main type-checking initiative across multiple transformers modules; PR 45425 is more of a targeted supporting typing patch.",
+        "best_pr_reason": "45425 is the more coherent, self-contained typing-infrastructure change, but it is not a duplicate of the broader type-checking sweep in 45415.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45415",
             "right": "pull_request:45425",
             "accept": false,
-            "reason": "They share typing-related files and some overlap in modeled code, but the changes are not the same concrete code-path fix. 45415 is a broad type-checking sweep; 45425 adds typing protocols and annotations for specific modeling_utils/PEFT-related interfaces. They could not plausibly be merged as one duplicate PR."
+            "reason": "Overlapping type-related edits and shared files are incidental; 45415 is a broad type-checking pass, while 45425 adds specific typing abstractions and targeted fixes, so they are not mergeable as one PR."
           }
         ]
       },
@@ -21852,40 +22074,40 @@
       "soft_pairs": [
         "pull_request:45513|pull_request:46084"
       ],
-      "prepared_review_unit_hash": "14f26e981def45b90f8857077e9a391a2b25a40e370d72111a1b27f8fcd3e0cc",
+      "prepared_review_unit_hash": "10bd2eb9b92c2c559d7b934ba8c4a2f7f98288795caeb38d12d9f80290f1fb81",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4858,
-        "estimated_input_tokens": 1215,
-        "estimated_eval_tokens": 2686
+        "serialized_chars": 4903,
+        "estimated_input_tokens": 1226,
+        "estimated_eval_tokens": 2708
       },
       "budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4858,
-        "estimated_input_tokens": 1215,
-        "estimated_eval_tokens": 2686
+        "serialized_chars": 4903,
+        "estimated_input_tokens": 1226,
+        "estimated_eval_tokens": 2708
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are both cache-related forward-pass fixes, but they target different model families and different bugs: Qwen3.5/Olmo hybrid linear-attention cached multi-token continuation vs. Mamba2 use_cache with seq_len > 1. They should not be deduplicated.",
-        "confidence": 0.96,
+        "summary": "These are both PRs about cached forward paths with sequence length > 1, but they target different model families and different concrete bugs. They should not be deduplicated.",
+        "confidence": 0.95,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #46084 is the cleaner canonical representative: it has an explicit linked issue, a narrow Mamba2-specific fix, and a focused diff around seq_len > 1 cached execution.",
+        "canonical_pr_reason": "PR #46084 is the better canonical PR for the cluster because it has an explicit issue target, a focused fix in one model, and the change is tightly scoped to a concrete cache-handling bug.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #46084 is the best single representative of a cache/seq_len>1 bug fix because it is tightly scoped, explicitly tracked, and clearly describes the affected code path.",
+        "best_pr_reason": "PR #46084 is the strongest standalone PR: it directly fixes a specific use_cache/seq_len>1 failure mode, references a tracking issue, and has a narrow, testable code change.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45513",
             "right": "pull_request:46084",
             "accept": false,
-            "reason": "Different underlying problems in different architectures. #45513 fixes Qwen3.5/related linear-attention cached multi-token forward behavior across several Qwen/Olmo files, while #46084 fixes Mamba2 use_cache handling when seq_len > 1. The shared theme is only general cache logic, not the same concrete bug or mergeable change."
+            "reason": "Different underlying problems in different architectures (Qwen3.5/OLMo hybrid linear attention vs Mamba2 state caching). They share a broad theme of cached multi-token forward bugs, but not the same code path or the same concrete fix."
           }
         ]
       },
@@ -21915,7 +22137,7 @@
       "soft_pairs": [
         "pull_request:45548|pull_request:45662"
       ],
-      "prepared_review_unit_hash": "6e2736714192df94eb32634e663a025309c1b6162f19ce87f1759a0529e0cfbc",
+      "prepared_review_unit_hash": "3730dcf06d58f19c49b830c89974bdd7b8423660362b289e4a6fde764838646a",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -21937,18 +22159,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Two separate PRs about expert-parallel loading, but they fix different code paths: one preserves distributed config / EP metadata during serialization and model init, while the other prevents expert parameters from being overwritten during distributed shard loading. They are related in theme, not duplicates.",
-        "confidence": 0.94,
+        "summary": "The two PRs are both about expert parallelism, but they fix different bugs in different loading paths: one is about preserving distributed_config/EP state during Accelerate + ZeRO-3 loading, the other is about preventing expert weights from being overwritten during FSDP2/rank-0 broadcast. They are related, but not duplicates.",
+        "confidence": 0.95,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No single canonical PR for this cluster: PR #45548 and PR #45662 address distinct bugs in EP loading rather than the same underlying fix.",
+        "canonical_pr_reason": "PR #45662 is the better representative for this cluster because it targets a concrete EP weight-loading corruption path with a direct fix in model loading and related MoE/tensor-parallel logic.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #45662 is the stronger representative because it targets a concrete failure mode in the loading path (expert weights being overwritten by rank-0 broadcast) and includes the more substantial, self-contained fix.",
+        "best_pr_reason": "PR #45662 is the strongest standalone change: it addresses a specific, user-visible correctness bug and the diff shows a direct fix in the affected code path rather than only plumbing state through configuration.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45548",
             "right": "pull_request:45662",
             "accept": false,
-            "reason": "Both mention expert parallelism and loading, but the bugs differ: #45548 fixes missing distributed_config/has_ep handling, while #45662 fixes post-shard parameter wrapping to avoid overwriting experts. Not the same change or code-path problem."
+            "reason": "Both mention EP/loading, but #45548 fixes serialization and EP detection for Accelerate/ZeRO-3, while #45662 fixes expert parameters being overwritten during FSDP2 broadcast. Different underlying bugs and not mergeable as one PR."
           }
         ]
       },
@@ -21978,12 +22200,12 @@
       "soft_pairs": [
         "pull_request:45639|pull_request:45645"
       ],
-      "prepared_review_unit_hash": "73d1888645b344b2e464eff62509dadddd6810d8cb6f74fbc8ac2e8eb0ded3c2",
+      "prepared_review_unit_hash": "ac2e9d880f243c401e87dc90d31cda7b1d6d7a00acfd22096e790345335f5641",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4589,
+        "serialized_chars": 4590,
         "estimated_input_tokens": 1148,
         "estimated_eval_tokens": 2552
       },
@@ -21991,7 +22213,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 4589,
+        "serialized_chars": 4590,
         "estimated_input_tokens": 1148,
         "estimated_eval_tokens": 2552
       },
@@ -22000,18 +22222,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs target the same xdist/captured_info collision problem and share the same affected testing code; PR 45645 is the more complete superset, while PR 45639 is the narrower core fix.",
-        "confidence": 0.92,
+        "summary": "Both PRs address the same xdist-related captured_info collision problem around `testing_utils.py`; PR 45645 is the broader end-to-end fix, while 45639 is a narrower subset.",
+        "confidence": 0.9,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45645 is the better canonical PR because it appears to subsume the core xdist-safe output-file fix and also updates CI/reporting consumers to handle multiple captured_info files, making it the more complete change set.",
+        "canonical_pr_reason": "PR 45645 is the best canonical PR because it contains the core xdist-safe output-file fix plus the CI/workflow changes needed to consume multiple `captured_info*.txt` files, making it the more complete representation of the bug fix.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45645 is the strongest representative because it addresses the full end-to-end artifact collision problem and preserves downstream debug-log collection, whereas PR 45639 only fixes the file-writing side.",
+        "best_pr_reason": "PR 45645 is the strongest representative of the cluster: it fixes the same concrete collision bug and also updates the reporting path that reads the artifacts, so it is more comprehensive than 45639.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45639",
             "right": "pull_request:45645",
             "accept": true,
-            "reason": "Accept: both PRs address the same underlying xdist collision in captured_info/debug logging, share the same testing files, and 45645 extends the same fix into CI consumers rather than introducing a different bug or subsystem."
+            "reason": "Accept: both target the same captured_info/xdist collision in the testing debug logs; 45645 extends the same fix with CI artifact handling, so they are the same underlying change family."
           }
         ]
       },
@@ -22044,7 +22266,7 @@
         "pull_request:45670|pull_request:46401",
         "pull_request:46320|pull_request:46401"
       ],
-      "prepared_review_unit_hash": "cbbf12c1e0298c33809f79e135ace65ba23fb411b0f21e7d1dae7b28003c3988",
+      "prepared_review_unit_hash": "93b17b824c62ebd199d4eef9c8feb6540a860498328ff704c2cace5aeeda6c58",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
@@ -22066,30 +22288,30 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are three separate pull requests, not duplicates: one auto-model registry nit for GLM-ASR, and two different attention-dispatch feature PRs for different model families. The two attn_implementation PRs are similar in theme but target distinct code paths and model sets.",
-        "confidence": 0.96,
+        "summary": "These PRs are not duplicates. One is an AutoModel mapping nit for GLM-ASR, while the other two add attention-implementation dispatch support to different model families (RecurrentGemma vs BLIP-2/InstructBLIP/InstructBLIPVideo). They share a broad theme but not the same concrete code path or mergeable change.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #46401 is the strongest canonical PR for the attn_implementation theme: it has the broadest scope and highest activity, and its changes are clearly centered on adding dispatch support across BLIP-2/InstructBLIP/InstructBLIPVideo.",
+        "canonical_pr_reason": "PR 46401 is the strongest representative of the attn_implementation work because it is the broadest concrete implementation across multiple related BLIP-family models, with corresponding model and test changes. It is still not a duplicate of the others, just the most substantial of the three.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #46401 is the best representative of the cluster because it is the most substantial and complete example of the shared attention-dispatch feature work. The other PRs are either unrelated or narrower and should not be grouped as duplicates.",
+        "best_pr_reason": "PR 46401 is the best single representative if one had to pick a canonical PR from this set, because it contains the most complete implementation and test coverage for a specific feature area. PR 46320 is similar in theme but targets a different model family; PR 45670 is unrelated.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45670",
             "right": "pull_request:46320",
             "accept": false,
-            "reason": "Different changes: one edits AutoModel registry entries for GLM-ASR, the other adds attn_implementation dispatch for RecurrentGemma. Same broad transformer area, but not the same bug or mergeable fix."
+            "reason": "Different changes: AutoModel registry addition for glmasr vs RecurrentGemma attn_implementation dispatch. No shared concrete bug or mergeable code path."
           },
           {
             "left": "pull_request:45670",
             "right": "pull_request:46401",
             "accept": false,
-            "reason": "Unrelated code paths: GLM-ASR auto-class mapping versus BLIP-2/InstructBLIP attention-dispatch support. Similar in being model plumbing, but not the same underlying change."
+            "reason": "Different scope and subsystem: a model mapping tweak vs BLIP-family attention dispatch support. Same broad repo area only, not the same fix."
           },
           {
             "left": "pull_request:46320",
             "right": "pull_request:46401",
             "accept": false,
-            "reason": "Both are attention-dispatch PRs, but they apply to different model families and implementations. They could not plausibly be merged into one concrete fix for the same bug."
+            "reason": "Both add attn_implementation support, but for different model families and distinct files/tests. They are similar feature work, not the same underlying change."
           }
         ]
       },
@@ -22119,7 +22341,7 @@
       "soft_pairs": [
         "pull_request:45720|pull_request:46259"
       ],
-      "prepared_review_unit_hash": "edb1354d1f92d3ca9a78ef1e6c43bbb10fabe4ed8671c71e19e2d4de17fe903c",
+      "prepared_review_unit_hash": "5b50eedc7805034b983560b5f86f3ce89227faff1ae76451d45ef7f629ebb8bf",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -22141,18 +22363,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are two different linter-rule PRs that happen to touch some of the same model files, but they fix unrelated code paths: one updates forward return-tuple handling (TRF003), the other refactors weight initialization patterns (TRF018).",
-        "confidence": 0.97,
+        "summary": "These two pull requests are not duplicates: they target different linter rules and different code-path changes. PR 45720 adds `@can_return_tuple` and removes explicit `return_dict` handling in several model `forward()` methods, while PR 46259 changes weight initialization logic for `_init_weights` across many model classes. The overlap in filenames is just broad model coverage, not the same fix.",
+        "confidence": 0.96,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 46259 is the broader, cleaner representative of its own lint-rule cleanup, but it is not the same change as PR 45720.",
+        "canonical_pr_reason": "If a representative PR is needed, 46259 is the better canonical choice because its scope is clearer and more self-contained: it implements TRF018 via `_init_weights` refactoring across model implementations. 45720 is also a legitimate standalone PR, but it is only a partial implementation of TRF003 and has a different code change pattern.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46259 is the more substantial standalone PR in the cluster, but it should not be treated as a duplicate of PR 45720 because the underlying fixes differ.",
+        "best_pr_reason": "46259 is the better overall representative of the cluster because its change intent is explicit in the title and diff, and it affects a consistent code path across models. However, it should not be merged with 45720; they fix different problems.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45720",
             "right": "pull_request:46259",
             "accept": false,
-            "reason": "Different linter rules and different code paths: TRF003 adds can_return_tuple/return_dict handling in forward methods, while TRF018 rewrites _init_weights logic. Shared model filenames are incidental."
+            "reason": "Different underlying changes: 45720 modifies forward-return behavior with `can_return_tuple`/`return_dict`, while 46259 rewrites `_init_weights` initialization logic for a different linter rule. Shared model filenames do not indicate the same bug or mergeable PR."
           }
         ]
       },
@@ -22248,7 +22470,7 @@
       "soft_pairs": [
         "pull_request:45932|pull_request:46100"
       ],
-      "prepared_review_unit_hash": "796bdb8c5dce04e976e7434e9f9433beb890fca0de7a27c0287bfac0b06cfc74",
+      "prepared_review_unit_hash": "99b0a9d2b637e281d581e833ce2fe9a3547155c49c2cbf3ee8a5ccdb70f7aba8",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -22270,18 +22492,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These two pull requests are not duplicates: one is a broad Grafana/CI instrumentation and model-code tweak PR, while the other is a ViT test-migration PR focused on test file updates. They do not appear to fix the same code-path problem or belong in a single merged PR.",
-        "confidence": 0.96,
+        "summary": "These are two unrelated open pull requests that only look similar at a high level: one is a code/CI change across model files and workflow secrets, while the other is a ViT test migration. They do not appear to fix the same concrete problem or belong in one merged PR.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 45932 is the better canonical PR candidate because it is the more substantial, code-affecting change set and represents a distinct concrete change, unlike the unrelated test-migration PR.",
+        "canonical_pr_reason": "PR #45932 is the better canonical PR because it is the broader, more substantive code change and has clear review activity; #46100 is a narrower test-only migration and does not represent the same change stream.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 45932 is the best representative of the cluster only because it is the more central/substantive PR; PR 46100 is unrelated and not a duplicate of it.",
+        "best_pr_reason": "PR #45932 is the best representative of this cluster because it touches the actual implementation/workflow surface and is more central than the test migration PR.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:45932",
             "right": "pull_request:46100",
             "accept": false,
-            "reason": "Different changes: 45932 touches CI/workflow plus multiple model/config files, while 46100 is a ViT test migration. They do not target the same underlying bug or code path and would not plausibly be merged into one PR."
+            "reason": "Reject: the PRs address different things. #45932 changes model code and CI workflow secrets; #46100 adds/migrates ViT tests. Shared broad area (transformers/ViT) is not enough to treat them as the same underlying fix."
           }
         ]
       },
@@ -22314,12 +22536,12 @@
         "pull_request:46073|pull_request:46196",
         "pull_request:46154|pull_request:46196"
       ],
-      "prepared_review_unit_hash": "3c45b3fcb5f035ac6b7a1a55d09da7101380951b1161d221a9301fc4a09688b1",
+      "prepared_review_unit_hash": "915790d9ac68beec66a591ea0c343c6e6a1e3ce6bbfe4917078156c1945e3d62",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 8400,
+        "serialized_chars": 8398,
         "estimated_input_tokens": 2100,
         "estimated_eval_tokens": 4456
       },
@@ -22327,7 +22549,7 @@
         "node_count": 3,
         "item_count": 3,
         "soft_pair_count": 3,
-        "serialized_chars": 8400,
+        "serialized_chars": 8398,
         "estimated_input_tokens": 2100,
         "estimated_eval_tokens": 4456
       },
@@ -22336,30 +22558,30 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Three PRs all add attn_implementation/SDPA dispatch support, but each targets a different model family (CTRL, T5, wav2vec2/audio stack). They are thematically similar, not duplicates of the same concrete change.",
-        "confidence": 0.86,
+        "summary": "These are three separate model-specific PRs enabling SDPA/attn_implementation dispatch in different architectures. They share the same feature theme, but they are not the same concrete change and should not be deduplicated.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #46196 is the strongest representative of the cluster: it covers the broadest set of affected model implementations, adds test coverage, and appears to be the most complete SDPA-dispatch adaptation among the three.",
+        "canonical_pr_reason": "Use PR #46196 as the representative because it is the broadest multi-model SDPA dispatch effort, but it is not a duplicate of the CTRL or T5 PRs.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #46196 is the best single exemplar for this cluster because it is the broadest and most fully developed change, while the other two are narrower, model-specific follow-ups.",
+        "best_pr_reason": "PR #46196 is the best cluster representative: it covers the largest set of affected models and includes the same SDPA-dispatch feature family as the others.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46073",
             "right": "pull_request:46154",
             "accept": false,
-            "reason": "Both add SDPA/attn_implementation support, but for different model code paths (CTRL vs T5). Same feature theme, not the same underlying bug/change."
+            "reason": "Both add attention-dispatch support, but they target different model code paths (CTRL vs T5). They are related feature work, not the same fix and would not be merged into one PR."
           },
           {
             "left": "pull_request:46073",
             "right": "pull_request:46196",
             "accept": false,
-            "reason": "CTRL dispatch support and wav2vec2/audio dispatch support are separate model-specific implementations; they are not mergeable into one concrete fix."
+            "reason": "CTRL and wav2vec2-family SDPA dispatch are separate model-specific implementations with different files and execution paths. Same general capability, not the same underlying change."
           },
           {
             "left": "pull_request:46154",
             "right": "pull_request:46196",
             "accept": false,
-            "reason": "T5 and wav2vec2/audio changes address different architectures and code paths, so this is shared feature work rather than a duplicate PR."
+            "reason": "T5 and wav2vec2-family dispatch support affect different architectures and code paths. They share the SDPA theme but are not duplicate PRs."
           }
         ]
       },
@@ -22389,12 +22611,12 @@
       "soft_pairs": [
         "pull_request:46126|pull_request:46269"
       ],
-      "prepared_review_unit_hash": "6df08e27b863869045f7330811f79f3efcdf94c90623c9782b23278474ad3e1a",
+      "prepared_review_unit_hash": "ba72bee9426b79b545b3014aa8501a4715f73c0057c470c720e26de0757c6ee4",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5570,
+        "serialized_chars": 5572,
         "estimated_input_tokens": 1393,
         "estimated_eval_tokens": 3042
       },
@@ -22402,7 +22624,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5570,
+        "serialized_chars": 5572,
         "estimated_input_tokens": 1393,
         "estimated_eval_tokens": 3042
       },
@@ -22411,18 +22633,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs touch distributed/EP-related code, but they are different changes: #46126 is a focused EP + Trainer integration feature, while #46269 is a much broader distributed branch base spanning many subsystems. They share files, but not the same concrete code-path problem or mergeable change set.",
-        "confidence": 0.91,
+        "summary": "Both PRs live in the distributed-training area, but they are not the same change: #46126 is a focused EP/Trainer integration feature, while #46269 is a broad distributed branch base touching many subsystems. The overlap in files is not enough to treat them as duplicates.",
+        "confidence": 0.9,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "#46269 is the broader, more representative distributed branch base PR; it spans the wider set of distributed integrations and is the better cluster anchor than the narrower EP-specific feature PR.",
+        "canonical_pr_reason": "#46126 is the more focused, concrete change with a specific code-path goal (EP routing/trainer integration). #46269 is a broad base branch spanning many distributed components and is too diffuse to serve as the canonical duplicate target.",
         "best_issue_reason": null,
-        "best_pr_reason": "#46269 is the best representative PR for the cluster because it subsumes a wider distributed foundation, whereas #46126 is a narrower follow-on feature slice.",
+        "best_pr_reason": "#46126 is the best representative of a single, coherent PR because it targets one feature area and has a clear implementation surface. #46269 is too large and cross-cutting to be a good canonical example.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46126",
             "right": "pull_request:46269",
             "accept": false,
-            "reason": "Shared distributed filenames are not enough: #46126 centers on EP routing and Trainer integration, while #46269 is a broad distributed base branch touching many unrelated areas. They do not look like the same underlying change and would not plausibly merge into one PR."
+            "reason": "Shared distributed filenames do not indicate the same underlying bug/change. One PR adds EP/Trainer integration, the other is a much broader distributed branch base; they would not plausibly be merged as one PR."
           }
         ]
       },
@@ -22484,190 +22706,190 @@
         "pull_request:46504|pull_request:46563",
         "pull_request:46432|pull_request:46563"
       ],
-      "prepared_review_unit_hash": "8fec697a9bf02cf539f12d24cd138f0d2145cfc2a98267a76a16e3bf5c9c6802",
+      "prepared_review_unit_hash": "3b2d1962ea007ba4571cf50fc9dac02842627cd26d45c0ac44f982aa42a9df89",
       "component_budget": {
         "node_count": 9,
         "item_count": 9,
         "soft_pair_count": 26,
-        "serialized_chars": 23358,
-        "estimated_input_tokens": 5840,
-        "estimated_eval_tokens": 11936
+        "serialized_chars": 23365,
+        "estimated_input_tokens": 5842,
+        "estimated_eval_tokens": 11940
       },
       "budget": {
         "node_count": 9,
         "item_count": 9,
         "soft_pair_count": 26,
-        "serialized_chars": 23358,
-        "estimated_input_tokens": 5840,
-        "estimated_eval_tokens": 11936
+        "serialized_chars": 23365,
+        "estimated_input_tokens": 5842,
+        "estimated_eval_tokens": 11940
       },
       "overflow_policy": "truncate_then_skip",
       "trimmed": false,
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Mostly unrelated workflow-maintenance PRs. The only apparent subcluster is the pr-ci-security-gate.yml trio, but those PRs change different behaviors/comments rather than the same underlying fix, so there are no true duplicates here.",
-        "confidence": 0.93,
+        "summary": "This cluster is mostly unrelated workflow-only PRs. Even the three pr-ci-security-gate changes touch different concerns (security checks, permissions/approval flow, and comments), so none should be merged as duplicates.",
+        "confidence": 0.91,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "No single canonical PR for deduplication: the set spans different workflow files and unrelated changes. If forced to pick the most substantive standalone change, 46554 is the clearest security-gate behavior update, but it should not absorb the others as duplicates.",
+        "canonical_pr_reason": "No single canonical PR stands out: the cluster spans several independent workflow edits. If forced to pick a representative, 46554 is the most substantive pr-ci-security-gate change, but it is not a duplicate of the others.",
         "best_issue_reason": null,
-        "best_pr_reason": "46554 is the strongest standalone PR in the set because it makes a concrete security-check behavior change, but it is not a duplicate of the other items.",
+        "best_pr_reason": "46554 is the best representative only because it has the clearest, self-contained security-gate logic change; however, the surrounding PRs are separate workflow tweaks rather than one underlying fix.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46138",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflow files and different changes; self-comment CI command piping is unrelated to pr-ci-security-gate logic."
+            "reason": "Different workflow files and different changes: self-comment CI message plumbing vs pr-ci-security-gate logic. Not the same underlying bug/change."
           },
           {
             "left": "pull_request:46138",
             "right": "pull_request:46558",
             "accept": false,
-            "reason": "Different workflow and unrelated behavior; not the same code-path problem."
+            "reason": "Unrelated workflow updates in different files; one changes CI comment input handling, the other adjusts PR CI security-gate permissions/approval flow."
           },
           {
             "left": "pull_request:46138",
             "right": "pull_request:46563",
             "accept": false,
-            "reason": "Different workflow and unrelated behavior; one updates self-comment CI, the other adjusts pr-ci-security-gate reuse/comments."
+            "reason": "Different file and purpose: self-comment CI invocation fix vs a comment-only reusable-workflow note in pr-ci-security-gate."
           },
           {
             "left": "pull_request:46138",
             "right": "pull_request:46504",
             "accept": false,
-            "reason": "Different workflow files and separate CI tweaks; not a single underlying change."
+            "reason": "Build CI docker image trigger handling is unrelated to self-comment CI stdin handling; not the same code-path problem."
           },
           {
             "left": "pull_request:46138",
             "right": "pull_request:46432",
             "accept": false,
-            "reason": "Different workflow files and separate fixes; no shared concrete bug/change."
+            "reason": "Different workflows and different concerns: self-comment CI messaging vs ssh-runner hardening/secret handling."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46503",
             "accept": false,
-            "reason": "SSH runner hardening is unrelated to benchmark_v2 trigger logic."
+            "reason": "Both are workflow edits, but one is SSH runner hardening and the other adds workflow_dispatch to benchmark_v2; not the same fix."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46501",
             "accept": false,
-            "reason": "Different workflow files and different fixes; not the same underlying change."
+            "reason": "Different workflow files and different changes: SSH runner secret handling vs anti-slop permission placement."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46502",
             "accept": false,
-            "reason": "SSH runner changes do not match reviewer-assignment workflow behavior."
+            "reason": "SSH runner hardening and assign-reviewers checkout ref selection solve unrelated problems."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflows and unrelated concerns; not a shared code-path issue."
+            "reason": "Security/hardening theme is too broad; ssh-runner changes are about remote-run safety, while 46554 changes PR CI security-gate behavior."
           },
           {
             "left": "pull_request:46432",
             "right": "pull_request:46558",
             "accept": false,
-            "reason": "Different workflows and unrelated concerns; no duplicate bug/change."
-          },
-          {
-            "left": "pull_request:46432",
-            "right": "pull_request:46563",
-            "accept": false,
-            "reason": "Different workflows and unrelated concerns; not mergeable as one fix."
+            "reason": "Different workflows and different fixes: SSH runner environment sanitization vs PR CI security-gate permissions and gating."
           },
           {
             "left": "pull_request:46501",
             "right": "pull_request:46503",
             "accept": false,
-            "reason": "Different workflow files and different CI settings changes; not the same underlying problem."
+            "reason": "Anti-slop permission scoping and benchmark_v2 dispatch support are unrelated workflow changes."
           },
           {
             "left": "pull_request:46501",
             "right": "pull_request:46502",
             "accept": false,
-            "reason": "Anti-slop permissions change is unrelated to assign-reviewers checkout ref change."
+            "reason": "Different workflows and different purposes: anti-slop permissions vs assign-reviewers checkout base ref."
           },
           {
             "left": "pull_request:46501",
             "right": "pull_request:46504",
             "accept": false,
-            "reason": "Different files and different workflow tweaks; no single bug/change."
+            "reason": "Anti-slop workflow permission change is unrelated to build-ci-docker-images trigger parsing."
           },
           {
             "left": "pull_request:46501",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflow files and separate concerns; not a duplicate pair."
+            "reason": "Same broad CI area, but not the same bug/change: anti-slop permissions vs PR CI security-gate logic."
           },
           {
             "left": "pull_request:46502",
             "right": "pull_request:46503",
             "accept": false,
-            "reason": "Different workflow files and unrelated CI behavior changes."
+            "reason": "Assign-reviewers base-ref checkout and benchmark_v2 workflow dispatch are unrelated."
           },
           {
             "left": "pull_request:46502",
             "right": "pull_request:46504",
             "accept": false,
-            "reason": "Reviewer assignment checkout ref fix is unrelated to build-ci-docker-images commit-message handling."
+            "reason": "Different workflows and separate fixes: reviewer assignment checkout behavior vs build-ci-docker-images trigger detection."
           },
           {
             "left": "pull_request:46502",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflows and unrelated changes; no shared concrete fix."
+            "reason": "No shared concrete issue: reviewer assignment checkout ref vs PR CI security-gate checks."
           },
           {
             "left": "pull_request:46503",
             "right": "pull_request:46504",
             "accept": false,
-            "reason": "Benchmark trigger addition is unrelated to build-ci-docker-images shell quoting fix."
+            "reason": "Benchmark_v2 dispatch support and build-ci-docker-images commit-message handling are separate workflow changes."
           },
           {
             "left": "pull_request:46503",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflows and different problems; not the same change."
+            "reason": "Benchmark execution triggers are unrelated to PR CI security-gate logic."
           },
           {
             "left": "pull_request:46504",
             "right": "pull_request:46554",
             "accept": false,
-            "reason": "Different workflows and unrelated CI behavior; not a duplicate."
+            "reason": "Different workflows and different code paths: build-ci-docker-images commit-message parsing vs PR CI security-gate checks."
           },
           {
             "left": "pull_request:46504",
             "right": "pull_request:46558",
             "accept": false,
-            "reason": "Different workflows and unrelated changes; no common underlying bug."
-          },
-          {
-            "left": "pull_request:46504",
-            "right": "pull_request:46563",
-            "accept": false,
-            "reason": "Different workflows and unrelated changes; not the same fix."
+            "reason": "Build-image tag trigger parsing is unrelated to PR CI security-gate permissions/approval changes."
           },
           {
             "left": "pull_request:46554",
             "right": "pull_request:46558",
             "accept": false,
-            "reason": "Same file, but different code paths: one changes security-check behavior, the other changes permissions/approval gating."
+            "reason": "Same file, but different fixes: 46554 adds checkout/diff-based security checks, while 46558 changes permissions and approval gating. Not the same underlying bug/change."
           },
           {
             "left": "pull_request:46554",
             "right": "pull_request:46563",
             "accept": false,
-            "reason": "Same workflow file, but one adds security checks while the other only adds explanatory comments around a reusable workflow call."
+            "reason": "46554 changes the security-check logic; 46563 only adds explanatory comments. These are not duplicate fixes."
           },
           {
             "left": "pull_request:46558",
             "right": "pull_request:46563",
             "accept": false,
-            "reason": "Same file, but different concerns: permissions/if-guard versus comment-only documentation."
+            "reason": "Both touch the same workflow file, but one alters permissions/approval flow and the other adds comments only; not a duplicate code-path fix."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Build CI docker image workflow update is unrelated to pr-ci-security-gate comment-only changes."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "SSH runner hardening is unrelated to pr-ci-security-gate comments; different workflow and different problem."
           }
         ]
       },
@@ -22697,7 +22919,7 @@
       "soft_pairs": [
         "pull_request:46152|pull_request:46207"
       ],
-      "prepared_review_unit_hash": "0e9364fbc753f10ec919bca2ab349e67072c702eda5559a5a16e6c946baac23f",
+      "prepared_review_unit_hash": "d3d2394f36e8bdee3dc84fe328af0f3d75c217bdf76da2c39dd47d9929fb5e2f",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -22719,18 +22941,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both PRs try to eliminate brittle source-file reads via `open()`, but they do so in different code paths and for different concrete behaviors: one refactors package-path resolution in `import_utils`, the other changes the attention-implementation capability check in `modeling_utils` and adds tests. They are related by theme, not duplicates.",
-        "confidence": 0.91,
+        "summary": "These PRs are thematically similar because both replace brittle file-open source inspection, but they touch different modules and different code paths, so they are not duplicates.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #46207 is the better canonical representative because it is a concrete, test-backed fix in a core runtime decision path (`_can_set_attn_implementation`) and clearly states the replacement mechanism.",
+        "canonical_pr_reason": null,
         "best_issue_reason": null,
-        "best_pr_reason": "PR #46207 is the strongest cluster representative: it fixes a specific bug, includes tests, and is more self-contained than the lower-level import utility refactor in #46152.",
+        "best_pr_reason": "PR 46207 is the better representative of the cluster theme: it makes a concrete source-inspection fix in _can_set_attn_implementation and includes tests.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46152",
             "right": "pull_request:46207",
             "accept": false,
-            "reason": "They share the broad idea of avoiding `open()` for source inspection, but they target different modules, different functions, and different runtime behaviors; they do not appear mergeable into one PR as the same underlying change."
+            "reason": "Different fixes in different modules: one changes import_utils resource lookup, the other changes modeling_utils source inspection for attention implementation. Same broad idea, not the same underlying bug or mergeable change."
           }
         ]
       },
@@ -22762,7 +22984,7 @@
         "pull_request:46436|pull_request:46474",
         "pull_request:46173|pull_request:46474"
       ],
-      "prepared_review_unit_hash": "71f73aab0b2cfd4b496379733532e170c7c27068fb3e8446b354d316f45d598c",
+      "prepared_review_unit_hash": "c963914ae25d9de77ab1282269240ca8a06d8ae74812b6346043918c4dceffe2",
       "component_budget": {
         "node_count": 3,
         "item_count": 3,
@@ -22784,30 +23006,30 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Two PRs are clear duplicates for the same StaticLayer type-hint fix (#45987); the third PR is a separate static-cache configuration fix in a different code path and should not be clustered with them.",
+        "summary": "PR 46173 and PR 46474 are duplicates: both make the same StaticLayer.get_seq_length type-hint change in cache_utils.py for issue #45987. PR 46436 is unrelated, fixing static-cache max length handling in generation/utils.py for a different issue.",
         "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR #46474 is the cleanest representative of the duplicate pair: it targets the exact StaticLayer.get_seq_length type-hint change, references the same issue, and has a straightforward descriptive title.",
+        "canonical_pr_reason": "PR 46173 is the clean canonical choice: it is the earliest exact-match fix for StaticLayer.get_seq_length on the same file and issue target (#45987).",
         "best_issue_reason": null,
-        "best_pr_reason": "PR #46474 best represents the cluster because it is the same concrete change as #46173 with a clearer title; #46436 is unrelated beyond the broad static-cache area.",
+        "best_pr_reason": "PR 46173 best represents the duplicate cluster because it matches PR 46474 almost exactly and is directly scoped to the type-hint fix, unlike PR 46436 which addresses a different cache-length bug.",
         "soft_edge_verdicts": [
           {
-            "left": "pull_request:46173",
+            "left": "pull_request:46436",
             "right": "pull_request:46474",
-            "accept": true,
-            "reason": "Same file, same issue target (#45987), and identical type-hint diff for StaticLayer.get_seq_length; these are the same underlying change."
+            "accept": false,
+            "reason": "Different bugs and code paths: 46436 changes static cache length preparation in generation/utils.py, while 46474 only adjusts StaticLayer.get_seq_length annotations in cache_utils.py."
           },
           {
-            "left": "pull_request:46436",
+            "left": "pull_request:46173",
             "right": "pull_request:46474",
-            "accept": false,
-            "reason": "Different bug and code path: #46436 changes generation cache sizing for max_cache_len, while #46474 only adjusts a StaticLayer return type annotation. Same subsystem mention is not enough to merge."
+            "accept": true,
+            "reason": "Same underlying change: both PRs make the same StaticLayer.get_seq_length return-type update in cache_utils.py and share the same issue target (#45987)."
           }
         ]
       },
       "evaluator_result": {
         "accept": true,
-        "feedback": "The analysis is grounded in the packet. PRs #46173 and #46474 have identical diffs, the same file, and the same explicit issue target (#45987), so treating them as duplicates is well supported. PR #46436 changes a different file and addresses a different issue target (#46424), so excluding it from the cluster is conservative and correct. No overstatement detected."
+        "feedback": "The summary is grounded in the packet: 46173 and 46474 have the same file, same explicit issue target, and identical diff for the StaticLayer.get_seq_length type-hint change. The 46436 vs 46474 verdict is also conservative, since they touch different files and target different issues. No evidence is overstated."
       },
       "evaluator_used": true,
       "retried": false,
@@ -22834,12 +23056,12 @@
       "soft_pairs": [
         "pull_request:46236|pull_request:46621"
       ],
-      "prepared_review_unit_hash": "6bf020fb625a54e182b3979810dd0219c0338cd289d18f9c5c33211c7050b973",
+      "prepared_review_unit_hash": "1275997b08282b88072e1cf0eaf5289d1798efd657d385bf53d4923c4131137f",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5921,
+        "serialized_chars": 5922,
         "estimated_input_tokens": 1481,
         "estimated_eval_tokens": 3218
       },
@@ -22847,7 +23069,7 @@
         "node_count": 2,
         "item_count": 2,
         "soft_pair_count": 1,
-        "serialized_chars": 5921,
+        "serialized_chars": 5922,
         "estimated_input_tokens": 1481,
         "estimated_eval_tokens": 3218
       },
@@ -22856,18 +23078,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Two separate documentation translation PRs: Bengali i18n docs and Chinese kernels docs. They are not the same underlying change.",
-        "confidence": 0.98,
+        "summary": "These are two unrelated documentation translation PRs: one adds Bengali docs for a large set of pages, the other adds a Chinese translation for a single kernels page. They do not describe the same underlying change and should not be deduplicated.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 46236 is the broader, older i18n docs translation and has an explicit issue link; PR 46621 is a separate Chinese translation for a different page.",
+        "canonical_pr_reason": "PR 46236 is the better canonical PR for this cluster because it is the larger, issue-linked localization change and more representative of the translation work here.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46236 is the most representative PR in the cluster because it is the larger translation sweep, but it is not a duplicate of PR 46621.",
+        "best_pr_reason": "PR 46236 is the strongest standalone PR in the set due to its explicit issue target, broader scope, and clearer project context; PR 46621 is a separate single-page translation.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46236",
             "right": "pull_request:46621",
             "accept": false,
-            "reason": "Different languages, different files, and different doc sections; they are separate translation PRs, not one underlying fix."
+            "reason": "Different languages and different documentation content: Bengali translation batch vs. Chinese kernels page. No shared files, targets, or concrete code/doc change to merge into one PR."
           }
         ]
       },
@@ -22897,7 +23119,7 @@
       "soft_pairs": [
         "pull_request:46310|pull_request:46546"
       ],
-      "prepared_review_unit_hash": "efa44aa5f6f70f570eff89e03ce14f5d80c67436b1648dd93bdd29495d1c388b",
+      "prepared_review_unit_hash": "d1d49e2ea32c41cdfc2b5524b8ad388dc948a5c313086b79419f8cbe98ff694f",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -22919,18 +23141,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "Both pull requests fix the same off-by-one bug in the streaming attention right-window boundary for Moonshine, with the second PR broadening the same fix pattern and adding tests.",
-        "confidence": 0.95,
+        "summary": "Both PRs fix the same off-by-one bug in the Moonshine streaming sliding-window mask: the right-side boundary uses `<` instead of `<=`. PR 46546 is the broader superset, adding the same Moonshine fix plus Gemma4 and tests; PR 46310 is the narrower Moonshine-only patch.",
+        "confidence": 0.97,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 46546 is the stronger canonical candidate: it fixes the same Moonshine streaming boundary bug, includes the shared Moonshine files, and also adds test coverage (plus the same fix in Gemma4).",
+        "canonical_pr_reason": "PR 46546 is the better canonical representative because it contains the Moonshine boundary fix, related Gemma4 adjustment, and a test update, making it the more complete change set.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46546 is the best overall representative because it contains the Moonshine fix plus validation, making it more complete than the smaller Moonshine-only patch in 46310.",
+        "best_pr_reason": "PR 46546 is the strongest overall fit: it addresses the same concrete off-by-one boundary bug as 46310, but with broader coverage and validation; 46310 is a subset of the same fix.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46310",
             "right": "pull_request:46546",
             "accept": true,
-            "reason": "Both PRs correct the same right-window boundary off-by-one in Moonshine streaming attention; 46546 is a superset with the same code-path fix and added tests, so they are duplicate enough to cluster."
+            "reason": "Same underlying off-by-one bug in the Moonshine streaming right-window mask; the code change is identical on the shared files, and 46546 includes the broader follow-up/test coverage."
           }
         ]
       },
@@ -22960,7 +23182,7 @@
       "soft_pairs": [
         "pull_request:46376|pull_request:46417"
       ],
-      "prepared_review_unit_hash": "40a5bf9e8d3fca2b34f87f741e5b136b65005c96b2889a76bc2d6204bb2a29ed",
+      "prepared_review_unit_hash": "b3ef4fb8a2aa6a0b77df8c82e0818c539736a1246e9a4b7cb174e22d9f36bf16",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -22982,18 +23204,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "The two pull requests are unrelated: one adds multimodal continuous batching support, while the other adds a new HunYuan VL model and its docs/tests. They do not appear to fix the same bug or implement the same change.",
-        "confidence": 0.97,
+        "summary": "These are not duplicates: one PR adds multimodal support to continuous batching infrastructure, while the other adds the HunYuan VL model, its docs, and tests. The soft similarity is superficial and not enough to merge them.",
+        "confidence": 0.98,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 46417 is the better representative PR in this cluster: it is a complete, non-draft model addition with substantial review activity, whereas PR 46376 is still a draft infrastructure change and addresses a different subsystem.",
+        "canonical_pr_reason": "PR 46417 is the better representative of the cluster because it is a more complete, self-contained feature addition with model code, docs, and tests; PR 46376 is a draft infrastructure change and is unrelated in scope.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46417 is the strongest standalone PR here because it is more self-contained and mature, with extensive review activity and no overlap in purpose with the other PR.",
+        "best_pr_reason": "PR 46417 is the strongest standalone PR: it has a full implementation surface and clearer deliverables, whereas PR 46376 is draft-stage and addresses a different subsystem.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46376",
             "right": "pull_request:46417",
             "accept": false,
-            "reason": "Reject: these PRs target different code paths and features. 46376 is about multimodal continuous batching internals; 46417 adds the HunYuan VL model, docs, processors, and tests. They are not mergeable as one change."
+            "reason": "Different underlying changes: continuous batching multimodal infrastructure vs. HunYuan VL model integration. No shared files or concrete code-path overlap, so this is not a duplicate."
           }
         ]
       },
@@ -23023,7 +23245,7 @@
       "soft_pairs": [
         "pull_request:46603|pull_request:46604"
       ],
-      "prepared_review_unit_hash": "0a0cf228ed91f846e0b75c748aa6bd60528489924b4ff71c276d75e98ca09a1f",
+      "prepared_review_unit_hash": "9d09d98b83e178507f3ee0af98294f159f6a76bdc1d280850edea00f4bb5eaaa",
       "component_budget": {
         "node_count": 2,
         "item_count": 2,
@@ -23045,18 +23267,18 @@
       "aggressively_trimmed": false,
       "split": false,
       "analyst_result": {
-        "summary": "These are related Mistral tokenizer work items but not the same change: one is a refactor that moves MistralConverter into the integrations package, while the other adds tekken tokenizer detection/conversion/save support. Shared files reflect the same subsystem, not the same underlying fix.",
-        "confidence": 0.93,
+        "summary": "Two open Mistral PRs overlap only at the package/file level; they target different changes. PR 46603 is a refactor that moves `MistralConverter` into `integrations/mistral/`, while PR 46604 adds tekken tokenizer detection/conversion/save utilities. They should not be merged as duplicates.",
+        "confidence": 0.94,
         "canonical_issue_reason": null,
-        "canonical_pr_reason": "PR 46604 is the more substantial, user-visible change and best represents the cluster\u2019s main Mistral tokenizer work; PR 46603 is a narrower packaging/refactor move.",
+        "canonical_pr_reason": "PR 46604 is the better canonical PR because it delivers the concrete tokenizer feature work; PR 46603 is a narrower code move/refactor that is not the same change.",
         "best_issue_reason": null,
-        "best_pr_reason": "PR 46604 is the best standalone representative because it introduces the concrete tokenizer functionality, whereas PR 46603 mainly reorganizes existing code.",
+        "best_pr_reason": "PR 46604 best represents the cluster since it contains the substantive Mistral tokenizer behavior change, whereas PR 46603 is mainly structural packaging cleanup.",
         "soft_edge_verdicts": [
           {
             "left": "pull_request:46603",
             "right": "pull_request:46604",
             "accept": false,
-            "reason": "Different underlying changes: 46603 moves MistralConverter into a new package location, while 46604 implements tekken tokenizer detection, conversion, and save utilities. They touch the same subsystem but are not the same fix and would not naturally collapse into one PR."
+            "reason": "Different underlying changes: one relocates `MistralConverter`, the other implements tekken tokenizer detection/conversion/save utilities. Shared Mistral files and subsystem are not enough to treat them as the same PR."
           }
         ]
       },
diff --git a/analysis/current/manifest.json b/analysis/current/manifest.json
index 9b9fb07ba30307d3e0d28ba1ba3967e8a4c54917..8dce81418ebd1afb114133bcf6cbf2fe1ad1b9cd 100644
--- a/analysis/current/manifest.json
+++ b/analysis/current/manifest.json
@@ -1,8 +1,8 @@
 {
-  "analysis_id": "hybrid-model-20260613t120022z",
+  "analysis_id": "hybrid-model-20260613t180123z",
   "archived_artifacts": {
-    "hybrid": "snapshots/20260613T120022Z/analysis-runs/hybrid-model-20260613t120022z/analysis-report-hybrid.json",
-    "hybrid_reviews": "snapshots/20260613T120022Z/analysis-runs/hybrid-model-20260613t120022z/analysis-report-hybrid.llm-reviews.json"
+    "hybrid": "snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.json",
+    "hybrid_reviews": "snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.llm-reviews.json"
   },
   "artifacts": {
     "hybrid": "analysis/current/analysis-report-hybrid.json",
@@ -10,9 +10,9 @@
   },
   "channel": "canonical",
   "model": null,
-  "published_at": "2026-06-13T12:08:50Z",
+  "published_at": "2026-06-13T18:06:05Z",
   "repo": "huggingface/transformers",
   "schema_version": 1,
-  "snapshot_id": "20260613T120022Z",
+  "snapshot_id": "20260613T180123Z",
   "variant": "hybrid"
 }
diff --git a/snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.json b/snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.json
new file mode 100644
index 0000000000000000000000000000000000000000..d2fb682d4bb171ec8dc77d70076b8c649e305fe8
--- /dev/null
+++ b/snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.json
@@ -0,0 +1,2150 @@
+{
+  "schema_version": "1.0",
+  "repo": "huggingface/transformers",
+  "snapshot_id": "20260613T180123Z",
+  "generated_at": "2026-06-13T18:06:03Z",
+  "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-45081-4",
+      "summary": "Cluster of 2 issues and 2 PRs centered on issue #45081.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 45081,
+      "canonical_pr_number": 45317,
+      "issue_numbers": [
+        45081,
+        45356
+      ],
+      "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-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": 0.797,
+          "size_similarity": 0.885,
+          "file_overlap": 0.75,
+          "area_overlap": 0.831,
+          "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": [
+                [
+                  730,
+                  737
+                ]
+              ]
+            },
+            {
+              "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": [
+                [
+                  720,
+                  726
+                ],
+                [
+                  1331,
+                  1337
+                ]
+              ]
+            },
+            {
+              "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-45987-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #45987.",
+      "status": "open",
+      "confidence": 0.75,
+      "canonical_issue_number": 45987,
+      "canonical_pr_number": 46173,
+      "issue_numbers": [
+        45987
+      ],
+      "pr_numbers": [
+        46173,
+        46474
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 46173,
+          "right_pr_number": 46474,
+          "code_similarity": 1.0,
+          "size_similarity": 1.0,
+          "file_overlap": 1.0,
+          "area_overlap": 1.0,
+          "patch_similarity": 1.0,
+          "shared_filenames": [
+            "src/transformers/cache_utils.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/cache_utils.py",
+              "left_ranges": [
+                [
+                  69,
+                  75
+                ],
+                [
+                  375,
+                  381
+                ]
+              ],
+              "right_ranges": [
+                [
+                  69,
+                  75
+                ],
+                [
+                  375,
+                  381
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "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": 0.494,
+          "size_similarity": 0.619,
+          "file_overlap": 0.5,
+          "area_overlap": 0.434,
+          "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": [
+                [
+                  3595,
+                  3602
+                ],
+                [
+                  3819,
+                  3853
+                ]
+              ],
+              "right_ranges": [
+                [
+                  3525,
+                  3558
+                ],
+                [
+                  3782,
+                  3788
+                ]
+              ]
+            },
+            {
+              "filename": "tests/utils/test_testing_utils.py",
+              "left_ranges": [
+                [
+                  1,
+                  86
+                ]
+              ],
+              "right_ranges": [
+                [
+                  1,
+                  114
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "cluster_id": "cluster-46305-3",
+      "summary": "Cluster of 1 issues and 2 PRs centered on issue #46305.",
+      "status": "open",
+      "confidence": 0.8,
+      "canonical_issue_number": 46305,
+      "canonical_pr_number": 46546,
+      "issue_numbers": [
+        46305
+      ],
+      "pr_numbers": [
+        46310,
+        46546
+      ],
+      "evidence_types": [
+        "closing_reference",
+        "shared_issue_target",
+        "soft_similarity"
+      ],
+      "pr_comparisons": [
+        {
+          "left_pr_number": 46310,
+          "right_pr_number": 46546,
+          "code_similarity": 0.566,
+          "size_similarity": 0.444,
+          "file_overlap": 0.5,
+          "area_overlap": 0.867,
+          "patch_similarity": 0.16,
+          "shared_filenames": [
+            "src/transformers/models/moonshine_streaming/modeling_moonshine_streaming.py",
+            "src/transformers/models/moonshine_streaming/modular_moonshine_streaming.py"
+          ],
+          "shared_file_areas": [
+            {
+              "filename": "src/transformers/models/moonshine_streaming/modeling_moonshine_streaming.py",
+              "left_ranges": [
+                [
+                  356,
+                  362
+                ]
+              ],
+              "right_ranges": [
+                [
+                  355,
+                  361
+                ]
+              ]
+            },
+            {
+              "filename": "src/transformers/models/moonshine_streaming/modular_moonshine_streaming.py",
+              "left_ranges": [
+                [
+                  269,
+                  275
+                ]
+              ],
+              "right_ranges": [
+                [
+                  269,
+                  275
+                ]
+              ]
+            }
+          ]
+        }
+      ]
+    },
+    {
+      "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": [
+    {
+      "cluster_id": "cluster-45081-4",
+      "canonical_issue_number": 45081,
+      "duplicate_issue_numbers": [
+        45356
+      ],
+      "reason": "Issues in cluster-45081-4 are treated as duplicates because they share closing_reference, shared_issue_target, soft_similarity evidence."
+    }
+  ],
+  "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-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-45081-4",
+      "canonical_pr_number": 45317,
+      "duplicate_pr_numbers": [
+        45086
+      ],
+      "target_issue_number": 45081,
+      "reason": "PRs in cluster-45081-4 are treated as duplicates because they converge on issue #45081 with closing_reference, shared_issue_target, soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-45561-3",
+      "canonical_pr_number": 45645,
+      "duplicate_pr_numbers": [
+        45639
+      ],
+      "target_issue_number": 45561,
+      "reason": "PRs in cluster-45561-3 are treated as duplicates because they converge on issue #45561 with closing_reference, shared_issue_target, soft_similarity evidence."
+    },
+    {
+      "cluster_id": "cluster-45647-3",
+      "canonical_pr_number": 45738,
+      "duplicate_pr_numbers": [
+        45737
+      ],
+      "target_issue_number": 45647,
+      "reason": "PRs in cluster-45647-3 are treated as duplicates because they converge on issue #45647 with closing_reference, shared_issue_target evidence."
+    },
+    {
+      "cluster_id": "cluster-45987-3",
+      "canonical_pr_number": 46173,
+      "duplicate_pr_numbers": [
+        46474
+      ],
+      "target_issue_number": 45987,
+      "reason": "PRs in cluster-45987-3 are treated as duplicates because they converge on issue #45987 with closing_reference, shared_issue_target evidence."
+    },
+    {
+      "cluster_id": "cluster-46305-3",
+      "canonical_pr_number": 46546,
+      "duplicate_pr_numbers": [
+        46310
+      ],
+      "target_issue_number": 46305,
+      "reason": "PRs in cluster-46305-3 are treated as duplicates because they converge on issue #46305 with closing_reference, shared_issue_target, soft_similarity evidence."
+    }
+  ],
+  "best_issue": {
+    "cluster_id": "cluster-41211-3",
+    "issue_number": 41211,
+    "reason": "Issue #41211 is the strongest global issue candidate because it is open, belongs to a cluster with 3 artifacts, and carries 6 discussion comments plus 6 inbound references.",
+    "score": 148.0
+  },
+  "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/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.llm-reviews.json b/snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.llm-reviews.json
new file mode 100644
index 0000000000000000000000000000000000000000..696c79d648ef284791b8f200224faf0a5361864a
--- /dev/null
+++ b/snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.llm-reviews.json
@@ -0,0 +1,23293 @@
+{
+  "schema_version": "1.0",
+  "repo": "huggingface/transformers",
+  "snapshot_id": "20260613T180123Z",
+  "generated_at": "2026-06-13T18:06:03Z",
+  "model": "gpt-5.4-mini?service_tier=flex",
+  "reviews": [
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 1,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:24643",
+        "issue:30277",
+        "issue:34634",
+        "issue:34928",
+        "issue:35545",
+        "issue:36010",
+        "issue:36331",
+        "issue:39290",
+        "issue:41720",
+        "issue:41762",
+        "issue:42915",
+        "issue:43039",
+        "issue:43425",
+        "issue:43541",
+        "issue:43596",
+        "issue:43638",
+        "issue:43716",
+        "issue:43827",
+        "issue:43828",
+        "issue:43866",
+        "issue:43975",
+        "issue:44292",
+        "issue:44322",
+        "issue:44457",
+        "issue:44509",
+        "issue:44512",
+        "issue:44560",
+        "issue:44805",
+        "issue:44918",
+        "issue:45081",
+        "issue:45161",
+        "issue:45237",
+        "issue:45464",
+        "issue:45507",
+        "issue:45685",
+        "issue:45800",
+        "issue:45823",
+        "issue:45859",
+        "issue:45874",
+        "issue:45901",
+        "issue:45902",
+        "issue:46015",
+        "issue:46165",
+        "issue:46294",
+        "issue:46304",
+        "issue:46402",
+        "issue:46468",
+        "issue:46515"
+      ],
+      "soft_pairs": [
+        "issue:44322|issue:45464",
+        "issue:45237|issue:45685",
+        "issue:43716|issue:45237",
+        "issue:43828|issue:45237",
+        "issue:45237|issue:46468",
+        "issue:45237|issue:45507",
+        "issue:30277|issue:35545",
+        "issue:35545|issue:42915",
+        "issue:35545|issue:41720",
+        "issue:45901|issue:46015",
+        "issue:44509|issue:44512",
+        "issue:35545|issue:36010",
+        "issue:44918|issue:46165",
+        "issue:44805|issue:45823",
+        "issue:43638|issue:44805",
+        "issue:43638|issue:46165",
+        "issue:43596|issue:44805",
+        "issue:44805|issue:46294",
+        "issue:44805|issue:45161",
+        "issue:46165|issue:46468",
+        "issue:44560|issue:46165",
+        "issue:43596|issue:46165",
+        "issue:34634|issue:35545",
+        "issue:30277|issue:43638",
+        "issue:43596|issue:43638",
+        "issue:43039|issue:45161",
+        "issue:43975|issue:44457",
+        "issue:44292|issue:44457",
+        "issue: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:44292|issue:45237",
+        "issue:45874|issue:46294",
+        "issue:39290|issue:45081",
+        "issue:45800|issue:46515",
+        "issue:45823|issue:45874",
+        "issue:43425|issue:44292",
+        "issue:34928|issue:41720",
+        "issue:43828|issue:45161",
+        "issue:43828|issue:44292",
+        "issue:43541|issue:43716",
+        "issue:43716|issue:46468",
+        "issue:43716|issue:45685",
+        "issue:44292|issue:44918",
+        "issue:44918|issue:45161",
+        "issue:43716|issue:44292",
+        "issue:45507|issue:45685",
+        "issue:43039|issue:44918",
+        "issue:41762|issue:43638",
+        "issue:43827|issue:44512",
+        "issue:43828|issue:45685",
+        "issue:44918|issue:45685",
+        "issue:43039|issue:43828",
+        "issue:30277|issue:36331",
+        "issue:34928|issue:36010",
+        "issue:43039|issue:44805",
+        "issue:45237|issue:46015",
+        "issue:43866|issue:46402",
+        "issue:45800|issue:46304"
+      ],
+      "prepared_review_unit_hash": "63b311923e869611acb498fecd32fbb40d5d294516e34f1ec039f40d1fc0af23",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 63,
+        "serialized_chars": 46525,
+        "estimated_input_tokens": 11632,
+        "estimated_eval_tokens": 23520
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous. The only clear duplicate subcluster is the DeepSpeed ZeRO-3 / BERT index-out-of-bounds reports, and there is one likely duplicate pair for the torch 2.4.1 compatibility regression. The rest are adjacent-sounding but distinct model-, hardware-, or docs-specific issues.",
+        "confidence": 0.72,
+        "canonical_issue_reason": "issue:43596 is the cleanest representative of the strongest true duplicate pair: it describes the ZeRO-3/BERT \"index 0 is out of bounds\" failure in a specific, reproducible way, and 43638 is the same bug with a slightly narrower non-pretrained-BERT framing.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43596 is the best global issue anchor because it captures a concrete, shareable failure mode that clearly duplicates another report; the broader set here does not form one coherent bug cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44322",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Both involve Qwen3.5, but one is a missing config attribute and the other is a streaming chat/completions failure; different code paths."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Different underlying problems: AMD GPU runtime failure vs MPS histogram op on Int; not the same bug."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Different models and failure modes: Mistral-3 dtype mismatch vs GPT-OSS AMD GPU breakage."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45237",
+            "accept": false,
+            "reason": "Torch autocast dtype mismatch in Phi-tiny-MoE is unrelated to GPT-OSS on AMD GPUs."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "One is AMD GPU support, the other is a broken Nemotron example; no shared concrete bug."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Different model families and symptoms; not plausibly one fix."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:35545",
+            "accept": false,
+            "reason": "DeepSpeed collective mismatch vs ONNX export error are unrelated."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "ModernBERT ONNX export and Qwen3Moe fine-grained FP8 failures are different issues."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "ModernBERT export failure and Qwen3 auto device mapping cuda assert do not share the same code path."
+          },
+          {
+            "left": "issue:45901",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "Table QA runtime crash vs outdated RAG docs; one is code, the other is documentation."
+          },
+          {
+            "left": "issue:44509",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are v5 docs cleanups, but they target different removed documentation topics and are not the same change."
+          },
+          {
+            "left": "issue:35545",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "ONNX export error and GenerationMixin import error are unrelated regressions."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:46165",
+            "accept": false,
+            "reason": "Different training issues: TRL SFT embedding unpacking vs resume/trust-remote-code behavior."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Both are device/masking-style failures, but one is a mask shape mismatch and the other is a Gemma4 PLE device placement bug."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Both mention index/shape errors under distributed training, but the models and triggers differ; not clearly one bug."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:46165",
+            "accept": false,
+            "reason": "BERT ZeRO-3 index error vs resume/trust-remote-code issue are unrelated."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Both are tensor/indexing failures, but 44805 is a different mask-shape mismatch in a different path."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "Mask shape mismatch and Qwen3.5 video token/feature mismatch are different multimodal bugs."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Generic shape/index error vs GPT-OSS MoE tensor-parallel issue; not the same underlying defect."
+          },
+          {
+            "left": "issue:46165",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "Training resume bug and broken example are unrelated."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:46165",
+            "accept": false,
+            "reason": "Video chat template StopIteration bug vs model training resume issue; different subsystems."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:46165",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO-3 BERT init failure is unrelated to model-resume/trust-remote-code behavior."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:35545",
+            "accept": false,
+            "reason": "BarkProcessor voice_preset bug has nothing to do with ModernBERT ONNX export."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Both involve DeepSpeed, but one is collective mismatch and the other is BERT zero-sized parameter indexing; different concrete failures."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:43638",
+            "accept": true,
+            "reason": "Same ZeRO-3/BERT index-0-out-of-bounds bug during initialization; 43638 is just a narrower variant of the same failure."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Liger Kernel cross_entropy being called is unrelated to GPT-OSS tensor-parallel failures."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:44457",
+            "accept": false,
+            "reason": "Incorrect detokenization and LoRA merge/save output mismatch are different bugs."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44457",
+            "accept": false,
+            "reason": "Qwen-3-8B NVFP4 runtime failure is unrelated to LoRA save/reload inconsistency."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Activation checkpoint/FSDP recomputation mismatch is a different distributed-training bug than the mask/index error in 44805."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:30277",
+            "accept": false,
+            "reason": "Both are DeepSpeed-related, but one is a weight dimensionality error and the other is a collective mismatch."
+          },
+          {
+            "left": "issue:45859",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Same model class, but the triggers differ: CPU offload incompatibility vs uninitialized inv_freq after meta->to_empty(cuda); not clearly the same bug."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Gemma3 config attribute regression and Mistral regex patch crash are unrelated."
+          },
+          {
+            "left": "issue:45800",
+            "right": "issue:46515",
+            "accept": true,
+            "reason": "Same torch 2.4.1 incompatibility report; the second issue is a shorter duplicate of the same compatibility problem."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:45874",
+            "accept": false,
+            "reason": "Device-map forward mismatch and expensive reverse embedding lookup are different Gemma4 problems."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "General Torch 2.10 incompatibility vs a specific Qwen-3 NVFP4 runtime error; not the same issue."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are docs issues tied to v5 removals, but they concern different removed features and docs paths."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Autocast dtype mismatch in Phi-tiny-MoE is unrelated to GPT-OSS tensor-parallel failure."
+          },
+          {
+            "left": "issue:30277",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "DeepSpeed collective mismatch vs CustomTrainer.compute_loss signature break are unrelated."
+          },
+          {
+            "left": "issue:34928",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "FSDP checkpointing mismatch and GenerationMixin import error do not share a bug root."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Liger Kernel cross-entropy issue and mask-shape mismatch are different failures."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "AMD GPU runtime failure vs outdated RAG examples; different artifact types and causes."
+          },
+          {
+            "left": "issue:43866",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Corrupted Ovis2 checkpoint and Sam3 text encoder weights not loading are distinct model-loading problems."
+          },
+          {
+            "left": "issue:45800",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "Both are torch/version compatibility reports, but one is tied to USE_DISTRIBUTED=0 and the other to a different v5.9.0 breakage."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:43716",
+            "accept": false,
+            "reason": "Mixtral grouped_mm tracing failure and Mistral-3 image dtype mismatch are unrelated model-specific bugs."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "Mistral-3 image dtype mismatch and a broken Nemotron example are different issues."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Mistral-3 image preprocessor dtype mismatch and MPS histogram-on-Int error are unrelated."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "Different models and symptoms: Mistral-3 dtype mismatch vs Qwen-3 NVFP4 runtime error."
+          },
+          {
+            "left": "issue:45507",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "GraniteMoEHybrid invalid method call and MPS histogram op failure are not the same bug."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Liger Kernel cross-entropy call and Qwen3.5 embedding unpacking under TRL are unrelated."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Gemma3 ZeRO-3 loading error and BERT ZeRO-3 init error both involve DeepSpeed, but the model/setup and failure mechanism differ."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Qwen-3 NVFP4 runtime error and Qwen3.5 SFT embedding unpacking are different training/runtime paths."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "TRL SFT embedding unpacking failure is unrelated to GPT-OSS only-TP behavior."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary stays grounded in the packet and is appropriately conservative. The two accepted duplicate pairs are well supported by the titles/body cues: 43596/43638 share the same ZeRO-3 BERT index-out-of-bounds failure, and 45800/46515 both report torch 2.4.1 incompatibility. The rest of the soft-edge rejections are cautious and do not overclaim duplication from weak textual similarity."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 2,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:36010",
+        "issue:39290",
+        "issue:41093",
+        "issue:41720",
+        "issue:41762",
+        "issue:43039",
+        "issue:43296",
+        "issue:43366",
+        "issue:43404",
+        "issue:43425",
+        "issue:43531",
+        "issue:43541",
+        "issue:43572",
+        "issue:43596",
+        "issue:43638",
+        "issue:43792",
+        "issue:43828",
+        "issue:43854",
+        "issue:43866",
+        "issue:43872",
+        "issue:43901",
+        "issue:43975",
+        "issue:44292",
+        "issue:44387",
+        "issue:44512",
+        "issue:44560",
+        "issue:44589",
+        "issue:44805",
+        "issue:44841",
+        "issue:44863",
+        "issue:44918",
+        "issue:45005",
+        "issue:45020",
+        "issue:45070",
+        "issue:45084",
+        "issue:45161",
+        "issue:45237",
+        "issue:45362",
+        "issue:45464",
+        "issue:45685",
+        "issue:45800",
+        "issue:45812",
+        "issue:45823",
+        "issue:45901",
+        "issue:45920",
+        "issue:46015",
+        "issue:46294",
+        "issue:46304"
+      ],
+      "soft_pairs": [
+        "issue:43296|issue:46304",
+        "issue:41762|issue:43596",
+        "issue:39290|issue:41720",
+        "issue:44589|issue:46304",
+        "issue:43901|issue:44512",
+        "issue:45823|issue:46294",
+        "issue:45237|issue:45901",
+        "issue:44841|issue:45823",
+        "issue:43039|issue:44560",
+        "issue:43975|issue:46015",
+        "issue:44863|issue:46015",
+        "issue:45020|issue:46304",
+        "issue:45812|issue:45920",
+        "issue:43866|issue:44863",
+        "issue:43872|issue:46304",
+        "issue:44292|issue:46015",
+        "issue:39290|issue:43531",
+        "issue:41093|issue:43638",
+        "issue:43792|issue:45901",
+        "issue:43854|issue:43866",
+        "issue:43425|issue:45070",
+        "issue:43596|issue:45800",
+        "issue:43975|issue:45901",
+        "issue:43792|issue:45070",
+        "issue:45362|issue:45464",
+        "issue:44292|issue:45901",
+        "issue:39290|issue:45823",
+        "issue:45464|issue:45685",
+        "issue:43541|issue:45685",
+        "issue:43541|issue:43828",
+        "issue:36010|issue:39290",
+        "issue:44841|issue:45084",
+        "issue:43404|issue:43866",
+        "issue:45800|issue:45823",
+        "issue:41093|issue:41762",
+        "issue:43296|issue:43366",
+        "issue:43572|issue:45070",
+        "issue:44805|issue:44918",
+        "issue:44387|issue:45005",
+        "issue:43541|issue:45161",
+        "issue:43828|issue:45464",
+        "issue:43366|issue:45464"
+      ],
+      "prepared_review_unit_hash": "37c5980aab730d0a5597efcc3864b14aa1a3dc11b939e73e9819096961b943ea",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 42,
+        "serialized_chars": 40034,
+        "estimated_input_tokens": 10009,
+        "estimated_eval_tokens": 20274
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous issue reports, not a single duplicate set. The strongest duplicate signal is the v5 AutoTokenizer token-ID regression reports, with a secondary likely duplicate around the ZeRO-3 index-out-of-bounds loading error.",
+        "confidence": 0.68,
+        "canonical_issue_reason": "issue:45920 is the broadest and most current report for the AutoTokenizer v5 regression, covering multiple model families rather than one specific model.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45920 best represents the underlying tokenizer regression because it aggregates failures across several models and is the most general open report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43296",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "Different failure modes and layers: vLLM/PaddleOCR-VL loading vs PyTorch build/distributed configuration."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:43596",
+            "accept": true,
+            "reason": "Same exact ZeRO-3 out-of-bounds index error during model loading/init; different models, but the concrete bug looks shared."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Different model-specific runtime errors: missing Gemma3 config attribute vs Qwen3 auto-device CUDA assert."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "Float8 storage lookup failure is unrelated to the distributed-compile support regression."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Both are documentation-related, but they discuss different APIs and different outdated text."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "Different bugs: Gemma4 device-map mismatch vs Qwen3.5 video chat-template feature/token mismatch."
+          },
+          {
+            "left": "issue:45237",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "Unrelated: GPT-OSS on AMD GPUs vs table-question-answering crash."
+          },
+          {
+            "left": "issue:44841",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Processor loading failure and device-map forward mismatch are different code paths."
+          },
+          {
+            "left": "issue:43039",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Liger kernel cross-entropy call vs Qwen3.5 video StopIteration are unrelated."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "Detokenization bug vs stale RAG docs are unrelated."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "NemotronH checkpoint loading failure is not a docs-example issue."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "Remote-code breakage and USE_DISTRIBUTED=0 compatibility are separate regressions."
+          },
+          {
+            "left": "issue:45812",
+            "right": "issue:45920",
+            "accept": true,
+            "reason": "Same underlying v5 tokenizer regression producing wrong token IDs across models; 45920 is the broader instance of the same bug."
+          },
+          {
+            "left": "issue:43866",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "Corrupted checkpoint report vs NemotronH implementation loading issue are not the same defect."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:46304",
+            "accept": false,
+            "reason": "bitsandbytes constructor incompatibility is unrelated to distributed build support."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:46015",
+            "accept": false,
+            "reason": "Qwen-3-8B-NVFP4 runtime error is unrelated to RAG documentation drift."
+          },
+          {
+            "left": "issue:39290",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Different model/config regressions: Gemma3 config attribute vs Qwen3-MoE sliding_window issue."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:43638",
+            "accept": false,
+            "reason": "Both are IndexErrors, but the reported contexts differ enough that the same root cause is not established."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "Whisper loading failure and table QA crash are unrelated."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "Unit-test load failure and corrupted Ovis2 checkpoint are different problems."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Torch-version incompatibility and PretrainedConfig pydantic-field regression are separate issues."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "ZeRO-3 loading IndexError and torch/transformers version incompatibility are unrelated."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "Detokenization bug and table QA crash do not share a concrete code path."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Whisper runtime failure is unrelated to the pydantic model-field regression."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Both involve Qwen3.5, but one is a chat crash and the other is a streaming API failure; not enough evidence of the same bug."
+          },
+          {
+            "left": "issue:45685",
+            "right": "issue:45901",
+            "accept": false,
+            "reason": "MoE histogram-on-MPS error is unrelated to table QA."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Mask-shape mismatch and Qwen3.5 embedding unpacking failures are different errors and code paths."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Int4 OOM regression and tied-weights translation-model issue are distinct."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Grouped_mm tracing error in Mixtral is not the same as GPT-OSS tensor-parallel-only failure."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Autocast dtype mismatch and streaming inference failure are different bugs."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "GGUF/gpt-oss support request is unrelated to Qwen3.5 streaming inference."
+          },
+          {
+            "left": "issue:45800",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Torch/transformers compatibility issue is unrelated to Gemma4 device-map mismatch."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:41762",
+            "accept": false,
+            "reason": "Both are index-related errors, but one is a mask/tensor shape mismatch and the other is ZeRO-3 empty-index loading; not the same concrete bug."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:43366",
+            "accept": false,
+            "reason": "PaddleOCR-VL loading in vLLM and GGUF gpt-oss support are unrelated."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Missing pad_token_idx and PretrainedConfig pydantic breakage are different regressions."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Mask-shape IndexError and template-node compilation failure are unrelated."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43866",
+            "accept": false,
+            "reason": "lm_head weight tying regression and corrupted checkpoint report are different issues."
+          },
+          {
+            "left": "issue:45800",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Different regressions in different subsystems: environment compatibility vs device placement."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: it correctly characterizes the cluster as mostly heterogeneous, and it identifies the strongest duplicate evidence around the v5 AutoTokenizer token-ID regression (45812/45920). The ZeRO-3 index-out-of-bounds pair is also a reasonable, conservative likely-duplicate call given the identical error and shared loading/init context. No soft-edge verdict appears overstated."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 3,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:36331",
+        "issue:39186",
+        "issue:39692",
+        "issue:40990",
+        "issue:41093",
+        "issue:41762",
+        "issue:42915",
+        "issue:43054",
+        "issue:43404",
+        "issue:43531",
+        "issue:43541",
+        "issue:43596",
+        "issue:43644",
+        "issue:43645",
+        "issue:43716",
+        "issue:43828",
+        "issue:43873",
+        "issue:43950",
+        "issue:43975",
+        "issue:43994",
+        "issue:44291",
+        "issue:44292",
+        "issue:44360",
+        "issue:44479",
+        "issue:44492",
+        "issue:44512",
+        "issue:44530",
+        "issue:44534",
+        "issue:44589",
+        "issue:44863",
+        "issue:44918",
+        "issue:44928",
+        "issue:44960",
+        "issue:45005",
+        "issue:45020",
+        "issue:45070",
+        "issue:45183",
+        "issue:45399",
+        "issue:45507",
+        "issue:45632",
+        "issue:45698",
+        "issue:45759",
+        "issue:45800",
+        "issue:45902",
+        "issue:46294",
+        "issue:46463",
+        "issue:46489",
+        "issue:46491"
+      ],
+      "soft_pairs": [
+        "issue:44291|issue:45070",
+        "issue:43531|issue:45070",
+        "issue:45020|issue:45800",
+        "issue:43716|issue:46294",
+        "issue:43541|issue:44589",
+        "issue:44479|issue:46294",
+        "issue:41093|issue:43596",
+        "issue:43054|issue:43994",
+        "issue:45759|issue:45902",
+        "issue:39186|issue:44918",
+        "issue:43716|issue:44918",
+        "issue:43645|issue:45399",
+        "issue:40990|issue:44960",
+        "issue:45632|issue:45698",
+        "issue:43404|issue:45005",
+        "issue:42915|issue:45399",
+        "issue:44530|issue:45005",
+        "issue:43873|issue:45005",
+        "issue:44360|issue:44512",
+        "issue:44863|issue:45005",
+        "issue:45183|issue:46463",
+        "issue:44918|issue:45507",
+        "issue:44492|issue:44512",
+        "issue:46489|issue:46491",
+        "issue:43975|issue:44292",
+        "issue:44292|issue:45507",
+        "issue:43950|issue:44534",
+        "issue:39692|issue:43054",
+        "issue:43644|issue:43950",
+        "issue:36331|issue:41762",
+        "issue:43828|issue:44928"
+      ],
+      "prepared_review_unit_hash": "6514a52aa1b3afc59732de644a7702d7806fa312b76f6998ec41e5eccba934cb",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 37034,
+        "estimated_input_tokens": 9259,
+        "estimated_eval_tokens": 18774
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most items are unrelated bug reports across different models/features. Only the non-persistent-buffer regression appears to be a true duplicate cluster; the rest are too model- or symptom-specific to merge.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "issue:43950 most clearly describes the underlying v5 regression affecting `register_buffer(persistent=False)` and `from_pretrained()`, and it subsumes the shorter duplicate phrasing in the related reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43950 is the strongest canonical issue because it states the concrete bug and regression scope most explicitly.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44291",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Both are v5 regressions, but one is an `init_empty_weights` TypeError and the other is a pydantic/config field break; different code paths."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45070",
+            "accept": false,
+            "reason": "Qwen3-MoE `sliding_window` behavior is unrelated to the `PretrainedConfig`/pydantic regression."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "One is a `remote_code` breakage, the other is a Torch/Transformers version incompatibility; not the same bug."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "Both involve multimodal input mismatches, but one is image preprocessor/model dtype mismatch and the other is video feature/token count mismatch."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "`grouped_mm` tracing failure and missing `Float8_e4m3fnStorage` are different failures with different triggers."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "Both mention video, but one is a v5.3 video-input regression across several models and the other is a specific Qwen3.5-9B chat template mismatch."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:43596",
+            "accept": false,
+            "reason": "Both are index errors, but one is a mask/tensor shape mismatch and the other is a ZeRO-3 empty-weight initialization issue."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Same model family, but one reports poor embedding quality while the other reports nonsensical output in AutoModel/pipeline; not the same concrete bug."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "One is about unwrapping `text_config` for `from_config`; the other is an uninitialized `inv_freq` buffer after meta materialization."
+          },
+          {
+            "left": "issue:39186",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "FSDP `'weight' must be 2-D` and TRL SFT unpacking errors are separate training-stack failures."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Image/video dtype mismatch and Qwen3.5 embedding unpacking are different model-path issues."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Defining custom models in notebooks is unrelated to the flash-attn2 fallback gating problem."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "High perplexity on gpt-oss-20b and a generic GLM5 issue do not describe the same failure."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Both involve `trust_remote_code`/custom modules, but one is a cache-path collision and the other loads the wrong module after save/load; different bugs."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Both mention tied weights, but they concern different models and different failure modes; not the same concrete regression."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 failure and flash-attn2 fallback gating are unrelated."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "PagedAttentionCache group-type crash is unrelated to tied-weight regressions."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Quantization/offloading behavior is a different problem from tied weights."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "A DSA/indexing discussion is unrelated to a docs typo about the removed run command."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "NemotronH checkpoint loading and tied-weight translation issues are separate defects."
+          },
+          {
+            "left": "issue:45183",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Both are validation/error-message problems, but one is a broad `input_ids`/`inputs_embeds` XOR issue and the other is a T5 decoder-input validation message."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "TRl SFT embedding unpacking and GraniteMoEHybrid invalid method calls are unrelated."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44512",
+            "accept": false,
+            "reason": "Cache-strategy typo and removed-doc-command mention are different documentation issues."
+          },
+          {
+            "left": "issue:46489",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different models and different round-trip failures."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:44292",
+            "accept": false,
+            "reason": "DeepSeek-Coder detokenization and Qwen-3-8B-NVFP4 runtime errors are unrelated."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "Different models and different failure modes; no shared underlying bug."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:44534",
+            "accept": true,
+            "reason": "These describe the same v5 regression corrupting non-persistent buffers; one is the detailed canonical report and the other is a shorter duplicate description."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43054",
+            "accept": false,
+            "reason": "SigLIP2 docs/example errors and SigLIP2 embedding-quality complaints are not the same bug."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43950",
+            "accept": true,
+            "reason": "Both report the same non-persistent-buffer corruption in Transformers 5.x; `issue:43950` gives the clearest concrete description."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41762",
+            "accept": false,
+            "reason": "CustomTrainer `num_items_in_batch` and Gemma3 ZeRO-3 empty-tensor index errors are unrelated."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:44928",
+            "accept": false,
+            "reason": "Both mention dtype/numerical issues, but one is an autocast mismatch and the other is an RLHF NaN/explosion caused by 3D position IDs and SDPA fallback."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is well grounded in the packet and stays conservative on the soft edges. The canonical duplicate cluster around non-persistent buffer corruption is supported by the titles/body descriptions, and the other pairings are appropriately rejected as distinct bug reports rather than overstated duplicates."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 4,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:24643",
+        "issue:28530",
+        "issue:38175",
+        "issue:41553",
+        "issue:41762",
+        "issue:42491",
+        "issue:42915",
+        "issue:42947",
+        "issue:43054",
+        "issue:43493",
+        "issue:43643",
+        "issue:43742",
+        "issue:43782",
+        "issue:43824",
+        "issue:43828",
+        "issue:43856",
+        "issue:43866",
+        "issue:43872",
+        "issue:43881",
+        "issue:43883",
+        "issue:43950",
+        "issue:43975",
+        "issue:44368",
+        "issue:44387",
+        "issue:44403",
+        "issue:44451",
+        "issue:44488",
+        "issue:44589",
+        "issue:44661",
+        "issue:44910",
+        "issue:44960",
+        "issue:44991",
+        "issue:45020",
+        "issue:45081",
+        "issue:45356",
+        "issue:45399",
+        "issue:45488",
+        "issue:45520",
+        "issue:45663",
+        "issue:45685",
+        "issue:45701",
+        "issue:45715",
+        "issue:45823",
+        "issue:45859",
+        "issue:45902",
+        "issue:45920",
+        "issue:46315",
+        "issue:46489"
+      ],
+      "soft_pairs": [
+        "issue:43872|issue:44589",
+        "issue:45920|issue:46489",
+        "issue:43054|issue:45701",
+        "issue:43866|issue:44403",
+        "issue:43950|issue:45902",
+        "issue:38175|issue:43054",
+        "issue:41553|issue:43883",
+        "issue:44910|issue:45859",
+        "issue:42947|issue:44387",
+        "issue:45663|issue:45902",
+        "issue:43054|issue:43493",
+        "issue:44403|issue:44451",
+        "issue:41553|issue:43881",
+        "issue:24643|issue:28530",
+        "issue:43975|issue:45520",
+        "issue:44403|issue:44488",
+        "issue:42947|issue:43856",
+        "issue:43643|issue:45020",
+        "issue:41762|issue:42915",
+        "issue:45685|issue:45823",
+        "issue:44991|issue:45399",
+        "issue:43824|issue:43872",
+        "issue:43782|issue:44960",
+        "issue:43881|issue:43883",
+        "issue:45020|issue:45823",
+        "issue:44661|issue:45020",
+        "issue:45081|issue:45399",
+        "issue:43828|issue:45823",
+        "issue:42491|issue:44368",
+        "issue:43742|issue:45399",
+        "issue:45020|issue:45356",
+        "issue:45715|issue:46315",
+        "issue:45488|issue:46489"
+      ],
+      "prepared_review_unit_hash": "a6eb8e25c9f4be8a82c2a0affeb7d23a2d9c108679f699a3f2ba4fb1261b32a7",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 37913,
+        "estimated_input_tokens": 9479,
+        "estimated_eval_tokens": 19214
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous: it mixes unrelated tokenizer regressions, model-loading failures, quantization/offload issues, and training/runtime bugs. I only see one plausible duplicate edge in the tokenizer-regression subgroup; the rest are too different to merge safely.",
+        "confidence": 0.74,
+        "canonical_issue_reason": "Issue 45920 is the best canonical issue because it is the broadest and most representative of the tokenizer-regression theme, covering multiple affected model families under the v5 tokenization breakage.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45920 is the strongest global issue candidate: it is open, high-impact, and serves as the clearest umbrella for the recurring wrong-tokenization regression seen in nearby reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43872",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Different low-level failures: bitsandbytes init arg incompatibility vs missing Float8 storage type."
+          },
+          {
+            "left": "issue:45920",
+            "right": "issue:46489",
+            "accept": true,
+            "reason": "Both report v5 tokenizer output/ID regressions on DeepSeek-family models and plausibly share the same tokenization code path."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "One is SigLIP2 embedding quality; the other is a broad tokenization-version change without the same concrete failure."
+          },
+          {
+            "left": "issue:43866",
+            "right": "issue:44403",
+            "accept": false,
+            "reason": "Checkpoint corruption is not the same as generic loading noise."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Both involve loading/materialization, but one is silent non-persistent buffer corruption and the other is a Qwen3 MoE uninitialized-memory bug."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43054",
+            "accept": false,
+            "reason": "Both concern SigLIP2 behavior, but one is zero probabilities and the other is degraded text embeddings; not the same bug."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "Bad tokenizer error messaging and Molmo tied-weights attribute failure are unrelated."
+          },
+          {
+            "left": "issue:44910",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Flash-Attention packed-sequence crash from 3D position_ids is unrelated to CPU offload incompatibility in Qwen3 MoE rotary embedding."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Gradient checkpointing not taking effect and int4 memory regression are different memory-related issues."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Gemma4 FSDP2 shared_kv_states KeyError is unrelated to Qwen3 MoE buffer materialization."
+          },
+          {
+            "left": "issue:43054",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Both are SigLIP2-related, but one is poor text embedding and the other is HF vs JAX implementation discrepancy."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Generic loading noise is not the same as a specific model load failure for ScandiBERT."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "AutoTokenizer Voxtral error messaging and glm-4v loading failure are distinct issues."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:28530",
+            "accept": false,
+            "reason": "DeepSpeed weight-dimension runtime error is unrelated to early-stopping metric configuration."
+          },
+          {
+            "left": "issue:43975",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Tokenizer detokenization regression and Python 3.13 flash_attn import error are unrelated."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Generic loading noise is not the same bug as a specific failure loading cjvt/sleng-bert."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "LoRA gradient-checkpointing behavior and Qwen3 MoE memory usage are different training problems."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Both touch remote_code, but one is missing fields in AutoConfig output and the other is broad model breakage; not the same concrete bug."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Gemma3 ZeRO-3 loading IndexError and Qwen3 MoE FP8 failure are unrelated model-specific errors."
+          },
+          {
+            "left": "issue:45685",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "MPS histogram op failure and Gemma4 device-mismatch forward failure are different backend issues."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Tokenizer-loading failure for a specific model and flash-attn2 fallback gating are unrelated."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43872",
+            "accept": false,
+            "reason": "Qwen2.5-VL import failure and bitsandbytes Int8Params incompatibility are unrelated."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "Qwen3VL weight_only loading error and GLM5 are different model-loading problems."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "GLM-4v load failure and Molmo missing all_tied_weights_keys are unrelated."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Remote_code breakage is broader and different from Gemma4 PLE device placement failure."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Tokenizer mapping-name bug and remote_code regression are different code paths."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "Mistral regex patch crash and flash-attn2 fallback blocking are unrelated tokenizer vs backend-selection issues."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Autocast dtype mismatch in Phi-tiny-MoE and Gemma4 device mismatch are different forward-path bugs."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "LoRA model compatibility across versions and tie_word_embeddings warning are not the same issue."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:45399",
+            "accept": false,
+            "reason": "MobileLLM key error is unrelated to flash-attn2 fallback gating."
+          },
+          {
+            "left": "issue:45020",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Remote_code breakage and Kimi-K2.5 tokenizer codec regression are different tokenizer/config problems."
+          },
+          {
+            "left": "issue:45715",
+            "right": "issue:46315",
+            "accept": false,
+            "reason": "Both are tokenizer performance regressions, but one is convert_ids_to_tokens and the other convert_tokens_to_ids; different hot paths."
+          },
+          {
+            "left": "issue:45488",
+            "right": "issue:46489",
+            "accept": false,
+            "reason": "Both involve DeepSeek-family tokenization regressions, but the concrete failures differ enough that they should not be merged without stronger evidence."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded overall. The cluster summary stays conservative and matches the packet: most pairs are genuinely heterogeneous, and the chosen canonical issue (45920) is the broadest open tokenizer-regression report in the set. The single accepted soft edge is cautiously worded and supported by the titles as a plausible shared tokenizer-path regression. No obvious overstatement."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 5,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:24643",
+        "issue:30064",
+        "issue:36331",
+        "issue:41093",
+        "issue:41553",
+        "issue:42915",
+        "issue:42947",
+        "issue:43257",
+        "issue:43366",
+        "issue:43404",
+        "issue:43531",
+        "issue:43596",
+        "issue:43638",
+        "issue:43716",
+        "issue:43792",
+        "issue:43827",
+        "issue:43828",
+        "issue:43854",
+        "issue:43872",
+        "issue:43940",
+        "issue:43950",
+        "issue:44155",
+        "issue:44220",
+        "issue:44291",
+        "issue:44479",
+        "issue:44488",
+        "issue:44509",
+        "issue:44530",
+        "issue:44560",
+        "issue:44589",
+        "issue:44683",
+        "issue:44805",
+        "issue:44863",
+        "issue:44928",
+        "issue:44991",
+        "issue:45305",
+        "issue:45313",
+        "issue:45325",
+        "issue:45584",
+        "issue:45588",
+        "issue:45663",
+        "issue:45823",
+        "issue:45834",
+        "issue:45902",
+        "issue:45923",
+        "issue:46402",
+        "issue:46515"
+      ],
+      "soft_pairs": [
+        "issue:43827|issue:44509",
+        "issue:43404|issue:43950",
+        "issue:43872|issue:44291",
+        "issue:42915|issue:43872",
+        "issue:43792|issue:44220",
+        "issue:36331|issue:41093",
+        "issue:43257|issue:45305",
+        "issue:42947|issue:45923",
+        "issue:41553|issue:44991",
+        "issue:43940|issue:45313",
+        "issue:43854|issue:44863",
+        "issue:44928|issue:45588",
+        "issue:41093|issue:44805",
+        "issue:44479|issue:44560",
+        "issue:41553|issue:44488",
+        "issue:24643|issue:30064",
+        "issue:43596|issue:46515",
+        "issue:44863|issue:46402",
+        "issue:44530|issue:44863",
+        "issue:43366|issue:43828",
+        "issue:43638|issue:46515",
+        "issue:43366|issue:43531",
+        "issue:45325|issue:45902",
+        "issue:45663|issue:45823",
+        "issue:43716|issue:44560",
+        "issue:44155|issue:45584",
+        "issue:43366|issue:44589",
+        "issue:44683|issue:46515",
+        "issue:43950|issue:45834"
+      ],
+      "prepared_review_unit_hash": "af39855888d449231a6fa348b2071189571d8b7c2058a21326bbe9ee9d1aec73",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 29,
+        "serialized_chars": 36073,
+        "estimated_input_tokens": 9019,
+        "estimated_eval_tokens": 18294
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is not one coherent duplicate set; it contains several separate subclusters. The clearest matches are around v5 docs cleanup, `_is_hf_initialized` loading failures, Whisper audio preprocessing, mask-shape indexing, and a Qwen VL video regression. Most other pairs are only loosely related and should stay separate.",
+        "confidence": 0.82,
+        "canonical_issue_reason": "No single canonical issue: the items split into multiple unrelated bug themes, so any one issue would misrepresent the cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No global best issue either; the strongest matches are in different subclusters rather than one shared underlying bug.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43827",
+            "right": "issue:44509",
+            "accept": true,
+            "reason": "Both are docs issues about stale pipeline-task references removed in v5."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43950",
+            "accept": false,
+            "reason": "`lm_head` weight tying is a different bug from non-persistent buffer corruption in `from_pretrained()`."
+          },
+          {
+            "left": "issue:43872",
+            "right": "issue:44291",
+            "accept": true,
+            "reason": "Both center on the same unexpected `_is_hf_initialized` loading/init argument error."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43872",
+            "accept": false,
+            "reason": "FP8 config failure and `_is_hf_initialized`/bitsandbytes incompatibility are different code paths."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:44220",
+            "accept": true,
+            "reason": "Both point to Whisper audio preprocessing / `_torch_extract_fbank_features()` preventing the model from running."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41093",
+            "accept": false,
+            "reason": "A trainer API kwarg mismatch is unrelated to a tensor mask-shape indexing error."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45305",
+            "accept": false,
+            "reason": "Weight loading under accelerate+deepspeed is not the same bug as gradient averaging with GAS."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "LoRA gradient-checkpointing ineffectiveness and a missing support flag on a trust_remote_code variant are different issues."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are tokenizer-loading problems, but they affect different models and different failure modes."
+          },
+          {
+            "left": "issue:43940",
+            "right": "issue:45313",
+            "accept": false,
+            "reason": "Both mention ZeRO-3 weight loading, but the model families and reported failure surfaces differ too much to merge confidently."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "These are separate model-loading failures for different architectures."
+          },
+          {
+            "left": "issue:44928",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "RLHF NaNs from SDPA/BF16 collapse are unrelated to a flash-attention crash on `s_aux=None`."
+          },
+          {
+            "left": "issue:41093",
+            "right": "issue:44805",
+            "accept": true,
+            "reason": "Same mask/index length mismatch error, just with different example sizes."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44560",
+            "accept": true,
+            "reason": "Both describe the same 5.3.0 video-input regression in the Qwen VL family."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "A bad Voxtral tokenizer error message is not the same bug as failing to load `cjvt/sleng-bert`."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "DeepSpeed training `'weight' must be 2-D` and void segmentation-map preprocessing are unrelated."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "ZeRO-3 init index errors are a different class of bug than general torch-version incompatibility."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "NemotronH checkpoint loading and Sam3 text-encoder weight loading are separate model-specific failures."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:44863",
+            "accept": false,
+            "reason": "PagedAttention group-type handling is unrelated to NemotronH checkpoint loading."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "GGUF architecture support and autocast dtype mismatch are different problems."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "A Zero-3 BERT init index error is not the same as a torch version compatibility issue."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "GPT-OSS GGUF support and Qwen3-MoE sliding-window behavior are unrelated."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45902",
+            "accept": false,
+            "reason": "Temporal position-id scaling and uninitialized rotary-embedding memory are different defects."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "FSDP2 shared-kv-state rebuilding and device-map placement mismatches are separate Gemma4 issues."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Image preprocessor dtype mismatch is not the same as a Qwen VL video StopIteration bug."
+          },
+          {
+            "left": "issue:44155",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Audio track leakage in batched inference is unrelated to empty-transcription generation failure."
+          },
+          {
+            "left": "issue:43366",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "GGUF gpt-oss support and missing Float8 storage handling are different concerns."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "Compiled flex_attention breakage on torch>=2.9 is not the same as torch 2.4.1 incompatibility."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Non-persistent buffer corruption and long-OCR-input index errors are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: it identifies several distinct subclusters rather than forcing a single duplicate set, and the listed strong matches align with the issue titles. The soft-edge verdicts are conservative overall, with the accepted pairs all plausibly describing the same bug family and the rejected pairs reasonably separated by model, code path, or failure mode."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 6,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:36010",
+        "issue:38175",
+        "issue:39692",
+        "issue:41720",
+        "issue:42491",
+        "issue:42617",
+        "issue:42915",
+        "issue:43317",
+        "issue:43366",
+        "issue:43550",
+        "issue:43596",
+        "issue:43632",
+        "issue:43638",
+        "issue:43644",
+        "issue:43645",
+        "issue:43646",
+        "issue:43720",
+        "issue:43824",
+        "issue:43957",
+        "issue:43994",
+        "issue:44451",
+        "issue:44488",
+        "issue:44534",
+        "issue:44560",
+        "issue:44661",
+        "issue:44756",
+        "issue:44805",
+        "issue:44877",
+        "issue:44910",
+        "issue:44918",
+        "issue:44960",
+        "issue:44977",
+        "issue:44991",
+        "issue:45030",
+        "issue:45276",
+        "issue:45313",
+        "issue:45357",
+        "issue:45507",
+        "issue:45685",
+        "issue:45812",
+        "issue:45823",
+        "issue:45834",
+        "issue:46459",
+        "issue:46468",
+        "issue:46489",
+        "issue:46512",
+        "issue:46515"
+      ],
+      "soft_pairs": [
+        "issue:43824|issue:43957",
+        "issue:44560|issue:45685",
+        "issue:42915|issue:43957",
+        "issue:43645|issue:43646",
+        "issue:44560|issue:44918",
+        "issue:43550|issue:44910",
+        "issue:42617|issue:43366",
+        "issue:36010|issue:42915",
+        "issue:45823|issue:46515",
+        "issue:44661|issue:45823",
+        "issue:45812|issue:46489",
+        "issue:44560|issue:44805",
+        "issue:38175|issue:43994",
+        "issue:43638|issue:44661",
+        "issue:43596|issue:44661",
+        "issue:43632|issue:45834",
+        "issue:44877|issue:45030",
+        "issue:44661|issue:44805",
+        "issue:43317|issue:44756",
+        "issue:38175|issue:39692",
+        "issue:43644|issue:44534",
+        "issue:42491|issue:43720",
+        "issue:41720|issue:42915",
+        "issue:42915|issue:44756",
+        "issue:46459|issue:46512",
+        "issue:42491|issue:45276",
+        "issue:44451|issue:44488",
+        "issue:44960|issue:44977",
+        "issue:45313|issue:45357",
+        "issue:45507|issue:46468",
+        "issue:44488|issue:44991",
+        "issue:43644|issue:43645"
+      ],
+      "prepared_review_unit_hash": "2ad7bcd7a987f9f277fe5b1d3c4af08d58f915d7e821b79efc8746dc08fb6f89",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36505,
+        "estimated_input_tokens": 9127,
+        "estimated_eval_tokens": 18510
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This is a mixed cluster of mostly unrelated Transformers bug reports. Only three soft pairs look like true duplicates: two SigLIP2/model-output reports, two non-persistent-buffer regression reports, and the custom-model-init pair. The rest are similar only at the subsystem or model-family level and should stay separate.",
+        "confidence": 0.63,
+        "canonical_issue_reason": "Issue 43644 is the clearest broad v5 regression in the set and has a near-exact duplicate in 44534; it is more general and higher-signal than the narrower follow-on reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "43644 is the best representative issue: concise, generic, and directly describes a concrete regression with strong duplicate evidence.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43824",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Both are import/load failures, but they point to different model surfaces and different failure modes; not the same code-path bug."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Completely different model/runtime issues: a Qwen3-VL video stop-iteration bug vs an MPS histogram error in MoE code."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Different models and different failures; similarity is only that both are Transformers v5-era regressions."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:43646",
+            "accept": true,
+            "reason": "Same underlying custom-model initialization regression; the notebook-specific report is a context variant of the more general breakage."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Different Qwen3.5-related failures with different triggers: video processing stop-iteration vs unpacking embeddings under TRL."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44910",
+            "accept": false,
+            "reason": "Both mention attention/compile failures, but the model, symptom, and code path differ materially."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:43366",
+            "accept": false,
+            "reason": "Unrelated topics: a 3d_parallel.py execution problem versus GGUF architecture support."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "ImportError for GenerationMixin is unrelated to a Qwen3Moe FP8 loading failure."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "Device-map forward mismatch is not the same bug as torch 2.4.1 incompatibility."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Tokenizer-mapping logic and Gemma4 device placement are different subsystems and failure modes."
+          },
+          {
+            "left": "issue:45812",
+            "right": "issue:46489",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different models and likely different tokenizer paths."
+          },
+          {
+            "left": "issue:44560",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Different symptoms and code paths; one is Qwen3-VL video handling, the other is a generic mask-shape indexing error."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43994",
+            "accept": true,
+            "reason": "Same SigLIP2 base model family and same underlying bad-output regression, just described with different symptoms."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "Zero3 Bert initialization failure is unrelated to tokenizer-mapping validation."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "Different bugs: ZeRO-3 Bert init index error vs add-new-model-like failure in tokenizer mappings."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "One is a global hf-initialization flag regression; the other is a Kosmos2.5 long-input indexing issue."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:45030",
+            "accept": false,
+            "reason": "Both involve config validation, but they target different model configs and different validation breakpoints."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Tokenizer-mapping validation is not the same as a generic mask/tensor shape mismatch."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Device-map dequantized loading/offload problem is different from disabling mmap to avoid OOM on Strix Halo."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:39692",
+            "accept": false,
+            "reason": "The documentation example has multiple errors, but that is not the same bug as the SigLIP2 zero-probability runtime issue."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44534",
+            "accept": true,
+            "reason": "Near-identical regression description: Transformers v5 fills non-persistent buffers with junk."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Different architectures and failures: qwen3_moe LoRA compatibility vs BitNet packed-weight loading."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Auto device mapping CUDA assert and FineGrainedFP8Config failure are distinct bugs."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "FP8 MoE failure and mmap/OOM on Strix Halo are unrelated."
+          },
+          {
+            "left": "issue:46459",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "Related to dtype handling, but one is ignored dtype on multimodal checkpoint loading and the other is stale config dtype after from_config."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "LoRA compatibility on hf5.x is not the same bug as resize_token_embeddings not updating Gemma4 embedding/output layers."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Same broad symptom of model loading failure, but the specific model/tokenizer failures are not shown to share a concrete code path."
+          },
+          {
+            "left": "issue:44960",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Different Qwen3.5/GLM5 attention-generation problems; not the same defect."
+          },
+          {
+            "left": "issue:45313",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "ZeRO-3 weight loading for language_model and save_pretrained visual-key regression are separate issues."
+          },
+          {
+            "left": "issue:45507",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "GraniteMoEHybrid invalid method calls and the Nemotron example breakage are unrelated."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are tokenizer/load complaints, but they concern different model families and different failure modes."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "One is a non-persistent-buffer regression; the other is custom-model initialization breakage. They are adjacent v5 issues, not the same bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and the rejections are conservative. The three accepted soft pairs are plausibly duplicate-level matches from the titles alone, and the canonical issue choice is well supported by the near-exact 43644/44534 pair."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 7,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "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:43708",
+        "issue:43854",
+        "issue:43856",
+        "issue:43873",
+        "issue:43931",
+        "issue:43957",
+        "issue:43994",
+        "issue:44075",
+        "issue:44155",
+        "issue:44164",
+        "issue:44291",
+        "issue:44292",
+        "issue:44387",
+        "issue:44402",
+        "issue:44488",
+        "issue:44756",
+        "issue:44863",
+        "issue:44910",
+        "issue:44912",
+        "issue:44960",
+        "issue:45276",
+        "issue:45292",
+        "issue:45335",
+        "issue:45362",
+        "issue:45406",
+        "issue:45488",
+        "issue:45588",
+        "issue:45663",
+        "issue:45685",
+        "issue:45704",
+        "issue:45759",
+        "issue:45812",
+        "issue:45823",
+        "issue:45859",
+        "issue:46050",
+        "issue:46468"
+      ],
+      "soft_pairs": [
+        "issue:43708|issue:44075",
+        "issue:39692|issue:43994",
+        "issue:43856|issue:43873",
+        "issue:43856|issue:44387",
+        "issue:43854|issue:43931",
+        "issue:43257|issue:43854",
+        "issue:42491|issue:45292",
+        "issue:42491|issue:43931",
+        "issue:44960|issue:46050",
+        "issue:45488|issue:45812",
+        "issue:45362|issue:45406",
+        "issue:44164|issue:44291",
+        "issue:45276|issue:45335",
+        "issue:44910|issue:45588",
+        "issue:43425|issue:43957",
+        "issue:44402|issue:44488",
+        "issue:41720|issue:44155",
+        "issue:43873|issue:44387",
+        "issue:43404|issue:43454",
+        "issue:42947|issue:45704",
+        "issue:39692|issue:44960",
+        "issue:45663|issue:45759",
+        "issue:44291|issue:44756",
+        "issue:36010|issue:39404",
+        "issue:45663|issue:45859",
+        "issue:44387|issue:44756",
+        "issue:44292|issue:44912",
+        "issue:44155|issue:45823",
+        "issue:43957|issue:44756",
+        "issue:43335|issue:45663",
+        "issue:42947|issue:43381",
+        "issue:44863|issue:46468",
+        "issue:45685|issue:46468",
+        "issue:45276|issue:45292"
+      ],
+      "prepared_review_unit_hash": "d7b48914154d2e334bd13749ec589b72e887490564886e03f3119570a052cec2",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 37950,
+        "estimated_input_tokens": 9488,
+        "estimated_eval_tokens": 19232
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most items are unrelated singleton issues. The only clear duplicate clusters are the resize_token_embeddings reports and, more weakly, the lm_head weight-tying reports in multimodal generation models.",
+        "confidence": 0.74,
+        "canonical_issue_reason": "issue:45292 is the most generic statement of the resize_token_embeddings bug, and the Gemma4/T5Gemma reports look like model-specific manifestations of the same underlying propagation failure.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45292 is the best global issue candidate because it is concise, broadly phrased, and captures the shared resize_token_embeddings regression without overfitting to one architecture.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43708",
+            "right": "issue:44075",
+            "accept": false,
+            "reason": "Different failures: checkpoint resume/max_steps calculation vs SGD argument handling."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Both mention SigLIP2-related problems, but one is a documentation example error and the other is a model/pipeline inference regression."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Both are memory-related, but one is Qwen3 MoE training inefficiency and the other is quantization/offloading behavior."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Different concrete bugs: training memory inefficiency vs increased CUDA reserved memory under int4 quantization."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different models and symptoms: GLM-4.7-Flash unit-test loading vs Qwen3-VL weight-shape mismatch."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Different code paths and models; one is Qwen3 MoE conversion with accelerate/deepspeed, the other is GLM-4.7-Flash loading."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:45292",
+            "accept": false,
+            "reason": "Qwen3 MoE HF4.x/HF5.x compatibility is unrelated to resize_token_embeddings not updating output embeddings."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different problems: LoRA/Qwen3 MoE compatibility vs Qwen3-VL loading shape mismatch."
+          },
+          {
+            "left": "issue:44960",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "GLM5 loading issue and save/load quantization regression are different bugs."
+          },
+          {
+            "left": "issue:45488",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but one is a pre-tokenizer override bug and the other is wrong token IDs for Granite models."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Different failures in different subsystems: chat crash on Qwen3.5 vs Gemma4Processor missing _tokenizer in serve."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Both involve loading/saving, but extra_state handling and init_empty_weights/_is_hf_initialized are separate bugs."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45335",
+            "accept": true,
+            "reason": "Both report resize_token_embeddings failing to propagate updated embeddings into model-specific submodules; same underlying bug class across architectures."
+          },
+          {
+            "left": "issue:44910",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Different flash-attention failures: packed-sequence misinterpretation vs s_aux=None AttributeError."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Torch version compatibility and meta-device model loading are unrelated."
+          },
+          {
+            "left": "issue:44402",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both involve tokenizer/vocab loading, but one is a specific vocab-size mismatch and the other is a different model load failure."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44155",
+            "accept": false,
+            "reason": "Different bugs: Qwen3 auto device mapping crash vs AudioFlamingo3 batched inference token leakage."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Quantization/offloading issues, but not the same concrete failure or code path."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43454",
+            "accept": true,
+            "reason": "Both describe the same underlying issue class: lm_head weights are not tied in model-specific generation configs, leading to incorrect generation."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Gradient checkpointing/LoRA behavior is unrelated to the apex RMSNorm memory leak."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44960",
+            "accept": false,
+            "reason": "SigLIP2 documentation/inference issues and GLM5 model-loading issues are unrelated."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Different composite-model bugs: FSDP2 shared_kv_states failure vs text_config unwrapping in AutoModelForCausalLM.from_config."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "init_empty_weights TypeError and mmap OOM mitigation are unrelated."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:39404",
+            "accept": false,
+            "reason": "ImportError for GenerationMixin and Whisper return_language pipeline regression are different issues."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Different Qwen/Gemma MoE behaviors; one is FSDP2 attention state rebuilding, the other is CPU offload incompatibility in rotary embedding."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Both can produce OOM, but the causes and fixes are different."
+          },
+          {
+            "left": "issue:44292",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Different quantization/load failures on different models and formats (NVFP4 vs MXFP4 fallback)."
+          },
+          {
+            "left": "issue:44155",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "AudioFlamingo3 track leakage and Gemma4 PLE device mismatch are unrelated."
+          },
+          {
+            "left": "issue:43957",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Meta-device loading regression is unrelated to Strix Halo mmap OOM."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer construction bug and Gemma-4 FSDP2 attention bug are unrelated."
+          },
+          {
+            "left": "issue:42947",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "Both mention gradient checkpointing, but one is a PEFT/LoRA ineffective checkpointing bug and the other is an eval-mode restriction."
+          },
+          {
+            "left": "issue:44863",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "NemotronH checkpoint-loading failure and the official Nemotron example being broken are not clearly the same bug."
+          },
+          {
+            "left": "issue:45685",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "MPS histogram support error is unrelated to Nemotron example loading."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45292",
+            "accept": true,
+            "reason": "Same core resize_token_embeddings regression: the updated embeddings are not reflected in the model's output-side embeddings."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "Grounded overall and conservative. The only accepted clusters are supported by the issue titles and look like the same bug class across models (resize_token_embeddings propagation; lm_head weight tying). The non-duplication calls are also appropriately cautious."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 8,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:39404",
+        "issue:42796",
+        "issue:43381",
+        "issue:43425",
+        "issue:43596",
+        "issue:43611",
+        "issue:43645",
+        "issue:43646",
+        "issue:43708",
+        "issue:43716",
+        "issue:43828",
+        "issue:43854",
+        "issue:43901",
+        "issue:43906",
+        "issue:43957",
+        "issue:43994",
+        "issue:44079",
+        "issue:44164",
+        "issue:44265",
+        "issue:44402",
+        "issue:44451",
+        "issue:44479",
+        "issue:44509",
+        "issue:44530",
+        "issue:44589",
+        "issue:44683",
+        "issue:44756",
+        "issue:44805",
+        "issue:44877",
+        "issue:44898",
+        "issue:44912",
+        "issue:44918",
+        "issue:44928",
+        "issue:45084",
+        "issue:45216",
+        "issue:45292",
+        "issue:45335",
+        "issue:45588",
+        "issue:45663",
+        "issue:45684",
+        "issue:45698",
+        "issue:45701",
+        "issue:45812",
+        "issue:45823",
+        "issue:46315",
+        "issue:46326",
+        "issue:46491"
+      ],
+      "soft_pairs": [
+        "issue:44479|issue:44877",
+        "issue:44164|issue:44756",
+        "issue:43381|issue:44928",
+        "issue:43854|issue:45216",
+        "issue:45588|issue:45663",
+        "issue:45684|issue:45698",
+        "issue:42796|issue:46326",
+        "issue:43828|issue:43957",
+        "issue:43646|issue:43708",
+        "issue:43645|issue:44164",
+        "issue:45292|issue:45335",
+        "issue:43425|issue:43828",
+        "issue:44402|issue:44451",
+        "issue:44589|issue:44756",
+        "issue:44683|issue:45823",
+        "issue:45812|issue:46315",
+        "issue:46315|issue:46491",
+        "issue:39404|issue:45701",
+        "issue:43425|issue:45823",
+        "issue:43611|issue:43645",
+        "issue:43425|issue:43716",
+        "issue:44530|issue:45084",
+        "issue:44479|issue:44805",
+        "issue:44589|issue:45084",
+        "issue:43906|issue:44530",
+        "issue:44265|issue:44898",
+        "issue:43425|issue:44918",
+        "issue:43901|issue:44509",
+        "issue:39404|issue:44912",
+        "issue:43994|issue:44079",
+        "issue:43596|issue:44683"
+      ],
+      "prepared_review_unit_hash": "0af629af899902b4e5de3d8db7fb6cfe2e314d4f06d124657ae4b7ef66f14647",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36554,
+        "estimated_input_tokens": 9139,
+        "estimated_eval_tokens": 18534
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a heterogeneous set of Transformers regression reports spanning tokenizers, loading, training, quantization, and docs. Only one soft pair looks like a true duplicate: the generic resize_token_embeddings bug and its t5gemma-specific instance.",
+        "confidence": 0.44,
+        "canonical_issue_reason": "issue:45292 is the broadest, model-agnostic formulation of the only clear duplicate pair, and it covers the same resize_token_embeddings/output_embeddings regression that issue:45335 narrows to t5gemma.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45292 is the best representative of the cluster\u2019s one concrete duplicate because it states the underlying bug in general terms rather than a single-model symptom.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44479",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Both are regressions, but one is a Qwen video-input issue and the other is a strict-config loading problem for granite_speech; different code paths."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "save/from_pretrained extra_state handling and Strix Halo mmap OOM are unrelated bugs."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:44928",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode is unrelated to Qwen3.5 RLHF NaNs from 3D position_ids/SDPA fallback."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "GLM-4.7-Flash unit-test loading and Qwen3.5 checkpoint save_pretrained corruption are different failure modes."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "Flash-attention s_aux=None crash and Gemma-4 FSDP2 shared_kv_states KeyError are separate code paths."
+          },
+          {
+            "left": "issue:45684",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "One is file permission propagation during save_pretrained; the other is loading the wrong custom module after save/load."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Both involve tokenizers, but BPE marker handling and continue_final_message behavior are distinct bugs."
+          },
+          {
+            "left": "issue:43828",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Autocast dtype mismatch for Phi-tiny-MoE is unrelated to meta-device loading failures."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:43708",
+            "accept": false,
+            "reason": "Custom model initialization and Trainer resume_from_checkpoint max_steps calculation are unrelated."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:44164",
+            "accept": false,
+            "reason": "Jupyter custom-model initialization and extra_state save/load handling do not describe the same bug."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45335",
+            "accept": true,
+            "reason": "These describe the same underlying resize_token_embeddings regression; 45335 is just the t5gemma-specific manifestation of the broader output_embeddings bug."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43828",
+            "accept": false,
+            "reason": "Torch 2.10 incompatibility is a compatibility claim, not the same as an autocast dtype mismatch bug."
+          },
+          {
+            "left": "issue:44402",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Tokenizer vocab-size mismatch and inability to load ScandiBERT are related broadly to tokenizers/models, but not the same concrete issue."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Float8 storage lookup failure and Strix Halo mmap OOM are different problems."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Compiled flex_attention on torch>=2.9 and Gemma4 PLE device mismatch are unrelated."
+          },
+          {
+            "left": "issue:45812",
+            "right": "issue:46315",
+            "accept": false,
+            "reason": "Wrong token IDs for Granite models is a correctness bug; convert_tokens_to_ids slowdown is a performance regression."
+          },
+          {
+            "left": "issue:46315",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "Tokenizer performance regression and CodeLlama round-trip spacing regression are different issues."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Whisper return_language pipeline behavior and version-dependent tokenization changes are not the same bug."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Torch 2.10 incompatibility and Gemma4 PLE device mismatch are distinct failures."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "base_model_prefix loading breakage and Jupyter custom-model initialization are related only at a high level, not the same code-path bug."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:43716",
+            "accept": false,
+            "reason": "Torch version incompatibility and Mistral-3 image preprocessor dtype mismatch are unrelated."
+          },
+          {
+            "left": "issue:44530",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "PagedAttentionCache linear_attention crash and non-template-node compilation error are different failures."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44805",
+            "accept": false,
+            "reason": "Qwen video-input regression and mask-shape IndexError are not the same bug."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Float8 storage missing and non-template-node compilation errors do not match."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "A reproduction of an external issue and a PagedAttentionCache crash are not enough to conclude the same underlying bug."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "torch.export failures with torch_compilable_check and Perceiver interpolation failure are different model/runtime issues."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44918",
+            "accept": false,
+            "reason": "Torch 2.10 compatibility and Qwen3.5 input-embedding unpacking in TRL are unrelated."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "Both are docs-related, but they describe different documentation problems and different APIs."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Whisper pipeline return_language behavior and MXFP4 quantization fallback are unrelated."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Siglip2 incorrect outputs and ModelOutput key assignment when previously None are different bugs."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Deepspeed zero3/BertModel index error and compiled flex_attention torch>=2.9 failure are separate issues."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The analysis stays grounded in the packet and remains conservative on the soft edges. The only accepted pair, issue:45292 / issue:45335, is reasonably treated as the broad vs. model-specific manifestation of the same resize_token_embeddings regression. The rest are appropriately rejected as distinct bugs based on the available titles and issue descriptions."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 9,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:32090",
+        "issue:34567",
+        "issue:36032",
+        "issue:39404",
+        "issue:40444",
+        "issue:40990",
+        "issue:42175",
+        "issue:42796",
+        "issue:42915",
+        "issue:43299",
+        "issue:43404",
+        "issue:43421",
+        "issue:43452",
+        "issue:43541",
+        "issue:43582",
+        "issue:43611",
+        "issue:43632",
+        "issue:43643",
+        "issue:43644",
+        "issue:43646",
+        "issue:43716",
+        "issue:43856",
+        "issue:43906",
+        "issue:43931",
+        "issue:44164",
+        "issue:44451",
+        "issue:44462",
+        "issue:44479",
+        "issue:44488",
+        "issue:44530",
+        "issue:44661",
+        "issue:44671",
+        "issue:44683",
+        "issue:44737",
+        "issue:44756",
+        "issue:44987",
+        "issue:44991",
+        "issue:45081",
+        "issue:45084",
+        "issue:45759",
+        "issue:45859",
+        "issue:46050",
+        "issue:46178",
+        "issue:46285",
+        "issue:46326",
+        "issue:46402",
+        "issue:46489"
+      ],
+      "soft_pairs": [
+        "issue:43299|issue:43931",
+        "issue:36032|issue:43452",
+        "issue:44671|issue:44991",
+        "issue:46050|issue:46326",
+        "issue:44164|issue:45759",
+        "issue:43906|issue:46326",
+        "issue:44991|issue:45081",
+        "issue:43856|issue:44756",
+        "issue:44164|issue:44479",
+        "issue:44661|issue:44683",
+        "issue:44488|issue:45081",
+        "issue:44488|issue:44987",
+        "issue:42796|issue:43404",
+        "issue:44462|issue:46489",
+        "issue:43611|issue:43644",
+        "issue:39404|issue:44671",
+        "issue:44737|issue:45859",
+        "issue:43716|issue:45084",
+        "issue:32090|issue:39404",
+        "issue:36032|issue:42915",
+        "issue:46285|issue:46402",
+        "issue:43582|issue:43632",
+        "issue:43643|issue:44530",
+        "issue:44451|issue:44987",
+        "issue:43541|issue:45084",
+        "issue:40990|issue:43421",
+        "issue:42915|issue:44530",
+        "issue:34567|issue:40444",
+        "issue:42175|issue:46178",
+        "issue:43644|issue:43646"
+      ],
+      "prepared_review_unit_hash": "c7a47004029087e2e03e4267f66311e813744294486fe5821794560aaf78ee15",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 30,
+        "serialized_chars": 36141,
+        "estimated_input_tokens": 9036,
+        "estimated_eval_tokens": 18328
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a heterogeneous mix of mostly unrelated Transformers issues, with a few broad v5 regression themes but no true duplicate pair that clearly shares the same concrete bug. No PRs are present.",
+        "confidence": 0.86,
+        "canonical_issue_reason": "issue:43646 is the broadest representative of the cluster\u2019s recurring v5 model-initialization/loading regressions, but the cluster is too mixed to treat as one duplicate family with high confidence.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43646 is the best single issue representative because it is the most general and could plausibly cover some of the nearby v5 initialization-related reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43299",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Both are Qwen3-VL/Qwen3-MoE loading failures, but one is a broad dev-version loading breakage and the other is a specific weight-shape mismatch; not the same concrete bug."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "T5 tokenizer special-token method conflict vs gguf_file loading failure; different code paths and failure modes."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "CamemBERT masked-LM prediction regression vs tokenizer load failure for a different model; unrelated bugs."
+          },
+          {
+            "left": "issue:46050",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Quantized save/load breakage vs mistral_common continue_final_message behavior; different components and symptoms."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "extra_state serialization issue vs from_config text_config unwrapping for composite Qwen configs; not the same bug."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "An isolated reproduction of an older issue vs a tokenizer-specific continue_final_message regression; no concrete overlap."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Tokenizer loading for EMBEDDIA/est-roberta vs a Mistral regex patch crash; separate failure points."
+          },
+          {
+            "left": "issue:43856",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Qwen3 MoE training memory inefficiency vs Strix Halo mmap OOM workaround; different root causes."
+          },
+          {
+            "left": "issue:44164",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Serialization of extra_state vs video-input regression in v5.3.0; unrelated subsystems."
+          },
+          {
+            "left": "issue:44661",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Tokenizer mapping maintenance bug vs torch 2.9 flex_attention compilation failure; different layers entirely."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Model/tokenizer loading failure for sleng-bert vs Mistral regex patch crash; not the same bug."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Both are loading failures, but for different models and different breakpoints in the stack; not a single underlying issue."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43404",
+            "accept": false,
+            "reason": "Tokenizer BPE marker handling vs untied lm_head weights in Mistral3; unrelated tokenizer vs model-weight bugs."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:46489",
+            "accept": false,
+            "reason": "AutoTokenizer ignoring tokenizer.json vs incorrect DeepSeek-Coder tokenizer output; different symptoms and likely different fixes."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43644",
+            "accept": false,
+            "reason": "Both are v5 loading regressions, but one is base_model_prefix loading and the other is non-persistent buffer initialization; too different to merge."
+          },
+          {
+            "left": "issue:39404",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Whisper return_language pipeline regression vs CamemBERT MLM prediction bug; unrelated model families and code paths."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Both involve device placement, but one is XLNet CPU arange creation and the other is Qwen rotary embedding CPU offload; different concrete bugs."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Mistral image/text dtype mismatch vs a template compilation TypeError; no shared failure mechanism."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:39404",
+            "accept": false,
+            "reason": "Trainer gpu broadcast NoneType error vs Whisper pipeline return_language regression; unrelated areas."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Tokenizer add_special_tokens conflict vs Qwen3 MoE FP8 config failure; different features and failures."
+          },
+          {
+            "left": "issue:46285",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "CLIPTextModel structure inconsistency vs Sam3Model text encoder weights not loading; different models and code paths."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:43632",
+            "accept": false,
+            "reason": "Apple Silicon allocator warmup TypeError vs _is_hf_initialized flag regression; distinct bugs."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "trust_remote_code config field loss vs PagedAttentionCache group-type crash on Qwen3.5; unrelated."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Loading a specific BERT model vs loading physical-intelligence/fast; same symptom class, different concrete issue."
+          },
+          {
+            "left": "issue:43541",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Torch Dynamo tracing failure in Mixtral grouped_mm vs template-node compilation TypeError; different code paths."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:43421",
+            "accept": false,
+            "reason": "High perplexity evaluation issue vs tokenizer post-processor update feature bug; unrelated."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44530",
+            "accept": false,
+            "reason": "Qwen3 MoE FP8 training failure vs PagedAttentionCache linear_attention crash; different subsystems and failures."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "TrainerState token counter not updating vs iterable multi-image Qwen2.5-VL finetuning failure; unrelated."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "Missing tensorflow backend in a torch-only install vs undefined tensorflow_text backend mapping; related ecosystem area but not the same bug."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43646",
+            "accept": false,
+            "reason": "Both are Transformers v5 regressions, but one is junk non-persistent buffers and the other is custom model initialization; too broad to treat as one 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": 21,
+      "review_unit_index": 10,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:28530",
+        "issue:34567",
+        "issue:36032",
+        "issue:40444",
+        "issue:42915",
+        "issue:43089",
+        "issue:43262",
+        "issue:43317",
+        "issue:43322",
+        "issue:43377",
+        "issue:43404",
+        "issue:43452",
+        "issue:43482",
+        "issue:43540",
+        "issue:43645",
+        "issue:43646",
+        "issue:43742",
+        "issue:43746",
+        "issue:43819",
+        "issue:43883",
+        "issue:43994",
+        "issue:44062",
+        "issue:44206",
+        "issue:44220",
+        "issue:44265",
+        "issue:44291",
+        "issue:44403",
+        "issue:44451",
+        "issue:44464",
+        "issue:44466",
+        "issue:44488",
+        "issue:44534",
+        "issue:44589",
+        "issue:44610",
+        "issue:44683",
+        "issue:44811",
+        "issue:44849",
+        "issue:44987",
+        "issue:44991",
+        "issue:45081",
+        "issue:45084",
+        "issue:45216",
+        "issue:45325",
+        "issue:45357",
+        "issue:45507",
+        "issue:45584",
+        "issue:45987",
+        "issue:46326"
+      ],
+      "soft_pairs": [
+        "issue:44811|issue:45584",
+        "issue:44206|issue:44220",
+        "issue:40444|issue:46326",
+        "issue:36032|issue:43645",
+        "issue:43883|issue:44488",
+        "issue:43452|issue:43482",
+        "issue:45084|issue:45507",
+        "issue:43317|issue:43482",
+        "issue:44987|issue:44991",
+        "issue:44464|issue:45987",
+        "issue:43262|issue:43377",
+        "issue:43746|issue:46326",
+        "issue:44403|issue:45081",
+        "issue:43646|issue:44534",
+        "issue:42915|issue:45084",
+        "issue:43645|issue:44534",
+        "issue:43883|issue:44991",
+        "issue:43404|issue:44466",
+        "issue:43994|issue:44610",
+        "issue:44849|issue:45325",
+        "issue:43377|issue:43819",
+        "issue:28530|issue:34567",
+        "issue:44451|issue:45081",
+        "issue:43322|issue:43540",
+        "issue:44987|issue:45081",
+        "issue:42915|issue:44291",
+        "issue:44220|issue:44683",
+        "issue:44265|issue:44610",
+        "issue:36032|issue:43742",
+        "issue:43089|issue:45987",
+        "issue:44062|issue:44987",
+        "issue:42915|issue:44589",
+        "issue:45216|issue:45357"
+      ],
+      "prepared_review_unit_hash": "14c039b3c5593b6380c8adaae06af50ea421e064f844696ba899d73db99b1031",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 37685,
+        "estimated_input_tokens": 9422,
+        "estimated_eval_tokens": 19100
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Most pairs are only broadly related and should stay separate. The clearest duplicates are the GGUF loading regression and the Qwen3.5 save_pretrained serialization regression.",
+        "confidence": 0.66,
+        "canonical_issue_reason": "issue:43452 is the broadest, most general statement of a concrete bug in the GGUF loading path, and it cleanly subsumes the more model-specific Qwen2.5-GGUF failure.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43452 is the best representative issue overall because it is model-agnostic, directly actionable, and describes the root loading-path regression rather than a single downstream manifestation.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44811",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Both involve Whisper, but the symptoms differ: batch decoding skip_special_tokens vs empty transcription after align_special_tokens."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Same audio area, but the bugs are different: unsupported `center` argument vs a separate `_torch_extract_fbank_features()` failure."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Different subsystems and failure modes: multimodal iterable dataset with multiple images vs tokenizer `continue_final_message` behavior."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "These are different problems: tokenizer `add_special_tokens` name conflict vs custom-model initialization in notebooks."
+          },
+          {
+            "left": "issue:43883",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Unrelated bugs: missing `all_tied_weights_keys` on Molmo vs inability to load a specific BERT model."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43482",
+            "accept": true,
+            "reason": "Same GGUF loading regression; the second issue is the Qwen2.5-specific manifestation of the same broken `gguf_file`/`from_pretrained` path."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45507",
+            "accept": false,
+            "reason": "No clear shared code-path problem; one is a compile/template-node error, the other is a GraniteMoEHybrid invalid-method call."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Both mention loading failures, but one is dequantized model offload and the other is GGUF/v5 loading; different code paths."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are loading regressions, but they affect different models and likely different tokenizer/model code paths."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45987",
+            "accept": false,
+            "reason": "Related to generation/caching, but one is compiled chunked generation output inconsistency and the other is a `StaticCache` return-type bug."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43377",
+            "accept": false,
+            "reason": "Both audio-related, but one is a sampling-rate default in chat template generation and the other is missing padding-mask support in MIMI encoding."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Different problems: PEFT local checkpoint loading for GraniteSpeech vs tokenizer message-finalization behavior."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Only broad tokenizer-loading overlap; the actual bugs are unrelated noise on load vs `_patch_mistral_regex` crashing."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Custom-model initialization regression is distinct from non-persistent buffer serialization corruption."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Different failures: FineGrainedFP8Config on Qwen3Moe vs template compilation error."
+          },
+          {
+            "left": "issue:43645",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Notebook/custom-model init issue is not the same as junk-filled non-persistent buffers."
+          },
+          {
+            "left": "issue:43883",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Separate model-loading failures with different roots; no evidence they share the same bug."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Both mention `lm_head.weight`, but one is untied weights in a specific model and the other is device-dependent serialization inconsistency."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Completely different bugs: SigLIP nonsensical outputs vs OmDet-Turbo processor image-size mismatch."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Different Qwen regressions affecting hidden-states vs RoPE position-id scaling."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Different audio-model bugs: padding-mask support vs DAC latent/forward mismatch due to missing STE."
+          },
+          {
+            "left": "issue:28530",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Training early-stopping metric selection and `num_input_tokens_seen` updating are unrelated trainer bugs."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both are tokenizer-load failures, but for different models and different failure mechanisms."
+          },
+          {
+            "left": "issue:43322",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Different multimodal loading/processing failures: Llava Next segfault vs Qwen3OmniMoe video-input ValueError."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both involve loading, but one is a model load regression and the other is a Mistral tokenizer regex patch crash."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Different initialization bugs: FineGrainedFP8Config on Qwen3Moe vs unexpected `_is_hf_initialized` with `init_empty_weights`."
+          },
+          {
+            "left": "issue:44220",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Audio feature extraction vs compiled flex_attention are unrelated."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Different export/compilation and processor-size issues; no shared underlying bug."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Tokenizer method-name conflict is unrelated to a key error when loading MobileLLM-125M."
+          },
+          {
+            "left": "issue:43089",
+            "right": "issue:45987",
+            "accept": false,
+            "reason": "Generation overhead/per-token syncs is a performance issue, not the same as a cache-length contract bug."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Different tokenizer-type error vs model load regression; only broad loading-related overlap."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "FineGrainedFP8Config failure and missing Float8 storage are different quantization/loading bugs."
+          },
+          {
+            "left": "issue:45216",
+            "right": "issue:45357",
+            "accept": true,
+            "reason": "Same Qwen3.5 `save_pretrained` serialization regression; 45357 is a more specific follow-up about visual encoder keys."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is conservative and grounded in the packet. The two accepted soft pairs are plausible same-root manifestations (GGUF loading and Qwen3.5 save_pretrained serialization), while the rejected pairs appropriately avoid overstating relatedness. The representative/canonical issue choice for issue:43452 is also reasonably supported by its broader, model-agnostic scope compared with the more specific Qwen2.5 follow-up."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 11,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:33666",
+        "issue:35707",
+        "issue:36032",
+        "issue:40444",
+        "issue:40990",
+        "issue:41553",
+        "issue:41720",
+        "issue:42175",
+        "issue:42796",
+        "issue:43296",
+        "issue:43317",
+        "issue:43381",
+        "issue:43421",
+        "issue:43425",
+        "issue:43482",
+        "issue:43525",
+        "issue:43531",
+        "issue:43582",
+        "issue:43611",
+        "issue:43632",
+        "issue:43746",
+        "issue:43792",
+        "issue:43818",
+        "issue:43856",
+        "issue:44079",
+        "issue:44291",
+        "issue:44403",
+        "issue:44462",
+        "issue:44479",
+        "issue:44534",
+        "issue:44568",
+        "issue:44589",
+        "issue:44756",
+        "issue:44811",
+        "issue:44829",
+        "issue:44877",
+        "issue:44991",
+        "issue:45081",
+        "issue:45325",
+        "issue:45479",
+        "issue:45538",
+        "issue:45542",
+        "issue:45663",
+        "issue:45812",
+        "issue:45923",
+        "issue:46315",
+        "issue:46402",
+        "issue:46491"
+      ],
+      "soft_pairs": [
+        "issue:44479|issue:45325",
+        "issue:43531|issue:44877",
+        "issue:36032|issue:41720",
+        "issue:44991|issue:46315",
+        "issue:43425|issue:44403",
+        "issue:43296|issue:43531",
+        "issue:43611|issue:44534",
+        "issue:43632|issue:44756",
+        "issue:40444|issue:45325",
+        "issue:45812|issue:46491",
+        "issue:43792|issue:44589",
+        "issue:42796|issue:43746",
+        "issue:44403|issue:44991",
+        "issue:33666|issue:40444",
+        "issue:43381|issue:45923",
+        "issue:44291|issue:44589",
+        "issue:43632|issue:44534",
+        "issue:44079|issue:45663",
+        "issue:40990|issue:44811",
+        "issue:43482|issue:45538",
+        "issue:43482|issue:43525",
+        "issue:35707|issue:40444",
+        "issue:43421|issue:44568",
+        "issue:43317|issue:43856",
+        "issue:42175|issue:45542",
+        "issue:43531|issue:44291",
+        "issue:41553|issue:43582",
+        "issue:43818|issue:46402",
+        "issue:44079|issue:44534",
+        "issue:44829|issue:45479",
+        "issue:41720|issue:45081",
+        "issue:44462|issue:45812"
+      ],
+      "prepared_review_unit_hash": "33059d63cd378d3a6659e0cecf4eb4174c8943103e2a7e84840676ad8fc54944",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37466,
+        "estimated_input_tokens": 9367,
+        "estimated_eval_tokens": 18990
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a heterogeneous mix of unrelated issues, with only a few weak thematic overlaps around tokenizer regressions or Transformers v5 behavior changes. None of the soft pairs look like the same underlying bug, so no merges should be made.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44479",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Both are vision-language regressions, but one is a broad video-input regression across several Qwen models while the other is a specific still-image rope-index scaling bug in Qwen2.5-VL."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior and GraniteSpeech strict-config loading are different code paths and different failure modes."
+          },
+          {
+            "left": "issue:36032",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Tokenizer method-name conflict vs CUDA assert during auto device mapping; no shared underlying bug."
+          },
+          {
+            "left": "issue:44991",
+            "right": "issue:46315",
+            "accept": false,
+            "reason": "Tokenizer loading failure for one repo/model versus a performance regression in slow tokenizer token-id conversion; unrelated fixes."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:44403",
+            "accept": false,
+            "reason": "Torch 2.10 compatibility break and loading noise are different issues."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "PaddleOCR-VL / vLLM load failure and Qwen3-MoE sliding_window are unrelated."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Base-model-prefix loading regression vs non-persistent buffer junk values; different bugs."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "_is_hf_initialized flag regression and Strix Halo mmap OOM are not the same code-path problem."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "IterableDataset multi-image finetuning failure is distinct from the Qwen2.5-VL still-image temporal position bug."
+          },
+          {
+            "left": "issue:45812",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but Granite token IDs and CodeLlama leading-space round-trip are separate tokenizer implementations and failure modes."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Whisper runtime failure and Float8 storage lookup error are unrelated."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Ministral BPE marker handling and GraniteSpeech PEFT checkpoint loading are different bugs."
+          },
+          {
+            "left": "issue:44403",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Generic loading noise vs tokenizer loading failure for EMBEDDIA/est-roberta; no duplicate evidence."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "Both involve Qwen VL training, but the former is multi-GPU training and the latter is multi-image IterableDataset failure; not the same defect."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Eval-mode gradient checkpointing restriction vs missing supports_gradient_checkpointing flag on a trust_remote_code model."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Unexpected _is_hf_initialized argument during init_empty_weights and missing Float8 storage object are distinct initialization/storage issues."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "Both are Transformers v5 regressions, but one is a flag initialization issue and the other is buffer serialization content; not mergeable into one fix."
+          },
+          {
+            "left": "issue:44079",
+            "right": "issue:45663",
+            "accept": false,
+            "reason": "ModelOutput key assignment bug and Gemma4 FSDP2 shared_kv_states rebuild bug are unrelated."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Perplexity anomaly on GPT-OSS and Whisper batch_decode skipping special tokens are different subsystems."
+          },
+          {
+            "left": "issue:43482",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Qwen2.5-GGUF loading in v5 and CLIPTokenizer model_max_length behavior are unrelated tokenizer/model-loading issues."
+          },
+          {
+            "left": "issue:43482",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Qwen2.5-GGUF loading failure and Llama4Config missing pad_token_id are different model/config regressions."
+          },
+          {
+            "left": "issue:35707",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "Progressive generation with inputs_embeds/past_key_values is unrelated to multi-image iterable dataset finetuning."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Runtime post-processor updates in TokenizersBackend and missing BOS/EOS addition in a specific tokenizer are separate tokenizer behaviors."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "Device-map dequantized-model offload failure and Qwen3 MoE training memory inefficiency are not the same bug."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Both are packaging/backend dependency problems, but one is missing TensorFlow in the torch extra and the other is tensorboard installed without TensorFlow; distinct fixes."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window issue and init_empty_weights argument regression do not share the same code path."
+          },
+          {
+            "left": "issue:41553",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "Bad AutoTokenizer error message for Voxtral and AppleSilicon caching_allocator_warmup TypeError are unrelated."
+          },
+          {
+            "left": "issue:43818",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Video-LLaVA temporal-attention/weight-sharing bug and Sam3Model text-encoder weight loading bug are unrelated model-specific regressions."
+          },
+          {
+            "left": "issue:44079",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "ModelOutput key assignment and non-persistent buffer junk are different internal model-state issues."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Degenerate training with flash_attention_3 and zero-loss single-label classification are both bad training outcomes but caused by different mechanisms."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "CUDA assert on Qwen3 auto device mapping and Mistral regex patch crash are unrelated."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Ignoring tokenizer.json and producing wrong Granite token IDs are both AutoTokenizer regressions, but they target different behavior and fixes."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 12,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:34634",
+        "issue:35707",
+        "issue:39692",
+        "issue:40444",
+        "issue:40990",
+        "issue:41720",
+        "issue:42175",
+        "issue:42796",
+        "issue:42915",
+        "issue:43262",
+        "issue:43296",
+        "issue:43377",
+        "issue:43454",
+        "issue:43482",
+        "issue:43550",
+        "issue:43606",
+        "issue:43611",
+        "issue:43632",
+        "issue:43644",
+        "issue:43646",
+        "issue:43701",
+        "issue:43746",
+        "issue:43749",
+        "issue:43792",
+        "issue:43818",
+        "issue:43819",
+        "issue:43873",
+        "issue:43881",
+        "issue:43957",
+        "issue:43994",
+        "issue:44079",
+        "issue:44265",
+        "issue:44291",
+        "issue:44466",
+        "issue:44479",
+        "issue:44488",
+        "issue:44877",
+        "issue:44898",
+        "issue:45072",
+        "issue:45325",
+        "issue:45357",
+        "issue:45381",
+        "issue:45491",
+        "issue:45584",
+        "issue:45701",
+        "issue:45800",
+        "issue:45823",
+        "issue:45834"
+      ],
+      "soft_pairs": [
+        "issue:43818|issue:45357",
+        "issue:43377|issue:45491",
+        "issue:43611|issue:43646",
+        "issue:43644|issue:44079",
+        "issue:44265|issue:45072",
+        "issue:42175|issue:43296",
+        "issue:43749|issue:44079",
+        "issue:40444|issue:45381",
+        "issue:43792|issue:43957",
+        "issue:45325|issue:45381",
+        "issue:43994|issue:44898",
+        "issue:45325|issue:45834",
+        "issue:43454|issue:44466",
+        "issue:42915|issue:43482",
+        "issue:43632|issue:45072",
+        "issue:42796|issue:45381",
+        "issue:44291|issue:44877",
+        "issue:43701|issue:45325",
+        "issue:43296|issue:45800",
+        "issue:44479|issue:45823",
+        "issue:43873|issue:45701",
+        "issue:43606|issue:44265",
+        "issue:43262|issue:43746",
+        "issue:40990|issue:45584",
+        "issue:34634|issue:35707",
+        "issue:35707|issue:41720",
+        "issue:43454|issue:43550",
+        "issue:39692|issue:40444",
+        "issue:43606|issue:43819",
+        "issue:43881|issue:44488",
+        "issue:41720|issue:45823"
+      ],
+      "prepared_review_unit_hash": "77a5be7a4e6932c058682d572ed2d791ad69524729bd4a4b6ecb4eb28e677b27",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 36975,
+        "estimated_input_tokens": 9244,
+        "estimated_eval_tokens": 18744
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a set of unrelated model- and regression-specific issues; only one soft pair looks plausibly like the same Qwen2.5-VL position-id bug. Most other pairs share broad subsystem names but not the same concrete failure mode.",
+        "confidence": 0.79,
+        "canonical_issue_reason": "issue:43611 is the broadest, most representative regression here: a concrete Transformers v5 loading/initialization breakage that could plausibly anchor nearby v5 model-loading reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43611 is the best global issue candidate because it describes a clear core loading regression with wider duplicate potential than the more model-specific reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43818",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different models and failures: Video-LLaVA temporal attention/weight sharing vs Qwen3.5 visual-encoder key serialization."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Both are batching/mask-related, but the concrete bugs differ: missing padding-mask support in MIMI vs NaNs from Gemma3 sliding-window all-padding handling."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43646",
+            "accept": false,
+            "reason": "Both are Transformers v5 regressions, but one is base_model_prefix loading and the other is custom model initialization; not the same code-path bug."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Non-persistent buffer serialization junk and ModelOutput key assignment are distinct internal state bugs."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Different failures: torch.export with torch_compilable_check vs bfloat16 dtype mismatch in inference."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43296",
+            "accept": false,
+            "reason": "Unrelated areas: backend packaging/deps vs audio processor sampling-rate defaulting."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "FSDP CPU RAM efficient loading and ModelOutput key handling are separate issues."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both involve Qwen VL, but one is multi-image training dataset handling and the other is video position-id computation."
+          },
+          {
+            "left": "issue:43792",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Whisper runtime failure is not the same as meta-device loading breakage."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45381",
+            "accept": true,
+            "reason": "Same likely Qwen2.5-VL multimodal position-id / rope-index regression in Transformers 5.3.0; one reports still-image temporal IDs and the other video vision_position_ids, which plausibly share the same code path."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Both are vision-model issues, but the concrete bugs and models differ substantially."
+          },
+          {
+            "left": "issue:45325",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Different models and symptoms: Qwen2.5-VL position IDs vs Kosmos2.5 OCR long-input indexing."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Both mention lm_head/tied weights, but one is missing weight tying causing bad generation and the other is device-dependent serialization inconsistency."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Different quantization/loading problems on different models."
+          },
+          {
+            "left": "issue:43632",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "_is_hf_initialized flag breakage and bfloat16 dtype mismatch are unrelated."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Tokenizer BPE marker handling and Qwen VL video position IDs are unrelated."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Unexpected _is_hf_initialized init arg and strict Granite Speech config loading are different v5 regressions."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Checkpoint resume key mismatch is not the same as Qwen2.5-VL position-id scaling."
+          },
+          {
+            "left": "issue:43296",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Different incompatibilities: PaddleOCR-VL loading in vLLM vs torch/transformers version mismatch."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Both are multimodal/device-related, but Qwen VL video regression and Gemma4 PLE device mismatch are different bugs."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Quantized offloading behavior and tokenization changes are unrelated."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44265",
+            "accept": false,
+            "reason": "CPU-offload device mismatch in Bark and torch.export failures with torch_compilable_check are not the same bug."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Audio chat-template sample-rate defaulting and GraniteSpeech checkpoint loading are separate issues."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Perplexity regression and Whisper empty-transcription failure do not share a concrete code path."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "BarkProcessor voice_preset and progressive generation with inputs_embeds/past_key_values are different features and failures."
+          },
+          {
+            "left": "issue:35707",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Progressive generation cache handling and Qwen3 device-map cudaErrorAssert are unrelated."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "lm_head weight tying and Bamba torch.compile SDPA failure are different issues."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:40444",
+            "accept": false,
+            "reason": "SigLIP2 docs/quantization errors and Qwen2.5-VL multi-image dataset failures are unrelated."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Bark CPU-offload mismatch and DAC from_latents/forward mismatch are different code paths."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Different model-loading failures on unrelated models."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Both involve device mapping, but the models and concrete failures differ enough that they are not duplicates."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary stays grounded in the packet and is appropriately conservative. The one accepted soft pair (issue:45325 / issue:45381) is plausibly the same Qwen2.5-VL position-id regression, and the rest are reasonably rejected as distinct model- or code-path-specific bugs. The broad canonical issue choice of issue:43611 is also well supported by the titles."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 13,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:26413",
+        "issue:28530",
+        "issue:36331",
+        "issue:40990",
+        "issue:41720",
+        "issue:42617",
+        "issue:42796",
+        "issue:42915",
+        "issue:42947",
+        "issue:43065",
+        "issue:43257",
+        "issue:43317",
+        "issue:43322",
+        "issue:43377",
+        "issue:43450",
+        "issue:43454",
+        "issue:43526",
+        "issue:43572",
+        "issue:43720",
+        "issue:43749",
+        "issue:43873",
+        "issue:43881",
+        "issue:43883",
+        "issue:44186",
+        "issue:44265",
+        "issue:44387",
+        "issue:44466",
+        "issue:44534",
+        "issue:44568",
+        "issue:44610",
+        "issue:44756",
+        "issue:44877",
+        "issue:44898",
+        "issue:44977",
+        "issue:44991",
+        "issue:45072",
+        "issue:45357",
+        "issue:45701",
+        "issue:45704",
+        "issue:45759",
+        "issue:45920",
+        "issue:45923",
+        "issue:46050",
+        "issue:46209",
+        "issue:46305",
+        "issue:46315",
+        "issue:46402"
+      ],
+      "soft_pairs": [
+        "issue:44186|issue:44265",
+        "issue:46305|issue:46315",
+        "issue:42796|issue:46209",
+        "issue:43377|issue:44610",
+        "issue:43454|issue:45072",
+        "issue:44568|issue:45920",
+        "issue:43257|issue:45923",
+        "issue:43377|issue:44898",
+        "issue:43317|issue:43873",
+        "issue:42796|issue:44977",
+        "issue:43454|issue:43720",
+        "issue:40990|issue:42947",
+        "issue:43572|issue:45759",
+        "issue:43322|issue:46050",
+        "issue:42617|issue:42915",
+        "issue:43749|issue:44534",
+        "issue:41720|issue:43322",
+        "issue:41720|issue:44977",
+        "issue:26413|issue:28530",
+        "issue:44466|issue:45357",
+        "issue:43526|issue:43883",
+        "issue:44387|issue:45704",
+        "issue:36331|issue:42617",
+        "issue:43881|issue:44991",
+        "issue:36331|issue:42915",
+        "issue:44877|issue:44991",
+        "issue:43450|issue:45701",
+        "issue:40990|issue:45701",
+        "issue:43065|issue:46402",
+        "issue:44186|issue:45072",
+        "issue:43873|issue:44756",
+        "issue:41720|issue:42617"
+      ],
+      "prepared_review_unit_hash": "09e854990c37b31f1b9c994fcea615df1e9fd31eb15f735436db02256509f304",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36762,
+        "estimated_input_tokens": 9191,
+        "estimated_eval_tokens": 18638
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is not a single duplicate cluster; it mixes many unrelated Transformers bugs and regressions across tokenization, loading/offloading, quantization, vision, and training utilities. None of the soft pairs look like the same underlying bug/change with enough specificity to merge.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No single issue is a safe canonical for the whole set because the reports span unrelated subsystems and failure modes. If forced to pick a representative, issue 45920 is the broadest v5 regression report, but it does not subsume the others.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44186",
+            "right": "issue:44265",
+            "accept": false,
+            "reason": "Both are crashes, but one is a LayoutLMv2 tokenizer/NER padding problem and the other is a torch.export/torch_compilable_check failure; different code paths and fixes."
+          },
+          {
+            "left": "issue:46305",
+            "right": "issue:46315",
+            "accept": false,
+            "reason": "MoonshineStreaming sliding-window mask off-by-one is unrelated to slow convert_tokens_to_ids in slow tokenizers."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Tokenizer BPE-marker handling and FP8 inference on B200 are unrelated bugs in different subsystems."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "MIMI encoder batch-vs-single output mismatch and OmDet-Turbo processor image-size mismatch are different model/component issues."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Tied lm_head weights serialization and bfloat16 dtype mismatches are unrelated failures."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45920",
+            "accept": false,
+            "reason": "Special-token insertion for one tokenizer is not the same concrete regression as broad wrong-token-id behavior across multiple tokenizers."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Qwen3 MoE weight conversion under accelerate+deepspeed and missing supports_gradient_checkpointing are different concerns."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Both involve batched inputs, but one is a MIMI encoder padding-mask bug and the other is a Perceiver image-size failure."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Device-map load failure with dequantized model offload is related to quantization/offloading broadly, but not the same concrete bug as the generic offloading report."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Tokenizer BPE-marker handling and Qwen3.5 flash-attention generation failure are unrelated."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "AyaVision tied-weight serialization and BitNet packed-weight unpacking during accelerate loading do not share the same underlying failure."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:42947",
+            "accept": false,
+            "reason": "High perplexity on a model/dataset and gradient checkpointing being ineffective are different bugs."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Missing pad_token_idx in StableLmConfig and unwrapping text_config for composite Qwen multimodal configs are distinct config-loading issues."
+          },
+          {
+            "left": "issue:43322",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Llava Next segmentation fault and sinq save/load breakage are different code paths."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "3d_parallel.py runtime failure and Qwen3Moe FineGrainedFP8Config failure are unrelated."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "FSDP CPU RAM efficient loading and non-persistent buffer junk initialization are different regressions."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43322",
+            "accept": false,
+            "reason": "Qwen3 auto device-map cudaErrorAssert and Llava Next loading segfault are separate model/loading failures."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Both mention Qwen, but one is an auto device-map load crash and the other is a flash-attention generation issue."
+          },
+          {
+            "left": "issue:26413",
+            "right": "issue:28530",
+            "accept": false,
+            "reason": "Resume-from-checkpoint epoch iterator exhaustion and early-stopping metric resolution are unrelated training-loop issues."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "lm_head.weight serialization depending on device and incorrect visual encoder keys are both save/load regressions, but not the same bug."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43883",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction and missing all_tied_weights_keys on Molmo are unrelated."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Int4 reserved-memory OOM and T5 apex FusedRMSNorm memory leak are different performance/memory problems."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "CustomTrainer compute_loss signature breakage and 3d_parallel.py execution failure are unrelated."
+          },
+          {
+            "left": "issue:43881",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "glm-4v-9b loading failure and tokenizer loading regression for est-roberta are different model/tokenizer problems."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:42915",
+            "accept": false,
+            "reason": "Unexpected num_items_in_batch in compute_loss and FineGrainedFP8Config failure do not share a code path."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Strict config rejection for granite_speech and tokenizer loading failure for est-roberta are distinct issues."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Batched video processor shape bug and version-dependent tokenization changes are unrelated."
+          },
+          {
+            "left": "issue:40990",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "High perplexity on one benchmark and tokenization changing across versions are not the same bug."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Sam3 pixel decoder dummy Conv2d and Sam3 text-encoder weights not loading affect different parts of the model."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Tokenizer crash on NER/batched truncation and bfloat16 dtype mismatch are separate failures."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Quantization offloading behavior and disabling mmap to avoid OOM are different memory-management issues."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "Qwen3 auto device-map crash and 3d_parallel.py 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": 21,
+      "review_unit_index": 14,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:28530",
+        "issue:30064",
+        "issue:33357",
+        "issue:33666",
+        "issue:34634",
+        "issue:35707",
+        "issue:36010",
+        "issue:36331",
+        "issue:41720",
+        "issue:42617",
+        "issue:42915",
+        "issue:43295",
+        "issue:43316",
+        "issue:43317",
+        "issue:43322",
+        "issue:43404",
+        "issue:43526",
+        "issue:43531",
+        "issue:43540",
+        "issue:43550",
+        "issue:43572",
+        "issue:43577",
+        "issue:43606",
+        "issue:43653",
+        "issue:43708",
+        "issue:43720",
+        "issue:43782",
+        "issue:43873",
+        "issue:43881",
+        "issue:44008",
+        "issue:44186",
+        "issue:44361",
+        "issue:44423",
+        "issue:44610",
+        "issue:44671",
+        "issue:44743",
+        "issue:44898",
+        "issue:44913",
+        "issue:45161",
+        "issue:45357",
+        "issue:45406",
+        "issue:45632",
+        "issue:45684",
+        "issue:45701",
+        "issue:45704",
+        "issue:45823",
+        "issue:46291"
+      ],
+      "soft_pairs": [
+        "issue:43577|issue:43708",
+        "issue:43550|issue:43720",
+        "issue:43404|issue:45357",
+        "issue:44008|issue:45823",
+        "issue:43550|issue:43606",
+        "issue:44186|issue:44610",
+        "issue:43295|issue:45701",
+        "issue:44610|issue:44898",
+        "issue:42915|issue:43317",
+        "issue:44423|issue:45406",
+        "issue:36010|issue:42617",
+        "issue:45632|issue:45684",
+        "issue:41720|issue:45161",
+        "issue:42915|issue:46291",
+        "issue:43531|issue:43572",
+        "issue:28530|issue:33666",
+        "issue:44671|issue:44743",
+        "issue:36331|issue:41720",
+        "issue:44186|issue:44361",
+        "issue:42915|issue:43322",
+        "issue:43526|issue:43881",
+        "issue:28530|issue:30064",
+        "issue:43316|issue:44913",
+        "issue:43540|issue:43782",
+        "issue:33357|issue:35707",
+        "issue:30064|issue:34634",
+        "issue:43873|issue:45704",
+        "issue:43550|issue:43653"
+      ],
+      "prepared_review_unit_hash": "64885f6f3d2d2e9f4904eee5e11addb4ae7c7be81964807d05149bac340144f4",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 28,
+        "serialized_chars": 35525,
+        "estimated_input_tokens": 8882,
+        "estimated_eval_tokens": 18020
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a loose embedding neighborhood of mostly unrelated Transformers issues: training loop regressions, tokenizer/processor bugs, multimodal/model-loading failures, and several device/offload/quantization problems. None of the soft pairs look like the same underlying bug or change.",
+        "confidence": 0.88,
+        "canonical_issue_reason": "If a single issue must anchor the cluster, issue 43873 is the broadest and most central runtime report (offloading/quantization) with multiple inbound references, but it is still not a duplicate of the other issues.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43873 is the best representative only in the sense that it is broad, active, and closer to the runtime/device-mapping subset than the other items; however, the cluster is too heterogeneous for a strong canonical issue.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43577",
+            "right": "issue:43708",
+            "accept": false,
+            "reason": "Different subsystems: model dtype propagation vs Trainer resume/max_steps accounting. No shared bug path."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Different failures: torch.compile/SDPA on Bamba vs packed-weight unpacking during accelerate loading for BitNet."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different model families and bugs: weight tying in Mistral3 vs incorrect visual encoder keys on save_pretrained for Qwen3.5."
+          },
+          {
+            "left": "issue:44008",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Both are Gemma-related but the symptoms differ: variable-name collision in forward vs device mismatch with device_map=auto."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43606",
+            "accept": false,
+            "reason": "Both mention device/runtime issues, but one is compile/SDPA on Bamba and the other is CPU offload mismatch on Bark; not the same bug."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Different components and root causes: LayoutLMv2Tokenizer NER/padding crash vs OmDet-Turbo processor size mismatch."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "One is a specific regression in processor.tokenizer/image-tokenizer handling; the other is a broad version-related tokenization complaint with no concrete shared code path."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Both are vision preprocessor issues, but one is fixed output size for OmDet-Turbo and the other is Perceiver interpolation handling; not the same defect."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43317",
+            "accept": false,
+            "reason": "Different problems: Qwen3Moe FineGrainedFP8Config failure vs device_map auto loading of a dequantized model."
+          },
+          {
+            "left": "issue:44423",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both are serving/multimodal processor crashes, but the concrete failures differ: str.to in continuous batching vs missing _tokenizer on Gemma4Processor."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42617",
+            "accept": false,
+            "reason": "Unrelated: GenerationMixin import path regression vs inability to run a 3d_parallel.py example/script."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:45684",
+            "accept": false,
+            "reason": "Different storage/copying bugs: trust_remote_code cache-path collision vs read-only permission propagation in save_pretrained/register_for_auto_class."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45161",
+            "accept": false,
+            "reason": "Different model/runtime issues: cudaErrorAssert on Qwen3 auto device mapping vs tensor-parallel not working for GPT-OSS MoE."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Completely different: Qwen3Moe FP8 model failure vs transformers.pipelines import failure caused by kernels version/revision requirements."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Different config bugs: Qwen3-MoE sliding_window behavior vs missing pad_token_idx in StableLmConfig after update."
+          },
+          {
+            "left": "issue:28530",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "Different domains: early stopping metric selection vs Qwen2-VL multi-GPU training."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Different models and symptoms: CamemBERT masked-LM predictions vs Qwen3.5 recurrent-state reset with cache."
+          },
+          {
+            "left": "issue:36331",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "Trainer API signature break vs Qwen3 device-mapping assertion; no shared code path."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Both are tokenizer/task crashes, but they involve different tokenizers and different failure mechanisms."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43322",
+            "accept": false,
+            "reason": "Different loading/runtime failures: FineGrainedFP8Config vs segmentation fault while loading Llava Next models."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "Different components: BeitImageProcessorFast label reduction bug vs glm-4v-9b model loading failure."
+          },
+          {
+            "left": "issue:28530",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "Different preprocessing/training issues: early stopping metric lookup vs void segmentation map processing."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Different config regression areas: Gemma3TextConfig API mismatch vs GPTNeoXConfig rotary_pct reload issue."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:43782",
+            "accept": false,
+            "reason": "Different multimodal model bugs: video input handling in Qwen3OmniMoe vs weight_only load error in Qwen3VL."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "Different problems: MacOS bus error loading a community CLIP model vs progressive generation with inputs_embeds/past_key_values."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:34634",
+            "accept": false,
+            "reason": "Different processor bugs: void segmentation maps vs Bark voice_preset support."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Both are runtime regressions, but one is quantization/offloading behavior and the other is a T5/apex RMSNorm memory leak; not the same bug."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Different model and bug class: torch.compile/SDPA on Bamba vs tokenizer special-token registration in BigBirdTokenizer."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 15,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:29159",
+        "issue:34567",
+        "issue:34634",
+        "issue:35707",
+        "issue:36032",
+        "issue:39692",
+        "issue:41720",
+        "issue:42490",
+        "issue:43329",
+        "issue:43381",
+        "issue:43388",
+        "issue:43421",
+        "issue:43531",
+        "issue:43606",
+        "issue:43653",
+        "issue:43720",
+        "issue:43746",
+        "issue:43844",
+        "issue:43927",
+        "issue:44062",
+        "issue:44077",
+        "issue:44368",
+        "issue:44387",
+        "issue:44451",
+        "issue:44492",
+        "issue:44509",
+        "issue:44514",
+        "issue:44545",
+        "issue:44589",
+        "issue:44610",
+        "issue:44625",
+        "issue:44898",
+        "issue:44918",
+        "issue:44936",
+        "issue:44987",
+        "issue:44991",
+        "issue:45042",
+        "issue:45059",
+        "issue:45072",
+        "issue:45357",
+        "issue:45406",
+        "issue:45464",
+        "issue:45479",
+        "issue:45588",
+        "issue:45701",
+        "issue:45759",
+        "issue:45859",
+        "issue:46402"
+      ],
+      "soft_pairs": [
+        "issue:29159|issue:36032",
+        "issue:43606|issue:43720",
+        "issue:44387|issue:45701",
+        "issue:43388|issue:45059",
+        "issue:44492|issue:44509",
+        "issue:41720|issue:43746",
+        "issue:43531|issue:45759",
+        "issue:45406|issue:45464",
+        "issue:42490|issue:43329",
+        "issue:45357|issue:46402",
+        "issue:43606|issue:43653",
+        "issue:34634|issue:41720",
+        "issue:34567|issue:35707",
+        "issue:43531|issue:44589",
+        "issue:44451|issue:44991",
+        "issue:45759|issue:45859",
+        "issue:44077|issue:44625",
+        "issue:44625|issue:45479",
+        "issue:45588|issue:45759",
+        "issue:44898|issue:45072",
+        "issue:44062|issue:44451",
+        "issue:44987|issue:45042",
+        "issue:44514|issue:44545",
+        "issue:39692|issue:44610",
+        "issue:44368|issue:45357",
+        "issue:44918|issue:44936",
+        "issue:43381|issue:43844",
+        "issue:43421|issue:43927"
+      ],
+      "prepared_review_unit_hash": "f71e2a47f9605ca52d9384d53b13b6d0d05264513d9c26ae554347cf9a3b3f35",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36482,
+        "estimated_input_tokens": 9121,
+        "estimated_eval_tokens": 18498
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a set of unrelated transformer/model bugs and docs issues; the only clear duplicate pair is the Qwen2_5_VLProcessor batched-input crash. No PRs are present.",
+        "confidence": 0.9,
+        "canonical_issue_reason": "issue:44545 is the best canonical issue because it is the clearer, more referenced instance of the exact duplicate pair with issue:44514 and describes the concrete crash succinctly.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44545 is the best representative issue overall since it is the only strong duplicate target in the set and has slightly more inbound reference evidence than its twin.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:29159",
+            "right": "issue:36032",
+            "accept": false,
+            "reason": "Both are tokenizer-related, but one is a slow/fast behavior inconsistency and the other is a T5 tokenizer load-time name conflict; different bugs and code paths."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Both involve model loading/offload failures, but Bark CPU offload device mismatch and BitNet packed-weight unpacking are distinct model-specific problems."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "One is an int4 quantization memory regression causing OOM; the other is a tokenization version-change report. Shared broad area only."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:45059",
+            "accept": false,
+            "reason": "Metric-gathering label truncation in a tuple is unrelated to SAM3 text/box odd behavior."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "A typo in cache strategy docs is not the same underlying issue as stale docs referencing removed pipeline tasks."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43746",
+            "accept": false,
+            "reason": "Different models and failures: Qwen3 device-mapping cuda assert versus GraniteSpeech PEFT local checkpoint loading."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior and Qwen multimodal config unwrapping are separate issues."
+          },
+          {
+            "left": "issue:45406",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "Gemma4Processor missing _tokenizer in serve is unrelated to Qwen3.5 streaming chat/completions failure."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Local-path slash import warning and multimodal token counting video-branch undefined names are unrelated."
+          },
+          {
+            "left": "issue:45357",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Both concern model weights/config for different families, but Qwen3.5 visual key saving and Sam3 text encoder loading are distinct bugs."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Bark CPU offload device mismatch is unrelated to BigBirdTokenizer mask token registration/decode behavior."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:41720",
+            "accept": false,
+            "reason": "BarkProcessor voice_preset failure and Qwen3 auto device mapping cuda assert are different models and failure modes."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:35707",
+            "accept": false,
+            "reason": "Trainer token-count tracking and progressive generation with inputs_embeds/past_key_values are not the same bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior is unrelated to a missing Float8 storage type."
+          },
+          {
+            "left": "issue:44451",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are version-dependent loading failures, but for different models and likely different tokenizer/config breakages."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Composite Qwen config unwrapping and Qwen3.5MoE rotary embedding CPU-offload compatibility are separate code paths."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44625",
+            "accept": false,
+            "reason": "PatchTSMixer post_init policy and Qwen3.5 num_labels propagation do not describe the same underlying change."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Config propagation for Qwen3.5 is different from the general single-label classification zero-loss bug."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Flash-attention s_aux handling for sink-less models is unrelated to Qwen multimodal config unwrapping."
+          },
+          {
+            "left": "issue:44898",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Perceiver interpolation at non-default resolution and bfloat16 dtype mismatches in other models are unrelated."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "AddedToken special-argument duplication is not the same as ScandiBERT loading failure."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "A model loading regression for one repo and a PIL backend torchvision requirement bug are different issues."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:44545",
+            "accept": true,
+            "reason": "Exact same title and described crash: Qwen2_5_VLProcessor.apply_chat_template fails on batched input with padding=False."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Both mention processor/model mismatches, but they involve different models and different concrete failures."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Tie_word_embeddings warning during LoRA fine-tuning is unrelated to incorrect visual encoder keys on save_pretrained."
+          },
+          {
+            "left": "issue:44918",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Unpacking Qwen3.5 input embeddings in TRL SFT trainer is not the same as trainer.evaluate() failing after trainer.train()."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "Gradient checkpointing eval-mode incompatibility and ZeRO-3 gradient growth are distinct training bugs."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Runtime post-processor updates for special tokens and DiaConfig save/load token-ID loss are separate tokenization/config issues."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: issue:44545 is the clearer duplicate target of issue:44514, and the inbound-reference comparison is accurate. The soft-edge rejections are conservative and consistent with the issue titles, and the note that no PRs are present is supported by the packet."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 16,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:30064",
+        "issue:36010",
+        "issue:39692",
+        "issue:41720",
+        "issue:42886",
+        "issue:43295",
+        "issue:43335",
+        "issue:43421",
+        "issue:43531",
+        "issue:43540",
+        "issue:43550",
+        "issue:43596",
+        "issue:43618",
+        "issue:43643",
+        "issue:43645",
+        "issue:43653",
+        "issue:43697",
+        "issue:43720",
+        "issue:43746",
+        "issue:43756",
+        "issue:43782",
+        "issue:43819",
+        "issue:43873",
+        "issue:43994",
+        "issue:44008",
+        "issue:44062",
+        "issue:44186",
+        "issue:44220",
+        "issue:44488",
+        "issue:44661",
+        "issue:44671",
+        "issue:44843",
+        "issue:44877",
+        "issue:44898",
+        "issue:45072",
+        "issue:45081",
+        "issue:45356",
+        "issue:45381",
+        "issue:45663",
+        "issue:45704",
+        "issue:45834",
+        "issue:45859",
+        "issue:45907",
+        "issue:45923",
+        "issue:46032",
+        "issue:46050",
+        "issue:46294",
+        "issue:46463"
+      ],
+      "soft_pairs": [
+        "issue:43746|issue:46050",
+        "issue:45704|issue:45923",
+        "issue:41720|issue:45704",
+        "issue:43531|issue:43645",
+        "issue:45704|issue:45859",
+        "issue:42886|issue:44843",
+        "issue:43782|issue:45381",
+        "issue:46032|issue:46463",
+        "issue:43421|issue:43653",
+        "issue:45663|issue:45923",
+        "issue:43994|issue:44671",
+        "issue:39692|issue:43873",
+        "issue:43746|issue:43873",
+        "issue:43596|issue:44220",
+        "issue:43421|issue:44186",
+        "issue:30064|issue:36010",
+        "issue:45081|issue:45356",
+        "issue:43720|issue:43819",
+        "issue:45356|issue:45907",
+        "issue:43653|issue:43720",
+        "issue:44220|issue:44661",
+        "issue:43540|issue:46294",
+        "issue:43295|issue:45834",
+        "issue:43550|issue:45072",
+        "issue:43618|issue:43697",
+        "issue:44062|issue:44488",
+        "issue:43550|issue:43819",
+        "issue:43531|issue:43643",
+        "issue:44186|issue:44898",
+        "issue:43653|issue:43819",
+        "issue:43540|issue:44008",
+        "issue:43643|issue:44877",
+        "issue:43335|issue:43756"
+      ],
+      "prepared_review_unit_hash": "66a3accc732090f24a7234524d1afbc51e2cae195de3f950beaa875ec4262060",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 37991,
+        "estimated_input_tokens": 9498,
+        "estimated_eval_tokens": 19252
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly a grab bag of unrelated Transformers regressions spanning tokenizers, config/loading, vision/video models, quantization, and offload behavior. Only the Mistral regex/tokenizer-loading pair looks like a plausible same-bug duplicate; the rest are distinct code paths.",
+        "confidence": 0.69,
+        "canonical_issue_reason": "issue:45081 is the clearest concrete bug report in the only near-duplicate area, centered on `_patch_mistral_regex` / `fix_mistral_regex` tokenizer loading failures.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44843 is the broadest representative of the cluster\u2019s recurring tokenizer/offline-loading regressions, so it is the best overall issue anchor.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43746",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Both are model-loading problems, but one is PEFT/local checkpoint loading and the other is quantized save/load corruption; different bugs."
+          },
+          {
+            "left": "issue:45704",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Unrelated: T5 RMSNorm memory leak versus a missing gradient-checkpointing flag on a trust_remote_code model."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Different subsystems and failures: Qwen3 device-mapping CUDA assert vs T5 RMSNorm memory leak."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "No shared code path: Qwen3-MoE sliding window bug vs custom model initialization in Jupyter notebooks."
+          },
+          {
+            "left": "issue:45704",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Different models and failures: T5 memory leak vs Qwen3_5Moe CPU-offload incompatibility."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:44843",
+            "accept": false,
+            "reason": "Both mention offline/tokenizer loading, but 42886 is generic cache loading while 44843 is a specific `_patch_mistral_regex`/model_info regression; not enough evidence they are the same bug."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both are VL/video-related, but one is a weight_only load error and the other is a video position-id mismatch; different fixes."
+          },
+          {
+            "left": "issue:46032",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Completely different: Mamba2 cache correctness vs T5 decoder-input error messaging."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Both involve special tokens, but one is runtime post-processor updates and the other is BigBird mask-token registration; not the same underlying bug."
+          },
+          {
+            "left": "issue:45663",
+            "right": "issue:45923",
+            "accept": false,
+            "reason": "Different areas: Gemma4 FSDP2 shared_kv_states bug vs missing supports_gradient_checkpointing on a remote-code model."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Both are output-quality regressions, but they affect different models and likely different root causes."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Docs/example mismatch plus quantization failure vs general offloading-with-quantization behavior; not the same bug."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Different mechanisms: PEFT local checkpoint loading vs quantization offloading behavior."
+          },
+          {
+            "left": "issue:43596",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "No overlap in functionality: DeepSpeed zero3/BertModel indexing vs audio fbank feature extraction."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44186",
+            "accept": false,
+            "reason": "Related only at a high level to tokenizers; runtime special-token updates and LayoutLMv2 NER/padding crashes are distinct."
+          },
+          {
+            "left": "issue:30064",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Image processor segmentation-map handling and a GenerationMixin import error are unrelated."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45356",
+            "accept": true,
+            "reason": "Both are about the same `_patch_mistral_regex` / `fix_mistral_regex` tokenizer path and its failures, so one PR could plausibly address both."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Different models and failures: BitNet packed-weight loading vs DAC latent reconstruction mismatch."
+          },
+          {
+            "left": "issue:45356",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Tokenizer regex/codec regression is unrelated to list_repo_files missing revision forwarding."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Special-token registration bug vs BitNet packed-weight loading bug; no shared root cause."
+          },
+          {
+            "left": "issue:44220",
+            "right": "issue:44661",
+            "accept": false,
+            "reason": "Audio feature extraction issue vs add-new-model-like / TOKENIZER_MAPPING_NAMES failure; different code paths."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:46294",
+            "accept": false,
+            "reason": "Both are video-related, but they describe different Qwen models and different failures (processing ValueError vs chat-template token/feature mismatch)."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Processor.tokenizer regression with images vs Kosmos2.5 long-OCR indexing error; unrelated."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Different areas and symptoms: torch.compile/SDPA on Bamba vs bfloat16 dtype mismatches in other models."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "CLIPOutput attentions regression and RTDetr output drift are different model-specific bugs."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "AddedToken keyword-collision error vs a separate tokenizer-loading issue; no clear same root cause."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Different model/runtime issues: Bamba torch.compile SDPA failure vs DAC latent mismatch."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding-window bug and AutoConfig missing-fields regression are unrelated."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Different models and code paths: LayoutLMv2 tokenizer crashes vs Perceiver image-classification resizing bug."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Tokenizer special-token registration vs DAC latent mismatch; unrelated."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44008",
+            "accept": false,
+            "reason": "Video-input failure in Qwen3OmniMoe vs Gemma3n variable-name collision; different bugs."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Both involve config loading, but one is missing fields in trust_remote_code AutoConfig and the other is strict config rejection for granite_speech; not the same bug."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43756",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer creation and Smollm3 RoPE-layer omission are unrelated model-configuration bugs."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative about the many unrelated issues. The only accepted soft-edge pair, issue:45081 and issue:45356, is plausibly a single Mistral/fix_mistral_regex tokenizer-regression fix. I don\u2019t see evidence of overclaiming."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 17,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:30064",
+        "issue:34634",
+        "issue:36010",
+        "issue:36331",
+        "issue:39692",
+        "issue:41720",
+        "issue:43322",
+        "issue:43388",
+        "issue:43421",
+        "issue:43540",
+        "issue:43577",
+        "issue:43582",
+        "issue:43606",
+        "issue:43644",
+        "issue:43716",
+        "issue:43749",
+        "issue:43782",
+        "issue:43818",
+        "issue:43825",
+        "issue:43950",
+        "issue:43994",
+        "issue:44008",
+        "issue:44186",
+        "issue:44442",
+        "issue:44448",
+        "issue:44479",
+        "issue:44485",
+        "issue:44509",
+        "issue:44671",
+        "issue:44805",
+        "issue:44811",
+        "issue:44821",
+        "issue:44843",
+        "issue:44936",
+        "issue:44977",
+        "issue:44991",
+        "issue:45072",
+        "issue:45081",
+        "issue:45084",
+        "issue:45183",
+        "issue:45381",
+        "issue:45479",
+        "issue:45584",
+        "issue:45685",
+        "issue:45834",
+        "issue:45987",
+        "issue:46032",
+        "issue:46512"
+      ],
+      "soft_pairs": [
+        "issue:30064|issue:36331",
+        "issue:41720|issue:44991",
+        "issue:44821|issue:45084",
+        "issue:44843|issue:45081",
+        "issue:43322|issue:45685",
+        "issue:39692|issue:45834",
+        "issue:43716|issue:44008",
+        "issue:43388|issue:46032",
+        "issue:43582|issue:46512",
+        "issue:44448|issue:44671",
+        "issue:45183|issue:45479",
+        "issue:44186|issue:44442",
+        "issue:43950|issue:45987",
+        "issue:45584|issue:45834",
+        "issue:43644|issue:43749",
+        "issue:44485|issue:44509",
+        "issue:34634|issue:36010",
+        "issue:43606|issue:45072",
+        "issue:44479|issue:45834",
+        "issue:43577|issue:43818",
+        "issue:44805|issue:44936",
+        "issue:43540|issue:45381",
+        "issue:43782|issue:44977",
+        "issue:43825|issue:43994",
+        "issue:43421|issue:44811"
+      ],
+      "prepared_review_unit_hash": "71d3178f7ac12eeb99da827b888c3818f5f775acfafc7a99a5f38c2429e67d1f",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 25,
+        "serialized_chars": 35560,
+        "estimated_input_tokens": 8890,
+        "estimated_eval_tokens": 18036
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous: the items cover unrelated bugs across different models, tokenizers, loading paths, and trainer utilities. No soft-edge pair looks like the same underlying bug/change, so I would not merge any of them as duplicates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43950 is the clearest standalone regression report: it names the exact broken behavior (`from_pretrained()` corrupting non-persistent buffers in v5), is self-contained, and is more actionable than the other mixed-symptom reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:30064",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Different subsystems and failure modes: image segmentation map processing vs trainer loss signature mismatch."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Unrelated models and bugs: Qwen3 auto device mapping CUDA assert vs tokenizer loading failure for a RoBERTa model."
+          },
+          {
+            "left": "issue:44821",
+            "right": "issue:45084",
+            "accept": false,
+            "reason": "Different code paths: remote image processor loading vs template compilation error."
+          },
+          {
+            "left": "issue:44843",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Both touch `_patch_mistral_regex`, but the bugs differ: offline `model_info()` call vs missing `backend_tokenizer` attribute crash."
+          },
+          {
+            "left": "issue:43322",
+            "right": "issue:45685",
+            "accept": false,
+            "reason": "Segfault while loading Llava Next is unrelated to an MPS histogram/int type error in MOE code."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Different models and symptoms: SigLIP2 doc example errors vs Kosmos2.5 long-OCR index error."
+          },
+          {
+            "left": "issue:43716",
+            "right": "issue:44008",
+            "accept": false,
+            "reason": "Different model families and root causes: dtype mismatch in Mistral-3 vs variable name collision in Gemma 3n."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:46032",
+            "accept": false,
+            "reason": "Trainer metric-gathering label truncation is unrelated to Mamba2 cache correctness."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "Apple Silicon warmup TypeError is unrelated to stale `config.dtype` after `from_config()`."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44671",
+            "accept": false,
+            "reason": "Both are v5 regressions, but they affect different models and likely different code paths."
+          },
+          {
+            "left": "issue:45183",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Input validation errors are unrelated to the zero-loss classification issue."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "Tokenizer padding/truncation crash is unrelated to AutoTokenizer failing to load a specific tokenizer class."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:45987",
+            "accept": false,
+            "reason": "Both are regression-style reports, but one is about corrupted buffers and the other about a return-type contract in cache length."
+          },
+          {
+            "left": "issue:45584",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Whisper empty-transcription generation bug is unrelated to Kosmos2.5 OCR index errors."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "General loading-buffer corruption and FSDP CPU RAM efficient loading are not the same concrete bug."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:44509",
+            "accept": false,
+            "reason": "GLM-5 RoPE implementation discussion is unrelated to documentation cleanup for removed pipeline tasks."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "BarkProcessor voice preset handling is unrelated to the `GenerationMixin` import error."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "CPU offload device mismatch in Bark is unrelated to dtype mismatches in SwitchTransformers/TimmWrapperModel."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Qwen video-input regression is unrelated to Kosmos2.5 long OCR indexing."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43818",
+            "accept": false,
+            "reason": "BLIP2 dtype propagation bug is unrelated to Video-LLaVA missing temporal attention / shared weights."
+          },
+          {
+            "left": "issue:44805",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Mask shape mismatch in indexing is unrelated to trainer evaluate-after-train behavior."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both involve Qwen video inputs, but the concrete failures differ too much to treat as the same bug."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Qwen3VL weight-only loading error is unrelated to flash-attention generation failure in Qwen3.5."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Pipeline task documentation/message mismatch is unrelated to SigLIP2 output correctness."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Runtime post-processor update for special tokens is unrelated to Whisper batch_decode ignoring skip_special_tokens."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 18,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:28218",
+        "issue:29159",
+        "issue:34634",
+        "issue:36331",
+        "issue:42175",
+        "issue:43262",
+        "issue:43526",
+        "issue:43577",
+        "issue:43582",
+        "issue:43653",
+        "issue:43697",
+        "issue:43720",
+        "issue:43742",
+        "issue:43749",
+        "issue:43782",
+        "issue:43819",
+        "issue:43825",
+        "issue:43827",
+        "issue:43906",
+        "issue:44060",
+        "issue:44291",
+        "issue:44368",
+        "issue:44561",
+        "issue:44589",
+        "issue:44610",
+        "issue:44671",
+        "issue:44857",
+        "issue:44912",
+        "issue:44991",
+        "issue:45003",
+        "issue:45020",
+        "issue:45042",
+        "issue:45072",
+        "issue:45081",
+        "issue:45084",
+        "issue:45325",
+        "issue:45372",
+        "issue:45406",
+        "issue:45538",
+        "issue:45701",
+        "issue:45987",
+        "issue:46192",
+        "issue:46209",
+        "issue:46285",
+        "issue:46291",
+        "issue:46294",
+        "issue:46326",
+        "issue:46463"
+      ],
+      "soft_pairs": [
+        "issue:44991|issue:46291",
+        "issue:46294|issue:46463",
+        "issue:44561|issue:45020",
+        "issue:43577|issue:43582",
+        "issue:44610|issue:45072",
+        "issue:43720|issue:45072",
+        "issue:28218|issue:29159",
+        "issue:45987|issue:46192",
+        "issue:43825|issue:43827",
+        "issue:43749|issue:44291",
+        "issue:44671|issue:45701",
+        "issue:42175|issue:45081",
+        "issue:43582|issue:44857",
+        "issue:43782|issue:45325",
+        "issue:34634|issue:36331",
+        "issue:45003|issue:46285",
+        "issue:43262|issue:43819",
+        "issue:43653|issue:44912",
+        "issue:43526|issue:43697",
+        "issue:43742|issue:45020",
+        "issue:45081|issue:45406",
+        "issue:45084|issue:45538",
+        "issue:44589|issue:45042",
+        "issue:43906|issue:46209",
+        "issue:45372|issue:46326",
+        "issue:44060|issue:44368"
+      ],
+      "prepared_review_unit_hash": "7ed7effcfbf9f7173b550e9b29dcb3e6f589fbcdcadc27176afb3fe0b58f4a3b",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 26,
+        "serialized_chars": 35797,
+        "estimated_input_tokens": 8950,
+        "estimated_eval_tokens": 18156
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items mostly describe distinct bugs across tokenization, loading, multimodal processors, dtype/quantization, and v5 regressions. None of the soft pairs look like the same concrete underlying issue strongly enough to merge.",
+        "confidence": 0.77,
+        "canonical_issue_reason": "issue:28218 is the most concrete and reproducible bug report in the set, with a specific tokenizer symptom and enough discussion to serve as a representative issue if one must be chosen.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:28218 is the strongest single issue candidate because it states a precise user-visible defect, making it more actionable than broader umbrella reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44991",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Different failures: tokenizer loading for a specific model vs pipeline import breaking on kernels' new version/revision requirement."
+          },
+          {
+            "left": "issue:46294",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Unrelated domains and symptoms: Qwen video chat template mismatch vs T5 decoder-input validation error message."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Both are remote-code/v5 regressions, but they point to different concrete breakages and code paths."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "Different subsystems: Blip2 dtype loading issue vs Apple Silicon allocator warmup TypeError."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Processor image-size mismatch vs bfloat16 dtype mismatch are unrelated bugs."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "BitNet packed-weight loading failure is a different code path from inference dtype mismatch issues."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:29159",
+            "accept": false,
+            "reason": "Related tokenizer area, but one is an added-token spacing bug and the other is a broader slow/fast inconsistency report; not clearly the same defect."
+          },
+          {
+            "left": "issue:45987",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Seq-length return-type bug vs return_tensors being ignored are separate API issues."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "One is an error message regression, the other is stale documentation after pipeline removals."
+          },
+          {
+            "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 handling."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "CamemBERT masked-LM prediction regression is not the same as a general tokenization-change report."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "TensorFlow backend packaging and Mistral regex patching are unrelated."
+          },
+          {
+            "left": "issue:43582",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Apple Silicon allocator warmup TypeError does not match the CUDA AMP image-loss crash."
+          },
+          {
+            "left": "issue:43782",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Different Qwen-VL problems: weight_only loading error vs rope/index scaling bug."
+          },
+          {
+            "left": "issue:34634",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "BarkProcessor voice preset handling and CustomTrainer compute_loss signature are unrelated."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "Unsafe sys.modules access and CLIP model structure inconsistency are different defects."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43819",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate defaulting and DAC latent reconstruction mismatch are unrelated."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "BigBird tokenizer special-token registration and MXFP4 quantization fallback are different bugs."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "BEiT fast processor label reduction and RTDetrV2 output differences are not the same underlying issue."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "KeyError while loading one model is distinct from generic remote_code breakage."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Mistral regex patch crash and Gemma4Processor missing _tokenizer are different attribute errors in different components."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Template compilation error and CLIPTokenizer max_length issue are unrelated."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Float8 storage lookup failure is unrelated to PIL backend unexpectedly requiring torchvision."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Isolated reproduction of an earlier issue and FP8 inference failure on B200 are different problems."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Both involve mistral_common, but one is an import error for Gemma4 processor loading and the other is a continue_final_message tokenizer behavior bug."
+          },
+          {
+            "left": "issue:44060",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Both mention Qwen tied-weights warnings, but they describe different warnings and 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": 21,
+      "review_unit_index": 19,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:36010",
+        "issue:36331",
+        "issue:38175",
+        "issue:39692",
+        "issue:42175",
+        "issue:42490",
+        "issue:42915",
+        "issue:43257",
+        "issue:43262",
+        "issue:43329",
+        "issue:43450",
+        "issue:43452",
+        "issue:43454",
+        "issue:43526",
+        "issue:43531",
+        "issue:43550",
+        "issue:43577",
+        "issue:43638",
+        "issue:43643",
+        "issue:43645",
+        "issue:43650",
+        "issue:43696",
+        "issue:43720",
+        "issue:43742",
+        "issue:43746",
+        "issue:43819",
+        "issue:43854",
+        "issue:43906",
+        "issue:43927",
+        "issue:43931",
+        "issue:44387",
+        "issue:44479",
+        "issue:44554",
+        "issue:44610",
+        "issue:44617",
+        "issue:44811",
+        "issue:44912",
+        "issue:44936",
+        "issue:44977",
+        "issue:44991",
+        "issue:45042",
+        "issue:45459",
+        "issue:45600",
+        "issue:45759",
+        "issue:45803",
+        "issue:46291",
+        "issue:46421",
+        "issue:46424"
+      ],
+      "soft_pairs": [
+        "issue:43329|issue:45600",
+        "issue:43526|issue:44554",
+        "issue:43550|issue:44912",
+        "issue:43638|issue:44936",
+        "issue:43452|issue:44991",
+        "issue:36010|issue:43645",
+        "issue:36010|issue:36331",
+        "issue:43819|issue:43927",
+        "issue:43257|issue:43329",
+        "issue:42175|issue:46291",
+        "issue:43531|issue:45042",
+        "issue:43720|issue:44610",
+        "issue:38175|issue:44610",
+        "issue:44387|issue:44617",
+        "issue:43746|issue:44811",
+        "issue:39692|issue:44912",
+        "issue:43854|issue:44977",
+        "issue:42490|issue:43257",
+        "issue:43643|issue:45759",
+        "issue:45459|issue:45803",
+        "issue:43650|issue:43906",
+        "issue:43262|issue:44811",
+        "issue:46421|issue:46424",
+        "issue:43329|issue:44479",
+        "issue:43257|issue:43696",
+        "issue:43329|issue:43450",
+        "issue:43257|issue:43931",
+        "issue:43454|issue:43577",
+        "issue:42915|issue:43742"
+      ],
+      "prepared_review_unit_hash": "4505e44028e4efa973d9686a977df5ac7933c72dc083e54dd6d138f06487eb61",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36624,
+        "estimated_input_tokens": 9156,
+        "estimated_eval_tokens": 18568
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly loose semantic similarity across unrelated Transformers issues, not true duplicates. The closest theme is the multimodal/video/Qwen subset, but even there the reports describe different failure modes and code paths, so I would not merge any pair.",
+        "confidence": 0.89,
+        "canonical_issue_reason": "issue:43329 is the clearest, most concrete bug report in the loose multimodal/video subset, with an explicit code-path description and likely duplicate-attracting symptoms; the rest are largely orthogonal.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43329 is the best representative issue here because it is specific, actionable, and tied to a clear failing branch, whereas most other items are either broader, packaging-related, or unrelated.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43329",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Video token-counting bug vs stale auto_mappings config references; different code paths and symptoms."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "Beit label reduction bug is unrelated to the MPS attention correctness issue."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "torch.compile SDPA failure in Bamba is unrelated to MXFP4 quantization fallback in git-oss."
+          },
+          {
+            "left": "issue:43638",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Deepspeed/ZeRO3 Bert index error vs trainer evaluate-after-train behavior; not the same bug."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "gguf_file loading breakage is a different loader issue than tokenizer loading for EMBEDDIA/est-roberta."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:43645",
+            "accept": false,
+            "reason": "GenerationMixin import error is unrelated to Jupyter custom model initialization in Transformers 5.0."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "ImportError for GenerationMixin is a different regression than CustomTrainer.compute_loss signature mismatch."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "DAC forward/from_latents mismatch and DiaConfig token-id persistence are unrelated model bugs."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Qwen3 MoE accelerate+deepspeed loading differs from a multimodal video token counting bug."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Packaging/backend installation issue is unrelated to pipelines import failing on kernels version/revision requirements."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window behavior and PIL image processor torchvision dependency are unrelated."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "BitNet packed-weights loading bug is not the same as OmDet processor/model input-size mismatch."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "SigLIP2 zero probabilities and OmDet 640px-vs-224px are different model/processor failures."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Quantization memory regression and Sam3Video OOM are both OOM-adjacent but not the same issue."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "PEFT local-checkpoint loading and Whisper batch_decode token handling are unrelated."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "Documentation example errors are not the same as a runtime quantization-loading fallback bug."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Model-load unit-test failure for GLM-4.7-Flash is different from Qwen3.5 flash-attention generation issues."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43257",
+            "accept": false,
+            "reason": "Trailing-slash local model warning is unrelated to Qwen3 MoE accelerate/deepspeed weight conversion."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Missing fields with trust_remote_code and AutoModelForCausalLM.from_config text_config unwrapping are different config-loading bugs."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Hidden tokenizer protobuf errors and bare except swallowing errors are related in spirit but not the same failure."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "The first item is not a substantive bug report; the second is a reproduction of another issue."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate default and Whisper skip_special_tokens handling are distinct processor behaviors."
+          },
+          {
+            "left": "issue:46421",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Both concern cache behavior, but one is lazy-init recompilation and the other is max_cache_length honoring; different fixes."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Both touch video/multimodal paths, but one is undefined helper usage in token counting while the other is a broader v5.3.0 video regression."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43696",
+            "accept": false,
+            "reason": "Qwen3 MoE loading/weight conversion bug is unrelated to GPT-oss-20b CUDA OOM."
+          },
+          {
+            "left": "issue:43329",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Undefined helper in video token counting is different from batched video processor shape output."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Qwen3 MoE loading under accelerate+deepspeed is not the same as Qwen3-VL weight-shape mismatch."
+          },
+          {
+            "left": "issue:43454",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "AyaVision lm_head tying bug and Blip2 dtype propagation bug are separate model-loading issues."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "FineGrainedFP8Config failure for Qwen3Moe is unrelated to MobileLLM key errors during loading."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 20,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:28282",
+        "issue:29127",
+        "issue:29159",
+        "issue:33357",
+        "issue:34567",
+        "issue:34689",
+        "issue:40990",
+        "issue:41720",
+        "issue:42175",
+        "issue:42915",
+        "issue:43317",
+        "issue:43404",
+        "issue:43425",
+        "issue:43452",
+        "issue:43525",
+        "issue:43540",
+        "issue:43577",
+        "issue:43582",
+        "issue:43606",
+        "issue:43697",
+        "issue:43742",
+        "issue:43749",
+        "issue:43824",
+        "issue:43927",
+        "issue:44062",
+        "issue:44291",
+        "issue:44479",
+        "issue:44488",
+        "issue:44560",
+        "issue:44568",
+        "issue:44756",
+        "issue:44821",
+        "issue:44849",
+        "issue:44912",
+        "issue:44987",
+        "issue:44991",
+        "issue:45042",
+        "issue:45127",
+        "issue:45137",
+        "issue:45356",
+        "issue:45357",
+        "issue:45381",
+        "issue:45405",
+        "issue:45538",
+        "issue:45701",
+        "issue:45859",
+        "issue:46291",
+        "issue:46512"
+      ],
+      "soft_pairs": [
+        "issue:43749|issue:44756",
+        "issue:43606|issue:45859",
+        "issue:44291|issue:45042",
+        "issue:42915|issue:44821",
+        "issue:43525|issue:46291",
+        "issue:34567|issue:40990",
+        "issue:29159|issue:34567",
+        "issue:43927|issue:44912",
+        "issue:44987|issue:45538",
+        "issue:44568|issue:45356",
+        "issue:42915|issue:45538",
+        "issue:43404|issue:45127",
+        "issue:43425|issue:46291",
+        "issue:43697|issue:45357",
+        "issue:43577|issue:46512",
+        "issue:42175|issue:43824",
+        "issue:45137|issue:45405",
+        "issue:28282|issue:34689",
+        "issue:43452|issue:43582",
+        "issue:43317|issue:45538",
+        "issue:33357|issue:34567",
+        "issue:44849|issue:45701",
+        "issue:43742|issue:44291",
+        "issue:44062|issue:44991",
+        "issue:28282|issue:29127",
+        "issue:41720|issue:43540",
+        "issue:44479|issue:45381",
+        "issue:43540|issue:44560",
+        "issue:43540|issue:44479",
+        "issue:42915|issue:43749",
+        "issue:42175|issue:44488"
+      ],
+      "prepared_review_unit_hash": "4a403f577ff54ee1f1a315c049b2604bb8e06945e6b4beee2284515d0b47c632",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 36989,
+        "estimated_input_tokens": 9248,
+        "estimated_eval_tokens": 18752
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a grab-bag of Transformers issues around loading, tokenization, quantization, offload, and model-specific regressions. The soft pairs mostly share broad subsystem keywords, but not the same concrete bug or fix path, so none are strong duplicates.",
+        "confidence": 0.72,
+        "canonical_issue_reason": "issue:45701 is the best umbrella representative because it is open, user-facing, and explicitly about version-dependent tokenization changes, which is the broadest theme present in the cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45701 is the most suitable issue to keep as the cluster representative, but only as a broad theme anchor; it is not a duplicate of the other issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43749",
+            "right": "issue:44756",
+            "accept": false,
+            "reason": "Different bugs: FSDP CPU RAM efficient loading vs mmap/OOM behavior on Strix Halo. Same general memory area, but not the same failure path."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "CPU offload is shared, but one is a bark-small device mismatch and the other is a Qwen3 rotary embedding incompatibility. Different model/code-path problems."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Both are loading/initialization regressions, but one is init_empty_weights/_is_hf_initialized and the other is PIL image processors incorrectly requiring torchvision. Not the same bug."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Qwen3Moe FineGrainedFP8Config failure is unrelated to AutoImageProcessor URL loading. Different features and code paths."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Llama4Config pad_token_id AttributeError is model config loading, while the other is pipelines import failing due to kernels version/revision requirements. Not the same issue."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:40990",
+            "accept": false,
+            "reason": "TrainerState token accounting and GPT-OSS perplexity are unrelated. One is training metrics, the other is evaluation quality."
+          },
+          {
+            "left": "issue:29159",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "Tokenizer slow/fast inconsistency is a tokenizer behavior bug; num_input_tokens_seen not updating is a TrainerState accounting bug. Different components."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:44912",
+            "accept": false,
+            "reason": "DiaConfig token IDs not persisting and gpt-oss MXFP4 quantization fallback are different model/config issues. No shared concrete code path."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "Loading a specific model failing on 5.1.0 is not the same as CLIPTokenizer's model_max_length being 10**30. Different symptom and fix surface."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "mdeberta add_special_tokens regression and Kimi-K2.5 codec/fix_mistral_regex regression are both tokenizer bugs, but they affect different tokenizers and mechanisms."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "FineGrainedFP8Config model failure is unrelated to CLIPTokenizer max_length behavior. Different bug families."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45127",
+            "accept": false,
+            "reason": "lm_head weight tying in Mistral3ForConditionalGeneration is distinct from LoRA merge collapse with extended vocabulary. Similar theme, different underlying defect."
+          },
+          {
+            "left": "issue:43425",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "Torch 2.10 compatibility and kernels import dependency errors are separate environment/dependency issues, not the same bug."
+          },
+          {
+            "left": "issue:43697",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "RTDetrV2 output differences in v5 and Qwen3.5 save_pretrained visual encoder key regression are model-specific regressions in different code paths."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "Blip2 dtype propagation bug is about loaded module dtypes, while from_config(dtype=...) leaving config.dtype stale is a different state-sync issue."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Tensorflow backend missing from pip install '.[torch]' is packaging, while Qwen2_5_VL import failure is an API/export problem. Not the same underlying bug."
+          },
+          {
+            "left": "issue:45137",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO3 deque error and MIN_PEFT_VERSION bump are unrelated. One is runtime failure, the other is dependency version management."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "AutoModel PyTorch ImportError and Llama 3.2 Vision loading breakage both mention loading, but they are different model/backend failures."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "gguf_file breaking AutoTokenizer/AutoModel loading and caching_allocator_warmup TypeError on Apple Silicon are unrelated failures."
+          },
+          {
+            "left": "issue:43317",
+            "right": "issue:45538",
+            "accept": false,
+            "reason": "device_map=auto dequantized model offload failure is unrelated to CLIPTokenizer model_max_length behavior."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "MacOS bus error on community CLIP model and TrainerState token counter not updating are not the same bug; only a loose version of 'trainer/model issue' overlap."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Qwen3.5 output_hidden_states regression and general tokenization changes are different model/output paths. Not a duplicate pair."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Key error loading MobileLLM-125M and init_empty_weights unexpected argument are both load-time errors, but for different root causes and models."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "AddedToken special-argument TypeError is a tokenizer construction bug; est-roberta tokenizer loading failure is a different compatibility issue."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "PyTorch missing ImportError and LayoutLMv3 box-content error are unrelated; one is environment dependency, the other is input validation/error messaging."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "Qwen3 auto device mapping cudaErrorAssert and Qwen3OmniMoe video processing ValueError are both Qwen-related but different subsystems and failure modes."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both concern video input regressions, but they target different versions/models and report different wrong internals (general regression vs vision_position_ids)."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44560",
+            "accept": false,
+            "reason": "Video processing ValueError in Qwen3OmniMoe and StopIteration in Qwen3-vl-embedding are distinct video-input bugs."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "Same broad video-input area, but different models and symptoms; not enough evidence for one concrete shared bug."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Qwen3Moe FineGrainedFP8Config and FSDP CPU RAM efficient loading are different advanced-loading failures with no clear shared root cause."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "TensorFlow backend packaging issue and inability to load a specific model are different bugs; one is install/backend selection, the other is model loading."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 21,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:24643",
+        "issue:27957",
+        "issue:28282",
+        "issue:29159",
+        "issue:34567",
+        "issue:36331",
+        "issue:38175",
+        "issue:43257",
+        "issue:43404",
+        "issue:43452",
+        "issue:43493",
+        "issue:43526",
+        "issue:43531",
+        "issue:43577",
+        "issue:43742",
+        "issue:43756",
+        "issue:43906",
+        "issue:43994",
+        "issue:44062",
+        "issue:44291",
+        "issue:44488",
+        "issue:44568",
+        "issue:44589",
+        "issue:44737",
+        "issue:44792",
+        "issue:44821",
+        "issue:44898",
+        "issue:44964",
+        "issue:44987",
+        "issue:45042",
+        "issue:45072",
+        "issue:45081",
+        "issue:45325",
+        "issue:45357",
+        "issue:45381",
+        "issue:45405",
+        "issue:45479",
+        "issue:45488",
+        "issue:45588",
+        "issue:45600",
+        "issue:45823",
+        "issue:45907",
+        "issue:45920",
+        "issue:45987",
+        "issue:46399",
+        "issue:46463",
+        "issue:46468",
+        "issue:46491"
+      ],
+      "soft_pairs": [
+        "issue:38175|issue:44898",
+        "issue:24643|issue:29159",
+        "issue:43493|issue:43994",
+        "issue:44987|issue:45907",
+        "issue:44291|issue:44821",
+        "issue:44568|issue:45488",
+        "issue:27957|issue:28282",
+        "issue:44737|issue:46463",
+        "issue:43577|issue:45072",
+        "issue:34567|issue:36331",
+        "issue:43257|issue:43756",
+        "issue:44062|issue:45081",
+        "issue:44792|issue:44964",
+        "issue:45042|issue:45081",
+        "issue:43906|issue:46468",
+        "issue:43742|issue:44589",
+        "issue:44737|issue:45588",
+        "issue:44568|issue:46491",
+        "issue:43756|issue:46399",
+        "issue:43756|issue:45325",
+        "issue:43531|issue:44821",
+        "issue:43404|issue:45987",
+        "issue:43526|issue:45479",
+        "issue:45823|issue:46463",
+        "issue:45405|issue:45600",
+        "issue:43452|issue:44488",
+        "issue:43906|issue:45381",
+        "issue:45357|issue:45920"
+      ],
+      "prepared_review_unit_hash": "b5a0039cea2f63c86a69a85df96d6b38403a265ff72c4f65054e9f07697390a0",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36511,
+        "estimated_input_tokens": 9128,
+        "estimated_eval_tokens": 18512
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a loose cluster of unrelated Transformers bug reports. A couple of pairs look like the same underlying model/tokenizer regression, but most edges are too broad to merge safely. No PRs are present in the input.",
+        "confidence": 0.55,
+        "canonical_issue_reason": "issue:45488 is the strongest concrete bug report in the set: it names a specific tokenizer regression in v5, gives a clear mechanism (ByteLevel pre-tokenizer overridden by Metaspace), and has broad user impact. It is the best available issue anchor, though the overall cluster is not a single clean duplicate set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45488 is the most actionable and specific issue; issue:45920 is broader, but 45488 provides the clearest root cause and is easier to use as a duplicate target for related tokenizer regressions.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:38175",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Different models and different failure modes; both are vision-related but not the same bug."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:29159",
+            "accept": false,
+            "reason": "DeepSpeed training error vs tokenizer inconsistency; unrelated code paths."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43994",
+            "accept": true,
+            "reason": "Both report SigLIP2 producing wrong outputs, one against the HF/JAX reference and one via AutoModel/pipeline; same underlying model implementation bug is plausible."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Model loading failure and repo-file revision forwarding are unrelated."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "`init_empty_weights` constructor argument bug vs image processor URL loading; different subsystems."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45488",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different tokenizer behaviors and model families; not clearly the same concrete bug."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Flash Attention 2 support vs missing PyTorch import; unrelated."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "XLNet device placement bug vs T5 error-message validation bug; unrelated."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Both concern dtype mismatches, but in different models and code paths; too broad to treat as duplicates."
+          },
+          {
+            "left": "issue:34567",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Trainer state accounting bug vs custom compute_loss signature change; different problems."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43756",
+            "accept": false,
+            "reason": "MoE weight conversion/loading issue vs RoPE layer/config mismatch; unrelated."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Different tokenizer crashes with different root causes."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Both are multimodal-model issues, but the symptoms and code paths do not match."
+          },
+          {
+            "left": "issue:45042",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Image-processor torchvision dependency issue vs Mistral tokenizer regex crash; unrelated."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:46468",
+            "accept": false,
+            "reason": "`43906` is only a reproduction of another issue; `46468` is a broken official Nemotron example, but the titles do not show the same concrete bug."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "Model load key error vs missing float8 storage type; unrelated."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "XLNet CPU tensor creation bug vs flash-attention `s_aux` crash; different code paths."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but one is BOS/EOS insertion and the other is CodeLlama round-trip spacing; not the same concrete change."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Different Qwen/SmolLM rotary-position issues with distinct symptoms and fixes."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Both touch RoPE/position ids, but the reported bugs are different model-specific failures."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding window issue vs image processor URL loading; unrelated."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45987",
+            "accept": false,
+            "reason": "LM head tying bug vs cache sequence-length return type; unrelated."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Image label reduction bug vs sequence-classification zero-loss bug; different layers of the stack."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Gemma4 device-map forward mismatch vs T5 decoder-input error message; unrelated."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Dependency version bump issue vs removed auto-mapping config reference; unrelated."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Both are loading failures, but one is gguf_file handling and the other is a specific model load regression; no clear shared root cause."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both involve Qwen-related models, but the reported bugs are different (reproduction vs video position IDs)."
+          },
+          {
+            "left": "issue:45357",
+            "right": "issue:45920",
+            "accept": false,
+            "reason": "Visual-encoder save_pretrained regression vs broad AutoTokenizer token-ID regression; different bugs."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary stays conservative and is grounded in the packet. It does not overclaim a single duplicate cluster, and the one accepted soft edge (SigLIP2-related issues) is reasonably supported by the titles. No PR-pair constraints apply because there are no PRs in the input."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 22,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:24643",
+        "issue:28282",
+        "issue:29127",
+        "issue:32090",
+        "issue:33357",
+        "issue:38175",
+        "issue:41762",
+        "issue:42898",
+        "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:44466",
+        "issue:44493",
+        "issue:44514",
+        "issue:44625",
+        "issue:44811",
+        "issue:44843",
+        "issue:44877",
+        "issue:44898",
+        "issue:44913",
+        "issue:44936",
+        "issue:45005",
+        "issue:45200",
+        "issue:45216",
+        "issue:45356",
+        "issue:45372",
+        "issue:45375",
+        "issue:45381",
+        "issue:45542",
+        "issue:45823",
+        "issue:46050",
+        "issue:46121",
+        "issue:46170",
+        "issue:46178",
+        "issue:46463",
+        "issue:46491"
+      ],
+      "soft_pairs": [
+        "issue:44625|issue:44877",
+        "issue:44913|issue:46121",
+        "issue:43720|issue:44898",
+        "issue:43697|issue:45216",
+        "issue:43643|issue:45375",
+        "issue:24643|issue:29127",
+        "issue:45200|issue:46463",
+        "issue:41762|issue:44936",
+        "issue:28282|issue:32090",
+        "issue:44493|issue:46050",
+        "issue:28282|issue:33357",
+        "issue:44466|issue:45005",
+        "issue:43452|issue:44843",
+        "issue:43819|issue:44811",
+        "issue:44936|issue:45823",
+        "issue:38175|issue:43295",
+        "issue:44361|issue:44442",
+        "issue:43688|issue:43749",
+        "issue:43653|issue:44361",
+        "issue:45372|issue:46170",
+        "issue:45542|issue:46178",
+        "issue:43452|issue:43577",
+        "issue:44514|issue:45381",
+        "issue:42898|issue:46491",
+        "issue:43295|issue:45356",
+        "issue:43531|issue:43742"
+      ],
+      "prepared_review_unit_hash": "e37a003fc6608ecb686801f1a7116e0bbb26d6b1102be56bb11d92724ee00124",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 26,
+        "serialized_chars": 35302,
+        "estimated_input_tokens": 8826,
+        "estimated_eval_tokens": 17908
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are a heterogeneous mix of unrelated bug reports (config loading, tokenizer regressions, training/DeepSpeed issues, backend import errors, and model-specific behavior changes). None of the soft pairs look like the same underlying bug/change, so this is not a duplicate cluster.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "No single issue is a strong canonical for the cluster; the reports are too diverse and do not describe one shared defect, so any canonical choice would be arbitrary.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "If one issue had to anchor triage, issue 44843 is the most concrete and actionable regression report, but it is still not representative of the rest of the set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44625",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Different model/config bugs: Qwen3.5 num_labels propagation vs granite_speech config strict-loading."
+          },
+          {
+            "left": "issue:44913",
+            "right": "issue:46121",
+            "accept": false,
+            "reason": "Both involve rope/config handling, but one is rotary_pct not persisting on reload while the other is a TypeError on JSON-loaded lists."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Unrelated models and code paths: BitNet accelerate loading vs Perceiver image classification/resolution handling."
+          },
+          {
+            "left": "issue:43697",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Different regressions: RTDetrV2 output drift in v5 vs Qwen3.5 checkpoint save_pretrained corruption."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both are config-field loss reports, but they involve different mechanisms and models; not the same concrete bug."
+          },
+          {
+            "left": "issue:24643",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "DeepSpeed weight-dimension training failure is unrelated to LayoutLMV3 box-validation message clarity."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:46463",
+            "accept": false,
+            "reason": "Different failures: Gemma 4 text-only token-type defaults vs T5 decoder input validation messaging."
+          },
+          {
+            "left": "issue:41762",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "ZeRO-3 loading index error is a different lifecycle problem from trainer.evaluate() failing after train()."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Missing PyTorch import vs NoneType passed to _gpu_broadcast_one are distinct errors."
+          },
+          {
+            "left": "issue:44493",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Unexpected position_id keys and quantized save/load breakage are different underlying problems."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "A missing-library ImportError is not the same as a MacOS bus error in CLIP loading."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45005",
+            "accept": false,
+            "reason": "Both mention tied weights, but one is device-dependent lm_head serialization and the other is a broader translation-model issue; too different to merge."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44843",
+            "accept": false,
+            "reason": "gguf_file loader breakage is unrelated to the unconditional model_info() offline-mode regression."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Different models and behaviors: DAC from_latents/forward mismatch vs Whisper batch_decode skip_special_tokens handling."
+          },
+          {
+            "left": "issue:44936",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Training/evaluation lifecycle failure is unrelated to Gemma4 device_map auto forward device mismatch."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "SigLIP2 zero-probability output is not the same bug as processor.tokenizer/image-passing regression."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "MLukeTokenizer task AttributeError and AutoTokenizer loading FastSpeech2ConformerTokenizer are different tokenizer failures."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Auxiliary-loss normalization in MoE models is unrelated to FSDP_CPU_RAM_EFFICIENT_LOADING."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "BigBird special-token registration and MLuke task errors are different tokenizer-specific issues."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Gemma 4 processor import failure due to mistral_common is unrelated to the Windows ROCm torch.distributed dependency issue."
+          },
+          {
+            "left": "issue:45542",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "Both are backend-related, but they concern different backends and different failure modes; not the same bug."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "gguf_file loading regressions are unrelated to Blip2 dtype propagation staying float32."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "apply_chat_template batching crash and qwen2.5-vl video position_ids are different bugs in the same family, not one underlying issue."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:46491",
+            "accept": false,
+            "reason": "clean_up_tokenization_spaces behavior changes are not the same as CodeLlama's leading-space encode/decode regression."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Custom processor.tokenizer regression and Kimi-K2.5 codec/warning issues are distinct."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window and MobileLLM key errors are unrelated model-loading/config 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": 21,
+      "review_unit_index": 23,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:29159",
+        "issue:32090",
+        "issue:33357",
+        "issue:33666",
+        "issue:36010",
+        "issue:36331",
+        "issue:41720",
+        "issue:43504",
+        "issue:43550",
+        "issue:43577",
+        "issue:43643",
+        "issue:43646",
+        "issue:43653",
+        "issue:43688",
+        "issue:43697",
+        "issue:43720",
+        "issue:43742",
+        "issue:43749",
+        "issue:43756",
+        "issue:43927",
+        "issue:44361",
+        "issue:44464",
+        "issue:44466",
+        "issue:44479",
+        "issue:44484",
+        "issue:44514",
+        "issue:44554",
+        "issue:44561",
+        "issue:44568",
+        "issue:44610",
+        "issue:44625",
+        "issue:44704",
+        "issue:44734",
+        "issue:44743",
+        "issue:44821",
+        "issue:44843",
+        "issue:44977",
+        "issue:44991",
+        "issue:45290",
+        "issue:45440",
+        "issue:45491",
+        "issue:45701",
+        "issue:45704",
+        "issue:46002",
+        "issue:46050",
+        "issue:46402",
+        "issue:46612"
+      ],
+      "soft_pairs": [
+        "issue:44484|issue:44704",
+        "issue:43577|issue:43697",
+        "issue:43653|issue:44568",
+        "issue:43646|issue:43749",
+        "issue:44464|issue:44743",
+        "issue:44843|issue:44991",
+        "issue:29159|issue:33357",
+        "issue:43688|issue:45704",
+        "issue:43756|issue:46402",
+        "issue:43653|issue:43927",
+        "issue:43550|issue:43927",
+        "issue:43688|issue:45440",
+        "issue:44734|issue:46612",
+        "issue:43688|issue:45701",
+        "issue:44479|issue:44625",
+        "issue:43550|issue:44361",
+        "issue:45491|issue:46002",
+        "issue:44464|issue:44977",
+        "issue:33666|issue:36010",
+        "issue:43742|issue:44821",
+        "issue:41720|issue:44464",
+        "issue:32090|issue:33666",
+        "issue:44514|issue:45290",
+        "issue:44466|issue:46050",
+        "issue:43504|issue:43550",
+        "issue:43643|issue:44561",
+        "issue:43504|issue:43720",
+        "issue:43577|issue:44554",
+        "issue:43504|issue:44610",
+        "issue:43720|issue:43927",
+        "issue:33666|issue:36331"
+      ],
+      "prepared_review_unit_hash": "dac06302d419380fbe4ed3d7fbc92efa6da77ee7a8579bc7b11eade89a3aab5a",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36527,
+        "estimated_input_tokens": 9132,
+        "estimated_eval_tokens": 18520
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is very heterogeneous: the items only share broad subsystem keywords like tokenizer, generation, save/load, or trust_remote_code. None of the soft pairs look like the same underlying bug/change, so I rejected all of them.",
+        "confidence": 0.36,
+        "canonical_issue_reason": "issue:45701 is the broadest and most general report in the set (a version-dependent tokenization regression) and is the best umbrella issue among otherwise unrelated reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45701 is the best overall representative because it is open, broadly phrased, and closest to the tokenizer/regression theme that appears in a few of the softer matches.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44484",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "Different bugs: save_pretrained shard-size question vs AutoProcessor kwargs propagation."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:43697",
+            "accept": false,
+            "reason": "Both are model correctness issues, but one is BLIP2 dtype handling and the other is RTDetrV2 output divergence; not the same code path."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both concern tokenizer special-token behavior, but on different tokenizers and different failures (mask token registration vs BOS/EOS insertion)."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:43749",
+            "accept": false,
+            "reason": "Unrelated areas: trust_remote_code config fields vs FSDP CPU RAM-efficient loading."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Both involve generation/cache behavior, but one is compiled forward chunked generation and the other is recurrent state reset with cache; not the same bug."
+          },
+          {
+            "left": "issue:44843",
+            "right": "issue:44991",
+            "accept": false,
+            "reason": "Both are tokenizer-loading regressions, but the failure modes differ (offline model_info call vs a model-specific tokenizer load breakage)."
+          },
+          {
+            "left": "issue:29159",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Tokenizer inconsistency vs a MacOS bus error on a CLIP model; no shared underlying defect."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45704",
+            "accept": false,
+            "reason": "Different model/runtime issues: auxiliary loss normalization vs an apex FusedRMSNorm memory leak in T5."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:46402",
+            "accept": false,
+            "reason": "Smollm3 RoPE layer count mismatch and Sam3Model text-encoder weights not loading are unrelated."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Special-token registration bug vs config losing custom token IDs after save/load; different mechanisms."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "torch.compile+SDPA failure in Bamba vs config/token ID persistence in DiaConfig; unrelated."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Aux loss normalization bug vs native DeepSeekV3MoE diverging from remote implementation; too broad to merge."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:46612",
+            "accept": false,
+            "reason": "Both are generation/cache issues, but one is KV-cache continuation indexing in serve and the other is beam-search cache reorder skipping in several models."
+          },
+          {
+            "left": "issue:43688",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "Auxiliary-loss normalization and tokenization version changes are unrelated."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44625",
+            "accept": false,
+            "reason": "Video-input regression vs num_labels propagation bug; different components and symptoms."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure in Bamba vs MLukeTokenizer AttributeError; no shared code path."
+          },
+          {
+            "left": "issue:45491",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Both mention NaNs, but the causes are different: sliding-window attention on mixed-length batches vs classification-head init/checkpointing."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Compiled-forward chunked generation issue vs flash-attention generation failure in Qwen3.5; not the same defect."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:36010",
+            "accept": false,
+            "reason": "Multi-GPU training problem vs missing GenerationMixin import; unrelated."
+          },
+          {
+            "left": "issue:43742",
+            "right": "issue:44821",
+            "accept": false,
+            "reason": "KeyError loading a model vs loading AutoImageProcessor from URL; different object types and failure modes."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:44464",
+            "accept": false,
+            "reason": "CUDA assert with auto device mapping vs compiled-forward generation inconsistency; no concrete overlap."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "Trainer broadcast TypeError vs general multi-GPU training support; not the same bug."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both use apply_chat_template, but one is batched padding=False behavior and the other is assistant tool-call handling; separate edge cases."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Serialization of tied lm_head weights vs quantized model save/load breakage; different persistence bugs."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Legacy field loading in BeitForSemanticSegmentation vs Bamba torch.compile SDPA failure; unrelated."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "Both involve trust_remote_code, but one is missing fields from AutoConfig and the other is a removed helper breaking remote-code models."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43720",
+            "accept": false,
+            "reason": "Legacy-field model loading vs BitNet packed-weight unpacking during accelerate loading; different model families and bugs."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:44554",
+            "accept": false,
+            "reason": "BLIP2 dtype persistence vs MPS attention correctness with mismatched head dims; no shared underlying bug."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Beit legacy-field load failure vs OmDet-Turbo processor/model input-size mismatch."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "BitNet packed-weight loading vs DiaConfig custom token IDs; unrelated persistence/loading bugs."
+          },
+          {
+            "left": "issue:33666",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "Multi-GPU training issue vs custom trainer signature mismatch; not the same defect."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 24,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:28218",
+        "issue:33357",
+        "issue:33666",
+        "issue:39692",
+        "issue:43257",
+        "issue:43262",
+        "issue:43278",
+        "issue:43295",
+        "issue:43299",
+        "issue:43334",
+        "issue:43452",
+        "issue:43493",
+        "issue:43525",
+        "issue:43540",
+        "issue:43572",
+        "issue:43575",
+        "issue:43606",
+        "issue:43646",
+        "issue:43688",
+        "issue:43701",
+        "issue:43720",
+        "issue:43825",
+        "issue:43844",
+        "issue:43927",
+        "issue:43994",
+        "issue:44361",
+        "issue:44451",
+        "issue:44466",
+        "issue:44488",
+        "issue:44554",
+        "issue:44568",
+        "issue:44589",
+        "issue:44625",
+        "issue:44843",
+        "issue:44849",
+        "issue:44964",
+        "issue:44977",
+        "issue:45357",
+        "issue:45405",
+        "issue:45520",
+        "issue:45584",
+        "issue:45812",
+        "issue:45823",
+        "issue:46050",
+        "issue:46192",
+        "issue:46209",
+        "issue:46365",
+        "issue:46489"
+      ],
+      "soft_pairs": [
+        "issue:43262|issue:43295",
+        "issue:43606|issue:43927",
+        "issue:44554|issue:46365",
+        "issue:44977|issue:46209",
+        "issue:43525|issue:43572",
+        "issue:43452|issue:44451",
+        "issue:43720|issue:44361",
+        "issue:43278|issue:43688",
+        "issue:45823|issue:46050",
+        "issue:43994|issue:46192",
+        "issue:43493|issue:43825",
+        "issue:39692|issue:43493",
+        "issue:44964|issue:45405",
+        "issue:44625|issue:45584",
+        "issue:28218|issue:33357",
+        "issue:44849|issue:46209",
+        "issue:39692|issue:43295",
+        "issue:43493|issue:43844",
+        "issue:43278|issue:43844",
+        "issue:44849|issue:45357",
+        "issue:43646|issue:44466",
+        "issue:33357|issue:33666",
+        "issue:44488|issue:44843",
+        "issue:44589|issue:45520",
+        "issue:43540|issue:44977",
+        "issue:44568|issue:46489",
+        "issue:43257|issue:43299",
+        "issue:43278|issue:43701",
+        "issue:43927|issue:44361",
+        "issue:43334|issue:43572",
+        "issue:43575|issue:45405",
+        "issue:44568|issue:45812"
+      ],
+      "prepared_review_unit_hash": "4b293cbea5cf59df19265c5d2b7f8f82921451590a92fd51ddbf77a7ff6c60e3",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37318,
+        "estimated_input_tokens": 9330,
+        "estimated_eval_tokens": 18916
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is semantically loose rather than a true duplicate set: most items are separate model/config/tokenizer regressions with different code paths. Under a strict duplicate policy, none of the soft edges should be merged.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "issue:45812 is the strongest single anchor: it describes a concrete, high-impact, silent tokenizer regression across a broad model family with clear symptomology and references. It is still only a loose representative, not a true canonical for the whole cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45812 is the best standalone issue to keep because it is specific, reproducible, and broadly user-facing, making it the most useful representative if one issue must anchor the group.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43262",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "Different regressions: audio chat-template sampling-rate default vs multimodal processor.tokenizer / image-handling breakage."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43927",
+            "accept": false,
+            "reason": "Unrelated subsystems: CPU offload device mismatch vs DiaConfig custom token IDs lost on save/load."
+          },
+          {
+            "left": "issue:44554",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Different code paths and failures: MPS attention correctness vs importing Trainer being blocked by datasets."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Both are inference bugs, but one is Qwen3.5 flash-attention behavior and the other is FP8/B200 support; not the same bug."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Both mention missing pad token fields, but they affect different config classes/models and appear to be separate regressions."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44451",
+            "accept": false,
+            "reason": "Generic load failures, but one is gguf_file handling and the other is a specific model load failure; too broad to merge."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "Packed-weight loading bug in BitNet is unrelated to MLukeTokenizer AttributeError."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43688",
+            "accept": false,
+            "reason": "Training/eval dtype drift is unrelated to auxiliary-loss normalization."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Different problems: Gemma4 device_map forwarding vs quantized model save/load corruption."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Wrong SigLIP2 outputs vs return_tensors being ignored are unrelated API bugs."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Model implementation discrepancy vs pipeline error-message wording; not the same underlying bug."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Both are SigLIP2-related, but one is docs/example errors and the other is implementation mismatch with JAX; separate issues."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Model loading failure vs dependency version bump; no shared code-path bug."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Qwen3.5 num_labels propagation and Whisper empty transcription handling are different regressions."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Tokenizer spacing bug and MacOS bus error are completely unrelated."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Qwen3.5 hidden-states behavior is unrelated to FP8/B200 inference support."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "SigLIP2 docs/model mismatch is unrelated to processor.tokenizer/image input regression."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "Implementation discrepancy in SigLIP2 is unrelated to ZeRO-3 gradient growth."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "Embedding dtype change and ZeRO-3 gradient instability are different issues."
+          },
+          {
+            "left": "issue:44849",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different Qwen3.5 bugs: hidden_states output vs incorrect visual encoder keys on save_pretrained."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Custom model initialization breakage is not the same as inconsistent lm_head serialization."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:33666",
+            "accept": false,
+            "reason": "MacOS bus error in CLIP loading is unrelated to Qwen2-VL multi-GPU training."
+          },
+          {
+            "left": "issue:44488",
+            "right": "issue:44843",
+            "accept": false,
+            "reason": "Both are model-loading complaints, but one is a specific model load failure and the other is an offline-mode regex patch bug."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "Float8 storage lookup failure is unrelated to flash_attn import mapping on Python 3.13."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Video-input parsing in Qwen3OmniMoe is different from Qwen3.5 flash-attention generation behavior."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:46489",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different models and symptoms; not the same bug."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:43299",
+            "accept": false,
+            "reason": "Accelerate+deepspeed weight conversion bug is distinct from Qwen3VL MoE loading failure."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Embedding dtype issue and resume_from_checkpoint key mismatch are unrelated."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:44361",
+            "accept": false,
+            "reason": "DiaConfig save/load corruption is unrelated to MLukeTokenizer task AttributeError."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43572",
+            "accept": false,
+            "reason": "Different missing-padding-token regressions in different configs; similar theme but not the same bug."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "TP OOM on Qwen2-57B-A14B is unrelated to a PEFT minimum-version bump."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Different tokenizer regressions with different failure modes; not clearly the same code-path problem."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 25,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:15354",
+        "issue:28218",
+        "issue:29159",
+        "issue:29942",
+        "issue:30333",
+        "issue:32090",
+        "issue:34567",
+        "issue:38175",
+        "issue:42490",
+        "issue:43065",
+        "issue:43334",
+        "issue:43493",
+        "issue:43504",
+        "issue:43572",
+        "issue:43606",
+        "issue:43653",
+        "issue:43696",
+        "issue:43701",
+        "issue:43824",
+        "issue:43844",
+        "issue:43873",
+        "issue:43994",
+        "issue:44112",
+        "issue:44291",
+        "issue:44423",
+        "issue:44464",
+        "issue:44493",
+        "issue:44554",
+        "issue:44589",
+        "issue:44734",
+        "issue:44792",
+        "issue:44964",
+        "issue:45084",
+        "issue:45200",
+        "issue:45229",
+        "issue:45405",
+        "issue:45406",
+        "issue:45446",
+        "issue:45520",
+        "issue:45584",
+        "issue:45588",
+        "issue:45759",
+        "issue:45823",
+        "issue:46050",
+        "issue:46121",
+        "issue:46285",
+        "issue:46369",
+        "issue:46424"
+      ],
+      "soft_pairs": [
+        "issue:43334|issue:45759",
+        "issue:43504|issue:43606",
+        "issue:43824|issue:44589",
+        "issue:44792|issue:45405",
+        "issue:43572|issue:44291",
+        "issue:43606|issue:44112",
+        "issue:44464|issue:46424",
+        "issue:45200|issue:45229",
+        "issue:44493|issue:46285",
+        "issue:43504|issue:43653",
+        "issue:45229|issue:45823",
+        "issue:44464|issue:45584",
+        "issue:43994|issue:44493",
+        "issue:43873|issue:46050",
+        "issue:45588|issue:46050",
+        "issue:29942|issue:30333",
+        "issue:43844|issue:44964",
+        "issue:43065|issue:46369",
+        "issue:29159|issue:32090",
+        "issue:42490|issue:43696",
+        "issue:45084|issue:45520",
+        "issue:44423|issue:44734",
+        "issue:28218|issue:34567",
+        "issue:44554|issue:45446",
+        "issue:15354|issue:38175",
+        "issue:45406|issue:46121",
+        "issue:43493|issue:43701"
+      ],
+      "prepared_review_unit_hash": "a7b6fd06974151ea6b92647e0cbb15a6553cdc3566f4704abb858dd2adbba094",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 35909,
+        "estimated_input_tokens": 8978,
+        "estimated_eval_tokens": 18212
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These issues are largely unrelated across tokenization, model loading, training, CI, attention, and multimodal inference. The soft-similarity pairs mostly share only broad subsystem vocabulary, not the same concrete bug/change, so none should be merged as duplicates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43334",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both touch Qwen multimodal configs, but one is a missing pad_token_id attribute in Qwen3-VL and the other is AutoModelForCausalLM.from_config not unwrapping text_config for Qwen 3.5/3.6; different failure modes and fixes."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43606",
+            "accept": false,
+            "reason": "43504 is about a legacy field preventing BeitForSemanticSegmentation preset loading; 43606 is a CPU offload device mismatch in suno/bark-small. Different models and code paths."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44589",
+            "accept": false,
+            "reason": "43824 is an ImportError for Qwen2_5_VLForConditionalGeneration; 44589 is a storage lookup failure for Float8_e4m3fnStorage. Unrelated import vs storage backend issues."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "44792 is a failing janus image-generation test; 45405 is an unreleased MIN_PEFT_VERSION bump. No shared underlying bug."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "43572 is about missing pad_token_idx in StableLmConfig after a 5.0 update; 44291 is init_empty_weights passing an unexpected _is_hf_initialized arg. Both are config-related but different breakages."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "43606 concerns offload device mismatch in bark-small; 44112 is a stale GraniteSpeech CI test. Same general area of device handling, but not the same defect."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "44464 reports inconsistent outputs with compiled forward during chunked generation; 46424 says max_cache_length is only honored for Quantized Cache. Different cache/compile behaviors."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "45200 is about mm_token_type_ids defaulting to zeros for text-only fine-tuning; 45229 is multi-GPU inference CUDA OOM. One is a shape/feature bug, the other is memory capacity."
+          },
+          {
+            "left": "issue:44493",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "44493 is an unexpected key with position id; 46285 is an inconsistent class structure after v5.6 involving .text_model removal. Not the same issue."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "43504 is a preset-load failure from a legacy field in BEiT; 43653 is BigBirdTokenizer mask token registration causing empty decode output. Different components and symptoms."
+          },
+          {
+            "left": "issue:45229",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "45229 is a multi-GPU CUDA OOM; 45823 is a device mismatch in Gemma4 PLE with device_map=auto. Memory exhaustion vs tensor placement bug."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "44464 is about compiled forward/chunked generation inconsistency; 45584 is Whisper failing on empty transcription after align_special_tokens. Different generation pipelines."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "43994 reports nonsensical outputs from SigLIP2 with AutoModel/pipeline; 44493 is a position-id key mismatch. Different model family and failure mode."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "43873 is an open quantization/offloading behavior issue; 46050 is quantized model save/load corruption with sinq. Both involve quantization, but not the same code-path defect."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "45588 crashes in flash_attention.py when s_aux is None for sink-less models; 46050 is broken quantized save_pretrained/from_pretrained behavior. Different subsystems."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:30333",
+            "accept": false,
+            "reason": "29942 is failing Flash Attention 2 tests; 30333 is MLflow not marking failed training jobs as finished. No underlying bug overlap."
+          },
+          {
+            "left": "issue:43844",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "43844 concerns gradient blow-up with HfDeepSpeedConfig + ZeRO-3; 44964 is loading Phi-4-multimodal-instruct on latest transformers. Different training vs model-loading issues."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "43065 is about a dummy nn.Conv2d in Sam3PixelDecoder; 46369 is memory overflow when calling sam3-litetext. Same model area, but one is architecture/code structure and the other is memory usage."
+          },
+          {
+            "left": "issue:29159",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "29159 is inconsistent slow vs fast tokenizer behavior; 32090 is a Trainer NoneType passed to _gpu_broadcast_one. Different components and fixes."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43696",
+            "accept": false,
+            "reason": "42490 is an import warning for local model paths ending with a slash; 43696 is CUDA OOM for GPT-oss-20b. Unrelated path parsing vs memory issue."
+          },
+          {
+            "left": "issue:45084",
+            "right": "issue:45520",
+            "accept": false,
+            "reason": "45084 is a compile-time TypeError about non template nodes; 45520 is a flash_attn import_utils KeyError on Python 3.13. Different compile/import problems."
+          },
+          {
+            "left": "issue:44423",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "44423 crashes in continuous-batching with multimodal models because a string is treated like a tensor; 44734 crashes in /v1/responses KV cache continuation due to wrong tensor indexing. Both are serve-time failures, but different concrete bugs."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "28218 is tokenizer spacing after added tokens; 34567 is TrainerState.num_input_tokens_seen not updating. Different tokenizer vs training-state issues."
+          },
+          {
+            "left": "issue:44554",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "44554 is an MPS attention correctness issue when value head dim differs from query; 45446 is a PyTorch version check bug for AuxRequest import in flex_attention. Different conditions and code paths."
+          },
+          {
+            "left": "issue:15354",
+            "right": "issue:38175",
+            "accept": false,
+            "reason": "15354 is torch.jit.script failing on GeneratorExp during export; 38175 is unexpected zero probabilities for siglip2-base-patch16-224. No meaningful overlap."
+          },
+          {
+            "left": "issue:45406",
+            "right": "issue:46121",
+            "accept": false,
+            "reason": "45406 is a Gemma4Processor missing _tokenizer attribute in transformers serve; 46121 is convert_rope_params_to_dict raising TypeError on a JSON-loaded list. Different processor vs rope-validation bugs."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "43493 is a SigLIP2 implementation discrepancy versus the original JAX model; 43701 is a resume_from_checkpoint key mismatch. Same training ecosystem only at a very broad level, 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": 21,
+      "review_unit_index": 26,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:28218",
+        "issue:35141",
+        "issue:37428",
+        "issue:38175",
+        "issue:39692",
+        "issue:42490",
+        "issue:42915",
+        "issue:43232",
+        "issue:43295",
+        "issue:43334",
+        "issue:43450",
+        "issue:43452",
+        "issue:43493",
+        "issue:43525",
+        "issue:43540",
+        "issue:43550",
+        "issue:43572",
+        "issue:43575",
+        "issue:43653",
+        "issue:43824",
+        "issue:43844",
+        "issue:43873",
+        "issue:43950",
+        "issue:43957",
+        "issue:44112",
+        "issue:44186",
+        "issue:44466",
+        "issue:44521",
+        "issue:44554",
+        "issue:44610",
+        "issue:44849",
+        "issue:44898",
+        "issue:44987",
+        "issue:45072",
+        "issue:45127",
+        "issue:45198",
+        "issue:45200",
+        "issue:45290",
+        "issue:45356",
+        "issue:45381",
+        "issue:45412",
+        "issue:45520",
+        "issue:45823",
+        "issue:46178",
+        "issue:46192",
+        "issue:46209",
+        "issue:46285",
+        "issue:46512"
+      ],
+      "soft_pairs": [
+        "issue:43540|issue:45200",
+        "issue:37428|issue:43957",
+        "issue:45072|issue:45198",
+        "issue:43295|issue:43450",
+        "issue:44186|issue:46192",
+        "issue:45200|issue:45823",
+        "issue:45823|issue:46192",
+        "issue:43295|issue:44610",
+        "issue:43232|issue:44849",
+        "issue:43295|issue:43873",
+        "issue:44521|issue:45290",
+        "issue:43572|issue:44987",
+        "issue:45520|issue:46178",
+        "issue:39692|issue:46192",
+        "issue:43450|issue:43575",
+        "issue:43572|issue:43957",
+        "issue:43550|issue:43844",
+        "issue:37428|issue:43824",
+        "issue:44554|issue:45356",
+        "issue:45381|issue:46209",
+        "issue:44466|issue:45127",
+        "issue:42490|issue:43493",
+        "issue:38175|issue:43493",
+        "issue:44898|issue:45412",
+        "issue:43653|issue:44112",
+        "issue:43950|issue:46285",
+        "issue:42915|issue:43824",
+        "issue:43452|issue:46512",
+        "issue:43334|issue:43525",
+        "issue:28218|issue:35141"
+      ],
+      "prepared_review_unit_hash": "aae047cf66a1fd397491d3fe8d82974624c5096854f867657e7e105abc401517",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 37047,
+        "estimated_input_tokens": 9262,
+        "estimated_eval_tokens": 18780
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is over-broad and spans many unrelated Transformers bugs; none of the soft pairs are safe duplicate merges, so there is no clear canonical issue or PR.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43540",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Different multimodal subsystems and failures: Qwen3OmniMoe video processing vs Gemma 4 text-only fine-tuning token-type defaults."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Both are loading-related, but one is a missing import symbol and the other is a meta-device model-loading regression; not the same bug."
+          },
+          {
+            "left": "issue:45072",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Different models and code paths: bfloat16 inference dtype mismatch vs Wav2Vec2 save_pretrained/tokenization failure."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Both involve multimodal processing, but one is processor.tokenizer/image passthrough regression and the other is batched video shape handling."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Tokenizer-related, but one is NER/batched padding crash and the other is return_tensors being ignored when text_kwargs is passed."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Gemma 4 token-type defaults and device_map=\"auto\" forward device mismatch are different problems."
+          },
+          {
+            "left": "issue:45823",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Device placement mismatch vs text_kwargs/return_tensors behavior; no shared underlying code-path."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Different processing bugs: images passed to tokenizer vs OmDet processor/model input-size mismatch."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44849",
+            "accept": false,
+            "reason": "Generation kwargs after sync_gpus vs Qwen3.5 output_hidden_states regression; unrelated."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Processor/tokenizer regression is unrelated to quantization offloading behavior."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Both are chat-template issues, but one is assistant_mask generation for multimodal inputs and the other is a tool-call/no-content crash."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Missing pad_token_idx in a config vs a model-loading failure for a specific repository; no clear shared bug."
+          },
+          {
+            "left": "issue:45520",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "Different backend/import problems: flash_attn on Python 3.13 vs tensorflow_text not present in BACKENDS_MAPPING."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Documentation/model-example errors are unrelated to return_tensors being ignored with text_kwargs."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Batched video-shape bug and tensor-parallel OOM on model load are unrelated."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:43957",
+            "accept": false,
+            "reason": "Config field regression and meta-device loading failure do not describe the same underlying issue."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43844",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure in Bamba is unrelated to gradient growth under ZeRO-3."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but for different missing symbols/modules and different root causes."
+          },
+          {
+            "left": "issue:44554",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "MPS attention correctness bug is unrelated to the Kimi-K2.5 tokenizer regression."
+          },
+          {
+            "left": "issue:45381",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Qwen2.5-VL video vision_position_ids bug and FP8 inference failure on B200 are different failures."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45127",
+            "accept": false,
+            "reason": "Both mention tied weights, but one is serialization inconsistency and the other is LoRA merge collapse after vocab extension."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Local-path trailing-slash import warning is unrelated to SigLIP2 HF/JAX parity discrepancies."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Same model family, but one is a specific zero-probability symptom and the other is a broader implementation-vs-JAX discrepancy; not clearly the same bug."
+          },
+          {
+            "left": "issue:44898",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Perceiver interpolation failure and RT-DETR memory release issue are unrelated."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Tokenizer special-token registration bug vs stale CI device-override test failure; unrelated."
+          },
+          {
+            "left": "issue:43950",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "Non-persistent buffer corruption in from_pretrained is unrelated to CLIPTextModel structure changes."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "FP8Config failure in Qwen3Moe is unrelated to a missing Qwen2_5_VL export import error."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "gguf_file loading breakage is unrelated to stale config.dtype after from_config(dtype=...)."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Both are missing pad_token_id errors, but on different config classes and model families; not enough to treat as one bug."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Tokenizer spacing after added tokens and post_init reinitializing output embeddings are entirely different issues."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 27,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:29942",
+        "issue:32090",
+        "issue:37428",
+        "issue:38175",
+        "issue:41720",
+        "issue:43262",
+        "issue:43278",
+        "issue:43335",
+        "issue:43388",
+        "issue:43425",
+        "issue:43450",
+        "issue:43526",
+        "issue:43550",
+        "issue:43606",
+        "issue:43611",
+        "issue:43653",
+        "issue:43819",
+        "issue:43824",
+        "issue:43827",
+        "issue:43856",
+        "issue:43901",
+        "issue:43931",
+        "issue:43975",
+        "issue:44112",
+        "issue:44206",
+        "issue:44442",
+        "issue:44514",
+        "issue:44521",
+        "issue:44561",
+        "issue:44610",
+        "issue:44792",
+        "issue:44857",
+        "issue:45083",
+        "issue:45198",
+        "issue:45200",
+        "issue:45325",
+        "issue:45412",
+        "issue:45520",
+        "issue:45563",
+        "issue:45584",
+        "issue:45593",
+        "issue:45759",
+        "issue:46170",
+        "issue:46192",
+        "issue:46209",
+        "issue:46234",
+        "issue:46285",
+        "issue:46512"
+      ],
+      "soft_pairs": [
+        "issue:43278|issue:43856",
+        "issue:43262|issue:43450",
+        "issue:43824|issue:43975",
+        "issue:43526|issue:46192",
+        "issue:45200|issue:45325",
+        "issue:43388|issue:43526",
+        "issue:44442|issue:45198",
+        "issue:44792|issue:45563",
+        "issue:43819|issue:44112",
+        "issue:38175|issue:45412",
+        "issue:41720|issue:46209",
+        "issue:44610|issue:44857",
+        "issue:44561|issue:46234",
+        "issue:43827|issue:43901",
+        "issue:45759|issue:46285",
+        "issue:44206|issue:45584",
+        "issue:44857|issue:45593",
+        "issue:43335|issue:43550",
+        "issue:43819|issue:45083",
+        "issue:45520|issue:46170",
+        "issue:37428|issue:43425",
+        "issue:45759|issue:46512",
+        "issue:43606|issue:45083",
+        "issue:29942|issue:32090",
+        "issue:44514|issue:44521",
+        "issue:43335|issue:43653",
+        "issue:43550|issue:44112",
+        "issue:43611|issue:43931"
+      ],
+      "prepared_review_unit_hash": "50f922e934734d1f9ad0db414ad801932dcda59782652c8279732ce0ed21a488",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 36257,
+        "estimated_input_tokens": 9065,
+        "estimated_eval_tokens": 18386
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly a set of unrelated bug reports across different models and subsystems. No PRs are present. The closest pair is the Qwen2.5-VL `apply_chat_template` reports, but the symptoms are still distinct enough to avoid deduping.",
+        "confidence": 0.72,
+        "canonical_issue_reason": "issue:44521 is the most central and broadly scoped report here: it is open, has the most inbound references/discussion, and concerns multimodal `apply_chat_template()` behavior rather than a narrow one-off failure.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44521 is the best representative issue for this cluster because it is active, well-referenced, and describes a user-facing multimodal template bug with wider impact than the other reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43278",
+            "right": "issue:43856",
+            "accept": false,
+            "reason": "Different problems: dtype drift in evaluation vs MoE training memory usage."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:43450",
+            "accept": false,
+            "reason": "Different processors and failures: audio sampling-rate default vs batched video shape."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43975",
+            "accept": false,
+            "reason": "Unrelated: import error for Qwen VL class vs DeepSeek detokenization bug."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Different bugs: Beit label reduction vs `return_tensors` being ignored."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "Different multimodal model issues with no shared code-path or symptom."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Different data handling bugs: metric gathering vs image processor label reduction."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "Different tokenizer/save-pretrained failures in different models."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Unrelated: Janus image-generation test failure vs stale generation warning."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Different scope: DAC latent mismatch vs GraniteSpeech CI device-override test."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "Different issues: zero probabilities in SigLIP2 vs RT-DETR memory release."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Both are hardware-specific inference failures, but on different models and likely different causes."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Different vision bugs: processor output size mismatch vs AMP crash in loss computation."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "Different missing utility symbols and different breakages."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Both are docs regressions, but they refer to different APIs and different documentation gaps."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "Different structural/config bugs in different model families."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Different audio/text bugs with no shared code-path."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Different detection-model issues: AMP crash vs auxiliary-loss omission."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Different model/config bug vs torch.compile+SDPA failure."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Different audio/model internals: latent reconstruction mismatch vs feature-length helper behavior."
+          },
+          {
+            "left": "issue:45520",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Both are import/runtime compatibility problems, but on different platforms and unrelated dependencies."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:43425",
+            "accept": false,
+            "reason": "Different compatibility failures: flash-attn import helper vs Torch 2.10 incompatibility."
+          },
+          {
+            "left": "issue:45759",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "Both involve `from_config`, but one is multimodal config unwrapping and the other is stale dtype metadata."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Unrelated: CPU offload device mismatch vs Qwen3 Omni feature-length helper behavior."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Different failures: Flash Attention 2 test breakage vs trainer broadcast NoneType error."
+          },
+          {
+            "left": "issue:44514",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Same general API, but one is a batched-input crash and the other is zero assistant masks; not clearly the same defect."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43653",
+            "accept": false,
+            "reason": "Different bugs: sparse-layer config creation vs tokenizer special-token registration."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Different failures: torch.compile/SDPA issue vs stale CI test."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Different model-loading problems with different triggers 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": 1,
+      "component_count": 21,
+      "review_unit_index": 28,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:26413",
+        "issue:28218",
+        "issue:29127",
+        "issue:31501",
+        "issue:32090",
+        "issue:34689",
+        "issue:35141",
+        "issue:37428",
+        "issue:39692",
+        "issue:42490",
+        "issue:43278",
+        "issue:43335",
+        "issue:43504",
+        "issue:43526",
+        "issue:43606",
+        "issue:43720",
+        "issue:43819",
+        "issue:43825",
+        "issue:43854",
+        "issue:44112",
+        "issue:44206",
+        "issue:44351",
+        "issue:44360",
+        "issue:44448",
+        "issue:44493",
+        "issue:44561",
+        "issue:44743",
+        "issue:44779",
+        "issue:44855",
+        "issue:44857",
+        "issue:45072",
+        "issue:45083",
+        "issue:45127",
+        "issue:45200",
+        "issue:45245",
+        "issue:45290",
+        "issue:45310",
+        "issue:45381",
+        "issue:45459",
+        "issue:45478",
+        "issue:45588",
+        "issue:45593",
+        "issue:45823",
+        "issue:45859",
+        "issue:46234",
+        "issue:46326",
+        "issue:46431",
+        "issue:46566"
+      ],
+      "soft_pairs": [
+        "issue:45290|issue:46326",
+        "issue:44448|issue:44743",
+        "issue:46431|issue:46566",
+        "issue:44855|issue:45083",
+        "issue:43335|issue:43504",
+        "issue:37428|issue:44351",
+        "issue:43278|issue:43825",
+        "issue:29127|issue:32090",
+        "issue:43335|issue:45083",
+        "issue:45200|issue:45588",
+        "issue:43720|issue:44112",
+        "issue:45127|issue:45245",
+        "issue:45588|issue:45859",
+        "issue:43819|issue:44493",
+        "issue:43335|issue:44112",
+        "issue:39692|issue:43526",
+        "issue:43335|issue:45593",
+        "issue:44206|issue:45823",
+        "issue:26413|issue:31501",
+        "issue:34689|issue:35141",
+        "issue:45310|issue:45478",
+        "issue:39692|issue:44493",
+        "issue:43606|issue:44857",
+        "issue:42490|issue:43854",
+        "issue:45459|issue:46234",
+        "issue:44448|issue:44779",
+        "issue:44857|issue:45072",
+        "issue:28218|issue:32090",
+        "issue:44360|issue:44561",
+        "issue:45290|issue:45381"
+      ],
+      "prepared_review_unit_hash": "e9590743cde7ab1ff71dd5871741553b1bf6ca4a4810f653538f0bb4b5a345d0",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36974,
+        "estimated_input_tokens": 9244,
+        "estimated_eval_tokens": 18744
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is mostly noise: the items span unrelated bugs across tokenizer, training, model-loading, and CI failures. The only clear duplicate pair is the repeated Qwen3.5 Moe from_pretrained error (issues 45310 and 45478).",
+        "confidence": 0.84,
+        "canonical_issue_reason": "Issue 45310 is the earlier and clearer report of the exact same Qwen3.5 Moe from_pretrained regression seen again in 45478.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45310 is the best representative because it is a concrete, self-contained bug report and has a near-identical duplicate in 45478.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45290",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Both involve chat/tokenizer behavior, but they describe different failure modes and likely different code paths."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Unrelated regressions in different models and subsystems; not the same bug."
+          },
+          {
+            "left": "issue:46431",
+            "right": "issue:46566",
+            "accept": false,
+            "reason": "A missing chat template error and a device-map/meta-weight issue are distinct problems."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "One is a Python import/parsing failure; the other is a feature-length helper behavior bug in a different model."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:43504",
+            "accept": false,
+            "reason": "Different models and different root causes; not mergeable as one fix."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but they concern different missing symbols and unrelated fixes."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Different symptom domains: dtype drift vs an incorrect pipeline error message."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "LayoutLMv3 error-message clarity and Trainer GPU broadcast TypeError are unrelated."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "No shared underlying bug beyond both touching model internals."
+          },
+          {
+            "left": "issue:45200",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Gemma text fine-tuning defaults and flash-attention auxiliary-state handling are unrelated."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Different CI/runtime failures in different code paths and models."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "LoRA/embedding collapse and category-count limits are unrelated defects."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "Both mention offload/attention-related failures, but they affect different modules and models."
+          },
+          {
+            "left": "issue:43819",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "A DAC forward/from_latents mismatch is unrelated to position-id key warnings."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Sparse-layer config bug and GraniteSpeech CI failure do not share the same root cause."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Documentation-example issues and an image-processor label bug are not the same underlying change."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "SwitchTransformers config construction and D-FINE auxiliary-loss behavior are separate issues."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Different models, different device/dtype symptoms, and different code paths."
+          },
+          {
+            "left": "issue:26413",
+            "right": "issue:31501",
+            "accept": false,
+            "reason": "Checkpoint resume failure and FineWeb startup failure are too vague and not clearly the same bug."
+          },
+          {
+            "left": "issue:34689",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Model-loading regression and embedding reinitialization on resize are different issues."
+          },
+          {
+            "left": "issue:45310",
+            "right": "issue:45478",
+            "accept": true,
+            "reason": "Exact same reported failure and same Qwen3.5 Moe from_pretrained bug; clear duplicate."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "A docs example problem and a position-id warning regression are unrelated."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Both are device/dtype mismatches, but they involve different models and different failure mechanisms."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Local-path import warning and GLM model-loading failure are not the same bug."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "Tokenizer protobuf fallback behavior and a missing utility symbol are unrelated."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both are v4/v5 behavior regressions, but they affect different tokenizers/models and likely different fixes."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Both are precision-related, but the failing code paths and model components are different."
+          },
+          {
+            "left": "issue:28218",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Tokenizer spacing behavior and Trainer broadcast TypeError are unrelated."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "A DSA indexer architectural comment and a v5 API removal breakage are not the same issue."
+          },
+          {
+            "left": "issue:45290",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Chat-template tool-call handling and Qwen2.5-VL video position IDs are distinct bugs."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The duplicate call on 45310/45478 is well grounded by the identical title and earlier timestamp, and the summary stays conservative about the rest of the cluster. The soft-edge rejections are appropriately cautious and do not overclaim shared root causes."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 29,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:10000",
+        "issue:15354",
+        "issue:25251",
+        "issue:27957",
+        "issue:29942",
+        "issue:32090",
+        "issue:33290",
+        "issue:33357",
+        "issue:34689",
+        "issue:35141",
+        "issue:37428",
+        "issue:38175",
+        "issue:40444",
+        "issue:41720",
+        "issue:42490",
+        "issue:43278",
+        "issue:43329",
+        "issue:43381",
+        "issue:43412",
+        "issue:43526",
+        "issue:43550",
+        "issue:43575",
+        "issue:43701",
+        "issue:43746",
+        "issue:43873",
+        "issue:43927",
+        "issue:43976",
+        "issue:44448",
+        "issue:44466",
+        "issue:44493",
+        "issue:44561",
+        "issue:44568",
+        "issue:44779",
+        "issue:44855",
+        "issue:44857",
+        "issue:44977",
+        "issue:44987",
+        "issue:45083",
+        "issue:45229",
+        "issue:45440",
+        "issue:45491",
+        "issue:45588",
+        "issue:45698",
+        "issue:45823",
+        "issue:45920",
+        "issue:46192",
+        "issue:46209",
+        "issue:46399"
+      ],
+      "soft_pairs": [
+        "issue:38175|issue:44493",
+        "issue:43701|issue:45083",
+        "issue:32090|issue:33357",
+        "issue:40444|issue:43526",
+        "issue:29942|issue:35141",
+        "issue:42490|issue:43575",
+        "issue:43550|issue:45083",
+        "issue:45920|issue:46192",
+        "issue:44448|issue:44977",
+        "issue:44855|issue:46399",
+        "issue:44977|issue:45229",
+        "issue:27957|issue:34689",
+        "issue:10000|issue:15354",
+        "issue:43278|issue:43329",
+        "issue:44857|issue:45588",
+        "issue:43746|issue:46209",
+        "issue:43927|issue:45698",
+        "issue:37428|issue:44987",
+        "issue:43278|issue:43381",
+        "issue:44561|issue:44779",
+        "issue:43873|issue:45229",
+        "issue:41720|issue:45229",
+        "issue:32090|issue:33290",
+        "issue:25251|issue:32090",
+        "issue:44448|issue:44466",
+        "issue:45588|issue:45823",
+        "issue:43976|issue:45440",
+        "issue:44568|issue:45491",
+        "issue:43412|issue:45588"
+      ],
+      "prepared_review_unit_hash": "157d6c9680e7a83d6d5dc45658a6ab63cc9df1e65813976be1de291900d0456a",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36525,
+        "estimated_input_tokens": 9132,
+        "estimated_eval_tokens": 18520
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "No soft pair appears to describe the same underlying bug/change; the set is a collection of unrelated issues across tokenizer, flash-attention, loading, training, and export paths. There are no PRs in the input. If a single representative issue must be chosen, #43873 is the broadest and most active report.",
+        "confidence": 0.59,
+        "canonical_issue_reason": "#43873 is the broadest open issue here, with multiple inbound references and discussion, and it describes a generic core-runtime problem rather than a model-specific one.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "#43873 is the best global representative because it is active, broadly framed, and has the strongest external signal among the listed issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:38175",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Different symptoms: zero probabilities in SigLIP2 vs unexpected position-id key warnings. Same general area only at a very high level."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "Checkpoint key mismatch during resume is unrelated to a qwen3_omni_moe helper-function behavior issue."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Trainer broadcast NoneType error and MacOS bus error are distinct failures with different triggers and code paths."
+          },
+          {
+            "left": "issue:40444",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "IterableDataset multi-image finetuning failure is unrelated to BeitImageProcessorFast reduce_labels behavior."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Flash Attention 2 test failures and token embedding reinitialization in post_init are different bugs."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Local-path trailing-slash import warning is unrelated to tensor-parallel OOM on loading Qwen2-57B."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:45083",
+            "accept": false,
+            "reason": "torch.compile/SDPA failure in Bamba is not the same as the qwen3_omni_moe feature-length helper issue."
+          },
+          {
+            "left": "issue:45920",
+            "right": "issue:46192",
+            "accept": false,
+            "reason": "Wrong token IDs in AutoTokenizer is a different bug from return_tensors being ignored when text_kwargs is passed."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Pegasus output regression and Qwen3.5 flash-attention generation failure are separate model-specific issues."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Both involve import/decorator parsing, but they affect different models and have different root causes."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Flash-attention generation corruption and multi-GPU CUDA OOM are unrelated failure modes."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "XLMRoberta Flash Attention 2 support and Llama 3.2 90B Vision loading breakage are distinct problems."
+          },
+          {
+            "left": "issue:10000",
+            "right": "issue:15354",
+            "accept": false,
+            "reason": "German DistilBertModel issue and GeneratorExp/torch.jit.script export failure are unrelated."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43329",
+            "accept": false,
+            "reason": "Embedding dtype drift in evaluate and undefined video-branch variables in multimodal token counting are different bugs."
+          },
+          {
+            "left": "issue:44857",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "float16 AMP crash in LwDetrImageLoss is unrelated to flash_attention.py crashing on s_aux=None."
+          },
+          {
+            "left": "issue:43746",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "PEFT local-checkpoint loading failure is not the same as FP8 inference failure on B200."
+          },
+          {
+            "left": "issue:43927",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Lost custom token IDs after save/load and wrong custom module resolution after save_pretrained are both serialization-related but not the same concrete bug."
+          },
+          {
+            "left": "issue:37428",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Missing _flash_supports_window_size import and failure loading physical-intelligence/fast are not clearly the same underlying issue."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "BF16-to-FP32 embedding dtype change is unrelated to gradient checkpointing being disallowed in eval mode."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Removal of is_torch_fx_available and DeepSeek tokenizer regression are unrelated v5 breakages."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Quantized offloading not working and Gemma4 multi-GPU OOM are both memory-related but not the same bug or code path."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Auto device mapping cudaErrorAssert and multi-GPU OOM are different runtime failures."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "NoneType GPU broadcast error and deepspeed Adafactor OOM are distinct training issues."
+          },
+          {
+            "left": "issue:25251",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Pipeline top_k nesting behavior and Trainer broadcast failure do not share a concrete code path."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:44466",
+            "accept": false,
+            "reason": "Different-output regression for Pegasus and lm_head.weight serialization inconsistency are related only loosely through v5 behavior changes."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "flash_attention.py s_aux=None crash and Gemma4 PLE device mismatch are separate Gemma-related bugs."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Python version incompatibility in Transformers 5.1.0 and DeepSeekV3MoE divergence from remote implementation are unrelated."
+          },
+          {
+            "left": "issue:44568",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Tokenizer BOS/EOS special-token regression and Gemma3 NaN embeddings under sliding-window attention are different issues."
+          },
+          {
+            "left": "issue:43412",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Executorch export failure for gemma3n and flash-attention attribute crash are different code paths."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 30,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:27957",
+        "issue:29127",
+        "issue:29942",
+        "issue:30064",
+        "issue:33357",
+        "issue:34567",
+        "issue:39692",
+        "issue:42490",
+        "issue:42915",
+        "issue:42994",
+        "issue:43065",
+        "issue:43066",
+        "issue:43278",
+        "issue:43381",
+        "issue:43421",
+        "issue:43575",
+        "issue:43611",
+        "issue:43653",
+        "issue:43701",
+        "issue:43756",
+        "issue:43873",
+        "issue:43994",
+        "issue:44186",
+        "issue:44206",
+        "issue:44265",
+        "issue:44361",
+        "issue:44442",
+        "issue:44462",
+        "issue:44479",
+        "issue:44743",
+        "issue:44871",
+        "issue:44977",
+        "issue:45198",
+        "issue:45245",
+        "issue:45372",
+        "issue:45406",
+        "issue:45412",
+        "issue:45440",
+        "issue:45459",
+        "issue:45491",
+        "issue:45563",
+        "issue:45588",
+        "issue:45593",
+        "issue:45712",
+        "issue:45803",
+        "issue:46121",
+        "issue:46369"
+      ],
+      "soft_pairs": [
+        "issue:44977|issue:45563",
+        "issue:42490|issue:43873",
+        "issue:45372|issue:45588",
+        "issue:45198|issue:45803",
+        "issue:33357|issue:39692",
+        "issue:29127|issue:34567",
+        "issue:43278|issue:43575",
+        "issue:43653|issue:44743",
+        "issue:45372|issue:45406",
+        "issue:44871|issue:44977",
+        "issue:43381|issue:43701",
+        "issue:43611|issue:46121",
+        "issue:44206|issue:44479",
+        "issue:42915|issue:42994",
+        "issue:44361|issue:45198",
+        "issue:44186|issue:45198",
+        "issue:43065|issue:45412",
+        "issue:43065|issue:43994",
+        "issue:29942|issue:30064",
+        "issue:45459|issue:45712",
+        "issue:45412|issue:46369",
+        "issue:43756|issue:45440",
+        "issue:45245|issue:45563",
+        "issue:43653|issue:44442",
+        "issue:43994|issue:45412",
+        "issue:45372|issue:45459",
+        "issue:44265|issue:45198",
+        "issue:27957|issue:29942",
+        "issue:43066|issue:44462",
+        "issue:29127|issue:33357",
+        "issue:45440|issue:45593",
+        "issue:43421|issue:45491"
+      ],
+      "prepared_review_unit_hash": "a273ac9d83eba9399289739683e4c5f27388a566a552b198a0162089c0b70762",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36680,
+        "estimated_input_tokens": 9170,
+        "estimated_eval_tokens": 18596
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are a heterogeneous mix of unrelated Transformers bugs across tokenizers, model loading, quantization, vision, and attention paths. The soft-similarity pairs mostly share a broad subsystem name, not the same underlying defect, so none should be merged as duplicates.",
+        "confidence": 0.61,
+        "canonical_issue_reason": "issue:43873 is the most active and broadly framed report in the set, with the most discussion and inbound references; it is the closest thing to a cluster anchor, though the overall set is not truly duplicate-tight.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43873 is the strongest representative issue here because it has the clearest ongoing user impact and the highest engagement among the listed issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44977",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Both involve generation behavior, but one is a Qwen3.5 flash-attention correctness bug and the other is a stale warning in paged generate(); different defects."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Local-path import warning and quantized offloading failures are separate loading/runtime issues with different triggers and fixes."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Gemma 4 processor import failure and flash_attention.py crashing on s_aux=None are unrelated code paths."
+          },
+          {
+            "left": "issue:45198",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Wav2Vec2 save/tokenization failure and a bare-except bug in FuyuBatchFeature are different tokenizer/data-handling problems."
+          },
+          {
+            "left": "issue:33357",
+            "right": "issue:39692",
+            "accept": false,
+            "reason": "MacOS CLIP bus error and SigLIP2 documentation errors are not the same bug; one is a runtime crash, the other is example/usage issues."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "LayoutLMv3 error-message clarity and TrainerState token counter updating are unrelated issues."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "A dtype change in evaluation and an OOM while loading Qwen2-57B with tensor parallelism are different failures."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration and Qwen3.5 recurrent-state reset are separate tokenizer vs model-state bugs."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Both mention Gemma 4, but one is an import dependency problem and the other is a missing _tokenizer attribute in serve; not the same underlying defect."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Gemma-3 EOS config mismatch and Qwen3.5 flash-attention generation failure are unrelated model-specific bugs."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode and resume_from_checkpoint key mismatch are different trainer behaviors."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:46121",
+            "accept": false,
+            "reason": "base_model_prefix loading breakage and JSON-loaded rope validation keys are unrelated config/serialization issues."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "LaserFeatureExtractor center-arg regression and Qwen video-input regression are separate version-specific breakages."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:42994",
+            "accept": false,
+            "reason": "FineGrainedFP8Config loading failure and quantized model saving failure are both quantization-related, but they affect different operations and code paths."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "MLukeTokenizer task AttributeError and Wav2Vec2 save/tokenization failure are distinct tokenizer bugs."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "LayoutLMv2 NER/padding crash and Wav2Vec2 save/tokenization failure do not share the same failure mode."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "A dummy Conv2d placeholder in Sam3PixelDecoder and an RT-DETR memory leak are unrelated."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Sam3PixelDecoder dummy layer and SigLIP2 nonsensical AutoModel/pipeline output are different problems."
+          },
+          {
+            "left": "issue:29942",
+            "right": "issue:30064",
+            "accept": false,
+            "reason": "Flash Attention 2 test failures and void segmentation-map processing are unrelated subsystem regressions."
+          },
+          {
+            "left": "issue:45459",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Tokenizer error masking when protobuf is absent and leftover dummy classes in dummy_pt_objects.py are unrelated maintenance bugs."
+          },
+          {
+            "left": "issue:45412",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "RT-DETR deletion leak and sam3-litetext memory overflow are different model-specific memory issues."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Smollm3 RoPE-layer mismatch and DeepseekV3MoE divergence from remote are different architecture/implementation discrepancies."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Category-count runtime limit and stale generate warning are unrelated."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration and AutoTokenizer loading FastSpeech2ConformerTokenizer are different tokenizer loading/registration issues."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:45412",
+            "accept": false,
+            "reason": "SigLIP2 incorrect outputs and RT-DETR memory release failure are unrelated model bugs."
+          },
+          {
+            "left": "issue:45372",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "Gemma 4 processor import breakage and protobuf-related tokenizer exception masking are different causes."
+          },
+          {
+            "left": "issue:44265",
+            "right": "issue:45198",
+            "accept": false,
+            "reason": "torch.export failure via torch_compilable_check and Wav2Vec2 save/tokenization failure are separate code paths."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:29942",
+            "accept": false,
+            "reason": "XLMRoberta Flash Attention 2 support and general Flash Attention 2 test failures are too broad to treat as the same concrete bug."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:44462",
+            "accept": false,
+            "reason": "Wrong tokenizer decoder type in v5 and AutoTokenizer ignoring tokenizer.json are different tokenizer loading semantics issues."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "LayoutLMv3 error-message wording and CLIP bus error on MacOS are unrelated."
+          },
+          {
+            "left": "issue:45440",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "DeepseekV3MoE divergence and D-FINE auxiliary-loss omission are different model implementation bugs."
+          },
+          {
+            "left": "issue:43421",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Runtime special-token post-processor updates and Gemma3 NaN embeddings with sliding-window attention 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": 21,
+      "review_unit_index": 31,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:27957",
+        "issue:29127",
+        "issue:33290",
+        "issue:34567",
+        "issue:34689",
+        "issue:35141",
+        "issue:42490",
+        "issue:42915",
+        "issue:43262",
+        "issue:43334",
+        "issue:43381",
+        "issue:43526",
+        "issue:43531",
+        "issue:43575",
+        "issue:43611",
+        "issue:43646",
+        "issue:43701",
+        "issue:43720",
+        "issue:43761",
+        "issue:43824",
+        "issue:43873",
+        "issue:43931",
+        "issue:43994",
+        "issue:44077",
+        "issue:44186",
+        "issue:44206",
+        "issue:44279",
+        "issue:44387",
+        "issue:44442",
+        "issue:44464",
+        "issue:44743",
+        "issue:44811",
+        "issue:44857",
+        "issue:44898",
+        "issue:44977",
+        "issue:45003",
+        "issue:45072",
+        "issue:45245",
+        "issue:45440",
+        "issue:45491",
+        "issue:45588",
+        "issue:45600",
+        "issue:45834",
+        "issue:45859",
+        "issue:46121",
+        "issue:46369",
+        "issue:46399"
+      ],
+      "soft_pairs": [
+        "issue:43526|issue:45834",
+        "issue:43575|issue:43701",
+        "issue:29127|issue:35141",
+        "issue:42490|issue:43381",
+        "issue:43994|issue:46369",
+        "issue:43575|issue:44387",
+        "issue:43526|issue:44186",
+        "issue:42915|issue:43931",
+        "issue:44442|issue:45072",
+        "issue:43824|issue:44279",
+        "issue:43526|issue:44857",
+        "issue:43334|issue:43531",
+        "issue:46121|issue:46399",
+        "issue:27957|issue:29127",
+        "issue:43761|issue:45859",
+        "issue:27957|issue:33290",
+        "issue:33290|issue:34689",
+        "issue:43262|issue:44206",
+        "issue:44977|issue:45245",
+        "issue:45588|issue:45600",
+        "issue:33290|issue:35141",
+        "issue:44464|issue:45491",
+        "issue:33290|issue:34567",
+        "issue:43381|issue:43873",
+        "issue:44743|issue:44811",
+        "issue:43646|issue:44077",
+        "issue:44442|issue:44898",
+        "issue:42490|issue:44743",
+        "issue:34689|issue:43611",
+        "issue:45003|issue:45440",
+        "issue:43720|issue:44442"
+      ],
+      "prepared_review_unit_hash": "f18fb9bce1a9addfb3d98b9390d982ae225fce5df381670ae0b4e9b43bed9c3a",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36306,
+        "estimated_input_tokens": 9077,
+        "estimated_eval_tokens": 18410
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a loose similarity bucket of unrelated Transformers issues. The soft edges are overwhelmingly false positives; no pair looks like the same underlying bug or change.",
+        "confidence": 0.95,
+        "canonical_issue_reason": "No single canonical issue stands out: the items span unrelated regressions across model loading, tokenizers, attention, quantization, and training behavior.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43873 is the strongest standalone anchor: it is open, broad in scope, and has the most discussion/inbound attention among the issues here.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43526",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Both are vision/OCR-related, but one is a BeitImageProcessorFast label-reduction bug and the other is a Kosmos2.5 OCR long-input index error."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "Different failure modes: tensor-parallel OOM on model load versus resume_from_checkpoint key mismatch."
+          },
+          {
+            "left": "issue:29127",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "LayoutLMv3 box-validation messaging and embedding resize reinitialization are unrelated bugs."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43381",
+            "accept": false,
+            "reason": "A path-with-slash import warning is unrelated to gradient checkpointing being disallowed in eval mode."
+          },
+          {
+            "left": "issue:43994",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Different models and symptoms: SigLIP2 bad outputs versus SAM3 LiteText memory overflow."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:44387",
+            "accept": false,
+            "reason": "Both mention OOM, but one is TP load OOM for Qwen2 and the other is int4 quantization reserved-memory growth."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44186",
+            "accept": false,
+            "reason": "Different processors/tokenizers and different crash surfaces; not the same underlying bug."
+          },
+          {
+            "left": "issue:42915",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "Qwen3Moe FP8 config failure and Qwen3-VL weight-shape mismatch are separate loading problems."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "Tokenizer loading failure and bfloat16 dtype mismatch are unrelated."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "A missing import for a specific class is not the same as a generic dependency issue."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "Beit image-label reduction and LwDetr loss AMP crashes are different code paths."
+          },
+          {
+            "left": "issue:43334",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Qwen3-VL missing pad_token_id and Qwen3-MoE sliding_window behavior are different regressions."
+          },
+          {
+            "left": "issue:46121",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "RoPE validation TypeError and Qwen3.5 kernel registration/decorator failure are unrelated."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "Flash Attention 2 support for XLM-Roberta is unrelated to LayoutLMv3 error messaging."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:45859",
+            "accept": false,
+            "reason": "CLIPVision hidden-states regression and Qwen3.5MoE CPU-offload incompatibility are unrelated."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "Flash-attention model support and adafactor/deepspeed OOM are not the same bug."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Both involve failures, but deepspeed OOM and Llama 3.2 Vision loading breakage have different causes."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44206",
+            "accept": false,
+            "reason": "Audio chat-template sample-rate defaulting and unsupported center arg in a feature extractor are distinct bugs."
+          },
+          {
+            "left": "issue:44977",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "Qwen3.5 flash-attention generation breakage is unrelated to the category-count limit error."
+          },
+          {
+            "left": "issue:45588",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Flash-attention sink-less crash and stale auto_mappings references are separate code paths."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Training OOM and output-embedding reinitialization after resize are unrelated."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45491",
+            "accept": false,
+            "reason": "Chunked generation inconsistency with compiled forward is not the same bug as Gemma3 NaNs from mixed-length sliding windows."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:34567",
+            "accept": false,
+            "reason": "OOM in deepspeed/adafactor is unrelated to TrainerState token-count tracking not updating."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode and quantization/offloading behavior are different feature regressions."
+          },
+          {
+            "left": "issue:44743",
+            "right": "issue:44811",
+            "accept": false,
+            "reason": "Qwen3_5 recurrent-state reset and Whisper batch_decode skip_special_tokens handling are unrelated."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "Both touch initialization semantics, but one is a broad custom-model init regression and the other is a patchtsmixer-specific post_init policy issue."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:44898",
+            "accept": false,
+            "reason": "Tokenizer loading and Perceiver image-classification interpolation are unrelated."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Import warnings from local path handling are unrelated to Qwen3_5 recurrent-state reset."
+          },
+          {
+            "left": "issue:34689",
+            "right": "issue:43611",
+            "accept": false,
+            "reason": "Both are model-loading regressions, but they affect different models and have different root causes."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Unsafe sys.modules access in modeling_utils and DeepseekV3MoE divergence from remote implementation are unrelated."
+          },
+          {
+            "left": "issue:43720",
+            "right": "issue:44442",
+            "accept": false,
+            "reason": "BitNet packed-weight loading under accelerate and tokenizer loading 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": 21,
+      "review_unit_index": 32,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:25251",
+        "issue:28282",
+        "issue:42175",
+        "issue:42617",
+        "issue:42796",
+        "issue:42907",
+        "issue:43232",
+        "issue:43316",
+        "issue:43388",
+        "issue:43493",
+        "issue:43525",
+        "issue:43550",
+        "issue:43575",
+        "issue:43643",
+        "issue:43701",
+        "issue:43723",
+        "issue:43749",
+        "issue:43761",
+        "issue:43827",
+        "issue:43873",
+        "issue:43931",
+        "issue:44062",
+        "issue:44077",
+        "issue:44079",
+        "issue:44188",
+        "issue:44190",
+        "issue:44442",
+        "issue:44464",
+        "issue:44479",
+        "issue:44496",
+        "issue:44855",
+        "issue:44871",
+        "issue:44913",
+        "issue:44987",
+        "issue:45137",
+        "issue:45216",
+        "issue:45290",
+        "issue:45479",
+        "issue:45698",
+        "issue:45759",
+        "issue:45803",
+        "issue:45907",
+        "issue:46178",
+        "issue:46234",
+        "issue:46291",
+        "issue:46421",
+        "issue:46506"
+      ],
+      "soft_pairs": [
+        "issue:43316|issue:43525",
+        "issue:42175|issue:46234",
+        "issue:44062|issue:45907",
+        "issue:43723|issue:45479",
+        "issue:43761|issue:44079",
+        "issue:44077|issue:44479",
+        "issue:43493|issue:43827",
+        "issue:43761|issue:44188",
+        "issue:44987|issue:46178",
+        "issue:43873|issue:44871",
+        "issue:43232|issue:43388",
+        "issue:43749|issue:43931",
+        "issue:43931|issue:45216",
+        "issue:44442|issue:45803",
+        "issue:43316|issue:45759",
+        "issue:44855|issue:45137",
+        "issue:42907|issue:45698",
+        "issue:43643|issue:44913",
+        "issue:43701|issue:46506",
+        "issue:44190|issue:44496",
+        "issue:44464|issue:46421",
+        "issue:42796|issue:45290",
+        "issue:42617|issue:43575",
+        "issue:25251|issue:28282",
+        "issue:46234|issue:46291",
+        "issue:43550|issue:43701"
+      ],
+      "prepared_review_unit_hash": "c9c27f15106afdb811394b423937334b7897c4fe034b36a2f9a0aaaa26697eec",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 26,
+        "serialized_chars": 35091,
+        "estimated_input_tokens": 8773,
+        "estimated_eval_tokens": 17802
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a loose similarity group of mostly unrelated issues across configs, tokenizers, generation, loading, docs, and quantization. None of the soft pairs look like the same underlying bug, so I would not merge any of them as duplicates.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "issue:43761 is one of the clearest, self-contained regression reports in the set and is better suited as a representative issue than the more generic or environment-specific reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43761 has a concrete, well-scoped failure mode (`output_hidden_states=True` returning `None`) that is easy to understand and triage.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43316",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Both are config/API mismatches, but one is about `Gemma3TextConfig` field parity and the other is a missing `pad_token_id` on `Llama4Config`; different models and different missing attributes."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "These concern different backend/import availability problems: one is about TensorFlow not being included with a torch install, the other about a missing `is_falcon_mamba_ssm_available` helper."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "`AddedToken(special=...)` duplication during tokenizer loading is unrelated to `list_repo_files` failing to forward `revision` in hub file listing."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Tokenizer loading in v5 and degenerate zero loss in sequence classification are unrelated subsystems and failure modes."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "One is a CLIP vision hidden-states regression, the other is a `ModelOutput` key assignment bug; same area of outputs but different code paths and symptoms."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "`patchtsmixer` optional `post_init` and a video input regression for Qwen variants are model-specific, unrelated issues."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:43827",
+            "accept": false,
+            "reason": "A SigLIP2 implementation discrepancy is not the same bug as stale docs referencing removed `pipeline()` APIs."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44188",
+            "accept": false,
+            "reason": "CLIP hidden-states regression and attention-kernel divergence under `torch.compile` affect different code paths and outputs."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "Both mention loading/hub behavior, but one is about a specific model failing on `transformers>=5.1.0` while the other is a backend registry error for `tensorflow_text`."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Quantization/offloading behavior is unrelated to the Gemma-3 EOS token config mismatch."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:43388",
+            "accept": false,
+            "reason": "A generation kwarg update issue after `sync_gpus` is unrelated to `gather_for_metrics` dropping tuple labels in the last batch."
+          },
+          {
+            "left": "issue:43749",
+            "right": "issue:43931",
+            "accept": false,
+            "reason": "FSDP CPU RAM efficient loading and Qwen3-VL weight-shape mismatch are different loading failures with different root causes."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both are model loading/saving issues, but one is a shape mismatch on load and the other is a checkpoint corruption/regression on save."
+          },
+          {
+            "left": "issue:44442",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "AutoTokenizer failing for one custom tokenizer is not the same as Fuyu swallowing exceptions in tensor conversion."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "These both involve config handling, but one is a discrepancy in `Gemma3TextConfig` APIs and the other is `AutoModelForCausalLM.from_config` not unwrapping `text_config` for multimodal Qwen configs."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "A Python 3.13 indentation/script parsing error in DebertaV2Model is unrelated to a DeepSpeed ZeRO3 deque underflow."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Saving dequantized Ministral/Devstral models and loading the wrong custom module after save/load are both persistence-related, but they are different failures."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Missing fields with `trust_remote_code=True` and `rotary_pct` resetting on reload are different config deserialization bugs."
+          },
+          {
+            "left": "issue:43701",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "A resume-from-checkpoint key mismatch is unrelated to CPU/CUDA result differences for PPDocLayoutV3."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Local dataset loading in an example script and an unrecognized custom model/config are unrelated."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:46421",
+            "accept": false,
+            "reason": "Chunked generation inconsistency with compiled forward and cache lazy initialization causing recompilation are distinct generation/runtime-performance bugs."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Tokenizer BPE marker handling and chat template crashes on tool-call assistant messages are different tokenizer/chat-template issues."
+          },
+          {
+            "left": "issue:42617",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Running `3d_parallel.py` and OOM loading a large Qwen model with tensor parallelism are both large-model execution problems, but not the same underlying bug."
+          },
+          {
+            "left": "issue:25251",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Top-k pipeline output nesting and an AutoModel PyTorch ImportError are completely unrelated."
+          },
+          {
+            "left": "issue:46234",
+            "right": "issue:46291",
+            "accept": false,
+            "reason": "A missing import-utils availability symbol and a kernels import requiring version/revision are different import/package integration issues."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "A torch.compile + SDPA failure in Bamba is unrelated to a checkpoint resume key mismatch."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 33,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:16998",
+        "issue:22355",
+        "issue:27957",
+        "issue:32090",
+        "issue:41950",
+        "issue:42175",
+        "issue:43065",
+        "issue:43295",
+        "issue:43316",
+        "issue:43388",
+        "issue:43404",
+        "issue:43441",
+        "issue:43452",
+        "issue:43475",
+        "issue:43540",
+        "issue:43572",
+        "issue:43575",
+        "issue:43606",
+        "issue:43646",
+        "issue:43650",
+        "issue:43688",
+        "issue:43701",
+        "issue:43716",
+        "issue:43742",
+        "issue:43761",
+        "issue:43825",
+        "issue:44077",
+        "issue:44162",
+        "issue:44479",
+        "issue:44496",
+        "issue:44514",
+        "issue:44534",
+        "issue:44610",
+        "issue:44617",
+        "issue:44792",
+        "issue:44871",
+        "issue:44913",
+        "issue:45584",
+        "issue:45632",
+        "issue:45803",
+        "issue:46050",
+        "issue:46170",
+        "issue:46178",
+        "issue:46202",
+        "issue:46365",
+        "issue:46369",
+        "issue:46424"
+      ],
+      "soft_pairs": [
+        "issue:27957|issue:32090",
+        "issue:43575|issue:43742",
+        "issue:43295|issue:43825",
+        "issue:43606|issue:43701",
+        "issue:16998|issue:22355",
+        "issue:43295|issue:44871",
+        "issue:43316|issue:43716",
+        "issue:44871|issue:45584",
+        "issue:42175|issue:43316",
+        "issue:45803|issue:46365",
+        "issue:44610|issue:46369",
+        "issue:44871|issue:46050",
+        "issue:45632|issue:46424",
+        "issue:44077|issue:44534",
+        "issue:43646|issue:44913",
+        "issue:44162|issue:44792",
+        "issue:43452|issue:44496",
+        "issue:43065|issue:44617",
+        "issue:41950|issue:43475",
+        "issue:45803|issue:46170",
+        "issue:43540|issue:44514",
+        "issue:43650|issue:46202",
+        "issue:43404|issue:43441",
+        "issue:43761|issue:44479",
+        "issue:43572|issue:46178",
+        "issue:43388|issue:43688"
+      ],
+      "prepared_review_unit_hash": "67e6e6e17f7e6cd5898e5b366cb9e2455d298470931ed209e7c7b083f5b0e0fd",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 26,
+        "serialized_chars": 34651,
+        "estimated_input_tokens": 8663,
+        "estimated_eval_tokens": 17582
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is a grab-bag of unrelated issues across different models, pipelines, configs, and training utilities. None of the soft pairs look like the same underlying bug or a mergeable duplicate PR case.",
+        "confidence": 0.92,
+        "canonical_issue_reason": "No single canonical duplicate stands out because these reports are largely unrelated; if forced to pick one, issue 43388 is the cleanest standalone bug report with a concrete repro and inbound references.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45584 is the strongest standalone triage target overall: it is a clear regression, has a concrete failure mode, and already has multiple inbound references.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:27957",
+            "right": "issue:32090",
+            "accept": false,
+            "reason": "Different failures in different code paths: FlashAttention support for XLMRoberta vs Trainer GPU broadcast NoneType error."
+          },
+          {
+            "left": "issue:43575",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "OOM while loading a 57B Qwen model is unrelated to a KeyError loading MobileLLM-125M."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Processor/tokenizer regression and pipeline translation messaging are different bugs in different areas."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:43701",
+            "accept": false,
+            "reason": "CPU offload device mismatch is unrelated to a resume_from_checkpoint key mismatch."
+          },
+          {
+            "left": "issue:16998",
+            "right": "issue:22355",
+            "accept": false,
+            "reason": "A model_max_length question for DeBERTa-V3 is unrelated to a missing transformers.onnx import."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Custom processor/tokenizer regression does not match Gemma eos_token_id configuration inconsistency."
+          },
+          {
+            "left": "issue:43316",
+            "right": "issue:43716",
+            "accept": false,
+            "reason": "Gemma3TextConfig API discrepancy and Mistral-3 dtype mismatch affect different models and mechanisms."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Gemma eos_token_id config mismatch is unrelated to Whisper empty-transcription failure after align_special_tokens."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43316",
+            "accept": false,
+            "reason": "Missing tensorflow in the torch extra is a packaging issue, not the Gemma3TextConfig API inconsistency."
+          },
+          {
+            "left": "issue:45803",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Fuyu tensor conversion exception handling is unrelated to importing Trainer failing because of datasets."
+          },
+          {
+            "left": "issue:44610",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "OmDet-Turbo processor image size mismatch and Sam3-litetext memory overflow are different model issues."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Gemma eos_token_id mismatch and sinq quantization save/load breakage do not share the same failure mode."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Remote-code cache path collisions are unrelated to max_cache_length being honored only for quantized cache."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44534",
+            "accept": false,
+            "reason": "PatchTSMixer post_init validation and non-persistent buffer junk in v5 are separate regression classes."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Custom model initialization breakage is not the same as GPTNeoX rotary_pct not persisting on reload."
+          },
+          {
+            "left": "issue:44162",
+            "right": "issue:44792",
+            "accept": false,
+            "reason": "ESM2 breakage affecting workflows and a janus test failure are not the same bug."
+          },
+          {
+            "left": "issue:43452",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "gguf_file loading failures and an unrecognized model_type config error are distinct loading problems."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Sam3PixelDecoder dummy Conv2d and Sam3Video CUDA OOM are different issues; one is code structure, the other capacity."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Video-classification pipeline processor lookup bug is unrelated to a missing fpn_position_embeddings attribute in Sam3 output."
+          },
+          {
+            "left": "issue:45803",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Bare except in Fuyu tensor conversion is unrelated to Windows ROCm import/type errors from torch.distributed."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "Qwen3OmniMoe video input ValueError and Qwen2_5_VLProcessor chat-template batching failure are different code paths."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "This is a placeholder/nonsensical issue, not the same as a Gemma pipeline roundtrip failure."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Mistral-3 lm_head tying and Ministral-3 FlashAttention failures are separate model-specific bugs."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44479",
+            "accept": false,
+            "reason": "CLIPVisionModel hidden_states regression is unrelated to the qwen video input regression in v5.3.0."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "StableLmConfig pad_token_idx regression and a tensorflow_text backend mapping error are unrelated."
+          },
+          {
+            "left": "issue:43388",
+            "right": "issue:43688",
+            "accept": false,
+            "reason": "Metric label truncation in gather_for_metrics is a different problem from auxiliary-loss normalization in OLMoE/GPT Oss."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 34,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:25251",
+        "issue:27957",
+        "issue:35141",
+        "issue:41720",
+        "issue:42490",
+        "issue:43065",
+        "issue:43335",
+        "issue:43504",
+        "issue:43525",
+        "issue:43531",
+        "issue:43644",
+        "issue:43673",
+        "issue:43761",
+        "issue:43827",
+        "issue:43931",
+        "issue:43994",
+        "issue:44060",
+        "issue:44190",
+        "issue:44242",
+        "issue:44479",
+        "issue:44534",
+        "issue:44610",
+        "issue:44617",
+        "issue:44734",
+        "issue:44743",
+        "issue:44792",
+        "issue:44855",
+        "issue:44913",
+        "issue:44933",
+        "issue:44964",
+        "issue:44977",
+        "issue:45127",
+        "issue:45200",
+        "issue:45245",
+        "issue:45362",
+        "issue:45381",
+        "issue:45406",
+        "issue:45542",
+        "issue:45563",
+        "issue:45584",
+        "issue:45588",
+        "issue:45600",
+        "issue:45632",
+        "issue:45698",
+        "issue:46234",
+        "issue:46369",
+        "issue:46421",
+        "issue:46617"
+      ],
+      "soft_pairs": [
+        "issue:45127|issue:45200",
+        "issue:45542|issue:46234",
+        "issue:43504|issue:45698",
+        "issue:43673|issue:44743",
+        "issue:44792|issue:44977",
+        "issue:43525|issue:43531",
+        "issue:44792|issue:45600",
+        "issue:44617|issue:45584",
+        "issue:44734|issue:45406",
+        "issue:44534|issue:44913",
+        "issue:44190|issue:44964",
+        "issue:35141|issue:42490",
+        "issue:45632|issue:46617",
+        "issue:45381|issue:46369",
+        "issue:43931|issue:44060",
+        "issue:44855|issue:45588",
+        "issue:43673|issue:46421",
+        "issue:43761|issue:44242",
+        "issue:45563|issue:45584",
+        "issue:43827|issue:43994",
+        "issue:44479|issue:44734",
+        "issue:43065|issue:44610",
+        "issue:43644|issue:44913",
+        "issue:44933|issue:45362",
+        "issue:41720|issue:45245",
+        "issue:25251|issue:27957",
+        "issue:43335|issue:44855"
+      ],
+      "prepared_review_unit_hash": "d2d2de85acd46b2a7c88141176a327ad2cc76cf421b8205a480e87e67682cb8d",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 35751,
+        "estimated_input_tokens": 8938,
+        "estimated_eval_tokens": 18132
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a similarity bucket of unrelated issues. The soft pairs share broad subsystem words (cache, loading, model init, generation), but none look like the same underlying bug/change closely enough to merge.",
+        "confidence": 0.39,
+        "canonical_issue_reason": "Issue 45698 is the strongest standalone candidate: it is open, has the most discussion/inbound references, and describes a concrete, user-facing `from_pretrained`/`save_pretrained` deserialization bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45698 is the best representative of the cluster because it is the most actively discussed and broadly impactful concrete bug among the listed issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45127",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Different subsystems and failure modes: LoRA vocab merge/tied embeddings vs Gemma4 multimodal token ids defaulting."
+          },
+          {
+            "left": "issue:45542",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "Unrelated backend-integration bug vs missing import utility symbol; only superficial TensorFlow/availability similarity."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:45698",
+            "accept": false,
+            "reason": "Different code paths: legacy field handling in BEiT preset loading vs wrong custom module resolution after save/load."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44743",
+            "accept": false,
+            "reason": "Both mention cache, but one is missing cache state in v5 chunked prefill while the other resets recurrent states in a specific modular model."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "Different models and bugs: janus image-generation test failure vs Qwen3.5 flash-attention generation issue."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43531",
+            "accept": false,
+            "reason": "Llama4 missing config attribute is unrelated to Qwen3-MoE sliding-window behavior."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Test failure for janus generation is unrelated to CI breakage from removed Sam3LiteText config mappings."
+          },
+          {
+            "left": "issue:44617",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "CUDA OOM in Sam3Video is not the same bug as Whisper failing on empty transcription after token alignment."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "KV-cache continuation tensor indexing crash is unrelated to Gemma4Processor missing `_tokenizer`."
+          },
+          {
+            "left": "issue:44534",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Both are reload/model-state issues, but one concerns non-persistent buffers and the other config serialization for `rotary_pct`."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different loading paths and targets: local dataset script vs multimodal model loading."
+          },
+          {
+            "left": "issue:35141",
+            "right": "issue:42490",
+            "accept": false,
+            "reason": "Embedding reinitialization on resize is unrelated to path normalization/import warning when loading local models."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "Both involve trust_remote_code/cache, but one is a leaf-directory collision and the other is a cache FileNotFoundError; not clearly the same bug."
+          },
+          {
+            "left": "issue:45381",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Different models and symptoms: Qwen2.5-VL vision position IDs vs Sam3-LiteText memory overflow."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44060",
+            "accept": false,
+            "reason": "Distinct Qwen variants with different failure classes: weight shape mismatch vs erroneous tied-weights warning."
+          },
+          {
+            "left": "issue:44855",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Python 3.13 decorator parsing/import error is unrelated to flash-attention crashing on `s_aux=None`."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:46421",
+            "accept": false,
+            "reason": "Both involve cache and chunked prefill, but one is missing cache initialization entirely while the other causes recompilation due to lazy init."
+          },
+          {
+            "left": "issue:43761",
+            "right": "issue:44242",
+            "accept": false,
+            "reason": "Different bugs in different models: hidden_states omission vs load-balancing loss not being added when router logits are disabled."
+          },
+          {
+            "left": "issue:45563",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Paged-generate warning cleanup is unrelated to Whisper empty-transcription generation failure."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Docs cleanup after pipeline removals is not the same as a model producing nonsensical outputs under AutoModel/pipeline."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Video-input regression in Qwen VL models is unrelated to serve-side KV-cache continuation indexing."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:44610",
+            "accept": false,
+            "reason": "Dummy Conv2d implementation detail is unrelated to OmDet-Turbo processor/model input-size mismatch."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Global v5 buffer-initialization regression is unrelated to GPTNeoX config values reverting on reload."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "Nonexistent import from image_utils is unrelated to Qwen3.5-35B chat crash."
+          },
+          {
+            "left": "issue:41720",
+            "right": "issue:45245",
+            "accept": false,
+            "reason": "Auto device-map CUDA assert is a model execution issue, not the same as category-cardinality limit overflow."
+          },
+          {
+            "left": "issue:25251",
+            "right": "issue:27957",
+            "accept": false,
+            "reason": "Top-k output nesting in pipelines is unrelated to XLMRoberta Flash Attention 2 support."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "SwitchTransformers sparse-layer creation and Python 3.13 import parsing are completely different bugs."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 35,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:36246",
+        "issue:38617",
+        "issue:41628",
+        "issue:42175",
+        "issue:42796",
+        "issue:43116",
+        "issue:43295",
+        "issue:43335",
+        "issue:43540",
+        "issue:43550",
+        "issue:43643",
+        "issue:43854",
+        "issue:44077",
+        "issue:44112",
+        "issue:44190",
+        "issue:44246",
+        "issue:44279",
+        "issue:44462",
+        "issue:44464",
+        "issue:44496",
+        "issue:44521",
+        "issue:44617",
+        "issue:44683",
+        "issue:44734",
+        "issue:44737",
+        "issue:44779",
+        "issue:44877",
+        "issue:44938",
+        "issue:44964",
+        "issue:45003",
+        "issue:45127",
+        "issue:45245",
+        "issue:45276",
+        "issue:45357",
+        "issue:45381",
+        "issue:45446",
+        "issue:45542",
+        "issue:45563",
+        "issue:45584",
+        "issue:45600",
+        "issue:45632",
+        "issue:45803",
+        "issue:45907",
+        "issue:46082",
+        "issue:46121",
+        "issue:46234",
+        "issue:46369",
+        "issue:46489"
+      ],
+      "soft_pairs": [
+        "issue:43550|issue:44190",
+        "issue:44464|issue:45563",
+        "issue:44617|issue:46369",
+        "issue:44077|issue:44877",
+        "issue:43335|issue:44190",
+        "issue:45127|issue:45357",
+        "issue:44521|issue:45381",
+        "issue:43540|issue:44734",
+        "issue:43116|issue:44190",
+        "issue:44779|issue:46489",
+        "issue:44112|issue:45600",
+        "issue:44737|issue:45803",
+        "issue:43295|issue:45276",
+        "issue:44938|issue:45542",
+        "issue:44462|issue:45907",
+        "issue:45276|issue:45584",
+        "issue:45632|issue:46234",
+        "issue:42796|issue:44521",
+        "issue:36246|issue:41628",
+        "issue:36246|issue:38617",
+        "issue:46082|issue:46121",
+        "issue:44246|issue:45003",
+        "issue:42175|issue:44279",
+        "issue:45245|issue:45584",
+        "issue:43643|issue:44496",
+        "issue:44964|issue:45600",
+        "issue:44112|issue:44964",
+        "issue:43854|issue:44964",
+        "issue:44683|issue:45446"
+      ],
+      "prepared_review_unit_hash": "fc2f5483003505445a74d4a35cfa810a6f5a6b75b6c8cbe82f72a4f8c70d15f5",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36506,
+        "estimated_input_tokens": 9127,
+        "estimated_eval_tokens": 18510
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous; the only pair that looks like the same underlying bug is the DeepSeek tokenizer regression. No pull requests are present in the items.",
+        "confidence": 0.71,
+        "canonical_issue_reason": "issue:46489 is the better canonical issue: it is the more specific DeepSeek-Coder v1 v5+ tokenizer regression and explicitly ties the problem to a gap in PR #44801, making it the clearest actionable duplicate target.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:46489 is the best representative issue for the cluster because it precisely identifies the failing tokenizer path and the version boundary, unlike the broader companion report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43550",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Different bug classes: torch.compile/SDPA model failure vs dataset loading in an example script."
+          },
+          {
+            "left": "issue:44464",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Unrelated generation issues: compiled forward inconsistency vs a stale warning about num_return_sequences."
+          },
+          {
+            "left": "issue:44617",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Same broad SAM3 area, but different concrete failures and likely different code paths/models."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44877",
+            "accept": false,
+            "reason": "Both are config-related, but they concern different models and different validation behavior."
+          },
+          {
+            "left": "issue:43335",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Completely different subsystems: SwitchTransformers config bug vs example data loading."
+          },
+          {
+            "left": "issue:45127",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different features and code paths: LoRA/vocab merging vs save_pretrained visual encoder keys."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "Both multimodal, but one is assistant mask generation and the other is video vision_position_ids; not the same bug."
+          },
+          {
+            "left": "issue:43540",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Different failures: Qwen3OmniMoe video input validation vs server KV-cache continuation indexing."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:44190",
+            "accept": false,
+            "reason": "Different example scripts and different symptoms; no shared underlying bug."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:46489",
+            "accept": true,
+            "reason": "Both report DeepSeek tokenizer corruption on transformers v5+, with the same underlying regression and overlapping fix scope."
+          },
+          {
+            "left": "issue:44112",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Both broke CI, but due to unrelated changes in GraniteSpeech tests vs removed Sam3LiteText auto-mapping references."
+          },
+          {
+            "left": "issue:44737",
+            "right": "issue:45803",
+            "accept": false,
+            "reason": "Different code paths: XLNet positional encoding device placement vs Fuyu tensor conversion exception handling."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "Both involve model/processor APIs, but one is a processor-tokenizer regression and the other is embedding resize propagation."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Import/startup issues, but one is Python 3.14 compatibility and the other is a missing TensorFlow backend dependency."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Repository/file listing behavior is unrelated to AutoTokenizer ignoring tokenizer.json."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Different functionality: token embedding resizing vs Whisper empty-transcription generation failure."
+          },
+          {
+            "left": "issue:45632",
+            "right": "issue:46234",
+            "accept": false,
+            "reason": "Unrelated issues: cache-path collision for trust_remote_code vs a missing import_utils availability helper."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Tokenizer BPE marker handling is not the same as multimodal assistant mask generation."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and different modules; not one concrete bug."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Different missing imports in different modules, with no shared code-path evidence."
+          },
+          {
+            "left": "issue:46082",
+            "right": "issue:46121",
+            "accept": false,
+            "reason": "Separate config-validation bugs: head_dim rejection vs rope validation type handling."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Both touch import behavior, but one is performance and the other is unsafe sys.modules access."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "Broad dependency/package issues, but not the same missing-backend problem."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Unrelated runtime/category limit error vs Whisper generation bug."
+          },
+          {
+            "left": "issue:43643",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Both involve config loading, but one is trust_remote_code field loss and the other is an unrecognized model/config shape."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Both are model-loading related, but they concern different models and different breakages."
+          },
+          {
+            "left": "issue:44112",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "CI test failure vs model loading regression; different artifacts and code paths."
+          },
+          {
+            "left": "issue:43854",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Different models and different failure modes; not a duplicate."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Same module area (flex_attention), but one is compiled flex_attention on torch>=2.9 and the other is a version-check/import gating bug."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: all items are issues (no PRs), the cluster is mostly heterogeneous, and the DeepSeek tokenizer pair is the only clearly related duplicate-style match. The soft-edge verdicts are conservative and the accepted pair is supported by the titles."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 36,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:27957",
+        "issue:28282",
+        "issue:32090",
+        "issue:35141",
+        "issue:41950",
+        "issue:42907",
+        "issue:43295",
+        "issue:43299",
+        "issue:43381",
+        "issue:43404",
+        "issue:43441",
+        "issue:43479",
+        "issue:43526",
+        "issue:43531",
+        "issue:43650",
+        "issue:43673",
+        "issue:43824",
+        "issue:43901",
+        "issue:43976",
+        "issue:44016",
+        "issue:44062",
+        "issue:44075",
+        "issue:44186",
+        "issue:44220",
+        "issue:44279",
+        "issue:44295",
+        "issue:44336",
+        "issue:44368",
+        "issue:44496",
+        "issue:44734",
+        "issue:44792",
+        "issue:44933",
+        "issue:44987",
+        "issue:45042",
+        "issue:45083",
+        "issue:45216",
+        "issue:45335",
+        "issue:45375",
+        "issue:45584",
+        "issue:45588",
+        "issue:45593",
+        "issue:45600",
+        "issue:45759",
+        "issue:46123",
+        "issue:46369",
+        "issue:46424",
+        "issue:46506",
+        "issue:46515"
+      ],
+      "soft_pairs": [
+        "issue:41950|issue:44016",
+        "issue:44186|issue:44734",
+        "issue:27957|issue:35141",
+        "issue:43404|issue:45216",
+        "issue:44295|issue:46506",
+        "issue:43650|issue:44220",
+        "issue:43673|issue:46424",
+        "issue:44016|issue:44220",
+        "issue:43479|issue:45375",
+        "issue:32090|issue:35141",
+        "issue:28282|issue:35141",
+        "issue:43531|issue:45600",
+        "issue:45083|issue:46506",
+        "issue:43976|issue:44987",
+        "issue:43295|issue:45335",
+        "issue:43526|issue:43901",
+        "issue:44496|issue:45042",
+        "issue:42907|issue:43441",
+        "issue:43673|issue:44336",
+        "issue:43381|issue:45593",
+        "issue:44016|issue:44062",
+        "issue:43441|issue:45588",
+        "issue:43901|issue:44496",
+        "issue:43976|issue:44279",
+        "issue:46123|issue:46369",
+        "issue:44792|issue:46424",
+        "issue:44496|issue:45759",
+        "issue:43824|issue:44933",
+        "issue:43299|issue:45216",
+        "issue:44279|issue:46515",
+        "issue:45335|issue:45584",
+        "issue:44075|issue:44368"
+      ],
+      "prepared_review_unit_hash": "cdfabc1bc0d819805fad51030f4171417bc0ff0eaa9e1cab5c19c1c371c671f2",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37244,
+        "estimated_input_tokens": 9311,
+        "estimated_eval_tokens": 18878
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous: it mixes unrelated issues across model loading, tokenization, generation, training, docs, and dependency problems. None of the soft pairs look like the same underlying bug or change.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No single issue is a good canonical representative because the cluster spans many distinct failure modes and subsystems rather than one duplicate report.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No best issue stands out as a cluster representative; the items are too diverse and do not converge on one concrete bug.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41950",
+            "right": "issue:44016",
+            "accept": false,
+            "reason": "Video-classification processor selection and a notebook syntax error are unrelated."
+          },
+          {
+            "left": "issue:44186",
+            "right": "issue:44734",
+            "accept": false,
+            "reason": "Tokenizer crashes in NER/padding are a different code path from serving KV-cache continuation indexing."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Flash Attention support for XLM-Roberta is unrelated to embedding reinitialization after resize."
+          },
+          {
+            "left": "issue:43404",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both involve weight tying/saving, but they affect different models and different failure modes; not the same bug."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "Reading position_ids after buffer registration is unrelated to CPU/CUDA divergence in PPDocLayoutV3."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "A vague data request is not the same as an audio feature extraction bug."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Both mention cache, but one is chunked prefill GenerationMixin cache absence and the other is cache_length handling for quantized cache."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Notebook syntax and `_torch_extract_fbank_features()` are unrelated."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Different config-initialization problems in different multimodal model families."
+          },
+          {
+            "left": "issue:32090",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Trainer broadcast TypeError is unrelated to token embedding resizing and post-init reinitialization."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Missing PyTorch import is an environment issue, not the same as embedding tying logic."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding-window behavior and auto_mappings/CI breakage are different issues."
+          },
+          {
+            "left": "issue:45083",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "Different model-specific numeric behavior problems with no shared code path."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Python-version compatibility and model-loading failure are unrelated."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "Both touch embeddings/API regressions, but one is processor/tokenizer behavior and the other is decoder.embed_tokens not updating."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Image-processor label reduction and text-classification docs mismatch are not the same bug."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45042",
+            "accept": false,
+            "reason": "Unrecognized model config and an unnecessary torchvision dependency are unrelated."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Saving dequantized models and FlashAttention inference failures are different code paths."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "Generation cache handling and ANSI logging output are unrelated."
+          },
+          {
+            "left": "issue:43381",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Gradient checkpointing in eval mode and missing auxiliary losses are different training behaviors."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44062",
+            "accept": false,
+            "reason": "Notebook syntax error and AddedToken argument duplication are unrelated."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Both mention flash attention, but they fail in different models and different code paths."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Docs wording for return_all_scores has nothing to do with model config import failure."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44279",
+            "accept": false,
+            "reason": "These are generic dependency/version complaints about different packages and symptoms."
+          },
+          {
+            "left": "issue:46123",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "A partial-batch pipeline bug is unrelated to a memory overflow in sam3-litetext."
+          },
+          {
+            "left": "issue:44792",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "A janus test failure is unrelated to cache_length semantics."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both are model-loading/config issues, but one is an unknown model_type error and the other is text_config unwrapping for composite configs."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Different import failures involving different modules and symbols."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both involve Qwen models, but one is a loading regression and the other is an incorrect save_pretrained checkpoint."
+          },
+          {
+            "left": "issue:44279",
+            "right": "issue:46515",
+            "accept": false,
+            "reason": "Different dependency incompatibilities: one is vague transformer dependency trouble, the other is torch 2.4.1 incompatibility."
+          },
+          {
+            "left": "issue:45335",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "Token embedding resize behavior is unrelated to Whisper generation after align_special_tokens."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "SGD optimizer args being ignored is unrelated to a tie_word_embeddings warning in ms-swift."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 37,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:38175",
+        "issue:39692",
+        "issue:41950",
+        "issue:42796",
+        "issue:42994",
+        "issue:43295",
+        "issue:43299",
+        "issue:43317",
+        "issue:43441",
+        "issue:43479",
+        "issue:43525",
+        "issue:43577",
+        "issue:43650",
+        "issue:43653",
+        "issue:43854",
+        "issue:43901",
+        "issue:43976",
+        "issue:44016",
+        "issue:44242",
+        "issue:44279",
+        "issue:44291",
+        "issue:44351",
+        "issue:44462",
+        "issue:44466",
+        "issue:44521",
+        "issue:44617",
+        "issue:44683",
+        "issue:44779",
+        "issue:45020",
+        "issue:45081",
+        "issue:45103",
+        "issue:45200",
+        "issue:45216",
+        "issue:45276",
+        "issue:45278",
+        "issue:45325",
+        "issue:45357",
+        "issue:45372",
+        "issue:45405",
+        "issue:45542",
+        "issue:45593",
+        "issue:45812",
+        "issue:45823",
+        "issue:45834",
+        "issue:46082",
+        "issue:46209",
+        "issue:46399",
+        "issue:46512"
+      ],
+      "soft_pairs": [
+        "issue:42796|issue:43650",
+        "issue:46082|issue:46399",
+        "issue:44279|issue:45405",
+        "issue:41950|issue:45020",
+        "issue:43577|issue:45216",
+        "issue:41950|issue:45823",
+        "issue:45103|issue:45372",
+        "issue:44351|issue:45278",
+        "issue:43295|issue:44779",
+        "issue:39692|issue:43901",
+        "issue:43299|issue:45357",
+        "issue:38175|issue:44617",
+        "issue:44683|issue:45542",
+        "issue:44242|issue:45200",
+        "issue:43976|issue:45405",
+        "issue:46082|issue:46512",
+        "issue:43650|issue:46209",
+        "issue:44779|issue:45812",
+        "issue:43299|issue:43854",
+        "issue:43479|issue:43525",
+        "issue:44521|issue:45325",
+        "issue:44466|issue:45276",
+        "issue:43653|issue:45593",
+        "issue:44016|issue:44291",
+        "issue:44462|issue:45081",
+        "issue:43441|issue:45834",
+        "issue:42994|issue:43317"
+      ],
+      "prepared_review_unit_hash": "a77a84e47e638eafe55d3da12a5edcd7060a8f7b731a387cc65c23c3a2c351cb",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 27,
+        "serialized_chars": 35914,
+        "estimated_input_tokens": 8979,
+        "estimated_eval_tokens": 18214
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is highly heterogeneous: tokenizer regressions, model-loading bugs, config defaults, docs issues, and unrelated infra/version problems. None of the soft pairs look like the same underlying bug or a single fixable code path, so I would reject all soft edges.",
+        "confidence": 0.56,
+        "canonical_issue_reason": "Issue 45812 is the best anchor available: it is a broad, high-signal tokenizer regression report with an explicit v4\u2192v5 behavior change and clear impact across a model family. That said, the overall cluster is not truly coherent.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "45812 is the most representative issue for a duplicate-triage anchor because it describes a concrete, user-visible regression in core tokenizer behavior rather than a narrow model-specific or docs-only problem.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42796",
+            "right": "issue:43650",
+            "accept": false,
+            "reason": "Tokenizer BPE-marker bug vs a vague 'ADD THE DATA' issue; no shared code path or symptom."
+          },
+          {
+            "left": "issue:46082",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Different models and failures: LlamaConfig head_dim validation vs Qwen3.5 rotary-registration crash."
+          },
+          {
+            "left": "issue:44279",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Unrelated dependency problems: generic transformers dependency issue vs PEFT minimum-version bump."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:45020",
+            "accept": false,
+            "reason": "Video pipeline processor lookup bug vs remote_code compatibility regressions; different subsystems and fixes."
+          },
+          {
+            "left": "issue:43577",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Blip2 dtype propagation bug vs Qwen3.5 save_pretrained checkpoint regression; not the same bug."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Pipeline processor selection bug vs Gemma4 device-map mismatch; different failure modes."
+          },
+          {
+            "left": "issue:45103",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "auto_docstring annotation parsing crash vs Gemma4 processor import failure; unrelated code paths."
+          },
+          {
+            "left": "issue:44351",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "A single missing import vs many import errors after upgrade; the latter is broader and not the same concrete issue."
+          },
+          {
+            "left": "issue:43295",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Custom processor/tokenizer regression vs Deepseek tokenizer output regression; both tokenizer-related but different bugs."
+          },
+          {
+            "left": "issue:39692",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Docs example errors for SigLIP2 vs pipeline docs mention behavior drift; different documentation issues."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Qwen3VL model loading failure vs Qwen3.5 save_pretrained key serialization bug; not the same fix."
+          },
+          {
+            "left": "issue:38175",
+            "right": "issue:44617",
+            "accept": false,
+            "reason": "Unexpected zero probabilities in SigLIP2 vs SAM3Video OOM; completely different symptoms."
+          },
+          {
+            "left": "issue:44683",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Torch flex_attention compilation bug vs TensorFlow backend/install issue; unrelated."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "MoE load-balancing loss omission vs Gemma4 mm_token_type_ids defaulting; different training/input semantics."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45405",
+            "accept": false,
+            "reason": "Python version support regression vs PEFT version pin mismatch; unrelated release/dependency problems."
+          },
+          {
+            "left": "issue:46082",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "LlamaConfig head_dim validation vs from_config dtype metadata staleness; both config-related but not the same bug."
+          },
+          {
+            "left": "issue:43650",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Nonsense/incomplete issue vs FP8 inference on B200; no meaningful overlap."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Deepseek-specific tokenizer regression vs Granite tokenizer-ID regression; both tokenizer bugs but different model families and code paths."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43854",
+            "accept": false,
+            "reason": "Qwen3VL loading regression vs GLM-4.7-Flash unit-test loading issue; different models and failures."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Phi4MultimodalConfig defaulting bug vs Llama4Config missing pad_token_id; both config issues but not the same defect."
+          },
+          {
+            "left": "issue:44521",
+            "right": "issue:45325",
+            "accept": false,
+            "reason": "apply_chat_template assistant-mask bug vs Qwen2.5-VL rope-index scaling bug; different multimodal code paths."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45276",
+            "accept": false,
+            "reason": "lm_head serialization inconsistency vs Gemma4 resize_token_embeddings behavior; not the same underlying change."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "BigBirdTokenizer special-token registration bug vs D-FINE auxiliary-loss omission; unrelated model components."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:44291",
+            "accept": false,
+            "reason": "Notebook syntax error vs init_empty_weights unexpected argument TypeError; unrelated."
+          },
+          {
+            "left": "issue:44462",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "AutoTokenizer ignoring tokenizer.json vs Mistral regex patch crash; both tokenizer loading, but different failure points and fixes."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Ministral FlashAttention failure vs Kosmos2.5 OCR index error; unrelated model-specific bugs."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43317",
+            "accept": false,
+            "reason": "Quantized saving failure vs dequantized offload loading failure; adjacent area, but distinct bugs 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": 1,
+      "component_count": 21,
+      "review_unit_index": 38,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:33290",
+        "issue:41628",
+        "issue:42175",
+        "issue:43441",
+        "issue:43525",
+        "issue:43643",
+        "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:44246",
+        "issue:44261",
+        "issue:44263",
+        "issue:44297",
+        "issue:44336",
+        "issue:44361",
+        "issue:44373",
+        "issue:44496",
+        "issue:44589",
+        "issue:44623",
+        "issue:44671",
+        "issue:44683",
+        "issue:44749",
+        "issue:44779",
+        "issue:45081",
+        "issue:45103",
+        "issue:45362",
+        "issue:45397",
+        "issue:45542",
+        "issue:45561",
+        "issue:45563",
+        "issue:45593",
+        "issue:45600",
+        "issue:45656",
+        "issue:45712",
+        "issue:45800",
+        "issue:45823",
+        "issue:45907",
+        "issue:46291",
+        "issue:46399"
+      ],
+      "soft_pairs": [
+        "issue:44361|issue:45593",
+        "issue:33290|issue:45656",
+        "issue:45081|issue:45907",
+        "issue:44261|issue:44263",
+        "issue:44623|issue:44749",
+        "issue:45600|issue:45712",
+        "issue:44623|issue:45397",
+        "issue:45542|issue:45800",
+        "issue:42175|issue:43976",
+        "issue:44016|issue:45823",
+        "issue:44671|issue:44779",
+        "issue:43901|issue:43994",
+        "issue:43441|issue:44683",
+        "issue:43723|issue:44162",
+        "issue:43673|issue:43761",
+        "issue:43525|issue:43881",
+        "issue:46291|issue:46399",
+        "issue:43901|issue:44246",
+        "issue:43824|issue:43976",
+        "issue:43525|issue:43643",
+        "issue:44188|issue:44336",
+        "issue:44589|issue:45362",
+        "issue:44297|issue:45397",
+        "issue:44336|issue:44373",
+        "issue:43784|issue:45542",
+        "issue:41628|issue:43976",
+        "issue:44246|issue:44496",
+        "issue:45561|issue:45563",
+        "issue:45103|issue:46399"
+      ],
+      "prepared_review_unit_hash": "1c2859bbd235fa200b0ed2faf0c685516e947fc9cfe9df45c5b2ce68774297bb",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36368,
+        "estimated_input_tokens": 9092,
+        "estimated_eval_tokens": 18440
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These look like a heterogeneous set of unrelated issues, not a duplicate cluster. None of the soft-edge pairs appear to describe the same underlying bug or change, so I would not merge any of them as duplicates.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "No single canonical issue: the items span unrelated bugs, regressions, docs, and dependency problems, so none of the soft edges are strong enough to represent one shared root cause.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No global best issue for duplication triage; the set is too mixed, and the more central-looking reports still do not subsume the others.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44361",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Tokenizer AttributeError vs D-FINE auxiliary-loss behavior; unrelated code paths."
+          },
+          {
+            "left": "issue:33290",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "OOM with Adafactor/DeepSpeed vs optimizer.step called twice; different bugs."
+          },
+          {
+            "left": "issue:45081",
+            "right": "issue:45907",
+            "accept": false,
+            "reason": "Mistral tokenizer regex crash vs list_repo_files revision forwarding; unrelated."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "MLA norm epsilon precision issue vs torch.split behavior in a different indexer."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "Missing saved processor files vs performance slowdown report; not the same issue."
+          },
+          {
+            "left": "issue:45600",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Removed config references in auto_mappings vs leftover dummy classes in dummy_pt_objects."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "processor.save_pretrained file omission vs Gemma-4 Zero3 loading failure."
+          },
+          {
+            "left": "issue:45542",
+            "right": "issue:45800",
+            "accept": false,
+            "reason": "Missing TF backend from tensorboard install vs torch/transformers version incompatibility."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "pip extras missing TensorFlow vs Python version support mismatch; different environments."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:45823",
+            "accept": false,
+            "reason": "Notebook syntax error vs Gemma4 device-map forward mismatch; unrelated."
+          },
+          {
+            "left": "issue:44671",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "CamemBERT MLM regression vs Deepseek tokenizer regression; different models."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:43994",
+            "accept": false,
+            "reason": "Docs mismatch for return_all_scores vs SigLIP2 bad outputs; unrelated."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "Ministral FlashAttention failure vs flex_attention on torch>=2.9; different concrete failures."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "AutoTokenizer loading issue vs ESM2 broken; different subsystems."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:43761",
+            "accept": false,
+            "reason": "Generation cache/chunked prefill bug vs CLIPVision hidden_states regression."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43881",
+            "accept": false,
+            "reason": "Llama4Config pad_token_id missing vs glm-4v-9b loading failure; not the same bug."
+          },
+          {
+            "left": "issue:46291",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Kernels import/version requirement vs missing decorator in rotary-pos-emb registration; different paths."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "Docs bug vs intermittent import slowness; unrelated."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "Qwen2.5-VL import error vs Python version compatibility issue; different causes."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "Missing pad_token_id vs AutoConfig missing fields with trust_remote_code; distinct config problems."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44336",
+            "accept": false,
+            "reason": "torch.compile attention divergence vs ANSI codes in loading_report; completely unrelated."
+          },
+          {
+            "left": "issue:44589",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "Float8 storage object error vs Qwen3.5 chat crash; no shared code path."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "tokenizer_class metadata mismatch vs Gemma-4 Zero3 loading issue; unrelated."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "ANSI formatting leak vs wrong docstring; not a duplicate bug."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Sentence-transformers import NameError vs missing TensorFlow backend error."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "Cannot import AutoImageProcessor vs Python 3.9/3.10 compatibility issue."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "Occasional slow import vs unrecognized model/config error; different failures."
+          },
+          {
+            "left": "issue:45561",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "xdist race on captured_info.txt vs stale num_return_sequences warning; unrelated."
+          },
+          {
+            "left": "issue:45103",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "auto_docstring annotations crash vs Qwen3.5 kernel decorator error; different subsystems."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 39,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:36246",
+        "issue:41950",
+        "issue:42491",
+        "issue:42757",
+        "issue:42886",
+        "issue:43257",
+        "issue:43262",
+        "issue:43299",
+        "issue:43352",
+        "issue:43475",
+        "issue:43526",
+        "issue:43531",
+        "issue:43550",
+        "issue:43596",
+        "issue:43673",
+        "issue:43901",
+        "issue:44162",
+        "issue:44220",
+        "issue:44242",
+        "issue:44246",
+        "issue:44295",
+        "issue:44336",
+        "issue:44351",
+        "issue:44485",
+        "issue:44496",
+        "issue:44561",
+        "issue:44734",
+        "issue:44749",
+        "issue:44779",
+        "issue:44857",
+        "issue:44871",
+        "issue:45003",
+        "issue:45137",
+        "issue:45200",
+        "issue:45250",
+        "issue:45292",
+        "issue:45362",
+        "issue:45375",
+        "issue:45397",
+        "issue:45561",
+        "issue:45593",
+        "issue:45759",
+        "issue:45859",
+        "issue:46132",
+        "issue:46369",
+        "issue:46399",
+        "issue:46424",
+        "issue:46506"
+      ],
+      "soft_pairs": [
+        "issue:44242|issue:45593",
+        "issue:44295|issue:45397",
+        "issue:42491|issue:43299",
+        "issue:43531|issue:45362",
+        "issue:43526|issue:45003",
+        "issue:41950|issue:44220",
+        "issue:44246|issue:44871",
+        "issue:44336|issue:44485",
+        "issue:43550|issue:46506",
+        "issue:36246|issue:42757",
+        "issue:44336|issue:46424",
+        "issue:36246|issue:44351",
+        "issue:42886|issue:46132",
+        "issue:44734|issue:46424",
+        "issue:36246|issue:43596",
+        "issue:44779|issue:45292",
+        "issue:45003|issue:45561",
+        "issue:45859|issue:46399",
+        "issue:43901|issue:45003",
+        "issue:43475|issue:46369",
+        "issue:44242|issue:44857",
+        "issue:43257|issue:45137",
+        "issue:44295|issue:44749",
+        "issue:43673|issue:44561",
+        "issue:44496|issue:45003",
+        "issue:43352|issue:45250",
+        "issue:44162|issue:45200",
+        "issue:45375|issue:45759",
+        "issue:43262|issue:44246"
+      ],
+      "prepared_review_unit_hash": "9306a39bfce22833965e0be842eb92d5724c157ece94e3373583a879c5c70223",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 29,
+        "serialized_chars": 36299,
+        "estimated_input_tokens": 9075,
+        "estimated_eval_tokens": 18406
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a loose mix of Transformers bug reports, mostly model- or subsystem-specific regressions. The soft pairs do not look like the same underlying bug/change, so all candidates are rejected.",
+        "confidence": 0.77,
+        "canonical_issue_reason": "issue:44561 is the broadest cross-cutting regression in the set: a v5 API removal breaking trust_remote_code models, which is more representative of the cluster than the narrower model-specific reports.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44561 best captures the cluster\u2019s general theme of v5 compatibility breakage across external models and is the most reusable as a cluster representative.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44242",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Both mention loss terms in model training, but they affect different models and different mechanics: router load-balancing vs D-FINE auxiliary losses."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "One is a buffer read/indexing problem, the other is a ZeRO3/from_pretrained issue on Gemma-4; no shared code path is evident."
+          },
+          {
+            "left": "issue:42491",
+            "right": "issue:43299",
+            "accept": false,
+            "reason": "Both involve Qwen3 MoE loading, but one is HF4.x to HF5.x conversion and the other is a Qwen3VL loading regression; not the same concrete bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45362",
+            "accept": false,
+            "reason": "These are different Qwen-related failures with different symptoms and execution paths: sliding-window handling vs chat/model crash."
+          },
+          {
+            "left": "issue:43526",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "BeitImageProcessorFast label reduction and unsafe sys.modules access are unrelated bugs in different layers."
+          },
+          {
+            "left": "issue:41950",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Both touch input feature handling, but one is a video-classification pipeline processor lookup bug and the other is a separate fbank feature extraction issue."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Import latency and Gemma eos_token_id configuration are unrelated; no common underlying defect."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "ANSI output formatting in loading_report and GLM-5 RoPE implementation are unrelated."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "Torch.compile/SDPA failure for Bamba and CPU/CUDA output differences for PPDocLayoutV3 are different model/runtime bugs."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are import errors, but they concern different missing symbols from different packages and different fixes."
+          },
+          {
+            "left": "issue:44336",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Console ANSI codes in loading_report and cache-length behavior are unrelated."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Missing Qwen2_5_VLImageProcessor and missing HybridCache are distinct missing-export issues with different symbols and likely different fixes."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Offline cache loading and AttentionInterface.register behavior are not the same bug."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "KV-cache continuation indexing in serve and max_cache_length handling are different cache-related problems."
+          },
+          {
+            "left": "issue:36246",
+            "right": "issue:43596",
+            "accept": false,
+            "reason": "A missing import symbol for Qwen VL and a DeepSpeed ZeRO3/BertModel index error are unrelated."
+          },
+          {
+            "left": "issue:44779",
+            "right": "issue:45292",
+            "accept": false,
+            "reason": "Tokenizer correctness regression and resize_token_embeddings not updating output embeddings are different code paths."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Unsafe sys.modules access and a pytest-xdist file race are unrelated."
+          },
+          {
+            "left": "issue:45859",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Both are Qwen3.5/rotary-related, but one is CPU-offload incompatibility and the other is a missing registration decorator; not the same concrete bug."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Documentation mismatch for TextClassificationPipeline and sys.modules access are unrelated."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "SAM 3 output attribute missing and SAM3 memory overflow are different failures in the same family, but not the same bug."
+          },
+          {
+            "left": "issue:44242",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "MoE load-balancing loss omission and LwDetrImageLoss AMP crash are unrelated."
+          },
+          {
+            "left": "issue:43257",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Both involve DeepSpeed/zero3-style environments, but one is Qwen3 MoE weight conversion and the other is an empty-deque index error."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "Buffer access error and upgrade-related slowdown are not the same issue."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44561",
+            "accept": false,
+            "reason": "Chunked-prefill cache missing and removal of is_torch_fx_available are separate regressions."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Unrecognized model config and unsafe sys.modules access are unrelated."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:45250",
+            "accept": false,
+            "reason": "A model-specific Flash Attention 2 unsupported error and a generic Flash Attention 2 issue title do not establish the same bug."
+          },
+          {
+            "left": "issue:44162",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "ESM2 breakage and Gemma-4 mm_token_type_ids defaulting are unrelated model-specific issues."
+          },
+          {
+            "left": "issue:45375",
+            "right": "issue:45759",
+            "accept": false,
+            "reason": "Both involve composite Qwen multimodal configs, but one is a dropped config field and the other is AutoModelForCausalLM.from_config unwrapping text_config; different defects."
+          },
+          {
+            "left": "issue:43262",
+            "right": "issue:44246",
+            "accept": false,
+            "reason": "Audio chat-template sampling-rate behavior and import performance 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": 21,
+      "review_unit_index": 40,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:42886",
+        "issue:43065",
+        "issue:43408",
+        "issue:43475",
+        "issue:43540",
+        "issue:43653",
+        "issue:43906",
+        "issue:43931",
+        "issue:44038",
+        "issue:44060",
+        "issue:44077",
+        "issue:44162",
+        "issue:44188",
+        "issue:44368",
+        "issue:44466",
+        "issue:44479",
+        "issue:44483",
+        "issue:44561",
+        "issue:44877",
+        "issue:44977",
+        "issue:44987",
+        "issue:44995",
+        "issue:45137",
+        "issue:45200",
+        "issue:45229",
+        "issue:45292",
+        "issue:45335",
+        "issue:45375",
+        "issue:45397",
+        "issue:45412",
+        "issue:45459",
+        "issue:45563",
+        "issue:45584",
+        "issue:45588",
+        "issue:45600",
+        "issue:45656",
+        "issue:45685",
+        "issue:45701",
+        "issue:45987",
+        "issue:46082",
+        "issue:46123",
+        "issue:46170",
+        "issue:46202",
+        "issue:46399",
+        "issue:46421",
+        "issue:46431",
+        "issue:46506"
+      ],
+      "soft_pairs": [
+        "issue:44877|issue:45375",
+        "issue:45292|issue:45584",
+        "issue:44038|issue:46506",
+        "issue:44561|issue:46399",
+        "issue:45292|issue:45701",
+        "issue:45137|issue:45600",
+        "issue:44466|issue:45335",
+        "issue:45412|issue:46123",
+        "issue:44987|issue:45459",
+        "issue:44479|issue:46170",
+        "issue:43475|issue:43540",
+        "issue:44060|issue:44077",
+        "issue:44995|issue:45200",
+        "issue:45397|issue:45656",
+        "issue:44368|issue:44977",
+        "issue:43653|issue:44162",
+        "issue:45685|issue:46431",
+        "issue:42886|issue:45229",
+        "issue:44995|issue:45588",
+        "issue:44877|issue:46170",
+        "issue:43065|issue:43408",
+        "issue:43906|issue:46202",
+        "issue:44368|issue:45563",
+        "issue:44483|issue:46202",
+        "issue:43931|issue:44368",
+        "issue:44188|issue:46202",
+        "issue:45987|issue:46082",
+        "issue:44188|issue:46421",
+        "issue:45584|issue:46170",
+        "issue:43065|issue:46123"
+      ],
+      "prepared_review_unit_hash": "0fba60c111f8498f9f7abcd3fc0d4c7351a3bb45e86b92764623153a5c2b95e2",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 30,
+        "serialized_chars": 36059,
+        "estimated_input_tokens": 9015,
+        "estimated_eval_tokens": 18286
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is highly heterogeneous: most items are unrelated model/config/runtime bugs, and none of the soft pairs look like the same concrete underlying issue. I rejected all soft duplicate candidates. The least-bad canonical issue is the broad open tokenizer report `issue:45701`, but overall the cluster does not form a strong duplicate group.",
+        "confidence": 0.69,
+        "canonical_issue_reason": "`issue:45701` is open and broadly about tokenizer behavior changing across versions, making it the closest umbrella for the few tokenizer-adjacent reports in this set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "`issue:45701` is the broadest and most reusable issue in the cluster; the others are too specific and mostly unrelated to serve as a global representative.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44877",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both concern config validation/strict loading, but they describe different models and different missing fields; not the same bug."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45584",
+            "accept": false,
+            "reason": "One is about `resize_token_embeddings`, the other about Whisper empty-transcription generation after token alignment; different code paths."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:46506",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe version regression versus PPDocLayoutV3 CPU/CUDA divergence; unrelated failures."
+          },
+          {
+            "left": "issue:44561",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "One is a removed helper breaking `trust_remote_code`; the other is a missing decorator on a Qwen3.5 kernel registration. Different fixes."
+          },
+          {
+            "left": "issue:45292",
+            "right": "issue:45701",
+            "accept": false,
+            "reason": "General tokenization/version behavior vs a specific embeddings-resize bug; not the same issue."
+          },
+          {
+            "left": "issue:45137",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "DeepSpeed ZeRO3 deque error and stale auto-mapping references are unrelated."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:45335",
+            "accept": false,
+            "reason": "Both mention tied/resized embeddings, but one is device-dependent serialization and the other is resize propagation to decoder embeddings."
+          },
+          {
+            "left": "issue:45412",
+            "right": "issue:46123",
+            "accept": false,
+            "reason": "Memory not released on delete versus partial-batch generation bookkeeping; different subsystems and symptoms."
+          },
+          {
+            "left": "issue:44987",
+            "right": "issue:45459",
+            "accept": false,
+            "reason": "Model load failure on transformers>=5.1.0 versus hidden tokenizer decode errors without protobuf; different root causes."
+          },
+          {
+            "left": "issue:44479",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Video-input regression in Qwen VL models versus Windows/ROCm import error from torch.distributed dependency; unrelated."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:43540",
+            "accept": false,
+            "reason": "A missing vision-encoder attribute versus a video-input ValueError in Qwen3OmniMoe; not the same bug."
+          },
+          {
+            "left": "issue:44060",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "Incorrect tied-weights warning and patchtsmixer `post_init` policy are different configuration/modeling issues."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Stale cache crash on second forward pass versus missing default `mm_token_type_ids`; different failure modes."
+          },
+          {
+            "left": "issue:45397",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Gemma Zero3 loading issue and duplicate optimizer stepping with DeepSpeed are distinct problems."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "A tied-embeddings warning in fine-tuning versus generation failure with flash-attention; different code paths."
+          },
+          {
+            "left": "issue:43653",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "Tokenizer special-token registration bug versus broad ESM2 breakage; not the same underlying defect."
+          },
+          {
+            "left": "issue:45685",
+            "right": "issue:46431",
+            "accept": false,
+            "reason": "MPS histogram/int kernel error versus missing chat template on a processor; unrelated."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:45229",
+            "accept": false,
+            "reason": "Offline cache tokenizer loading and Gemma multi-GPU CUDA OOM are unrelated."
+          },
+          {
+            "left": "issue:44995",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Second-forward stale cache crash versus flash-attention `s_aux=None` AttributeError; different code paths."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Strict config loading for Granite Speech versus Windows/ROCm import-time dependency failure; unrelated."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:43408",
+            "accept": false,
+            "reason": "Dummy Conv2d in `Sam3PixelDecoder` and sampler/tracker model-type warning are related to SAM3 but not the same concrete bug."
+          },
+          {
+            "left": "issue:43906",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "Isolated reproduction of an external issue versus Gemma pipeline roundtrip failure on `\u2581`; different reported behavior."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "A repeated tied-weights warning and a stale `num_return_sequences` warning in paged generate; different warnings and fixes."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "Chat completions request rejection versus Gemma token roundtrip failure; different APIs and symptoms."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Weight-shape mismatch on model load versus tied-embeddings warning during fine-tuning; unrelated."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "Torch.compile attention-kernel divergence versus Gemma token roundtrip failure; unrelated."
+          },
+          {
+            "left": "issue:45987",
+            "right": "issue:46082",
+            "accept": false,
+            "reason": "Cache sequence-length return-type bug versus LlamaConfig head_dim validation; different components."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:46421",
+            "accept": false,
+            "reason": "Compiled attention divergence versus cache lazy-init recompilation; both performance/runtime related but not the same bug."
+          },
+          {
+            "left": "issue:45584",
+            "right": "issue:46170",
+            "accept": false,
+            "reason": "Whisper empty-transcription failure versus Windows/ROCm import error; no shared code path."
+          },
+          {
+            "left": "issue:43065",
+            "right": "issue:46123",
+            "accept": false,
+            "reason": "SAM3 pixel-decoder placeholder layer versus MaskGenerationPipeline batching bug; 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": 21,
+      "review_unit_index": 41,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:28282",
+        "issue:30990",
+        "issue:41628",
+        "issue:42371",
+        "issue:42673",
+        "issue:42994",
+        "issue:43232",
+        "issue:43475",
+        "issue:43482",
+        "issue:43525",
+        "issue:43643",
+        "issue:43646",
+        "issue:43673",
+        "issue:43704",
+        "issue:43723",
+        "issue:43784",
+        "issue:43873",
+        "issue:44038",
+        "issue:44162",
+        "issue:44188",
+        "issue:44297",
+        "issue:44351",
+        "issue:44361",
+        "issue:44623",
+        "issue:44749",
+        "issue:44869",
+        "issue:44871",
+        "issue:44933",
+        "issue:44945",
+        "issue:44964",
+        "issue:45081",
+        "issue:45092",
+        "issue:45103",
+        "issue:45137",
+        "issue:45230",
+        "issue:45310",
+        "issue:45356",
+        "issue:45375",
+        "issue:45381",
+        "issue:45397",
+        "issue:45406",
+        "issue:45478",
+        "issue:45656",
+        "issue:45753",
+        "issue:46132",
+        "issue:46202",
+        "issue:46326",
+        "issue:46399"
+      ],
+      "soft_pairs": [
+        "issue:44297|issue:44623",
+        "issue:44361|issue:44749",
+        "issue:44749|issue:45656",
+        "issue:43646|issue:44162",
+        "issue:44749|issue:45310",
+        "issue:44623|issue:45656",
+        "issue:43232|issue:43673",
+        "issue:44871|issue:46132",
+        "issue:42994|issue:43482",
+        "issue:43475|issue:43643",
+        "issue:45230|issue:45397",
+        "issue:44869|issue:45356",
+        "issue:28282|issue:30990",
+        "issue:44038|issue:44623",
+        "issue:41628|issue:44933",
+        "issue:43475|issue:45381",
+        "issue:45375|issue:45478",
+        "issue:42371|issue:45092",
+        "issue:45310|issue:45375",
+        "issue:42994|issue:43525",
+        "issue:44964|issue:45137",
+        "issue:43784|issue:46399",
+        "issue:43723|issue:46326",
+        "issue:45406|issue:46202",
+        "issue:43232|issue:44188",
+        "issue:43873|issue:44945",
+        "issue:41628|issue:44351",
+        "issue:42673|issue:43704",
+        "issue:44945|issue:45753",
+        "issue:45081|issue:45103"
+      ],
+      "prepared_review_unit_hash": "c6a6b8e9eedc272a45da0a97fda4432f75fd07313e6389f53ba6ad07b7121d5e",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36500,
+        "estimated_input_tokens": 9125,
+        "estimated_eval_tokens": 18506
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The set is mostly heterogeneous transformer issues, with only a couple of clear duplicate pairs; most soft matches are related by subsystem or release, not the same concrete bug.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "issue:42673 is the strongest canonical issue: it is a concrete, well-scoped bug report with a clear duplicate counterpart (issue:43704) and enough discussion to serve as the representative thread.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:42673 is the best overall issue candidate because it cleanly captures a specific reproducible bug and is an exact duplicate target, unlike the other items which are largely unrelated.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44297",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Both are save_pretrained problems, but one is tokenizer metadata and the other is processor file output; different bugs."
+          },
+          {
+            "left": "issue:44361",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "Tokenizer AttributeError and a data-filtering slowdown after upgrade are unrelated issues."
+          },
+          {
+            "left": "issue:44749",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Performance regression in filtering vs double optimizer stepping under deepspeed; not the same code path."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:44162",
+            "accept": false,
+            "reason": "Missing fields from AutoConfig with trust_remote_code is not the same as ESM2 being broken."
+          },
+          {
+            "left": "issue:44749",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "A filtering slowdown and a Qwen3.5 Moe from_pretrained failure are distinct bugs."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "save_pretrained missing files and optimizer-step duplication are unrelated."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "Both are generation-related, but one is update_model_kwargs after sync_gpus and the other is missing cache in chunked_prefill."
+          },
+          {
+            "left": "issue:44871",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Gemma EOS config mismatch and AttentionInterface.register behavior changes are different problems."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43482",
+            "accept": false,
+            "reason": "Quantized model saving failure vs Qwen2.5-GGUF loading failure are opposite directions and different failures."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:43643",
+            "accept": false,
+            "reason": "SAM 3 video encoder attribute error is unrelated to missing config fields from trust_remote_code."
+          },
+          {
+            "left": "issue:45230",
+            "right": "issue:45397",
+            "accept": false,
+            "reason": "Generic bug report vs gemma-4 ZeRO3 from_pretrained failure; no shared underlying defect."
+          },
+          {
+            "left": "issue:44869",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Whisper timestamp decode crash and Kimi tokenizer regression are different tokenizer paths."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:30990",
+            "accept": false,
+            "reason": "PyTorch not installed import error and a loading hang are not the same bug."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe/v5 incompatibility and processor save_pretrained missing files are distinct."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Both concern imports, but they name different missing symbols and likely different regressions."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:45381",
+            "accept": false,
+            "reason": "A missing encoder attribute and wrong video position ids are different model bugs."
+          },
+          {
+            "left": "issue:45375",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Missing vision-config field and a from_pretrained error for Qwen3.5 Moe are related only at a broad family level."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "TF32 settings warning is unrelated to InternVL2 meta-initialization incompatibility."
+          },
+          {
+            "left": "issue:45310",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Both involve Qwen3.5-family models, but one is a from_pretrained error and the other is a dropped config field."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Quantized saving failure and missing pad_token_id in Llama4Config are different failure modes."
+          },
+          {
+            "left": "issue:44964",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "Phi-4 multimodal loading and DeepSpeed ZeRO3 deque underflow are unrelated."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:46399",
+            "accept": false,
+            "reason": "Sentence-transformers import NameError and Qwen3.5 kernel registration error are unrelated."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "AutoTokenizer loading in v5 and continue_final_message with mistral_common are different tokenizer issues."
+          },
+          {
+            "left": "issue:45406",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "Gemma4Processor missing _tokenizer and Gemma pipeline roundtrip failure are not the same bug."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44188",
+            "accept": false,
+            "reason": "Generation kwargs update after sync_gpus is different from attention-kernel divergence under torch.compile."
+          },
+          {
+            "left": "issue:43873",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Quantization offloading and pipeline-parallelism output correctness are distinct issues."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but they involve different missing exports and different symbols."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:43704",
+            "accept": true,
+            "reason": "Exact same VRAM leak title and same Qwen3ForCausalLM/multiple dataloader threads bug."
+          },
+          {
+            "left": "issue:44945",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Incorrect output under pipeline parallelism and an infinite loop on a specific image are different model behaviors."
+          },
+          {
+            "left": "issue:45310",
+            "right": "issue:45478",
+            "accept": true,
+            "reason": "Same title and same Qwen3.5 Moe from_pretrained failure; clear duplicate pair."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The verdicts are grounded in the packet and stay conservative overall. The two accepted duplicate pairs are strongly supported by identical titles, and the canonical/best issue choice for 42673 is reasonable given the clear duplicate counterpart and discussion depth. No mergeability concerns apply here."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 42,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:30333",
+        "issue:30990",
+        "issue:33290",
+        "issue:35141",
+        "issue:36010",
+        "issue:41628",
+        "issue:42175",
+        "issue:42222",
+        "issue:42831",
+        "issue:42898",
+        "issue:42994",
+        "issue:43232",
+        "issue:43278",
+        "issue:43377",
+        "issue:43479",
+        "issue:43504",
+        "issue:43723",
+        "issue:43784",
+        "issue:43824",
+        "issue:43874",
+        "issue:43976",
+        "issue:44016",
+        "issue:44038",
+        "issue:44075",
+        "issue:44079",
+        "issue:44222",
+        "issue:44242",
+        "issue:44496",
+        "issue:44568",
+        "issue:44623",
+        "issue:44734",
+        "issue:44779",
+        "issue:44792",
+        "issue:44869",
+        "issue:44933",
+        "issue:45092",
+        "issue:45103",
+        "issue:45137",
+        "issue:45200",
+        "issue:45310",
+        "issue:45561",
+        "issue:45600",
+        "issue:45712",
+        "issue:45812",
+        "issue:46077",
+        "issue:46123",
+        "issue:46231"
+      ],
+      "soft_pairs": [
+        "issue:42994|issue:43824",
+        "issue:44222|issue:45310",
+        "issue:44038|issue:45310",
+        "issue:36010|issue:41628",
+        "issue:43232|issue:44792",
+        "issue:42831|issue:44222",
+        "issue:43278|issue:45137",
+        "issue:42175|issue:44933",
+        "issue:44496|issue:45561",
+        "issue:42175|issue:45712",
+        "issue:43504|issue:45561",
+        "issue:45561|issue:45600",
+        "issue:43976|issue:44933",
+        "issue:30333|issue:33290",
+        "issue:30990|issue:35141",
+        "issue:46077|issue:46123",
+        "issue:42898|issue:44568",
+        "issue:43377|issue:44242",
+        "issue:44623|issue:45310",
+        "issue:42222|issue:45200",
+        "issue:43479|issue:44869",
+        "issue:43479|issue:44079",
+        "issue:43874|issue:45092",
+        "issue:44016|issue:46231",
+        "issue:42898|issue:45812",
+        "issue:43784|issue:43824",
+        "issue:43723|issue:44568",
+        "issue:42898|issue:44779",
+        "issue:44734|issue:45103",
+        "issue:44075|issue:46231"
+      ],
+      "prepared_review_unit_hash": "b94e6232306c3243121c3961d34da357cca870bbd7daa296f0aa960496b9ba05",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 30,
+        "serialized_chars": 36121,
+        "estimated_input_tokens": 9031,
+        "estimated_eval_tokens": 18318
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The cluster is very heterogeneous: most items are separate v5 regressions, import failures, model-specific loading bugs, or unrelated runtime/training issues. None of the soft pairs look like the same underlying bug/change with a mergeable shared fix.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42994",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Different failures: quantized model saving vs missing Qwen2.5-VL import symbol."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Both involve MoE/Qwen, but one is FP8 save_pretrained and the other is a from_pretrained load error; different code paths."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Different Qwen variants and different breakages; not the same concrete bug."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "Both are import errors, but for unrelated symbols and likely unrelated export problems."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:44792",
+            "accept": false,
+            "reason": "Generation KV-cache bug vs janus image-generation test failure; no clear shared root cause."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "FP8 accuracy regression is distinct from FP8 MoE save_pretrained failure."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:45137",
+            "accept": false,
+            "reason": "BF16/FP32 dtype drift is unrelated to a DeepSpeed ZeRO3 deque IndexError."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Package backend/dependency issue vs missing import from image_utils; unrelated."
+          },
+          {
+            "left": "issue:44496",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Model config loading failure vs pytest-xdist file race; unrelated."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Both touch packaging, but one is dependency inclusion and the other is dummy class cleanup/leakage; not the same bug."
+          },
+          {
+            "left": "issue:43504",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "BEiT legacy-field loading bug vs test-worker file race; unrelated."
+          },
+          {
+            "left": "issue:45561",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Captured-info race in testing utils vs stale auto_mappings references; different files and failure modes."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44933",
+            "accept": false,
+            "reason": "Python version compatibility issue vs missing image_utils import; unrelated."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:33290",
+            "accept": false,
+            "reason": "MLflow reporting on failed training jobs vs Deepspeed/Adafactor OOM; not the same problem."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "Loading hang in Sentence Transformers vs token embedding resize/post_init reinit bug; unrelated."
+          },
+          {
+            "left": "issue:46077",
+            "right": "issue:46123",
+            "accept": false,
+            "reason": "Gemma4 attention-mask construction vs pipeline final-batch handling; different code paths."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Both are tokenizer regressions in v5, but one is cleanup-space behavior and the other is add_special_tokens/BOS-EOS handling; not the same bug."
+          },
+          {
+            "left": "issue:43377",
+            "right": "issue:44242",
+            "accept": false,
+            "reason": "MIMI padding-mask mismatch vs MoE router load-balancing loss omission; unrelated."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "processor.save_pretrained file omissions vs Qwen3.5 MoE from_pretrained error; different operations."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "VitPose model breakage vs Gemma4 mm_token_type_ids defaulting; unrelated model families and fixes."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44869",
+            "accept": false,
+            "reason": "Config default initialization bug vs Whisper word-timestamp decode crash; unrelated."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44079",
+            "accept": false,
+            "reason": "Config None-handling bug vs ModelOutput key assignment bug; different components."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "ImageProcessor method missing vs remote-code checkpoint meta-init incompatibility; unrelated."
+          },
+          {
+            "left": "issue:44016",
+            "right": "issue:46231",
+            "accept": false,
+            "reason": "Notebook syntax error vs docs parameter-name mismatch; not the same issue."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:45812",
+            "accept": false,
+            "reason": "Different tokenizer regressions affecting different models and APIs; not mergeable as one fix."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import-time failures, but for different missing symbols and different packages."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "AutoTokenizer v5 loading issue vs add_special_tokens token insertion regression; different failure modes."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both affect tokenizers in v5, but they concern different models and different incorrect behaviors."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Serve/KV-cache tensor indexing crash vs auto-docstring crash under future annotations; unrelated."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:46231",
+            "accept": false,
+            "reason": "Optimizer argument handling bug vs documentation typo; 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": 21,
+      "review_unit_index": 43,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:22355",
+        "issue:27957",
+        "issue:28282",
+        "issue:30333",
+        "issue:34689",
+        "issue:35141",
+        "issue:36331",
+        "issue:41628",
+        "issue:42757",
+        "issue:42796",
+        "issue:42831",
+        "issue:42898",
+        "issue:43122",
+        "issue:43295",
+        "issue:43475",
+        "issue:43479",
+        "issue:43644",
+        "issue:43723",
+        "issue:44060",
+        "issue:44206",
+        "issue:44222",
+        "issue:44263",
+        "issue:44295",
+        "issue:44297",
+        "issue:44373",
+        "issue:44448",
+        "issue:44484",
+        "issue:44485",
+        "issue:44623",
+        "issue:44933",
+        "issue:44977",
+        "issue:45071",
+        "issue:45092",
+        "issue:45103",
+        "issue:45125",
+        "issue:45245",
+        "issue:45310",
+        "issue:45356",
+        "issue:45381",
+        "issue:45405",
+        "issue:45656",
+        "issue:45712",
+        "issue:46082",
+        "issue:46132",
+        "issue:46209",
+        "issue:46305",
+        "issue:46431"
+      ],
+      "soft_pairs": [
+        "issue:44295|issue:45656",
+        "issue:42898|issue:45356",
+        "issue:42898|issue:43295",
+        "issue:27957|issue:30333",
+        "issue:30333|issue:34689",
+        "issue:45405|issue:45712",
+        "issue:44060|issue:45125",
+        "issue:44933|issue:45712",
+        "issue:44222|issue:44297",
+        "issue:30333|issue:36331",
+        "issue:30333|issue:35141",
+        "issue:44297|issue:45656",
+        "issue:22355|issue:28282",
+        "issue:42831|issue:44623",
+        "issue:44448|issue:46132",
+        "issue:44222|issue:44623",
+        "issue:43479|issue:43644",
+        "issue:42796|issue:43723",
+        "issue:45245|issue:46132",
+        "issue:43475|issue:46305",
+        "issue:41628|issue:42757",
+        "issue:44297|issue:45310",
+        "issue:45071|issue:46082",
+        "issue:45381|issue:46305",
+        "issue:46209|issue:46431",
+        "issue:44373|issue:44485",
+        "issue:43122|issue:46209",
+        "issue:44263|issue:45092",
+        "issue:43723|issue:44977",
+        "issue:44222|issue:44484",
+        "issue:44206|issue:45103"
+      ],
+      "prepared_review_unit_hash": "4d13eefc96978d6740c0fd55cb3c691ad02d3e416092c6d0cf3c8c9968a3867a",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 31,
+        "serialized_chars": 36263,
+        "estimated_input_tokens": 9066,
+        "estimated_eval_tokens": 18388
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is largely heterogeneous: most items are unrelated bugs across tokenization, model loading, config validation, save_pretrained, FP8, and MLFlow. None of the soft pairs look like the same underlying bug/change, so I would not merge any of them as duplicates.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "issue:43122 is the most general and representative report in the only coherent subtheme here (tokenization regressions across versions), while the rest are mostly unrelated defects.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43122 is the best representative issue for this mixed cluster because it captures the broad tokenization-regression theme more generally than the other tokenizer-specific reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44295",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Buffer access after registration vs deepspeed optimizer-step duplication; different subsystems and failure modes."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:45356",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but one is clean_up_tokenization_spaces behavior and the other is codec handling/warning behavior in a specific tokenizer."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:43295",
+            "accept": false,
+            "reason": "clean_up_tokenization_spaces regression and processor.tokenizer/image-passing regression are different code paths."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:30333",
+            "accept": false,
+            "reason": "Flash Attention 2 support for XLMRoberta vs MLFlow reporting after training failure are unrelated."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Failed-job MLFlow reporting vs model loading regression for Llama 3.2 Vision are different bugs."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "PEFT version pin mismatch vs leftover dummy classes in repo cleanup; no shared underlying bug."
+          },
+          {
+            "left": "issue:44060",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "Both touch Qwen3.5 MoE, but tied-weights warning and missing _tp_plan are separate defects."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45712",
+            "accept": false,
+            "reason": "Missing image_utils import and leftover dummy classes are unrelated import/repo hygiene issues."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "Both involve save_pretrained, but FP8 MoE serialization and tokenizer_class metadata mismatch are different problems."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:36331",
+            "accept": false,
+            "reason": "MLFlow failure reporting vs Trainer compute_loss API change are unrelated."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:35141",
+            "accept": false,
+            "reason": "MLFlow reporting on failed jobs vs embedding reinitialization in post_init are unrelated."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "Tokenizer config metadata bug vs deepspeed optimizer-step duplication are unrelated."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:28282",
+            "accept": false,
+            "reason": "Both are import errors, but missing transformers.onnx and missing PyTorch are different root causes."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "FP8 accuracy issue vs missing files in processor.save_pretrained are different kinds of bugs."
+          },
+          {
+            "left": "issue:44448",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Pegasus output regression across versions and AttentionInterface.register behavior change are distinct issues."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Both mention save_pretrained, but FP8 model serialization and processor file emission are not the same bug."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:43644",
+            "accept": false,
+            "reason": "Multimodal config default initialization vs non-persistent buffer junk filling are unrelated."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43723",
+            "accept": false,
+            "reason": "Tokenizer BPE-marker handling and AutoTokenizer loading failures are different bugs."
+          },
+          {
+            "left": "issue:45245",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Category cardinality limit error vs attention registration API behavior change are unrelated."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:46305",
+            "accept": false,
+            "reason": "Missing SAM3 attribute vs Moonshine streaming mask off-by-one are different model-specific defects."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Missing imports from transformers vs huggingface_hub are separate dependency/export issues."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Tokenizer save metadata mismatch vs Qwen3.5 MoE from_pretrained error are different failures."
+          },
+          {
+            "left": "issue:45071",
+            "right": "issue:46082",
+            "accept": false,
+            "reason": "PretrainedConfig type-checking regression vs LlamaConfig head_dim validation are distinct config-path issues."
+          },
+          {
+            "left": "issue:45381",
+            "right": "issue:46305",
+            "accept": false,
+            "reason": "Qwen2.5-VL vision_position_ids bug and Moonshine sliding-window mask bug are unrelated."
+          },
+          {
+            "left": "issue:46209",
+            "right": "issue:46431",
+            "accept": false,
+            "reason": "FP8 inference on B200 vs missing chat template in processor are unrelated."
+          },
+          {
+            "left": "issue:44373",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Wrong docstring and RoPE implementation discussion are not the same underlying bug."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Tokenization regression across versions vs FP8 inference failure on B200 are unrelated."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "torch.split return-value issue in GLM MoE indexer vs remote-code checkpoint meta-init incompatibility are different defects."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "AutoTokenizer loading failure and Qwen3.5 flash-attention generation issue are unrelated."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "FP8 MoE save_pretrained bug vs max_shard_size question are not the same bug."
+          },
+          {
+            "left": "issue:44206",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Feature extractor center-arg crash vs auto_docstring annotations crash 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": 21,
+      "review_unit_index": 44,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:22355",
+        "issue:27957",
+        "issue:28218",
+        "issue:30990",
+        "issue:38175",
+        "issue:39401",
+        "issue:41628",
+        "issue:42757",
+        "issue:42831",
+        "issue:42907",
+        "issue:43232",
+        "issue:43408",
+        "issue:43479",
+        "issue:43644",
+        "issue:43723",
+        "issue:43824",
+        "issue:44038",
+        "issue:44190",
+        "issue:44261",
+        "issue:44263",
+        "issue:44279",
+        "issue:44295",
+        "issue:44360",
+        "issue:44368",
+        "issue:44373",
+        "issue:44483",
+        "issue:44568",
+        "issue:44623",
+        "issue:44734",
+        "issue:44829",
+        "issue:44871",
+        "issue:44913",
+        "issue:45092",
+        "issue:45278",
+        "issue:45310",
+        "issue:45357",
+        "issue:45406",
+        "issue:45753",
+        "issue:45834",
+        "issue:46002",
+        "issue:46050",
+        "issue:46082",
+        "issue:46305",
+        "issue:46328",
+        "issue:46402",
+        "issue:46421",
+        "issue:46424"
+      ],
+      "soft_pairs": [
+        "issue:43232|issue:46424",
+        "issue:44734|issue:46421",
+        "issue:44190|issue:46082",
+        "issue:43232|issue:45834",
+        "issue:44295|issue:44623",
+        "issue:43723|issue:44871",
+        "issue:44263|issue:44360",
+        "issue:45092|issue:45357",
+        "issue:44279|issue:46328",
+        "issue:44261|issue:45092",
+        "issue:30990|issue:38175",
+        "issue:42907|issue:46050",
+        "issue:43644|issue:46305",
+        "issue:44263|issue:44373",
+        "issue:42831|issue:45310",
+        "issue:44829|issue:46002",
+        "issue:44483|issue:45406",
+        "issue:42831|issue:44038",
+        "issue:41628|issue:43824",
+        "issue:22355|issue:27957",
+        "issue:43479|issue:44913",
+        "issue:43408|issue:46402",
+        "issue:42831|issue:45753",
+        "issue:22355|issue:28218",
+        "issue:42757|issue:45278",
+        "issue:39401|issue:44568",
+        "issue:44368|issue:46305",
+        "issue:27957|issue:30990"
+      ],
+      "prepared_review_unit_hash": "2c5e86cc4c3cf508f5c9132f85a12675051c4ebf53edb5d53304d24a5e784131",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 28,
+        "serialized_chars": 35268,
+        "estimated_input_tokens": 8817,
+        "estimated_eval_tokens": 17890
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "The soft pairs are mostly unrelated issues across distinct bugs (generation, cache, tokenizer, config, model loading, imports, and serve/API failures). None look like the same underlying defect closely enough to merge as duplicates.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "issue:45278 is the broadest representative of the import-regression subset, covering the shared symptom of broken imports after the 5.x upgrade.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the clearest cluster representative: high-level, high-impact, and broad enough to subsume the other import-error reports better than the more specific issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43232",
+            "right": "issue:46424",
+            "accept": false,
+            "reason": "Generation kwargs after sync_gpus vs cache max_length semantics; different code paths and symptoms."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:46421",
+            "accept": false,
+            "reason": "KV-cache indexing crash in /v1/responses vs lazy cache init causing recompilation; not the same bug."
+          },
+          {
+            "left": "issue:44190",
+            "right": "issue:46082",
+            "accept": false,
+            "reason": "Dataset loading script failure vs LlamaConfig head_dim validation; unrelated."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:45834",
+            "accept": false,
+            "reason": "Generation bookkeeping bug vs Moonshine streaming mask off-by-one; distinct defects."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Reading a buffer registered position_ids vs processor.save_pretrained missing files; different problems."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44871",
+            "accept": false,
+            "reason": "Tokenizer loading regression vs eos_token_id configuration mismatch; not the same issue."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Both touch the DSA indexer, but one is a torch.split return bug and the other is a missing ReLU; separate fixes."
+          },
+          {
+            "left": "issue:45092",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different models and paths: remote-code meta-init incompatibility vs save_pretrained saving wrong visual encoder keys."
+          },
+          {
+            "left": "issue:44279",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "Generic dependency issue vs a padding performance regression; no common underlying bug."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Layernorm epsilon precision issue vs remote-code meta initialization incompatibility; unrelated."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:38175",
+            "accept": false,
+            "reason": "Loading hang in Sentence Transformers vs zero probabilities in SigLIP2; different model-specific failures."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "Both involve save/load around quantization, but one is dequantized model saving and the other is sinq roundtrip breakage; not the same concrete bug."
+          },
+          {
+            "left": "issue:43644",
+            "right": "issue:46305",
+            "accept": false,
+            "reason": "Non-persistent buffer junk vs streaming mask off-by-one; unrelated."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "Runtime indexing/splitting bug vs docstring text issue; not a duplicate."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "FineGrainedFP8 accuracy degradation vs Qwen3.5 MoE from_pretrained error; different failure modes."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "FlashAttention3 training degeneration vs NaNs on checkout/init; distinct symptoms and likely causes."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45406",
+            "accept": false,
+            "reason": "Serve request rejection vs Gemma4Processor missing _tokenizer crash; both in serve, but different code paths."
+          },
+          {
+            "left": "issue:42831",
+            "right": "issue:44038",
+            "accept": false,
+            "reason": "FP8 accuracy problem vs Qwen3-VL-Moe bug; not the same underlying defect."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and likely different fixes; too broad to merge."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:27957",
+            "accept": false,
+            "reason": "ONNX module import error vs Flash Attention 2 on XLMRoberta; unrelated."
+          },
+          {
+            "left": "issue:43479",
+            "right": "issue:44913",
+            "accept": false,
+            "reason": "Different config persistence bugs affecting different fields and models."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Missing huggingface_hub symbol vs broad transformers import failures after upgrade; related theme, not the same defect."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44568",
+            "accept": false,
+            "reason": "Wrong offset_mapping vs add_special_tokens not adding BOS/EOS; separate tokenizer bugs."
+          },
+          {
+            "left": "issue:44368",
+            "right": "issue:46305",
+            "accept": false,
+            "reason": "Config warning about tie_word_embeddings vs streaming window off-by-one; unrelated."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:30990",
+            "accept": false,
+            "reason": "Flash Attention 2 model issue vs Sentence Transformers loading hang; different subsystems and symptoms."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 45,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:30990",
+        "issue:33357",
+        "issue:39401",
+        "issue:41628",
+        "issue:42371",
+        "issue:42796",
+        "issue:42886",
+        "issue:42907",
+        "issue:42994",
+        "issue:43066",
+        "issue:43116",
+        "issue:43122",
+        "issue:43232",
+        "issue:43408",
+        "issue:43550",
+        "issue:43576",
+        "issue:43646",
+        "issue:43756",
+        "issue:43784",
+        "issue:43937",
+        "issue:43976",
+        "issue:44038",
+        "issue:44060",
+        "issue:44117",
+        "issue:44188",
+        "issue:44261",
+        "issue:44263",
+        "issue:44295",
+        "issue:44297",
+        "issue:44360",
+        "issue:44373",
+        "issue:44485",
+        "issue:44488",
+        "issue:44521",
+        "issue:44625",
+        "issue:44704",
+        "issue:44829",
+        "issue:44945",
+        "issue:44964",
+        "issue:45200",
+        "issue:45278",
+        "issue:45290",
+        "issue:45593",
+        "issue:45600",
+        "issue:46002",
+        "issue:46132",
+        "issue:46326",
+        "issue:46369"
+      ],
+      "soft_pairs": [
+        "issue:42886|issue:43066",
+        "issue:43408|issue:46369",
+        "issue:42796|issue:43122",
+        "issue:43232|issue:45593",
+        "issue:44188|issue:44373",
+        "issue:42886|issue:42994",
+        "issue:43550|issue:46132",
+        "issue:43576|issue:43784",
+        "issue:43646|issue:46002",
+        "issue:44261|issue:44373",
+        "issue:42371|issue:44263",
+        "issue:44295|issue:44297",
+        "issue:44117|issue:44704",
+        "issue:42907|issue:44945",
+        "issue:30990|issue:33357",
+        "issue:43756|issue:44060",
+        "issue:41628|issue:45278",
+        "issue:39401|issue:44488",
+        "issue:43976|issue:44945",
+        "issue:43784|issue:45600",
+        "issue:42907|issue:44964",
+        "issue:42907|issue:43937",
+        "issue:44360|issue:44485",
+        "issue:44704|issue:45290",
+        "issue:43122|issue:44521",
+        "issue:39401|issue:44521",
+        "issue:43122|issue:46326",
+        "issue:44625|issue:44829",
+        "issue:44038|issue:44297",
+        "issue:43116|issue:45200"
+      ],
+      "prepared_review_unit_hash": "4c9b085dbe632a056f8a98aa54fcfc6f4f77b491bd0fe83e365d105931b63d11",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 30,
+        "serialized_chars": 36196,
+        "estimated_input_tokens": 9049,
+        "estimated_eval_tokens": 18354
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a loose collection of unrelated transformer regressions and API breakages; none of the soft pairs look like true duplicates of the same concrete bug.",
+        "confidence": 0.56,
+        "canonical_issue_reason": "issue:45278 is the broadest umbrella report in the set and best represents the import-regression slice of the cluster; the others are narrower, model-specific, or unrelated.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45278 is the most representative issue to anchor triage because it captures a broad user-facing regression class rather than a single niche failure mode.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42886",
+            "right": "issue:43066",
+            "accept": false,
+            "reason": "Offline tokenizer cache loading and wrong decoder type in v5 are different failure modes."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "A model-type warning and a memory overflow on a different model are not the same bug."
+          },
+          {
+            "left": "issue:42796",
+            "right": "issue:43122",
+            "accept": false,
+            "reason": "Both concern tokenization, but one is BPE-marker handling for a specific model while the other is a version-to-version tokenization drift report."
+          },
+          {
+            "left": "issue:43232",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Generation kwargs after sync_gpus and missing auxiliary losses are unrelated code paths."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "A torch.compile attention-kernel divergence is unrelated to a docstring issue."
+          },
+          {
+            "left": "issue:42886",
+            "right": "issue:42994",
+            "accept": false,
+            "reason": "Tokenizer offline cache loading and quantized model saving are separate subsystems."
+          },
+          {
+            "left": "issue:43550",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "torch.compile SDPA failure and AttentionInterface.register behavior change are different problems."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:43784",
+            "accept": false,
+            "reason": "The broken env CLI and the sentence-transformers import NameError are distinct regressions."
+          },
+          {
+            "left": "issue:43646",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Custom model initialization breakage and NaNs after checkout/init are not the same concrete defect."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44373",
+            "accept": false,
+            "reason": "Missing rms_norm_eps in MLA layernorm is unrelated to a docstring correction."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "TF32 settings behavior and torch.split return values in a specific indexer are unrelated."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "Reading position_ids as a buffer and tokenizer_class mismatch on save_pretrained are different bugs."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "A tokenizer mapping None assumption and AutoProcessor kwargs forwarding are separate API issues."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Failed dequantized model saving and wrong pipeline-parallel output are different code paths."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:33357",
+            "accept": false,
+            "reason": "Sentence-transformers loading hang and a MacOS bus error are not the same failure."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44060",
+            "accept": false,
+            "reason": "RoPE layer count mismatch and tied-weights warning misbinding are unrelated model-specific bugs."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "The latter is a broad import-regression report, but the former is one specific missing import; not enough to treat as the same concrete bug."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44488",
+            "accept": false,
+            "reason": "Wrong offset mapping and failure to load a specific model are unrelated tokenizer/model issues."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Python version support and pipeline-parallel output correctness are unrelated."
+          },
+          {
+            "left": "issue:43784",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Importing sentence-transformers and stale auto_mappings references are different breakages."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "Dequantized-model save failure and multimodal model load failure are unrelated."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43937",
+            "accept": false,
+            "reason": "Quantized saving and invalid generation config are separate issues."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "A missing ReLU in a DSA indexer and a RoPE implementation discussion are not the same bug."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Passing kwargs to cached_file and crashing on tool-call assistant messages are different APIs."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Tokenization drift across versions and zeroed assistant masks for multimodal inputs are different behaviors."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44521",
+            "accept": false,
+            "reason": "Wrong offset mapping and assistant-mask generation are unrelated tokenizer features."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:46326",
+            "accept": false,
+            "reason": "Version-dependent tokenization differences and continue_final_message behavior are distinct bugs."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "num_labels propagation in Qwen3.5 and flash_attention_3 training degeneration are unrelated."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44297",
+            "accept": false,
+            "reason": "A Qwen3-VL-Moe bug and tokenizer save metadata mismatch are not the same failure."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:45200",
+            "accept": false,
+            "reason": "Empty multi-label example output and missing mm_token_type_ids defaults are different issues."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 46,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:22355",
+        "issue:29127",
+        "issue:29159",
+        "issue:36296",
+        "issue:38617",
+        "issue:39401",
+        "issue:42175",
+        "issue:42371",
+        "issue:42673",
+        "issue:42757",
+        "issue:42898",
+        "issue:42907",
+        "issue:43066",
+        "issue:43116",
+        "issue:43122",
+        "issue:43278",
+        "issue:43408",
+        "issue:43452",
+        "issue:43475",
+        "issue:43493",
+        "issue:43611",
+        "issue:43824",
+        "issue:43874",
+        "issue:43937",
+        "issue:43976",
+        "issue:43992",
+        "issue:44038",
+        "issue:44188",
+        "issue:44222",
+        "issue:44295",
+        "issue:44351",
+        "issue:44355",
+        "issue:44485",
+        "issue:44514",
+        "issue:44704",
+        "issue:44779",
+        "issue:44938",
+        "issue:44945",
+        "issue:45072",
+        "issue:45216",
+        "issue:45278",
+        "issue:45446",
+        "issue:45478",
+        "issue:45753",
+        "issue:46002",
+        "issue:46231",
+        "issue:46328",
+        "issue:46614"
+      ],
+      "soft_pairs": [
+        "issue:39401|issue:46614",
+        "issue:42907|issue:43452",
+        "issue:43408|issue:43475",
+        "issue:44038|issue:44295",
+        "issue:44038|issue:44222",
+        "issue:43824|issue:44351",
+        "issue:45216|issue:45478",
+        "issue:22355|issue:29159",
+        "issue:43122|issue:44514",
+        "issue:42757|issue:44351",
+        "issue:36296|issue:43116",
+        "issue:43278|issue:44945",
+        "issue:42175|issue:44938",
+        "issue:43493|issue:45753",
+        "issue:22355|issue:29127",
+        "issue:43937|issue:44485",
+        "issue:44188|issue:46328",
+        "issue:38617|issue:45278",
+        "issue:42898|issue:46614",
+        "issue:42673|issue:45072",
+        "issue:43066|issue:44779",
+        "issue:43611|issue:46002",
+        "issue:43116|issue:43493",
+        "issue:43992|issue:44704",
+        "issue:44355|issue:46231",
+        "issue:42371|issue:43874",
+        "issue:44038|issue:45478",
+        "issue:43976|issue:45446"
+      ],
+      "prepared_review_unit_hash": "e7c69f665bc2187113560cf23e5af02f3a68626345c5f4fb401a2d9f1912c9e1",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 35686,
+        "estimated_input_tokens": 8922,
+        "estimated_eval_tokens": 18100
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This is a heterogeneous cluster of unrelated Transformers issues, not a single duplicate bug thread. All soft edges look like false positives, so none are accepted.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "No single canonical issue stands out: the set mixes unrelated import errors, tokenizer regressions, model-save/load bugs, backend/version problems, and training/runtime issues. If forced to pick the closest thematic anchor, 44779 is the strongest tokenizer-regression example, but it is not cluster-wide canonical.",
+        "canonical_pr_reason": "No PRs are present in this cluster.",
+        "best_issue_reason": "44779 is the closest representative of the largest coherent subtheme (tokenizer regressions), but it still does not unify the rest of the cluster, so it is only a weak global anchor.",
+        "best_pr_reason": "No PRs are present in this cluster.",
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:39401",
+            "right": "issue:46614",
+            "accept": false,
+            "reason": "Both are tokenizer-adjacent, but one is Qwen3 offset mapping and the other is Wav2Vec2 phonemizer language state; different components and root causes."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "Save/dequantization failure for Ministrals/Devstrals is unrelated to gguf AutoTokenizer/AutoModel loading breakage."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Both are SAM 3 related, but one is a model-type warning while the other is a missing attribute crash in the video encoder."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44295",
+            "accept": false,
+            "reason": "Qwen3-VL-Moe load/bug report and a position_ids buffer read error are unrelated code paths."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Both mention MoE, but one is a Qwen3-VL-Moe issue and the other is an FP8 save_pretrained bug; not the same defect."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "These are both import errors, but for different missing symbols from different modules, so they are not the same bug."
+          },
+          {
+            "left": "issue:45216",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both involve Qwen3.5 checkpoints, but one is save_pretrained checkpoint corruption and the other is a from_pretrained loading error."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:29159",
+            "accept": false,
+            "reason": "Missing transformers.onnx and tokenizer slow/fast inconsistencies are unrelated issues."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "A version-to-version tokenization regression and a batched chat-template padding crash are different failures in different APIs."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but they concern different removed/moved symbols and do not look like one concrete defect."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43116",
+            "accept": false,
+            "reason": "Tensor parallel training bug and multi-label classification example returning empty results are unrelated."
+          },
+          {
+            "left": "issue:43278",
+            "right": "issue:44945",
+            "accept": false,
+            "reason": "Embedding dtype drift in eval is a different problem from incorrect pipeline-parallel output."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "Missing TF backend from an extra install and Python 3.14 load failure are different environment/package issues."
+          },
+          {
+            "left": "issue:43493",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "SigLIP2 implementation discrepancy and Qwen3_5 infinite-loop-on-image are unrelated model-specific bugs."
+          },
+          {
+            "left": "issue:22355",
+            "right": "issue:29127",
+            "accept": false,
+            "reason": "An ONNX import failure and a LayoutLMv3 error-message clarity issue are not the same underlying bug."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "GenerationConfig validation failure and GLM-5 RoPE implementation discussion are separate GLM-5 issues."
+          },
+          {
+            "left": "issue:44188",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "torch.compile attention-kernel divergence and a SequenceFeatureExtractor.pad performance problem are unrelated."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "The latter is an umbrella report for many import errors after upgrade, but the former is one specific missing import; not enough to treat as the same defect."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:46614",
+            "accept": false,
+            "reason": "A v5 tokenization-space behavior change and a stale phonemizer backend state bug are unrelated."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:45072",
+            "accept": false,
+            "reason": "VRAM leakage in multi-threaded dataloader use and dtype mismatches in bfloat16 inference are different runtime problems."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both are tokenizer regressions in v5, but one is decoder-type selection and the other is incorrect Deepseek tokenization output."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Base-model-prefix loading breakage and NaNs in classification heads are unrelated model-loading/training failures."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:43493",
+            "accept": false,
+            "reason": "Multi-label classification example output and SigLIP2 implementation mismatch are different model/use-case problems."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:44704",
+            "accept": false,
+            "reason": "UMT5Encoder missing embed_tokens.weight and AutoProcessor kwargs forwarding are unrelated loading issues."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:46231",
+            "accept": false,
+            "reason": "Compiled Python file errors and a documentation parameter-name mismatch are not the same issue."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:43874",
+            "accept": false,
+            "reason": "TF32 API-settings guidance and a GLM46V image-patch counting AttributeError are unrelated."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both mention Qwen3 family MoE models, but they describe different versions and different failure modes (bug vs from_pretrained error)."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Python-version compatibility failure and a flex_attention import version-check bug 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": 21,
+      "review_unit_index": 47,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:33453",
+        "issue:38617",
+        "issue:39401",
+        "issue:41628",
+        "issue:42222",
+        "issue:42673",
+        "issue:42796",
+        "issue:43064",
+        "issue:43116",
+        "issue:43441",
+        "issue:43450",
+        "issue:43673",
+        "issue:43825",
+        "issue:43827",
+        "issue:43906",
+        "issue:43937",
+        "issue:43986",
+        "issue:44075",
+        "issue:44117",
+        "issue:44291",
+        "issue:44380",
+        "issue:44410",
+        "issue:44466",
+        "issue:44625",
+        "issue:44779",
+        "issue:44829",
+        "issue:44936",
+        "issue:44977",
+        "issue:45071",
+        "issue:45081",
+        "issue:45083",
+        "issue:45216",
+        "issue:45310",
+        "issue:45341",
+        "issue:45356",
+        "issue:45357",
+        "issue:45405",
+        "issue:45446",
+        "issue:45478",
+        "issue:45561",
+        "issue:45563",
+        "issue:45588",
+        "issue:46002",
+        "issue:46077",
+        "issue:46178",
+        "issue:46328",
+        "issue:46512",
+        "issue:46614"
+      ],
+      "soft_pairs": [
+        "issue:44625|issue:46002",
+        "issue:44075|issue:44936",
+        "issue:45357|issue:45478",
+        "issue:43116|issue:43825",
+        "issue:42222|issue:43441",
+        "issue:45083|issue:45446",
+        "issue:42673|issue:42796",
+        "issue:43064|issue:43673",
+        "issue:43937|issue:45563",
+        "issue:43827|issue:45446",
+        "issue:33453|issue:39401",
+        "issue:44380|issue:46077",
+        "issue:45356|issue:46614",
+        "issue:44117|issue:45588",
+        "issue:45071|issue:46614",
+        "issue:43673|issue:46328",
+        "issue:44291|issue:45071",
+        "issue:42673|issue:43906",
+        "issue:39401|issue:44779",
+        "issue:44466|issue:46002",
+        "issue:43450|issue:43986",
+        "issue:43937|issue:44977",
+        "issue:45405|issue:45446",
+        "issue:44410|issue:45310",
+        "issue:38617|issue:41628",
+        "issue:45341|issue:46328",
+        "issue:45341|issue:45561",
+        "issue:44829|issue:45216",
+        "issue:45071|issue:46512",
+        "issue:44117|issue:45081",
+        "issue:42222|issue:46178"
+      ],
+      "prepared_review_unit_hash": "3ff66aa0327b256421ed1a3bcc137bf660e6f1a353b03c48f4b638f9abe189ba",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 31,
+        "serialized_chars": 37136,
+        "estimated_input_tokens": 9284,
+        "estimated_eval_tokens": 18824
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are mostly unrelated bugs across tokenizers, model loading, training, generation, and docs. The soft-similarity scores are driven by shared Transformers vocabulary, not the same underlying defect, so no duplicate pairs are strong enough to merge.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "No single duplicate target emerges from the set; if one must be chosen, issue 45356 is the most concrete and well-scoped regression report.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 45356 is the best representative because it is specific, reproducible, and points to a focused tokenizer regression with a plausible code-path root cause.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44625",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Both concern classification heads, but one is about num_labels propagation and the other about NaNs after checkout/init; different bugs."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:44936",
+            "accept": false,
+            "reason": "Optimizer-argument handling and trainer evaluate/train interaction are separate code paths and failure modes."
+          },
+          {
+            "left": "issue:45357",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Same model family, but save_pretrained visual keys and from_pretrained load failure are not the same concrete bug."
+          },
+          {
+            "left": "issue:43116",
+            "right": "issue:43825",
+            "accept": false,
+            "reason": "Example script multi-label results and pipeline translation support messaging are unrelated issues."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:43441",
+            "accept": false,
+            "reason": "Different subsystems and failures: vitpose model breakage vs flash-attention failure for Ministral-3."
+          },
+          {
+            "left": "issue:45083",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "One is a qwen3_omni_moe helper behavior issue; the other is a flex_attention version-check import bug."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:42796",
+            "accept": false,
+            "reason": "VRAM leak in Qwen3 causal LM and tokenizer BPE-marker handling are unrelated."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "Training optimizer-state corruption and generation cache/chunked_prefill are different problems."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:45563",
+            "accept": false,
+            "reason": "Invalid GenerationConfig and stale generate warning are distinct generation issues."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Docs still mentioning pipeline() and a PyTorch version check bug are unrelated."
+          },
+          {
+            "left": "issue:33453",
+            "right": "issue:39401",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but they affect different models and different tokenizer behaviors."
+          },
+          {
+            "left": "issue:44380",
+            "right": "issue:46077",
+            "accept": false,
+            "reason": "Attention scaling ignored and bidirectional masks built causally are separate attention bugs."
+          },
+          {
+            "left": "issue:45356",
+            "right": "issue:46614",
+            "accept": false,
+            "reason": "Both involve tokenization/phonemization, but they are different models and different statefulness bugs."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "TOKENIZER_MAPPING_NAMES None handling and flash_attention s_aux=None crash are unrelated."
+          },
+          {
+            "left": "issue:45071",
+            "right": "issue:46614",
+            "accept": false,
+            "reason": "PretrainedConfig type checking and stale phonemizer language state have no shared code-path."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "Generation cache handling and SequenceFeatureExtractor.pad performance are unrelated."
+          },
+          {
+            "left": "issue:44291",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Both are v5 compatibility regressions, but one is init_empty_weights argument handling and the other is PretrainedConfig type checking."
+          },
+          {
+            "left": "issue:42673",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "The second is an isolated reproduction reference, but the titles do not indicate the same concrete bug."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:44779",
+            "accept": false,
+            "reason": "Both are tokenizer regressions, but for different models and outputs; not the same underlying defect."
+          },
+          {
+            "left": "issue:44466",
+            "right": "issue:46002",
+            "accept": false,
+            "reason": "Serialization inconsistency and NaNs in classification heads are different failure modes."
+          },
+          {
+            "left": "issue:43450",
+            "right": "issue:43986",
+            "accept": false,
+            "reason": "Batched video processor shape bug and missing-torchvision crash are distinct video-processing issues."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44977",
+            "accept": false,
+            "reason": "GenerationConfig validation and flash-attention generation failure are different."
+          },
+          {
+            "left": "issue:45405",
+            "right": "issue:45446",
+            "accept": false,
+            "reason": "Dependency version bump policy and a backend import version check are unrelated."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Missing qwen3next projections and Qwen3.5 MoE from_pretrained error are different model-specific bugs."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:41628",
+            "accept": false,
+            "reason": "Both are import errors, but they concern different symbols and modules."
+          },
+          {
+            "left": "issue:45341",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "A testing_utils bug and a feature-extractor pad inefficiency are unrelated."
+          },
+          {
+            "left": "issue:45341",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "Both touch testing utils, but one is vague and the other is a specific xdist race; not enough to treat as the same bug."
+          },
+          {
+            "left": "issue:44829",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Training degeneration with flash_attention_3 and incorrect checkpoint serialization are separate issues."
+          },
+          {
+            "left": "issue:45071",
+            "right": "issue:46512",
+            "accept": false,
+            "reason": "dtype type-checking breakage and stale config.dtype after from_config are different problems."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:45081",
+            "accept": false,
+            "reason": "Tokenizer mapping fallback and Mistral regex patch crash are unrelated."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "Vitpose model breakage and tensorflow_text backend mapping error 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": 21,
+      "review_unit_index": 48,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:27957",
+        "issue:38617",
+        "issue:39401",
+        "issue:41628",
+        "issue:42175",
+        "issue:42222",
+        "issue:42371",
+        "issue:42490",
+        "issue:42617",
+        "issue:42673",
+        "issue:43012",
+        "issue:43066",
+        "issue:43097",
+        "issue:43122",
+        "issue:43352",
+        "issue:43408",
+        "issue:43577",
+        "issue:43673",
+        "issue:43742",
+        "issue:43874",
+        "issue:43931",
+        "issue:43937",
+        "issue:44075",
+        "issue:44220",
+        "issue:44261",
+        "issue:44263",
+        "issue:44410",
+        "issue:44483",
+        "issue:44484",
+        "issue:44485",
+        "issue:44493",
+        "issue:44496",
+        "issue:44877",
+        "issue:44933",
+        "issue:45071",
+        "issue:45278",
+        "issue:45310",
+        "issue:45341",
+        "issue:45362",
+        "issue:45464",
+        "issue:45478",
+        "issue:45542",
+        "issue:45656",
+        "issue:46209",
+        "issue:46231",
+        "issue:46298",
+        "issue:46328",
+        "issue:46617"
+      ],
+      "soft_pairs": [
+        "issue:42617|issue:45341",
+        "issue:43122|issue:46298",
+        "issue:43408|issue:43742",
+        "issue:45362|issue:46328",
+        "issue:42371|issue:44261",
+        "issue:44877|issue:46231",
+        "issue:42490|issue:42673",
+        "issue:42371|issue:46617",
+        "issue:44484|issue:45310",
+        "issue:38617|issue:44220",
+        "issue:39401|issue:46209",
+        "issue:43937|issue:44496",
+        "issue:44933|issue:45278",
+        "issue:44410|issue:45478",
+        "issue:42222|issue:45542",
+        "issue:44075|issue:45656",
+        "issue:43066|issue:43122",
+        "issue:42175|issue:42222",
+        "issue:43097|issue:44493",
+        "issue:44263|issue:44485",
+        "issue:43012|issue:43577",
+        "issue:43673|issue:45341",
+        "issue:43874|issue:44261",
+        "issue:41628|issue:45071",
+        "issue:45071|issue:45542",
+        "issue:43931|issue:44410",
+        "issue:27957|issue:43352",
+        "issue:44483|issue:45464"
+      ],
+      "prepared_review_unit_hash": "8f965681ad9268ecf8f348bcaae804d56bfe5d506ffdd674454cef28356ae571",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 28,
+        "serialized_chars": 35430,
+        "estimated_input_tokens": 8858,
+        "estimated_eval_tokens": 17972
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These items are overwhelmingly unrelated, spanning distinct model-specific bugs, import regressions, API warnings, and docs/test issues. Only one soft pair looks like a true duplicate: the v5 tokenizer regression around decoder type / changed tokenization.",
+        "confidence": 0.91,
+        "canonical_issue_reason": "If forced to pick a canonical issue, issue:43066 is the best anchor for the only credible duplicate pair because it states the concrete root cause (wrong tokenizer decoder type in v5) rather than just the downstream symptom.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43066 is the most actionable representative of the only near-duplicate cluster; 43122 reads like the user-visible manifestation of the same tokenizer regression.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:42617",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Different domains: 3D parallel execution failure vs a small testing_utils bug."
+          },
+          {
+            "left": "issue:43122",
+            "right": "issue:46298",
+            "accept": false,
+            "reason": "Tokenizer behavior regression vs noisy logging; no shared underlying bug."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:43742",
+            "accept": false,
+            "reason": "SAM3 tracker/model warning and MobileLLM key error are unrelated model-loading problems."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "Qwen chat crash vs SequenceFeatureExtractor.pad performance issue; unrelated code paths."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44261",
+            "accept": false,
+            "reason": "TF32 warning about new API settings is not the same as the MLA rms_norm_eps precision bug."
+          },
+          {
+            "left": "issue:44877",
+            "right": "issue:46231",
+            "accept": false,
+            "reason": "Strict config loading failure and a docs parameter-name mismatch are different issues."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:42673",
+            "accept": false,
+            "reason": "Local-path import warning and VRAM leak in multithreaded dataloading are unrelated."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "TF32 API warning and trust_remote_code cache FileNotFoundError do not share a code-path."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Shard-size default question and Qwen3.5 MoE from_pretrained failure are unrelated."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:44220",
+            "accept": false,
+            "reason": "Missing configuration_utils symbol import and audio fbank extraction bug are separate regressions."
+          },
+          {
+            "left": "issue:39401",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Qwen3 tokenizer offset mapping bug and FP8 inference failure on B200 are unrelated."
+          },
+          {
+            "left": "issue:43937",
+            "right": "issue:44496",
+            "accept": false,
+            "reason": "GLM-5 generation config validation and OLMo-Hybrid model recognition are different loader issues."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "A missing image_utils import and a broad upgrade-related import regression are not the same concrete bug."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Missing qwen3next projections and Qwen3.5 MoE from_pretrained error are different model failures."
+          },
+          {
+            "left": "issue:42222",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "VitPose breakage and tensorboard-without-TF backend errors are unrelated."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:45656",
+            "accept": false,
+            "reason": "SGD args ignored and deepspeed causing double optimizer steps are distinct optimizer bugs."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:43122",
+            "accept": true,
+            "reason": "Both report the same v5 tokenizer regression: wrong decoder type / changed tokenization with the same tokenizer."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:42222",
+            "accept": false,
+            "reason": "Missing TensorFlow in the torch extra and broken vitpose models are unrelated."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:44493",
+            "accept": false,
+            "reason": "Removed tie_embeddings flag and unexpected position-id keys are different compatibility issues."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "torch.split return handling in GLM MoE indexer and RoPE implementation concerns are not the same bug."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43577",
+            "accept": false,
+            "reason": "PyTorch float32 warning during compile and Blip2 dtype propagation are separate dtype issues."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Generation cache missing in v5 chunked_prefill and a testing_utils bug are unrelated."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:44261",
+            "accept": false,
+            "reason": "Missing image patch-count method and MLA layernorm epsilon precision are unrelated."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "AutoImageProcessor import error and PreTrainedModel type-checking regression are different API breakages."
+          },
+          {
+            "left": "issue:45071",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Type-checking regression and TF backend installation issue are unrelated."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44410",
+            "accept": false,
+            "reason": "Qwen3-VL weight-shape mismatch and qwen3next missing projections are different model-loading failures."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:43352",
+            "accept": false,
+            "reason": "Both mention Flash Attention 2, but they concern different models and different support failures."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45464",
+            "accept": false,
+            "reason": "General chat/completions API rejection and a Qwen3.5 streaming inference failure are not the same concrete issue."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The review is conservative and grounded in the packet. The only accepted soft pair, issue:43066 with issue:43122, is a plausible same-regression duplicate based on the titles, and the rest are reasonably rejected as unrelated or at least not clearly duplicates. No obvious overstatement."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 49,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:38617",
+        "issue:42757",
+        "issue:42890",
+        "issue:43064",
+        "issue:43097",
+        "issue:43441",
+        "issue:43452",
+        "issue:43502",
+        "issue:43531",
+        "issue:43673",
+        "issue:43873",
+        "issue:43874",
+        "issue:43931",
+        "issue:43986",
+        "issue:43992",
+        "issue:44038",
+        "issue:44062",
+        "issue:44117",
+        "issue:44261",
+        "issue:44263",
+        "issue:44315",
+        "issue:44351",
+        "issue:44360",
+        "issue:44410",
+        "issue:44484",
+        "issue:44485",
+        "issue:44704",
+        "issue:44829",
+        "issue:44933",
+        "issue:44936",
+        "issue:44987",
+        "issue:45003",
+        "issue:45103",
+        "issue:45341",
+        "issue:45372",
+        "issue:45375",
+        "issue:45440",
+        "issue:45478",
+        "issue:45588",
+        "issue:45753",
+        "issue:46050",
+        "issue:46123",
+        "issue:46172",
+        "issue:46231",
+        "issue:46298",
+        "issue:46365",
+        "issue:46369"
+      ],
+      "soft_pairs": [
+        "issue:44263|issue:46172",
+        "issue:44360|issue:46172",
+        "issue:44484|issue:45478",
+        "issue:43441|issue:44829",
+        "issue:43931|issue:44038",
+        "issue:43097|issue:43452",
+        "issue:45003|issue:45341",
+        "issue:43986|issue:45341",
+        "issue:45753|issue:46369",
+        "issue:44933|issue:46231",
+        "issue:43874|issue:44263",
+        "issue:44261|issue:44485",
+        "issue:44117|issue:44987",
+        "issue:43673|issue:46123",
+        "issue:43531|issue:45341",
+        "issue:43986|issue:44484",
+        "issue:45440|issue:46172",
+        "issue:43064|issue:44315",
+        "issue:38617|issue:42757",
+        "issue:43097|issue:45588",
+        "issue:43986|issue:45372",
+        "issue:43097|issue:46050",
+        "issue:44410|issue:45375",
+        "issue:44038|issue:44410",
+        "issue:38617|issue:44351",
+        "issue:43502|issue:43992",
+        "issue:42890|issue:46298",
+        "issue:44062|issue:44117",
+        "issue:43097|issue:43873",
+        "issue:44936|issue:46365",
+        "issue:44704|issue:45103",
+        "issue:43986|issue:44117"
+      ],
+      "prepared_review_unit_hash": "38d6ad199b29cdbaa7fa43c87c7aae0920cf8f53a10263cde4026e5fa2b10368",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 47,
+        "item_count": 47,
+        "soft_pair_count": 32,
+        "serialized_chars": 36527,
+        "estimated_input_tokens": 9132,
+        "estimated_eval_tokens": 18520
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is heterogeneous: it mixes unrelated import errors, model-loading regressions, training bugs, attention/kernel issues, docs, and test-only problems. None of the soft edges look like the same underlying bug or change.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "No clear canonical duplicate: the items span distinct failure modes and code paths, so there isn\u2019t one issue that meaningfully represents the whole cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "If one issue must anchor the cluster, issue:43097 is the broadest release-regression style report and the closest thing to a generic upstream tracker, but it still does not unify these unrelated reports.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44263",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "Different bugs: DSA indexer return handling vs DeepSeek-v4 transpose logic."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "Unrelated failure modes: missing ReLU in DSA indexer vs DeepSeek-v4 tensor transpose."
+          },
+          {
+            "left": "issue:44484",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Max shard size default question is unrelated to Qwen3.5 MoE load errors."
+          },
+          {
+            "left": "issue:43441",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Both involve flash attention, but one is a model load crash and the other is degenerate training on a different code path."
+          },
+          {
+            "left": "issue:43931",
+            "right": "issue:44038",
+            "accept": false,
+            "reason": "Related model family, but different concrete bugs and different variants; not the same issue."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43452",
+            "accept": false,
+            "reason": "API removal/tie_embeddings issue is unrelated to gguf_file loading failure."
+          },
+          {
+            "left": "issue:45003",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Unsafe sys.modules access is unrelated to a testing_utils bug."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Video-model loading without torchvision is unrelated to the testing_utils report."
+          },
+          {
+            "left": "issue:45753",
+            "right": "issue:46369",
+            "accept": false,
+            "reason": "Different models and symptoms: infinite loop on Qwen3_5 vs memory overflow on sam3-litetext."
+          },
+          {
+            "left": "issue:44933",
+            "right": "issue:46231",
+            "accept": false,
+            "reason": "Import error from image_utils is unrelated to a documentation parameter-name mismatch."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:44263",
+            "accept": false,
+            "reason": "Different GLM components and bugs; missing image patch method is not the same as torch.split indexing behavior."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Precision/config epsilon discussion is unrelated to GLM-5 RoPE implementation."
+          },
+          {
+            "left": "issue:44117",
+            "right": "issue:44987",
+            "accept": false,
+            "reason": "Tokenizer mapping None vs a specific model-loading failure are not the same concrete bug."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:46123",
+            "accept": false,
+            "reason": "Generation cache missing during chunked prefill is different from the pipeline is_last batching bug."
+          },
+          {
+            "left": "issue:43531",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Qwen3-MoE sliding_window issue is unrelated to testing_utils."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Missing torchvision during processor loading has nothing to do with save_pretrained shard size."
+          },
+          {
+            "left": "issue:45440",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "DeepSeek implementation divergence is broader and distinct from the DeepSeek-v4 transpose bug."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44315",
+            "accept": false,
+            "reason": "Trainer/FSDP optimizer-state corruption is unrelated to Liger Kernel not applying with model_init."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and different upstream API surfaces."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:45588",
+            "accept": false,
+            "reason": "Tie-embeddings removal is unrelated to flash_attention s_aux handling."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:45372",
+            "accept": false,
+            "reason": "Torchvision dependency crash is unrelated to ReasoningEffort import failure."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:46050",
+            "accept": false,
+            "reason": "API removal regression is unrelated to sinq quantization save/load breakage."
+          },
+          {
+            "left": "issue:44410",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Different Qwen3-family config/architecture issues; not the same bug."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44410",
+            "accept": false,
+            "reason": "Both mention Qwen3 variants, but one is a model load bug and the other is missing projections in a different architecture."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Separate import-name failures from different modules and symbols."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:43992",
+            "accept": false,
+            "reason": "Offline-mode network leakage is unrelated to missing embed_tokens weights on load."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:46298",
+            "accept": false,
+            "reason": "Test flakiness due to missing set_seed is unrelated to a noisy logging message."
+          },
+          {
+            "left": "issue:44062",
+            "right": "issue:44117",
+            "accept": false,
+            "reason": "AddedToken keyword conflict is a different tokenizer bug than TOKENIZER_MAPPING_NAMES returning None."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:43873",
+            "accept": false,
+            "reason": "Tie-embeddings removal and quantization/offloading behavior are unrelated."
+          },
+          {
+            "left": "issue:44936",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Trainer.evaluate after train is a runtime bug; import failure until uninstalling datasets is a different dependency issue."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:45103",
+            "accept": false,
+            "reason": "Passing kwargs to cached_file and auto_docstring future-annotations crash are unrelated."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44117",
+            "accept": false,
+            "reason": "Processor loading without torchvision and tokenizer mapping None are separate problems."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 50,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:36683",
+        "issue:38617",
+        "issue:42175",
+        "issue:42490",
+        "issue:42757",
+        "issue:42907",
+        "issue:43012",
+        "issue:43064",
+        "issue:43066",
+        "issue:43097",
+        "issue:43208",
+        "issue:43352",
+        "issue:43502",
+        "issue:43526",
+        "issue:43550",
+        "issue:43582",
+        "issue:43824",
+        "issue:43825",
+        "issue:43906",
+        "issue:43976",
+        "issue:43986",
+        "issue:43992",
+        "issue:44016",
+        "issue:44038",
+        "issue:44077",
+        "issue:44162",
+        "issue:44222",
+        "issue:44230",
+        "issue:44368",
+        "issue:44380",
+        "issue:44483",
+        "issue:44484",
+        "issue:44514",
+        "issue:44655",
+        "issue:44683",
+        "issue:44938",
+        "issue:45478",
+        "issue:45479",
+        "issue:45584",
+        "issue:45600",
+        "issue:45712",
+        "issue:45753",
+        "issue:46132",
+        "issue:46172",
+        "issue:46178",
+        "issue:46298",
+        "issue:46365",
+        "issue:46421"
+      ],
+      "soft_pairs": [
+        "issue:44380|issue:46421",
+        "issue:44938|issue:46178",
+        "issue:43986|issue:44016",
+        "issue:43208|issue:43550",
+        "issue:43012|issue:43550",
+        "issue:42907|issue:44655",
+        "issue:42907|issue:45753",
+        "issue:43012|issue:44368",
+        "issue:43825|issue:44655",
+        "issue:43992|issue:44484",
+        "issue:45712|issue:46298",
+        "issue:43066|issue:43906",
+        "issue:43352|issue:44380",
+        "issue:43097|issue:46298",
+        "issue:43352|issue:43976",
+        "issue:44162|issue:46172",
+        "issue:44380|issue:46132",
+        "issue:45584|issue:46365",
+        "issue:44077|issue:46365",
+        "issue:44222|issue:45478",
+        "issue:44380|issue:45479",
+        "issue:36683|issue:43582",
+        "issue:43064|issue:44380",
+        "issue:43012|issue:44683",
+        "issue:43208|issue:43526",
+        "issue:42175|issue:46365",
+        "issue:42757|issue:43824",
+        "issue:42490|issue:43012",
+        "issue:38617|issue:43502",
+        "issue:44483|issue:44514",
+        "issue:45600|issue:46298",
+        "issue:44038|issue:44230"
+      ],
+      "prepared_review_unit_hash": "d8323a86c358c0f0fef3ab66a783b26e1014448ff8d76fb0d36f731639aeeb7f",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 32,
+        "serialized_chars": 37069,
+        "estimated_input_tokens": 9268,
+        "estimated_eval_tokens": 18792
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This set is mostly a grab bag of unrelated standalone bugs. None of the soft-similarity pairs look like the same concrete underlying issue, so I did not merge any of them as duplicates.",
+        "confidence": 0.92,
+        "canonical_issue_reason": "No single issue is a clear duplicate target; the items span unrelated bugs across imports, attention/backends, saving/loading, training, and model-specific failures.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No issue cleanly represents the whole set as a canonical duplicate; the cluster is too heterogeneous to pick one representative bug.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44380",
+            "right": "issue:46421",
+            "accept": false,
+            "reason": "Different bugs: GPT-2 attention scaling backend behavior vs cache lazy initialization causing recompilation."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:46178",
+            "accept": false,
+            "reason": "Unrelated failures: Python 3.14 import/load issue vs invalid backend mapping for tensorflow_text."
+          },
+          {
+            "left": "issue:43986",
+            "right": "issue:44016",
+            "accept": false,
+            "reason": "Different domains: AutoProcessor/torchvision crash vs a notebook syntax error."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "Different model bugs: xLSTM training blockers vs Bamba torch.compile/SDPA failure."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:43550",
+            "accept": false,
+            "reason": "A PyTorch warning about bfloat16 precision is not the same as a Bamba compile failure."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Both involve saving, but one is dequantized model saving and the other is Pipeline.save_pretrained."
+          },
+          {
+            "left": "issue:42907",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "Saving dequantized models and Qwen3_5 image infinite-loop behavior are unrelated."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:44368",
+            "accept": false,
+            "reason": "Different concerns: compile-time precision warning vs tie_word_embeddings config warning."
+          },
+          {
+            "left": "issue:43825",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Pipeline task messaging and Pipeline serialization are different bugs."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:44484",
+            "accept": false,
+            "reason": "Loading missing weights is unrelated to the default max_shard_size question."
+          },
+          {
+            "left": "issue:45712",
+            "right": "issue:46298",
+            "accept": false,
+            "reason": "Repository dummy classes/check_repo failure is not the same as a noisy logging message."
+          },
+          {
+            "left": "issue:43066",
+            "right": "issue:43906",
+            "accept": false,
+            "reason": "No clear same bug: tokenizer decoder type regression vs an isolated reproduction of another issue."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Different models and failures: Nemotron FlashAttention support vs GPT-2 attention scaling/backend handling."
+          },
+          {
+            "left": "issue:43097",
+            "right": "issue:46298",
+            "accept": false,
+            "reason": "A removed config field is unrelated to a logging-message annoyance."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "Model-specific FlashAttention support issue vs a Python version compatibility problem."
+          },
+          {
+            "left": "issue:44162",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "ESM2 training/runtime breakage and a DeepSeek tensor transpose issue are different code paths."
+          },
+          {
+            "left": "issue:44380",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Backend config being ignored is not the same as AttentionInterface.register altering function behavior."
+          },
+          {
+            "left": "issue:45584",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Whisper empty transcription failure is unrelated to importing Trainer with datasets installed."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "patchtsmixer post_init validation and Trainer import conflicts are unrelated."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "FP8 save_pretrained for MoE models is not the same as a from_pretrained regression in transformers>=5.4.0."
+          },
+          {
+            "left": "issue:44380",
+            "right": "issue:45479",
+            "accept": false,
+            "reason": "Attention scaling/backend behavior is unrelated to sequence-classification zero-loss degeneracy."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43582",
+            "accept": false,
+            "reason": "Gemma3Config vocab_size missing and Apple Silicon allocator TypeError are different bugs."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "FSDP2/PEFT optimizer-state mismatch is unrelated to GPT-2 attention scaling handling."
+          },
+          {
+            "left": "issue:43012",
+            "right": "issue:44683",
+            "accept": false,
+            "reason": "A warning about bfloat16 precision is not the same as compiled flex_attention failing on torch>=2.9."
+          },
+          {
+            "left": "issue:43208",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "xLSTM training blockers and BeitImageProcessorFast reduce_labels behavior are different problems."
+          },
+          {
+            "left": "issue:42175",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "TensorFlow backend packaging issue and Trainer import failure are unrelated."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "Both are import errors, but for different symbols and libraries, so not the same bug."
+          },
+          {
+            "left": "issue:42490",
+            "right": "issue:43012",
+            "accept": false,
+            "reason": "Trailing-slash local model import warning is unrelated to a bfloat16 compile warning."
+          },
+          {
+            "left": "issue:38617",
+            "right": "issue:43502",
+            "accept": false,
+            "reason": "A missing internal import and unwanted network API calls are different failure modes."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:44514",
+            "accept": false,
+            "reason": "OpenAI chat-completions API rejection and Qwen2_5_VLProcessor batch padding crash are different code paths."
+          },
+          {
+            "left": "issue:45600",
+            "right": "issue:46298",
+            "accept": false,
+            "reason": "Broken auto_mappings references and noisy logging are not the same underlying issue."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "Both mention Qwen3-VL-Moe/fp8, but one is a transformers bug report and the other is fp8 inference support; not 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": 21,
+      "review_unit_index": 51,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:36010",
+        "issue:36683",
+        "issue:42757",
+        "issue:42890",
+        "issue:42898",
+        "issue:42994",
+        "issue:43010",
+        "issue:43064",
+        "issue:43066",
+        "issue:43299",
+        "issue:43352",
+        "issue:43475",
+        "issue:43525",
+        "issue:43611",
+        "issue:43618",
+        "issue:43756",
+        "issue:43761",
+        "issue:43867",
+        "issue:43901",
+        "issue:43937",
+        "issue:43976",
+        "issue:43992",
+        "issue:44075",
+        "issue:44077",
+        "issue:44355",
+        "issue:44380",
+        "issue:44483",
+        "issue:44485",
+        "issue:44492",
+        "issue:44625",
+        "issue:44655",
+        "issue:44855",
+        "issue:44938",
+        "issue:45071",
+        "issue:45216",
+        "issue:45276",
+        "issue:45278",
+        "issue:45290",
+        "issue:45310",
+        "issue:45375",
+        "issue:45561",
+        "issue:45593",
+        "issue:45698",
+        "issue:46077",
+        "issue:46082",
+        "issue:46172",
+        "issue:46202",
+        "issue:46365"
+      ],
+      "soft_pairs": [
+        "issue:45698|issue:46172",
+        "issue:43618|issue:44380",
+        "issue:43618|issue:43761",
+        "issue:44483|issue:45290",
+        "issue:43299|issue:43867",
+        "issue:42994|issue:44655",
+        "issue:43756|issue:44485",
+        "issue:36683|issue:43475",
+        "issue:42898|issue:43066",
+        "issue:44625|issue:46365",
+        "issue:36010|issue:42757",
+        "issue:45276|issue:46077",
+        "issue:43352|issue:43937",
+        "issue:43611|issue:44655",
+        "issue:43901|issue:44655",
+        "issue:36683|issue:44077",
+        "issue:43867|issue:45216",
+        "issue:43352|issue:45593",
+        "issue:43976|issue:44938",
+        "issue:44625|issue:46077",
+        "issue:46077|issue:46082",
+        "issue:36683|issue:43525",
+        "issue:43618|issue:45216",
+        "issue:36683|issue:44625",
+        "issue:43992|issue:45310",
+        "issue:43010|issue:45278",
+        "issue:43064|issue:44492",
+        "issue:44075|issue:46365",
+        "issue:44355|issue:44855",
+        "issue:44938|issue:45071",
+        "issue:43867|issue:45375",
+        "issue:43867|issue:46202",
+        "issue:36683|issue:43064",
+        "issue:42890|issue:45561"
+      ],
+      "prepared_review_unit_hash": "2ad0bbd3b31d0f30c4839a0c0126beafae8d9e5a9e6b6c27f629dca7bf0cc5ed",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 34,
+        "serialized_chars": 37506,
+        "estimated_input_tokens": 9377,
+        "estimated_eval_tokens": 19010
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a heterogeneous set of Transformers v5 regression/compatibility reports. The CLIPVision pair is the only near-match, but the packet only supports a cautious similarity call, not a firm duplicate merge.",
+        "confidence": 0.67,
+        "canonical_issue_reason": "issue:43761 is the strongest representative issue in the cluster because it states a concrete v5 CLIPVision regression with an explicit model/API symptom; issue:43618 is related, but the available metadata does not justify collapsing them into one bug.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43761 is the best overall issue anchor because it is the most specific regression report in the otherwise mixed cluster and most clearly describes the affected code path.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45698",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "Different bugs: custom-module save/load roundtripping versus a DeepSeek-v4 tensor transpose issue."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44380",
+            "accept": false,
+            "reason": "Both are regression-style model issues, but they affect different models and paths: CLIPVision outputs versus GPT2 attention scaling under SDPA/FlashAttention."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:43761",
+            "accept": false,
+            "reason": "These are related CLIPVision v5 output regressions, but one is about attentions assignment and the other about hidden_states; the packet does not prove they are the same underlying bug or fix."
+          },
+          {
+            "left": "issue:44483",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Different layers and symptoms: chat/completions request acceptance versus apply_chat_template crashing on tool-call messages."
+          },
+          {
+            "left": "issue:43299",
+            "right": "issue:43867",
+            "accept": false,
+            "reason": "Both are loading failures, but for different models and different failure modes."
+          },
+          {
+            "left": "issue:42994",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Both mention save_pretrained, but quantized-model saving and Pipeline serialization are different code paths."
+          },
+          {
+            "left": "issue:43756",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Different model families and implementation details; not the same RoPE bug."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43475",
+            "accept": false,
+            "reason": "Unrelated missing-attribute bugs in different model/config objects."
+          },
+          {
+            "left": "issue:42898",
+            "right": "issue:43066",
+            "accept": false,
+            "reason": "Both concern tokenizer behavior in v5, but cleanup-space semantics and decoder type selection are separate issues."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "Different scope: Qwen3.5 config propagation versus importing Trainer being blocked by datasets installation."
+          },
+          {
+            "left": "issue:36010",
+            "right": "issue:42757",
+            "accept": false,
+            "reason": "Both are import errors, but they involve different modules and breakages."
+          },
+          {
+            "left": "issue:45276",
+            "right": "issue:46077",
+            "accept": false,
+            "reason": "Same model family, but different functionality: resizing embeddings versus bidirectional attention masks."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:43937",
+            "accept": false,
+            "reason": "Different problems: FlashAttention support gating versus invalid GenerationConfig for GLM-5."
+          },
+          {
+            "left": "issue:43611",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Different serialization/loading paths: base_model_prefix loading regression versus Pipeline save_pretrained."
+          },
+          {
+            "left": "issue:43901",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "Docs mismatch for TextClassificationPipeline return_all_scores is unrelated to Pipeline save_pretrained behavior."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "Different bug classes: missing model config attribute versus optimizer-argument handling."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "Both concern checkpoint/load behavior, but one is a load failure with sorted state_dict and the other is incorrect saved checkpoint content."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:45593",
+            "accept": false,
+            "reason": "Different model-specific bugs: Flash Attention 2 support versus auxiliary losses being omitted."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44938",
+            "accept": false,
+            "reason": "Both are Python-version compatibility reports, but they target different versions and likely different causes."
+          },
+          {
+            "left": "issue:44625",
+            "right": "issue:46077",
+            "accept": false,
+            "reason": "Same broad ecosystem only; one is num_labels propagation and the other is attention-mask construction."
+          },
+          {
+            "left": "issue:46077",
+            "right": "issue:46082",
+            "accept": false,
+            "reason": "Different config validation problems for different models and parameters."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43525",
+            "accept": false,
+            "reason": "Distinct missing-config-attribute issues in different models."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:45216",
+            "accept": false,
+            "reason": "CLIPVision output regression and Qwen3.5 save_pretrained regression are unrelated."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:44625",
+            "accept": false,
+            "reason": "Different config propagation bugs in different model families."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Both involve from_pretrained, but the model families and failure symptoms differ."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:45278",
+            "accept": false,
+            "reason": "Cache update semantics and many import errors after upgrade are unrelated."
+          },
+          {
+            "left": "issue:43064",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "FSDP2/PEFT optimizer-state corruption and a typo in cache strategies are unrelated."
+          },
+          {
+            "left": "issue:44075",
+            "right": "issue:46365",
+            "accept": false,
+            "reason": "SGD argument handling is unrelated to importing Trainer being blocked by datasets."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44855",
+            "accept": false,
+            "reason": "Both are import-related, but one is compiled-file execution errors and the other is a Python 3.13 JIT parsing/IndentationError issue."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Python 3.14 load failure and PretrainedConfig type-checking regression are different issues."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Sorted-state_dict load failure is unrelated to Qwen3_5MoeVisionConfig silently dropping a field."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:46202",
+            "accept": false,
+            "reason": "Checkpoint load ordering failure and Gemma pipeline roundtrip tokenization failure are different bugs."
+          },
+          {
+            "left": "issue:36683",
+            "right": "issue:43064",
+            "accept": false,
+            "reason": "A missing config attribute and distributed training optimizer-state corruption are not the same underlying bug."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:45561",
+            "accept": false,
+            "reason": "One is flaky model test seeding, the other is an xdist race on captured_info.txt in testing utils."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The cluster summary is mostly grounded, but it overstates evidence in a few places. In particular, the claim that issue:43618 is \"clearly reproducible\" is not supported by the packet, and the accepted duplicate verdict for issue:43618 vs issue:43761 is a bit stronger than the available metadata alone justifies. The rest of the soft-edge rejections look conservative, but I would not accept this as written."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 52,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:28282",
+        "issue:30333",
+        "issue:30990",
+        "issue:31795",
+        "issue:33453",
+        "issue:36296",
+        "issue:41669",
+        "issue:42548",
+        "issue:42890",
+        "issue:43010",
+        "issue:43065",
+        "issue:43232",
+        "issue:43502",
+        "issue:43575",
+        "issue:43576",
+        "issue:43618",
+        "issue:43723",
+        "issue:43824",
+        "issue:43867",
+        "issue:43874",
+        "issue:43976",
+        "issue:43992",
+        "issue:44062",
+        "issue:44077",
+        "issue:44112",
+        "issue:44222",
+        "issue:44246",
+        "issue:44351",
+        "issue:44355",
+        "issue:44360",
+        "issue:44387",
+        "issue:44485",
+        "issue:44556",
+        "issue:44734",
+        "issue:44749",
+        "issue:44857",
+        "issue:44961",
+        "issue:44964",
+        "issue:45125",
+        "issue:45357",
+        "issue:45375",
+        "issue:45440",
+        "issue:45478",
+        "issue:45600",
+        "issue:45850",
+        "issue:46132",
+        "issue:46285",
+        "issue:46617"
+      ],
+      "soft_pairs": [
+        "issue:43010|issue:45850",
+        "issue:41669|issue:44246",
+        "issue:36296|issue:43575",
+        "issue:43867|issue:45357",
+        "issue:43010|issue:43976",
+        "issue:43992|issue:45478",
+        "issue:43618|issue:44077",
+        "issue:28282|issue:33453",
+        "issue:43618|issue:46285",
+        "issue:43618|issue:44857",
+        "issue:44360|issue:45440",
+        "issue:42890|issue:44964",
+        "issue:43502|issue:44062",
+        "issue:42890|issue:44112",
+        "issue:42890|issue:45600",
+        "issue:43723|issue:44556",
+        "issue:44387|issue:44556",
+        "issue:43874|issue:46617",
+        "issue:44222|issue:45125",
+        "issue:44734|issue:46617",
+        "issue:43010|issue:43065",
+        "issue:43576|issue:46132",
+        "issue:43502|issue:44485",
+        "issue:43824|issue:46617",
+        "issue:45125|issue:45478",
+        "issue:44355|issue:44749",
+        "issue:30333|issue:33453",
+        "issue:42548|issue:44351",
+        "issue:45125|issue:45375",
+        "issue:43010|issue:43232",
+        "issue:43502|issue:44222",
+        "issue:43502|issue:44961",
+        "issue:30990|issue:31795"
+      ],
+      "prepared_review_unit_hash": "c4dfd665dc5b929d65a1e9803108a709f89c4692c3cef6674ed85df3fbe0f4da",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 37136,
+        "estimated_input_tokens": 9284,
+        "estimated_eval_tokens": 18824
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a loose mix of Transformers v5 regressions, import/load failures, and a few unrelated model-specific bugs. Most pairs are not true duplicates; only the import-time slowdown reports look like the same underlying issue.",
+        "confidence": 0.68,
+        "canonical_issue_reason": "issue:45850 is the broadest umbrella report in the set ('Since 5.0 version it breaks too many models'), so it best represents the cluster\u2019s general v5-regression theme, even though many items are actually distinct bugs.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:45850 is the most general and highest-level issue here, making it the best single anchor for this mixed regression cluster.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43010",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "Different problems: one is about decorating cache/layer update methods with no_grad, the other is a broad complaint about many models breaking in v5."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44246",
+            "accept": true,
+            "reason": "Both describe Transformers import taking too long; 41669 points to import-* overhead as the likely root cause of the general slowdown reported in 44246."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43575",
+            "accept": false,
+            "reason": "Both mention tensor parallelism, but one is a generic training bug and the other is a specific OOM when loading a large Qwen model with tp."
+          },
+          {
+            "left": "issue:43867",
+            "right": "issue:45357",
+            "accept": false,
+            "reason": "Different failure modes: sorted state_dict loading vs incorrect visual encoder keys in save_pretrained."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43976",
+            "accept": false,
+            "reason": "Unrelated: cache/no_grad update behavior versus Python-version compatibility breakage."
+          },
+          {
+            "left": "issue:43992",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Both are from_pretrained issues, but they affect different models and likely different code paths/cause sets."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44077",
+            "accept": false,
+            "reason": "One is a CLIPOutput attribute regression, the other is a patchtsmixer post_init policy issue."
+          },
+          {
+            "left": "issue:28282",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "ImportError for missing PyTorch is unrelated to a tokenizer-loading regression."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:46285",
+            "accept": false,
+            "reason": "Both involve model structure/attributes, but they are different CLIP API regressions and not the same bug."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44857",
+            "accept": false,
+            "reason": "CLIP attentions assignment and LwDetr float16 AMP crash are separate defects."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:45440",
+            "accept": false,
+            "reason": "Completely different subsystems: DSA indexer/ReLU versus DeepSeek MoE implementation divergence."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44964",
+            "accept": false,
+            "reason": "A flaky integration test is not the same as a model-loading failure."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44062",
+            "accept": false,
+            "reason": "Local-files-only networking behavior and AddedToken argument handling are unrelated."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44112",
+            "accept": false,
+            "reason": "Different CI/test failures with different causes and targets."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:45600",
+            "accept": false,
+            "reason": "Test seeding flakiness is unrelated to auto_mappings referencing removed configs."
+          },
+          {
+            "left": "issue:43723",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Both are load regressions after v5, but one is tokenizer loading and the other is checkpoint/model reload compatibility."
+          },
+          {
+            "left": "issue:44387",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Int4 CUDA memory regression is a different bug from reload failures after upgrading versions."
+          },
+          {
+            "left": "issue:43874",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "A missing image-processor method and a trust_remote_code cache FileNotFoundError are unrelated."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "FP8 save_pretrained behavior and missing tensor-parallel plan are different issues."
+          },
+          {
+            "left": "issue:44734",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "KV-cache indexing crash in serve is unrelated to cache lookup FileNotFoundError."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43065",
+            "accept": false,
+            "reason": "Different model internals: no_grad decoration for cache updates versus a dummy Conv2d in Sam3PixelDecoder."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:46132",
+            "accept": false,
+            "reason": "Broken env command and changed AttentionInterface.register behavior are separate API issues."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Local-files-only network leakage is unrelated to GLM-5 RoPE implementation."
+          },
+          {
+            "left": "issue:43824",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "Missing Qwen2_5_VL export is not the same as a cache FileNotFoundError when loading trust_remote_code."
+          },
+          {
+            "left": "issue:45125",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Missing _tp_plan for tensor parallelism is different from a Qwen3.5 from_pretrained regression."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44749",
+            "accept": false,
+            "reason": "Compiled-Python-file errors and slow filtering after upgrade are unrelated."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "MLflow reporting on training failure has nothing to do with tokenizer loading regression."
+          },
+          {
+            "left": "issue:42548",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Both are import errors, but they are for different exported symbols and likely different root causes."
+          },
+          {
+            "left": "issue:45125",
+            "right": "issue:45375",
+            "accept": false,
+            "reason": "Missing tensor-parallel plan and missing vision-config field are not the same bug."
+          },
+          {
+            "left": "issue:43010",
+            "right": "issue:43232",
+            "accept": false,
+            "reason": "Different generation/cache path issues with different symptoms and fixes."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44222",
+            "accept": false,
+            "reason": "Network suppression and FP8 save_pretrained are unrelated."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "'racoon' gives no evidence of the same local_files_only network bug."
+          },
+          {
+            "left": "issue:30990",
+            "right": "issue:31795",
+            "accept": false,
+            "reason": "Sentence Transformers hanging on load is unrelated to documentation confusion in model.forward."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet and stays conservative about the soft edges. The choice of issue:45850 as the umbrella issue is reasonable given its broad v5-regression wording, and the only accepted pair (issue:41669 vs issue:44246) is well-supported by the import-slowdown overlap. The other verdicts appropriately avoid overclaiming duplicates."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 53,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:27957",
+        "issue:30333",
+        "issue:31795",
+        "issue:33453",
+        "issue:34689",
+        "issue:36296",
+        "issue:38617",
+        "issue:41628",
+        "issue:42548",
+        "issue:42757",
+        "issue:43299",
+        "issue:43475",
+        "issue:43526",
+        "issue:43576",
+        "issue:43673",
+        "issue:43824",
+        "issue:43827",
+        "issue:43901",
+        "issue:44038",
+        "issue:44222",
+        "issue:44230",
+        "issue:44273",
+        "issue:44315",
+        "issue:44351",
+        "issue:44355",
+        "issue:44360",
+        "issue:44393",
+        "issue:44485",
+        "issue:44492",
+        "issue:44556",
+        "issue:44623",
+        "issue:44655",
+        "issue:44908",
+        "issue:44938",
+        "issue:44961",
+        "issue:45003",
+        "issue:45071",
+        "issue:45125",
+        "issue:45230",
+        "issue:45362",
+        "issue:45468",
+        "issue:45542",
+        "issue:45753",
+        "issue:45850",
+        "issue:46143",
+        "issue:46144",
+        "issue:46209",
+        "issue:46617"
+      ],
+      "soft_pairs": [
+        "issue:43299|issue:44556",
+        "issue:43827|issue:44961",
+        "issue:30333|issue:36296",
+        "issue:44355|issue:45071",
+        "issue:44355|issue:44655",
+        "issue:33453|issue:34689",
+        "issue:27957|issue:33453",
+        "issue:42757|issue:43576",
+        "issue:43576|issue:45071",
+        "issue:44315|issue:44360",
+        "issue:43576|issue:43824",
+        "issue:31795|issue:38617",
+        "issue:44961|issue:45468",
+        "issue:44351|issue:46617",
+        "issue:44485|issue:44961",
+        "issue:43576|issue:44938",
+        "issue:44908|issue:44961",
+        "issue:44222|issue:44230",
+        "issue:46143|issue:46144",
+        "issue:36296|issue:43901",
+        "issue:41628|issue:42548",
+        "issue:36296|issue:43526",
+        "issue:43475|issue:46617",
+        "issue:44315|issue:46209",
+        "issue:45362|issue:45850",
+        "issue:43673|issue:44492",
+        "issue:44623|issue:45230",
+        "issue:44273|issue:45542",
+        "issue:44038|issue:45125",
+        "issue:44393|issue:45071",
+        "issue:44273|issue:45003",
+        "issue:44393|issue:45753",
+        "issue:43576|issue:44556"
+      ],
+      "prepared_review_unit_hash": "fa16d6bba93336a8a25b514a0d6e4a90970af1aaba4e17e6c59c9ce0ce4e2603",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 33,
+        "serialized_chars": 36576,
+        "estimated_input_tokens": 9144,
+        "estimated_eval_tokens": 18544
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is mostly a grab bag of unrelated Transformers issues, with a loose subtheme around v5 regressions, model loading/import breakages, and docs mismatches. None of the soft pairs look like the same concrete bug or change, so I would not merge any of them as duplicates.",
+        "confidence": 0.77,
+        "canonical_issue_reason": "issue:45850 is the broadest umbrella for the repeated v5-regression complaints in this cluster, though it is very vague and not an ideal tracker.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44556 is the most concrete representative of the v5 model-loading breakage theme and is easier to act on than the broader complaint-style issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43299",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "Both mention v5 loading breakage, but they concern different models and failure modes; not the same underlying bug."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "The first is a docs mismatch about pipeline removal; the second is unrelated and too vague to match."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "MLflow reporting on failed training jobs and tensor-parallel training bugs are unrelated."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Compiled-file import/runtime errors and PreTrainedConfig type-checking regressions are different problems."
+          },
+          {
+            "left": "issue:44355",
+            "right": "issue:44655",
+            "accept": false,
+            "reason": "One is about running compiled Python files; the other is about saving Pipeline objects."
+          },
+          {
+            "left": "issue:33453",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Both are loading regressions, but tokenizer loading and Llama 3.2 vision model loading are different code paths."
+          },
+          {
+            "left": "issue:27957",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Flash Attention support for XLM-Roberta is unrelated to tokenizer-loading regression reports."
+          },
+          {
+            "left": "issue:42757",
+            "right": "issue:43576",
+            "accept": false,
+            "reason": "An import error from huggingface_hub is not the same as the transformers env CLI being broken."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "CLI/env breakage and config type-checking regressions are different issues."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44360",
+            "accept": false,
+            "reason": "Liger kernel application and a DSA indexer ReLU discussion have no overlap."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:43824",
+            "accept": false,
+            "reason": "A broken env command is not the same as a Qwen2.5-VL import error."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:38617",
+            "accept": false,
+            "reason": "Documentation confusion and an import error for layer_type_validation are unrelated."
+          },
+          {
+            "left": "issue:44961",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "The title 'racoon' provides no evidence of the same bug as Gemma-4 audio positional encoding."
+          },
+          {
+            "left": "issue:44351",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "Missing public import and cache FileNotFoundError are different failure modes."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "GLM-5 RoPE implementation discussion is unrelated to the unrelated placeholder issue."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:44961",
+            "accept": false,
+            "reason": "Scheduler kwargs handling has nothing to do with the unrelated placeholder issue."
+          },
+          {
+            "left": "issue:44222",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "Both involve FP8/MoE, but one is save_pretrained behavior and the other is inference support; not the same bug."
+          },
+          {
+            "left": "issue:46143",
+            "right": "issue:46144",
+            "accept": false,
+            "reason": "Both are RoFormer-related, but one is missing kwargs passthrough while the other is attention-interface usage; related but not the same concrete defect."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43901",
+            "accept": false,
+            "reason": "Tensor-parallel training and TextClassificationPipeline docs are unrelated."
+          },
+          {
+            "left": "issue:41628",
+            "right": "issue:42548",
+            "accept": false,
+            "reason": "These are different missing-symbol import errors affecting different APIs."
+          },
+          {
+            "left": "issue:36296",
+            "right": "issue:43526",
+            "accept": false,
+            "reason": "Tensor-parallel training and BeitImageProcessorFast label reduction are unrelated."
+          },
+          {
+            "left": "issue:43475",
+            "right": "issue:46617",
+            "accept": false,
+            "reason": "A SAM 3 Video attribute error and a trust_remote_code cache FileNotFoundError are different problems."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:46209",
+            "accept": false,
+            "reason": "Liger kernel application and FP8 inference on B200 are different issues."
+          },
+          {
+            "left": "issue:45362",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "A specific Qwen3.5 chat crash is not the same as a broad complaint that v5 breaks too many models."
+          },
+          {
+            "left": "issue:43673",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "A missing GenerationMixin cache path and a typo in cache-strategy docs are unrelated."
+          },
+          {
+            "left": "issue:44623",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "A concrete save_pretrained file-missing bug does not match an empty 'bug report' placeholder."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:45542",
+            "accept": false,
+            "reason": "Lazy loading problems and a TensorFlow backend import issue are not the same bug."
+          },
+          {
+            "left": "issue:44038",
+            "right": "issue:45125",
+            "accept": false,
+            "reason": "Both mention Qwen3/MoE, but one is a broad v5 bug report and the other is specifically missing _tp_plan for tensor parallelism."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Qwen3-VL 2D bbox output errors are unrelated to PreTrainedConfig type-checking regressions."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:45003",
+            "accept": false,
+            "reason": "Lazy loading failures and unsafe sys.modules access may be adjacent, but they are not clearly the same concrete bug."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45753",
+            "accept": false,
+            "reason": "A 2D bounding-box output issue and an infinite-loop-on-specific-image issue are different failure modes."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44556",
+            "accept": false,
+            "reason": "The env CLI regression and checkpoint reload regression are both v5-related, but they are different code paths and symptoms."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 54,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:31795",
+        "issue:33453",
+        "issue:34689",
+        "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:43976",
+        "issue:44077",
+        "issue:44230",
+        "issue:44246",
+        "issue:44273",
+        "issue:44295",
+        "issue:44297",
+        "issue:44315",
+        "issue:44351",
+        "issue:44360",
+        "issue:44393",
+        "issue:44485",
+        "issue:44492",
+        "issue:44556",
+        "issue:44655",
+        "issue:44704",
+        "issue:44829",
+        "issue:44861",
+        "issue:44908",
+        "issue:45071",
+        "issue:45092",
+        "issue:45230",
+        "issue:45278",
+        "issue:45310",
+        "issue:45341",
+        "issue:45468",
+        "issue:45478",
+        "issue:45850",
+        "issue:46018",
+        "issue:46095",
+        "issue:46172",
+        "issue:46218",
+        "issue:46328"
+      ],
+      "soft_pairs": [
+        "issue:43525|issue:45850",
+        "issue:41669|issue:43673",
+        "issue:44556|issue:45092",
+        "issue:44230|issue:46095",
+        "issue:43572|issue:45850",
+        "issue:45278|issue:45850",
+        "issue:44492|issue:45850",
+        "issue:42371|issue:44492",
+        "issue:41669|issue:46218",
+        "issue:42890|issue:44908",
+        "issue:43408|issue:44315",
+        "issue:43352|issue:44393",
+        "issue:43576|issue:44908",
+        "issue:44246|issue:44273",
+        "issue:43976|issue:44393",
+        "issue:44704|issue:44908",
+        "issue:44295|issue:45230",
+        "issue:44908|issue:45071",
+        "issue:44297|issue:45230",
+        "issue:41669|issue:44351",
+        "issue:44492|issue:46328",
+        "issue:41669|issue:45341",
+        "issue:43976|issue:45850",
+        "issue:44393|issue:45478",
+        "issue:43976|issue:46218",
+        "issue:44077|issue:44861",
+        "issue:43827|issue:44908",
+        "issue:31795|issue:33453",
+        "issue:44315|issue:44829",
+        "issue:43576|issue:44273",
+        "issue:44360|issue:45468",
+        "issue:44655|issue:44861",
+        "issue:45230|issue:45310",
+        "issue:44273|issue:44908",
+        "issue:43576|issue:46218",
+        "issue:43519|issue:44485",
+        "issue:43519|issue:44908",
+        "issue:41669|issue:43704",
+        "issue:46018|issue:46172",
+        "issue:31795|issue:34689"
+      ],
+      "prepared_review_unit_hash": "776a3a48c954b27e567bc41caae327cb8c7613961936dce5b002d0c16b1c948a",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 48,
+        "item_count": 48,
+        "soft_pair_count": 40,
+        "serialized_chars": 39071,
+        "estimated_input_tokens": 9768,
+        "estimated_eval_tokens": 19792
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "Mostly a loose cluster of v5-era regressions and import/model-loading complaints. Only one soft edge looks like the same underlying issue: lazy loading/import-time slowdown. The rest are too broad or target different concrete bugs.",
+        "confidence": 0.66,
+        "canonical_issue_reason": "issue:44273 is the best concrete anchor for the import/lazy-loading regression theme and is the clearest representative of the only plausible duplicate pair.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:44273 has the most specific bug description among the candidates while still covering the shared lazy-loading/import symptom; broader v5 regression reports are too vague to serve as a good canonical issue.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43525",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "Both are v5-related, but one is a specific missing config attribute and the other is a broad complaint about many models breaking."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:43673",
+            "accept": false,
+            "reason": "One is import-star/lazy-loading performance, the other is a chunked-prefill cache regression; different code paths."
+          },
+          {
+            "left": "issue:44556",
+            "right": "issue:45092",
+            "accept": false,
+            "reason": "Both concern checkpoint loading after v5 changes, but they describe different model families and different failure modes."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:46095",
+            "accept": false,
+            "reason": "FP8 inference support for Qwen3-VL/Qwen3.5 is unrelated to a DeepSeek-V4 weight conversion question."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "A missing StableLm config field is a concrete config bug, not the same as a generic v5 breakage report."
+          },
+          {
+            "left": "issue:45278",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "General import errors after upgrade are too broad and not the same specific bug as the generic 'too many models' complaint."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "A typo in cache strategies is unrelated to a broad v5 regression report."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:44492",
+            "accept": false,
+            "reason": "TF32 API guidance and a cache-strategy typo are different issues entirely."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "One is import-time bloat from import-star usage, the other is a missing module import error for a specific model."
+          },
+          {
+            "left": "issue:42890",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "A flaky integration test due to missing set_seed is unrelated to a scheduler kwargs bug."
+          },
+          {
+            "left": "issue:43408",
+            "right": "issue:44315",
+            "accept": false,
+            "reason": "Model-type mismatch warning and Liger kernel not applied are different model-init problems."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:44393",
+            "accept": false,
+            "reason": "Flash-Attn-2 support gating for Nemotron is unrelated to Qwen3-VL bounding-box errors."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "A broken env command and an inverse_sqrt scheduler kwargs bug are different subsystems and symptoms."
+          },
+          {
+            "left": "issue:44246",
+            "right": "issue:44273",
+            "accept": true,
+            "reason": "Both point to the same import/lazy-loading regression: import transformers is slow because lazy loading is not behaving properly."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:44393",
+            "accept": false,
+            "reason": "Python version compatibility and a Qwen3-VL output bug are unrelated."
+          },
+          {
+            "left": "issue:44704",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "AutoProcessor kwargs forwarding and scheduler kwargs handling are different APIs and failures."
+          },
+          {
+            "left": "issue:44295",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "A position_ids buffer access bug is not the same as an empty placeholder bug report."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:45071",
+            "accept": false,
+            "reason": "Scheduler kwargs omission and PretrainedConfig type-checking breakage are separate regressions."
+          },
+          {
+            "left": "issue:44297",
+            "right": "issue:45230",
+            "accept": false,
+            "reason": "Tokenizer save metadata mismatch is not the same as a generic bug report."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:44351",
+            "accept": false,
+            "reason": "Import-star performance overhead and 'cannot import HybridCache' are both import-related but not the same concrete failure."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:46328",
+            "accept": false,
+            "reason": "A cache-strategy typo and a SequenceFeatureExtractor.pad performance issue are unrelated."
+          },
+          {
+            "left": "issue:41669",
+            "right": "issue:45341",
+            "accept": false,
+            "reason": "Import performance regression and a testing_utils bug do not match."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "Python version support and generic model breakage are not the same issue."
+          },
+          {
+            "left": "issue:44393",
+            "right": "issue:45478",
+            "accept": false,
+            "reason": "Two Qwen3.5/Qwen3-VL reports, but one is bbox hallucination and the other is a from_pretrained error; different bugs."
+          },
+          {
+            "left": "issue:43976",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Python compatibility and a missing model import are unrelated."
+          },
+          {
+            "left": "issue:44077",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Optional post_init on patchtsmixer and tied-weight-key AttributeError are different config/model-loading bugs."
+          },
+          {
+            "left": "issue:43827",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Docs still referencing pipeline() and scheduler kwargs handling are unrelated."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Documentation confusion in model.forward and tokenizer-loading regression are different topics."
+          },
+          {
+            "left": "issue:44315",
+            "right": "issue:44829",
+            "accept": false,
+            "reason": "Applying Liger Kernel in model_init and flash_attention_3 training degradation are distinct issues."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:44273",
+            "accept": false,
+            "reason": "Broken env command and lazy-loading/import slowdown are both v5-related but not the same bug."
+          },
+          {
+            "left": "issue:44360",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "A DSA indexer activation bug and Gemma-4 audio positional encoding are unrelated."
+          },
+          {
+            "left": "issue:44655",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Saving Pipeline objects and tied-weight-key attribute handling are different code paths."
+          },
+          {
+            "left": "issue:45230",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "Generic bug report versus a specific Qwen3.5 Moe from_pretrained error."
+          },
+          {
+            "left": "issue:44273",
+            "right": "issue:44908",
+            "accept": false,
+            "reason": "Lazy loading/import slowdown and scheduler kwargs omission are unrelated."
+          },
+          {
+            "left": "issue:43576",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Broken env command and missing Qwen3_5ForCausalLM import are different failures."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Qwen3VL timestamp calculation and GLM-5 RoPE implementation are different model-specific math bugs."
+          },
+          {
+            "left": "issue:46018",
+            "right": "issue:46172",
+            "accept": false,
+            "reason": "DeepSeek-V4 gating logic and a DeepSeek-V4 tensor transpose issue are not the same bug."
+          },
+          {
+            "left": "issue:31795",
+            "right": "issue:34689",
+            "accept": false,
+            "reason": "Documentation confusion in forward args and a Llama 3.2 model-loading breakage are unrelated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary stays grounded in the packet and is conservative about the soft edges. The one accepted pair (44246/44273) is a reasonable lazy-loading/import regression match, and the other verdicts avoid overstating similarity."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 1,
+      "component_count": 21,
+      "review_unit_index": 55,
+      "review_unit_count": 55,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10000-467",
+      "nodes": [
+        "issue:30333",
+        "issue:31515",
+        "issue:31795",
+        "issue:33453",
+        "issue:36296",
+        "issue:37934",
+        "issue:38617",
+        "issue:42371",
+        "issue:42913",
+        "issue:42936",
+        "issue:43066",
+        "issue:43240",
+        "issue:43298",
+        "issue:43352",
+        "issue:43502",
+        "issue:43519",
+        "issue:43525",
+        "issue:43572",
+        "issue:43606",
+        "issue:43618",
+        "issue:43668",
+        "issue:43704",
+        "issue:43717",
+        "issue:43813",
+        "issue:44031",
+        "issue:44230",
+        "issue:44261",
+        "issue:44263",
+        "issue:44351",
+        "issue:44485",
+        "issue:44492",
+        "issue:44623",
+        "issue:44861",
+        "issue:44908",
+        "issue:44938",
+        "issue:44998",
+        "issue:45290",
+        "issue:45310",
+        "issue:45468",
+        "issue:45850",
+        "issue:46018",
+        "issue:46095",
+        "issue:46144",
+        "issue:46218"
+      ],
+      "soft_pairs": [
+        "issue:31515|issue:31795",
+        "issue:43704|issue:44485",
+        "issue:44485|issue:45468",
+        "issue:42371|issue:43704",
+        "issue:45850|issue:46218",
+        "issue:43502|issue:43519",
+        "issue:43519|issue:44623",
+        "issue:43519|issue:43704",
+        "issue:44263|issue:45468",
+        "issue:43298|issue:46095",
+        "issue:43606|issue:45468",
+        "issue:44261|issue:45468",
+        "issue:44351|issue:46218",
+        "issue:43352|issue:46218",
+        "issue:42913|issue:43066",
+        "issue:44908|issue:44998",
+        "issue:44492|issue:44998",
+        "issue:43618|issue:44998",
+        "issue:44998|issue:45850",
+        "issue:44938|issue:44998",
+        "issue:44861|issue:45290",
+        "issue:43525|issue:44861",
+        "issue:31515|issue:38617",
+        "issue:43572|issue:44861",
+        "issue:44230|issue:44623",
+        "issue:30333|issue:31515",
+        "issue:31515|issue:36296",
+        "issue:44230|issue:45310",
+        "issue:31515|issue:33453",
+        "issue:43298|issue:44230",
+        "issue:43717|issue:46144",
+        "issue:42936|issue:46018",
+        "issue:43668|issue:44031",
+        "issue:42936|issue:43240",
+        "issue:43240|issue:46144",
+        "issue:42936|issue:44031",
+        "issue:42936|issue:46144",
+        "issue:42936|issue:43813",
+        "issue:43240|issue:43668",
+        "issue:43668|issue:43813",
+        "issue:37934|issue:43668",
+        "issue:42936|issue:43668"
+      ],
+      "prepared_review_unit_hash": "953c3ff77207b23ae4798a487e22fdd0587f3d6f22911f6c6de125e53b57aba9",
+      "component_budget": {
+        "node_count": 467,
+        "item_count": 467,
+        "soft_pair_count": 1714,
+        "serialized_chars": 889302,
+        "estimated_input_tokens": 222326,
+        "estimated_eval_tokens": 444908
+      },
+      "budget": {
+        "node_count": 44,
+        "item_count": 44,
+        "soft_pair_count": 42,
+        "serialized_chars": 37129,
+        "estimated_input_tokens": 9283,
+        "estimated_eval_tokens": 18822
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "This cluster is a grab-bag of unrelated issues across tokenizers, configs, model-specific bugs, and documentation. I do not see any soft pair that appears to be the same underlying bug/change, so all soft edges should be rejected.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "issue:42371 is the most active and broadly framed substantive report in the set, so it is the best anchor for this mixed cluster.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:42371 is the clearest standalone issue here with strong discussion activity; it is representative only as an anchor, not because it duplicates the others.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:31515",
+            "right": "issue:31795",
+            "accept": false,
+            "reason": "Different problems: slow checkpoint loading vs confusing forward-argument docs."
+          },
+          {
+            "left": "issue:43704",
+            "right": "issue:44485",
+            "accept": false,
+            "reason": "Different code paths: VRAM leak in Qwen3ForCausalLM vs GLM-5 RoPE implementation."
+          },
+          {
+            "left": "issue:44485",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "Unrelated models and bugs: GLM-5 RoPE vs Gemma-4 audio positional encoding."
+          },
+          {
+            "left": "issue:42371",
+            "right": "issue:43704",
+            "accept": false,
+            "reason": "TF32 API guidance is unrelated to a Qwen3 VRAM leak."
+          },
+          {
+            "left": "issue:45850",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Generic version-breakage complaint vs import failure for Qwen3_5ForCausalLM."
+          },
+          {
+            "left": "issue:43502",
+            "right": "issue:43519",
+            "accept": false,
+            "reason": "Different issues: unwanted network access vs Qwen3VL timestamp calculation."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "Different processor bugs: timestamp math vs save_pretrained missing files."
+          },
+          {
+            "left": "issue:43519",
+            "right": "issue:43704",
+            "accept": false,
+            "reason": "Qwen3VL timestamp bug is unrelated to multi-thread VRAM leakage."
+          },
+          {
+            "left": "issue:44263",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "Different models and failures: torch.split indexing vs Gemma-4 audio positional encoding."
+          },
+          {
+            "left": "issue:43298",
+            "right": "issue:46095",
+            "accept": false,
+            "reason": "Continuous batching audio support is unrelated to a DeepSeek-V4 weight-conversion question."
+          },
+          {
+            "left": "issue:43606",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "CPU offload device mismatch is unrelated to Gemma-4 audio positional encoding."
+          },
+          {
+            "left": "issue:44261",
+            "right": "issue:45468",
+            "accept": false,
+            "reason": "Different architecture bugs: MLA layernorm epsilon config vs Gemma-4 audio positional encoding."
+          },
+          {
+            "left": "issue:44351",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Different import failures for different symbols/models; not the same missing-module bug."
+          },
+          {
+            "left": "issue:43352",
+            "right": "issue:46218",
+            "accept": false,
+            "reason": "Flash-Attention support gap is unrelated to Qwen3_5 import failure."
+          },
+          {
+            "left": "issue:42913",
+            "right": "issue:43066",
+            "accept": false,
+            "reason": "Both tokenizer-related, but one is a version behavior regression and the other is a decoder type mismatch."
+          },
+          {
+            "left": "issue:44908",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Scheduler kwargs bug is unrelated to the generic 'Unemployment' spam issue."
+          },
+          {
+            "left": "issue:44492",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Cache-strategy typo is unrelated to the spam/non-bug issue."
+          },
+          {
+            "left": "issue:43618",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "CLIPOutput attentions regression is unrelated to the spam/non-bug issue."
+          },
+          {
+            "left": "issue:44998",
+            "right": "issue:45850",
+            "accept": false,
+            "reason": "Spam/non-bug issue vs broad version-regression complaint; not the same artifact."
+          },
+          {
+            "left": "issue:44938",
+            "right": "issue:44998",
+            "accept": false,
+            "reason": "Python 3.14 load failure is unrelated to the spam/non-bug issue."
+          },
+          {
+            "left": "issue:44861",
+            "right": "issue:45290",
+            "accept": false,
+            "reason": "Tied-weight key handling is unrelated to apply_chat_template crashing on tool-call assistant messages."
+          },
+          {
+            "left": "issue:43525",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "Missing pad_token_id in Llama4Config is unrelated to tied-weight key handling."
+          },
+          {
+            "left": "issue:43572",
+            "right": "issue:44861",
+            "accept": false,
+            "reason": "StableLmConfig pad_token_idx regression is unrelated to tied-weight key handling."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:44623",
+            "accept": false,
+            "reason": "FP8 support request is unrelated to processor.save_pretrained missing files."
+          },
+          {
+            "left": "issue:30333",
+            "right": "issue:31515",
+            "accept": false,
+            "reason": "MLFlow reporting on failed training jobs is unrelated to slow from_pretrained checkpoint loading."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:36296",
+            "accept": false,
+            "reason": "Checkpoint loading slowness is unrelated to a tensor-parallel training bug."
+          },
+          {
+            "left": "issue:44230",
+            "right": "issue:45310",
+            "accept": false,
+            "reason": "FP8 inference support request is unrelated to Qwen3.5 Moe from_pretrained errors."
+          },
+          {
+            "left": "issue:31515",
+            "right": "issue:33453",
+            "accept": false,
+            "reason": "Slow checkpoint loading is unrelated to tokenizer-loading regression."
+          },
+          {
+            "left": "issue:43298",
+            "right": "issue:44230",
+            "accept": false,
+            "reason": "Continuous batching audio limitation is unrelated to FP8 support for Qwen3-VL/Qwen3.5 moe."
+          },
+          {
+            "left": "issue:43717",
+            "right": "issue:46144",
+            "accept": false,
+            "reason": "Mamba init_weights behavior is unrelated to RoFormer attention-interface usage."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:46018",
+            "accept": false,
+            "reason": "Mask2former ignore_value bug is unrelated to DeepSeek-V4 shared-expert gating."
+          },
+          {
+            "left": "issue:43668",
+            "right": "issue:44031",
+            "accept": false,
+            "reason": "Incorrect type hint in ModernBERTConfig is unrelated to tokenizer regex warnings."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43240",
+            "accept": false,
+            "reason": "Mask2former ignore_value handling is unrelated to kwargs not being passed to loss calculation."
+          },
+          {
+            "left": "issue:43240",
+            "right": "issue:46144",
+            "accept": false,
+            "reason": "Loss-function kwargs propagation is unrelated to RoFormer attention-interface usage."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:44031",
+            "accept": false,
+            "reason": "Mask2former ignore_value bug is unrelated to tokenizer regex warnings."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:46144",
+            "accept": false,
+            "reason": "Different models and failures: Mask2former ignore_value vs RoFormer attention interface."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43813",
+            "accept": false,
+            "reason": "Mask2former ignore_value bug is unrelated to a typo in quantization-operation naming."
+          },
+          {
+            "left": "issue:43240",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Loss kwargs propagation is unrelated to a ModernBERT type-hint issue."
+          },
+          {
+            "left": "issue:43668",
+            "right": "issue:43813",
+            "accept": false,
+            "reason": "ModernBERT type-hint bug is unrelated to the quantization-operation typo."
+          },
+          {
+            "left": "issue:37934",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Llama4TextL2Norm/RMS norm question is unrelated to a ModernBERT type-hint bug."
+          },
+          {
+            "left": "issue:42936",
+            "right": "issue:43668",
+            "accept": false,
+            "reason": "Mask2former ignore_value bug is unrelated to ModernBERT norm_eps typing."
+          }
+        ]
+      },
+      "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": 21,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-10519-3",
+      "nodes": [
+        "issue:10519",
+        "issue:11005",
+        "issue:11307"
+      ],
+      "soft_pairs": [
+        "issue:10519|issue:11307",
+        "issue:10519|issue:11005",
+        "issue:11005|issue:11307"
+      ],
+      "prepared_review_unit_hash": "b541324b38ab4fcf0a92def373c31fb93bc811030c3955944264a08ec68c6a03",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 2876,
+        "estimated_input_tokens": 719,
+        "estimated_eval_tokens": 1694
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 2876,
+        "estimated_input_tokens": 719,
+        "estimated_eval_tokens": 1694
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are three unrelated issues: token truncation behavior, ReduceLROnPlateau-like training logic, and Wav2Vec2 word boundary offsets. No duplicate cluster is supported by the titles or evidence.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "issue:11307 is the most discussion-heavy and concrete user-facing problem in the set, so it is the best representative issue if one must be chosen.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:11307 has the strongest signal of active demand and a clearly scoped feature request, making it the best standalone issue in this group.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:10519",
+            "right": "issue:11307",
+            "accept": false,
+            "reason": "Different problem domains: tokenizer truncation direction vs Wav2Vec2 time offsets. No evidence they describe the same underlying bug or change."
+          },
+          {
+            "left": "issue:10519",
+            "right": "issue:11005",
+            "accept": false,
+            "reason": "Unrelated features: truncation strategy vs learning-rate scheduler behavior. They do not share a concrete code path or fix."
+          },
+          {
+            "left": "issue:11005",
+            "right": "issue:11307",
+            "accept": false,
+            "reason": "One is about optimizer scheduling, the other about speech model word alignment offsets. These are clearly distinct requests."
+          }
+        ]
+      },
+      "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": 21,
+      "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": "55e00ac334874ab10e1f73fac65208fffd8b2916f06ecaf466a0ad76e2895bd2",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 2478,
+        "estimated_input_tokens": 620,
+        "estimated_eval_tokens": 1496
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 2478,
+        "estimated_input_tokens": 620,
+        "estimated_eval_tokens": 1496
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The cluster is only loosely related by Hugging Face/TensorFlow model-loading behavior, but the three issues describe different problems: tokenizer mismatch, private-model auth/error messaging, and a generic load failure for bert-base-uncased. They should not be merged as duplicates.",
+        "confidence": 0.83,
+        "canonical_issue_reason": "issue:16618 is the most concrete and broadly representative user-facing failure in the cluster, but it is still a distinct bug from the other two issues.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:16618 has the clearest, most actionable bug report centered on a specific model-loading failure, making it the best standalone issue in the set.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:13244",
+            "right": "issue:16618",
+            "accept": false,
+            "reason": "Different underlying problems: tokenizer behavior differs from TensorFlow code, while the other reports a model-loading failure. Same ecosystem, not the same bug."
+          },
+          {
+            "left": "issue:15980",
+            "right": "issue:16618",
+            "accept": false,
+            "reason": "One is about a misleading auth/download error for private models; the other is a failure loading bert-base-uncased. Related area, but 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": 4,
+      "component_count": 21,
+      "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": "e6e6a9b42589bf568c2d2a15cbb13553db00fb1d0d63b9d1a83b1c4b7847cbc7",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1590,
+        "estimated_input_tokens": 398,
+        "estimated_eval_tokens": 1052
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1590,
+        "estimated_input_tokens": 398,
+        "estimated_eval_tokens": 1052
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are thematically similar community-contribution issues, but they request different features for different code paths: Flash Attention 2 support vs GGUF support. I would not treat them as duplicates or merge them into one canonical issue.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:26350",
+            "right": "issue:33260",
+            "accept": false,
+            "reason": "Not the same underlying request: one is about extending Flash Attention 2 support to more architectures, the other is about adding GGUF support to more architectures. Similar wording and subsystem area, but different concrete changes."
+          }
+        ]
+      },
+      "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": 21,
+      "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-27088-8",
+      "nodes": [
+        "issue:27088",
+        "issue:28936",
+        "issue:38435",
+        "issue:43834",
+        "issue:44365",
+        "issue:44370",
+        "issue:46222",
+        "issue:46226"
+      ],
+      "soft_pairs": [
+        "issue:27088|issue:44370",
+        "issue:27088|issue:44365",
+        "issue:27088|issue:43834",
+        "issue:27088|issue:38435",
+        "issue:27088|issue:28936",
+        "issue:28936|issue:43834",
+        "issue:28936|issue:44365",
+        "issue:28936|issue:38435",
+        "issue:28936|issue:44370",
+        "issue:38435|issue:44370",
+        "issue:38435|issue:44365",
+        "issue:38435|issue:43834",
+        "issue:38435|issue:46222",
+        "issue:38435|issue:46226",
+        "issue:43834|issue:44370",
+        "issue:43834|issue:44365",
+        "issue:43834|issue:46222",
+        "issue:43834|issue:46226",
+        "issue:44365|issue:44370",
+        "issue:44365|issue:46222",
+        "issue:44365|issue:46226",
+        "issue:44370|issue:46222",
+        "issue:44370|issue:46226",
+        "issue:46222|issue:46226"
+      ],
+      "prepared_review_unit_hash": "71be268594f0e555ad6c8c89af31187feceea8163a4e60ea85066e5a288da172",
+      "component_budget": {
+        "node_count": 8,
+        "item_count": 8,
+        "soft_pair_count": 24,
+        "serialized_chars": 13523,
+        "estimated_input_tokens": 3381,
+        "estimated_eval_tokens": 7018
+      },
+      "budget": {
+        "node_count": 8,
+        "item_count": 8,
+        "soft_pair_count": 24,
+        "serialized_chars": 13523,
+        "estimated_input_tokens": 3381,
+        "estimated_eval_tokens": 7018
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster is mostly heterogeneous i18n translation-request issues. The concrete language issues (Turkish, Spanish, Romanian) are not duplicates of each other. The only plausible duplicate subgroups are the three identical generic i18n template issues and the identical Thai-title pair.",
+        "confidence": 0.78,
+        "canonical_issue_reason": "issue:43834 is the best representative of the larger identical-title i18n template subgroup; it is the earliest of that set and matches the repeated placeholder pattern most directly.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43834 is the strongest issue representative in the cluster because it anchors the duplicate template subgroup and is clearer than the language-specific requests for summarization.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:27088",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Different concrete request (Turkish) vs generic placeholder template; same subsystem but not the same issue."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Turkish translation request is distinct from the generic placeholder issue."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:43834",
+            "accept": false,
+            "reason": "Turkish request and placeholder template are not the same underlying issue."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:38435",
+            "accept": false,
+            "reason": "Different target languages (Turkish vs Romanian), so not duplicates."
+          },
+          {
+            "left": "issue:27088",
+            "right": "issue:28936",
+            "accept": false,
+            "reason": "Different target languages (Turkish vs Spanish), so not duplicates."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:43834",
+            "accept": false,
+            "reason": "Spanish translation request is separate from the generic template issue."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Spanish request and placeholder template are not the same underlying issue."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:38435",
+            "accept": false,
+            "reason": "Different target languages (Spanish vs Romanian), so not duplicates."
+          },
+          {
+            "left": "issue:28936",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Spanish request is distinct from the generic placeholder issue."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:44370",
+            "accept": false,
+            "reason": "Romanian translation request is separate from the generic placeholder issue."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:44365",
+            "accept": false,
+            "reason": "Romanian request and placeholder template are not the same underlying issue."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:43834",
+            "accept": false,
+            "reason": "Romanian request is not the same as the generic template issue."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:46222",
+            "accept": false,
+            "reason": "Romanian translation request is unrelated to the Thai-text issue."
+          },
+          {
+            "left": "issue:38435",
+            "right": "issue:46226",
+            "accept": false,
+            "reason": "Romanian translation request is unrelated to the Thai-text issue."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:44370",
+            "accept": true,
+            "reason": "Same exact placeholder title and near-identical metadata/body size; looks like the same templated issue duplicated."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:44365",
+            "accept": true,
+            "reason": "Same exact placeholder title and near-identical metadata/body size; appears to be the same issue copied/reopened."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:46222",
+            "accept": false,
+            "reason": "Different titles and language content; not the same underlying issue."
+          },
+          {
+            "left": "issue:43834",
+            "right": "issue:46226",
+            "accept": false,
+            "reason": "Different titles and language content; not the same underlying issue."
+          },
+          {
+            "left": "issue:44365",
+            "right": "issue:44370",
+            "accept": true,
+            "reason": "Same exact placeholder title with very similar body size and timing; likely duplicates of one templated issue."
+          },
+          {
+            "left": "issue:44365",
+            "right": "issue:46222",
+            "accept": false,
+            "reason": "Placeholder i18n template and Thai-text issue are unrelated."
+          },
+          {
+            "left": "issue:44365",
+            "right": "issue:46226",
+            "accept": false,
+            "reason": "Placeholder i18n template and Thai-text issue are unrelated."
+          },
+          {
+            "left": "issue:44370",
+            "right": "issue:46222",
+            "accept": false,
+            "reason": "Placeholder i18n template and Thai-text issue are unrelated."
+          },
+          {
+            "left": "issue:44370",
+            "right": "issue:46226",
+            "accept": false,
+            "reason": "Placeholder i18n template and Thai-text issue are unrelated."
+          },
+          {
+            "left": "issue:46222",
+            "right": "issue:46226",
+            "accept": true,
+            "reason": "Identical Thai title and same body size strongly suggest the same issue duplicated."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The cluster summary is grounded in the packet and stays appropriately conservative. The negative duplicate calls between distinct language requests are supported by the titles, while the two duplicate subgroups are limited to the near-identical template issues and the exact Thai-title pair. Representative selection for issue:43834 is also defensible given the earliest timestamp within that subgroup."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 6,
+      "component_count": 21,
+      "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-36817-16",
+      "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",
+        "issue:46133"
+      ],
+      "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",
+        "issue:45322|issue:46133"
+      ],
+      "prepared_review_unit_hash": "e1a4d3bf5e8a2e46c14df8bf5da736ef1e14f7033bb95c94683adda9130d1271",
+      "component_budget": {
+        "node_count": 16,
+        "item_count": 16,
+        "soft_pair_count": 36,
+        "serialized_chars": 21658,
+        "estimated_input_tokens": 5415,
+        "estimated_eval_tokens": 11086
+      },
+      "budget": {
+        "node_count": 16,
+        "item_count": 16,
+        "soft_pair_count": 36,
+        "serialized_chars": 21658,
+        "estimated_input_tokens": 5415,
+        "estimated_eval_tokens": 11086
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These issues are mostly broad model-addition requests, but they do not describe the same underlying bug or change. The similarity is driven by repeated \u201cadd support for model X\u201d wording, not duplicate concrete implementation work.",
+        "confidence": 0.93,
+        "canonical_issue_reason": "No true duplicate exists in this cluster. If one issue must serve as a representative, issue 43671 is the best fit because it has the highest discussion activity and sits near the center of the model-support requests, but it is still a distinct Qwen3-TTS request.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43671 is the strongest cluster representative, but only in a loose thematic sense; it is not a canonical duplicate target because the set spans unrelated models and modalities.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:36817",
+            "right": "issue:41211",
+            "accept": false,
+            "reason": "EuroBERT and DEIMv2 are different model families with separate additions."
+          },
+          {
+            "left": "issue:36817",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "EuroBERT config support is unrelated to Qwen3-TTS support."
+          },
+          {
+            "left": "issue:36817",
+            "right": "issue:36879",
+            "accept": false,
+            "reason": "EuroBERT and RF-DETR are entirely different architectures and tasks."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:41211",
+            "accept": false,
+            "reason": "RF-DETR and DEIMv2 are distinct model requests, not the same fix."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "RF-DETR vision detection work is unrelated to LFM2.5 Audio support."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "Object-detection model support is not the same as sparse-transformer support."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "RF-DETR and Qwen3-ASR target different models and code paths."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "RF-DETR and Qwen3-TTS are unrelated model-support requests."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "NeoBERT and Qwen3-ASR are different model additions."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "NeoBERT support and weight-sparse transformer support are separate changes."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "NeoBERT and LFM2.5 Audio do not describe the same implementation."
+          },
+          {
+            "left": "issue:36879",
+            "right": "issue:37015",
+            "accept": false,
+            "reason": "RF-DETR and NeoBERT are unrelated architectures."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:42503",
+            "accept": false,
+            "reason": "NeoBERT and ModernVBERT are distinct model requests."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "MiniCPM3 support is separate from Qwen3-ASR support."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "MiniCPM3 and weight-sparse transformer requests are different code paths."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:42503",
+            "accept": false,
+            "reason": "MiniCPM3 and ModernVBERT are different model architectures."
+          },
+          {
+            "left": "issue:41115",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "MiniCPM3 and Voxtral TTS are unrelated model additions."
+          },
+          {
+            "left": "issue:41211",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "DEIMv2 and Qwen3-TTS are different models and implementations."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "ModernVBERT and Qwen3-ASR are separate model-support requests."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:43671",
+            "accept": false,
+            "reason": "ModernVBERT and Qwen3-TTS are not the same change."
+          },
+          {
+            "left": "issue:42503",
+            "right": "issue:42738",
+            "accept": false,
+            "reason": "ModernVBERT addition is not the same as the broader BERT-with-RoPE request."
+          },
+          {
+            "left": "issue:42738",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "A generic BERT-RoPE request is unrelated to Voxtral TTS support."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:43837",
+            "accept": false,
+            "reason": "TTS and ASR are related but distinct model-support changes."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "Qwen3-TTS and LFM2.5 Audio are different audio-model requests."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "TTS support and sparse-transformer support are separate changes."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "Qwen3-TTS and Voxtral TTS are different models, not duplicates."
+          },
+          {
+            "left": "issue:43671",
+            "right": "issue:45177",
+            "accept": false,
+            "reason": "Qwen3-TTS and DeepSeek-OCR2 target different modalities and code paths."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:43909",
+            "accept": false,
+            "reason": "ASR support and LFM2.5 Audio support are distinct requests."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45322",
+            "accept": false,
+            "reason": "Qwen3-ASR and EUPE are unrelated model additions."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:44121",
+            "accept": false,
+            "reason": "ASR support is not the same as adding a sparse transformer."
+          },
+          {
+            "left": "issue:43909",
+            "right": "issue:44205",
+            "accept": false,
+            "reason": "Audio-model support and SAM3-LiteText are unrelated requests."
+          },
+          {
+            "left": "issue:44121",
+            "right": "issue:45322",
+            "accept": false,
+            "reason": "Weight-sparse transformer support and EUPE are different model changes."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "Qwen3-ASR and Voxtral TTS are separate audio-model integrations."
+          },
+          {
+            "left": "issue:37015",
+            "right": "issue:45027",
+            "accept": false,
+            "reason": "NeoBERT and Voxtral TTS are unrelated additions."
+          },
+          {
+            "left": "issue:43837",
+            "right": "issue:45177",
+            "accept": false,
+            "reason": "ASR support and DeepSeek-OCR2 support are different model requests."
+          },
+          {
+            "left": "issue:45322",
+            "right": "issue:46133",
+            "accept": false,
+            "reason": "EUPE and TIPSv2 are distinct vision-language model requests."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 7,
+      "component_count": 21,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-38740-2",
+      "nodes": [
+        "issue:38740",
+        "issue:46443"
+      ],
+      "soft_pairs": [
+        "issue:38740|issue:46443"
+      ],
+      "prepared_review_unit_hash": "f6349d9a843c45d69be5113b6dcffc72c5956080c8203fd422ddfebacc01dfd6",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1505,
+        "estimated_input_tokens": 377,
+        "estimated_eval_tokens": 1010
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1505,
+        "estimated_input_tokens": 377,
+        "estimated_eval_tokens": 1010
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two items are unrelated: one is a docs request to add `pruna` as an optimization framework, and the other is a Vision RoPE refactor. No duplicate relationship is evident.",
+        "confidence": 0.98,
+        "canonical_issue_reason": "Issue 46443 is the stronger canonical issue because it is an open technical work item, whereas 38740 is a closed documentation request.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 46443 is the best representative issue for the cluster because it is the only substantive code/refactor task; 38740 is documentation-only and unrelated.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:38740",
+            "right": "issue:46443",
+            "accept": false,
+            "reason": "Reject: the titles and contexts are completely different\u2014docs for an optimization framework vs. a Vision RoPE refactor. They do not describe 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": 8,
+      "component_count": 21,
+      "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": "adf33fd04a4a8acd07ef0414c8e15d232614cb99bb9cd171ff6d5f550612653b",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1478,
+        "estimated_input_tokens": 370,
+        "estimated_eval_tokens": 996
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1478,
+        "estimated_input_tokens": 370,
+        "estimated_eval_tokens": 996
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two issues are unrelated: one is about a decoding setting, the other about adding DeepSeek v3.2 support. They should not be clustered as duplicates.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:41084",
+            "right": "issue:43037",
+            "accept": false,
+            "reason": "Different underlying requests. \"Set Block Decoding\" sounds like a decoding/configuration behavior change, while \"DeepSeek v3.2 support\" is a model compatibility/support request. No shared code-path or concrete duplicate indication."
+          }
+        ]
+      },
+      "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": 21,
+      "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": "655fd1b956b72e24ecbf18bdadb5f7242c799a2dee8120f088ccc388e76e32ae",
+      "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": "The two issues are related to Trainer internals, but they describe different underlying problems: `remove_unused_column` customizability versus FLOPS estimation for custom trainers. They should not be treated as duplicates.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:43598 is the cleaner representative issue: it is a focused Trainer customization request with a concrete behavior to revisit.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43598",
+            "right": "issue:43600",
+            "accept": false,
+            "reason": "Different code paths and bugs: one concerns input-column pruning customizability, the other concerns FLOPS counting for custom trainers."
+          }
+        ]
+      },
+      "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": 21,
+      "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": "bc7e54ab32a10bb52814540bc17a286152b2e72de3a1b9e6f6dffa1b657cd14d",
+      "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 two closed issues are about different Qwen model-support requests in different auto classes, so they are not duplicates.",
+        "confidence": 0.96,
+        "canonical_issue_reason": "Issue #43684 is the broader representative request: it asks for new model registration in AutoModel and AutoModelForConditionalGeneration.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #43684 is the best issue to represent the cluster because it describes a general registration gap rather than a narrower, class-specific feature request.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43684",
+            "right": "issue:44405",
+            "accept": false,
+            "reason": "Different underlying change requests: Qwen3-Omni registration for AutoModel/AutoModelForConditionalGeneration vs AutoModelForSequenceClassification support for Qwen3.5. They target different model families and different auto classes."
+          }
+        ]
+      },
+      "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": 21,
+      "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": "f195a834418fd77de8504b8b8a6dc7b92b89673e1332bbe5dd9be47c9f182218",
+      "component_budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4267,
+        "estimated_input_tokens": 1067,
+        "estimated_eval_tokens": 2390
+      },
+      "budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 6,
+        "serialized_chars": 4267,
+        "estimated_input_tokens": 1067,
+        "estimated_eval_tokens": 2390
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster appears to contain one unrelated issue titled \"Fett\" and a likely duplicate set of three issues titled \"huggingface\". The huggingface issues look like the same report; Fett does not.",
+        "confidence": 0.83,
+        "canonical_issue_reason": "Issue 43845 is the earliest of the near-identical \"huggingface\" reports and is the best representative of that duplicate set.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 43845 is the clearest cluster anchor: earliest created among the matching issues and aligned with the repeated title/body pattern.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:43835",
+            "right": "issue:43845",
+            "accept": false,
+            "reason": "Different titles and no shared targets/files; similarity is too weak to treat as the same underlying issue."
+          },
+          {
+            "left": "issue:43835",
+            "right": "issue:43846",
+            "accept": false,
+            "reason": "Different titles and no concrete evidence of the same bug; likely unrelated."
+          },
+          {
+            "left": "issue:43835",
+            "right": "issue:43859",
+            "accept": false,
+            "reason": "Different titles and no shared technical context; insufficient basis for duplication."
+          },
+          {
+            "left": "issue:43845",
+            "right": "issue:43846",
+            "accept": true,
+            "reason": "Same title, very similar minimal bodies, and created back-to-back; likely duplicate reports of the same issue."
+          },
+          {
+            "left": "issue:43845",
+            "right": "issue:43859",
+            "accept": true,
+            "reason": "Same title and clearly repeated report pattern; likely the same underlying issue."
+          },
+          {
+            "left": "issue:43846",
+            "right": "issue:43859",
+            "accept": true,
+            "reason": "Same title and highly similar minimal content; likely duplicate reports of the same issue."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: 43835 is the outlier by title and low similarity, while 43845/43846/43859 form the plausible duplicate set. The canonical selection of 43845 is conservative and correctly based on earliest creation among the matching huggingface issues. The soft-edge acceptances are also sufficiently hedged given the available evidence."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 12,
+      "component_count": 21,
+      "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": "ff7f0a4fe8beaf51a4f1d0c44986e89f7fefae8a8893a0bb2af7986a4787ea9c",
+      "component_budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 9,
+        "serialized_chars": 5788,
+        "estimated_input_tokens": 1447,
+        "estimated_eval_tokens": 3150
+      },
+      "budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 9,
+        "serialized_chars": 5788,
+        "estimated_input_tokens": 1447,
+        "estimated_eval_tokens": 3150
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "All items are low-signal spam or placeholder issues with no actionable bug/change description. There is no defensible canonical representative for a real duplicate cluster.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:44371",
+            "right": "issue:44572",
+            "accept": false,
+            "reason": "Both are '<spam>' placeholders, but there is no underlying issue to unify."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "Same spam placeholder pattern; no concrete shared bug or change."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44418",
+            "accept": false,
+            "reason": "Titles are different spam/placeholder text and contain no actionable problem statement."
+          },
+          {
+            "left": "issue:44371",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "Different spam-like titles with no evidence they report the same real issue."
+          },
+          {
+            "left": "issue:44418",
+            "right": "issue:44572",
+            "accept": false,
+            "reason": "One is generic documentation spam, the other is a blank spam placeholder; not the same underlying report."
+          },
+          {
+            "left": "issue:44418",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "No concrete bug/change description is present in either issue."
+          },
+          {
+            "left": "issue:44572",
+            "right": "issue:44573",
+            "accept": false,
+            "reason": "Identical spam titles do not establish a shared underlying bug or change."
+          },
+          {
+            "left": "issue:44572",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "The Google LLC title is unrelated to the blank spam placeholder."
+          },
+          {
+            "left": "issue:44573",
+            "right": "issue:44937",
+            "accept": false,
+            "reason": "The issues are both low-signal, but not clearly the same underlying report."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 13,
+      "component_count": 21,
+      "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": "364e0db265ee9e53cf3bdbd23da77e22d6ef12210624bcb697d25194fc54a09c",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1646,
+        "estimated_input_tokens": 412,
+        "estimated_eval_tokens": 1080
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1646,
+        "estimated_input_tokens": 412,
+        "estimated_eval_tokens": 1080
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The pair looks like duplicate issue reports for the same Energy/N6 Arithmetic proposal: the titles are identical and the bodies are very similar, with no distinct files, targets, or code-path evidence separating them.",
+        "confidence": 0.83,
+        "canonical_issue_reason": "Issue #45145 is the better canonical thread because it was created first and already has an inbound reference, making it the more established reference point for this duplicate topic.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #45145 is the best issue to aggregate on because it is earliest and already linked from elsewhere; #45151 does not add enough distinguishing information to justify preferring it.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45145",
+            "right": "issue:45151",
+            "accept": true,
+            "reason": "Accept as duplicate issues: the titles are identical, the bodies are near-matches in length and topic, and there is no evidence of different scope or targets."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": false,
+        "feedback": "The duplicate/canonicalization direction is plausible because the titles are identical and the bodies are similarly sized, but the packet only provides a soft_similarity signal and explicitly marks the pair as deterministic_accept=false. The writeup is too strong in saying they are 'effectively the same issue' and that there is 'no evidence of distinct scope.' Also, the 'marginally more discussion context' claim is not supported: issue 45145 has 0 discussion activity while 45151 has 1. The best/canonical issue rationale should lean on earlier creation plus the inbound reference, and the confidence should be more conservative."
+      },
+      "evaluator_used": true,
+      "retried": true,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 14,
+      "component_count": 21,
+      "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": 15,
+      "component_count": 21,
+      "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": "24557ba0273a1765734bcb08ecf223b7e4c7c1d9580fc90360a4597f9f6822ad",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1672,
+        "estimated_input_tokens": 418,
+        "estimated_eval_tokens": 1092
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1672,
+        "estimated_input_tokens": 418,
+        "estimated_eval_tokens": 1092
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are two separate 4-bit inference bugs in different code paths: one in Gemma4 audio modules failing on uint8 weights via torch.finfo(), and one in BitsAndBytesConfig where llm_int8_skip_modules unintentionally removes the default lm_head exclusion, triggering an AssertionError.",
+        "confidence": 0.96,
+        "canonical_issue_reason": "Issue #45674 is the better canonical issue because it states a clearer, more general configuration bug with an explicit failure mode and likely broader applicability across 4-bit inference users.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue #45674 is the strongest representative of its cluster because it describes the root config behavior directly and is not limited to a single model family like Gemma4 audio modules.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45672",
+            "right": "issue:45674",
+            "accept": false,
+            "reason": "They describe different underlying failures: a Gemma4 audio-module uint8/torch.finfo() TypeError versus a BitsAndBytesConfig lm_head exclusion/AssertionError regression. Same broad area, but not the same bug or change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 16,
+      "component_count": 21,
+      "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": "e7a5f39f8ea7428d1fcb741982642597e163792ef3a065d8e2646a4de01c2429",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1476,
+        "estimated_input_tokens": 369,
+        "estimated_eval_tokens": 994
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1476,
+        "estimated_input_tokens": 369,
+        "estimated_eval_tokens": 994
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two issues are unrelated standalone reports consisting only of different personal names and dates; there is no evidence they describe the same underlying bug or request.",
+        "confidence": 0.99,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Neither issue is a good canonical representative because they do not appear to be duplicates of each other or part of the same problem.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45762",
+            "right": "issue:45763",
+            "accept": false,
+            "reason": "Different subjects (two different names and birth dates) with no shared description, target, or evidence of 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": 17,
+      "component_count": 21,
+      "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": "d5fa7897b8928411f85e72e04c5cfd189c9cba01770ce39ac48872caf95cb3f1",
+      "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 only person-name/date titles and no body or shared technical context. The similarity signal looks superficial, so they should not be clustered as duplicates.",
+        "confidence": 0.94,
+        "canonical_issue_reason": "Neither issue provides any technical description or linkage; if forced to choose, issue #45764 is no better than #45765, but there is no meaningful canonical issue here.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "No issue is a strong representative because both lack substance and do not indicate the same underlying bug or request.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45764",
+            "right": "issue:45765",
+            "accept": false,
+            "reason": "Only superficial title similarity (names/dates); no body, shared targets, or technical evidence that they describe the same underlying problem."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 18,
+      "component_count": 21,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-45810-2",
+      "nodes": [
+        "issue:45810",
+        "issue:46295"
+      ],
+      "soft_pairs": [
+        "issue:45810|issue:46295"
+      ],
+      "prepared_review_unit_hash": "8de23704017197615329fedd3c07bfd668d8a0da21993b3ced760ee8cad20b3f",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1542,
+        "estimated_input_tokens": 386,
+        "estimated_eval_tokens": 1028
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1542,
+        "estimated_input_tokens": 386,
+        "estimated_eval_tokens": 1028
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues are unrelated: one requests a new Qwen3_5 token-classification/value-model head, while the other asks for FP8 training support. No duplicate relationship is evident.",
+        "confidence": 0.99,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": "Issue 46295 is the broader, cleaner request and is the better representative issue if one must be chosen, but it is not a duplicate of issue 45810.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45810",
+            "right": "issue:46295",
+            "accept": false,
+            "reason": "Different feature areas and code paths: Qwen3_5 token classification/value model support vs FP8 training support. They do not describe 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": 19,
+      "component_count": 21,
+      "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": "75635bec1deef7f0cc5df73ef669ed319ea1d621093032b83f1d1a98a0dea6a6",
+      "component_budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 8,
+        "serialized_chars": 5729,
+        "estimated_input_tokens": 1433,
+        "estimated_eval_tokens": 3122
+      },
+      "budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 8,
+        "serialized_chars": 5729,
+        "estimated_input_tokens": 1433,
+        "estimated_eval_tokens": 3122
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "There are two clear duplicate pairs: 45870/45871 and 45872/45873. Issue 45869 is a different request (dataset bundle) and should not be merged with the example or scraping-service issues.",
+        "confidence": 0.97,
+        "canonical_issue_reason": "Issue 45870 is a clean canonical representative for the exact duplicate pair 45870/45871: same title/body intent, and it has slightly more activity than 45871.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "45870 is the best overall issue representative in this cluster because it is an exact duplicate target with the most signal/engagement among the near-identical issues.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:45869",
+            "right": "issue:45870",
+            "accept": false,
+            "reason": "Different artifacts: dataset sale/sample request vs an example for financial time series prediction. Same broad domain, but not the same underlying issue."
+          },
+          {
+            "left": "issue:45869",
+            "right": "issue:45871",
+            "accept": false,
+            "reason": "Different artifacts: dataset sale/sample request vs an example for financial time series prediction. Overlap is thematic only."
+          },
+          {
+            "left": "issue:45870",
+            "right": "issue:45871",
+            "accept": true,
+            "reason": "Exact duplicate: identical title and effectively the same issue text/intent, with no meaningful distinguishing details."
+          },
+          {
+            "left": "issue:45870",
+            "right": "issue:45872",
+            "accept": false,
+            "reason": "Both mention financial time series/DataForge, but one is an example request and the other is a scraping-service integration request; different underlying change/request."
+          },
+          {
+            "left": "issue:45870",
+            "right": "issue:45873",
+            "accept": false,
+            "reason": "Same broad topic, but different concrete request: example prediction vs scraping-service integration. Not the same issue."
+          },
+          {
+            "left": "issue:45871",
+            "right": "issue:45872",
+            "accept": false,
+            "reason": "Same broad topic, but one is an example request and the other is a scraping-service integration request; not a duplicate."
+          },
+          {
+            "left": "issue:45871",
+            "right": "issue:45873",
+            "accept": false,
+            "reason": "The requests differ materially: example prediction workflow versus scraping-service integration. Not the same underlying issue."
+          },
+          {
+            "left": "issue:45872",
+            "right": "issue:45873",
+            "accept": true,
+            "reason": "Exact duplicate: identical title and same request, with no visible differences in the provided metadata."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: 45870/45871 and 45872/45873 are the only clearly near-identical pairs by title/body length, and 45869 is materially different. The soft-edge rejections are conservative overall, and the choice of 45870 as the canonical representative is supported by its slightly higher discussion activity."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 20,
+      "component_count": 21,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46439-2",
+      "nodes": [
+        "issue:46439",
+        "issue:46441"
+      ],
+      "soft_pairs": [
+        "issue:46439|issue:46441"
+      ],
+      "prepared_review_unit_hash": "defc721e66577ae0a07ca97c19dd5215e8f2496c56fa2d9686dc71da2d078fce",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1636,
+        "estimated_input_tokens": 409,
+        "estimated_eval_tokens": 1074
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1636,
+        "estimated_input_tokens": 409,
+        "estimated_eval_tokens": 1074
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both issues concern static caching with linear-attention models, but they describe different failure modes: one is shape corruption during `early_initialization`, the other is a Qwen3.5 `KeyError` for `linear_attention`. I\u2019d keep them separate unless a maintainer confirms they share the same root cause.",
+        "confidence": 0.77,
+        "canonical_issue_reason": "issue:46439 is the better representative because it states the broader static-cache/linear-attention failure mode and includes the concrete mechanism (`early_initialization` causing wrong shapes).",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:46439 is the strongest standalone report: it is more general, names the affected code path, and is less model-specific than the Qwen3.5-only crash report.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46439",
+            "right": "issue:46441",
+            "accept": false,
+            "reason": "Related area, but not clearly the same bug: one report is shape corruption from `early_initialization`, the other is a `KeyError: 'linear_attention'` on Qwen3.5. Different observed failures likely need different fixes."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "issue",
+      "component_index": 21,
+      "component_count": 21,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "cache",
+      "cache_hit": true,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46551-2",
+      "nodes": [
+        "issue:46551",
+        "issue:46552"
+      ],
+      "soft_pairs": [
+        "issue:46551|issue:46552"
+      ],
+      "prepared_review_unit_hash": "9e6277ad101b1563cb31f2d0a440396c92c16fbdffdd817754aa8461794534c5",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1707,
+        "estimated_input_tokens": 427,
+        "estimated_eval_tokens": 1110
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 1707,
+        "estimated_input_tokens": 427,
+        "estimated_eval_tokens": 1110
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two issues are duplicate reports of the exact same Wav2Vec2CTCTokenizer v5 special-token regression; the open issue is the better tracking target.",
+        "confidence": 0.99,
+        "canonical_issue_reason": "issue:46552 is the better canonical issue because it is open, has inbound references, and reports the same bug as the closed duplicate.",
+        "canonical_pr_reason": null,
+        "best_issue_reason": "issue:46552 is the best issue overall since it is the active tracker for the same concrete tokenizer bug.",
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "issue:46551",
+            "right": "issue:46552",
+            "accept": true,
+            "reason": "Identical title and body describing the same tokenizer special-token leak; this is a clear duplicate pair."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The assessment is well grounded in the packet: the two issues have identical titles, identical body length, and only soft similarity evidence, so treating them as duplicates is conservative and reasonable. The choice of issue:46552 as the canonical tracking target is also supported by its open state and inbound references. No unsupported stronger claim stands out."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 1,
+      "component_count": 37,
+      "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-39895-14",
+      "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:46048",
+        "pull_request:46267"
+      ],
+      "soft_pairs": [
+        "pull_request:39895|pull_request:43424",
+        "pull_request:41855|pull_request:43424",
+        "pull_request:41855|pull_request:43823",
+        "pull_request:42668|pull_request:43823",
+        "pull_request:43251|pull_request:43823",
+        "pull_request:43251|pull_request:44827",
+        "pull_request:43424|pull_request:44827",
+        "pull_request:43424|pull_request:43823",
+        "pull_request:43823|pull_request:45655",
+        "pull_request:43823|pull_request:43995",
+        "pull_request:43995|pull_request:45213",
+        "pull_request:43995|pull_request:44827",
+        "pull_request:44827|pull_request:45170",
+        "pull_request:44827|pull_request:46267",
+        "pull_request:44827|pull_request:46048",
+        "pull_request:44827|pull_request:45655",
+        "pull_request:44827|pull_request:45213",
+        "pull_request:43823|pull_request:45213",
+        "pull_request:43251|pull_request:43254"
+      ],
+      "prepared_review_unit_hash": "b00fe57e94a2c0594d257c899e348af4097121c7273a9ce43b9375424fd19076",
+      "component_budget": {
+        "node_count": 14,
+        "item_count": 14,
+        "soft_pair_count": 19,
+        "serialized_chars": 38082,
+        "estimated_input_tokens": 9521,
+        "estimated_eval_tokens": 19298
+      },
+      "budget": {
+        "node_count": 14,
+        "item_count": 14,
+        "soft_pair_count": 19,
+        "serialized_chars": 38082,
+        "estimated_input_tokens": 9521,
+        "estimated_eval_tokens": 19298
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster is mostly unrelated pull requests. The only strong duplicate is the pair of loss-function PRs that both add `weight`/`label_smoothing` support to `fixed_cross_entropy`; the rest are separate model additions, docs, or refactors that should stay distinct.",
+        "confidence": 0.92,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #43251 is the better canonical representative for the `fixed_cross_entropy` kwargs fix: it explicitly targets issue #43240 and carries the more complete discussion/history, while #43254 is a near-identical smaller variant of the same code-path change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #43251 is the best single PR in this cluster because it is the clearest, issue-linked implementation of the shared loss-function bug fix and can subsume the duplicate variant (#43254).",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:39895",
+            "right": "pull_request:43424",
+            "accept": false,
+            "reason": "Both are unrelated feature/test PRs; they do not fix the same code path or change."
+          },
+          {
+            "left": "pull_request:41855",
+            "right": "pull_request:43424",
+            "accept": false,
+            "reason": "Different tokenizer utility vs executorch export test; same broad library, not the same bug/change."
+          },
+          {
+            "left": "pull_request:41855",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Mistral tokenizer methods and MobileLLM model addition are unrelated."
+          },
+          {
+            "left": "pull_request:42668",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Processor robustness changes are not the same underlying work as adding a new model."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Loss-function kwargs support and MobileLLM model addition are unrelated."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "A cross-entropy helper fix and Mistral4 test/model work are different code paths."
+          },
+          {
+            "left": "pull_request:43424",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "Executorch dynamic-shape export testing and Mistral4 fixes are separate concerns."
+          },
+          {
+            "left": "pull_request:43424",
+            "right": "pull_request:43823",
+            "accept": false,
+            "reason": "Dynamic-shape export tests do not overlap with model addition work."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:45655",
+            "accept": false,
+            "reason": "New MobileLLM model code and config-resolution ordering fix are unrelated."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:43995",
+            "accept": false,
+            "reason": "Model addition and Falcon output-collection refactor are not the same change."
+          },
+          {
+            "left": "pull_request:43995",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "Falcon refactoring is unrelated to the model-creation skill PR."
+          },
+          {
+            "left": "pull_request:43995",
+            "right": "pull_request:44827",
+            "accept": false,
+            "reason": "Falcon output-interface refactor and Mistral4 fixes are different model-specific changes."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45170",
+            "accept": false,
+            "reason": "Mistral4 test fixes and layrnorm/layernorm renaming are different bugs."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:46267",
+            "accept": false,
+            "reason": "Mistral4 test fixes and chat parsing updates do not share the same concrete code path."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:46048",
+            "accept": false,
+            "reason": "Mistral4 test fixes and TDT loss kernel integration are unrelated."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45655",
+            "accept": false,
+            "reason": "Mistral4 work and config-resolution fix are unrelated."
+          },
+          {
+            "left": "pull_request:44827",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "They only overlap on a broad auto/modeling area; the actual changes are unrelated."
+          },
+          {
+            "left": "pull_request:43823",
+            "right": "pull_request:45213",
+            "accept": false,
+            "reason": "New model implementation and model-creation skill scaffolding are not the same artifact or bug."
+          },
+          {
+            "left": "pull_request:43251",
+            "right": "pull_request:43254",
+            "accept": true,
+            "reason": "Both PRs implement the same `fixed_cross_entropy` kwargs support in the same file for the same issue (#43240); one is a near-duplicate variant of the other."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The cluster summary is grounded in the packet and stays conservative: the only accepted duplicate pair, PR #43251/#43254, is a clear same-file same-issue `fixed_cross_entropy` kwargs fix and looks mergeable into a single PR. The other verdicts are appropriately negative and supported by distinct titles/files; I don\u2019t see any overstatement that would require rejection."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 2,
+      "component_count": 37,
+      "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": "91c9c06ec8e714322654fca35759a57d59906006a232854bc83b7f7d2387b357",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 7226,
+        "estimated_input_tokens": 1807,
+        "estimated_eval_tokens": 3870
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 7226,
+        "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 are related to VibeVoice, but they are not duplicates: #40546 adds the core VibeVoice integration, while #42781 adds a separate realtime/acoustic-tokenizer extension on top of that base work.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #40546 is the foundational VibeVoice implementation and the broader representative of this cluster; #42781 is a later, narrower realtime extension rather than the same change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "#40546 is the best cluster representative because it introduces the main model integration and core plumbing, whereas #42781 focuses on a distinct realtime variant and tokenizer stack.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:40546",
+            "right": "pull_request:42781",
+            "accept": false,
+            "reason": "Reject: these are adjacent VibeVoice features, not the same underlying fix or change. #40546 implements the base model; #42781 adds realtime/acoustic-tokenizer support and additional generation/runtime changes. Shared filenames come from the same model family, but the code paths and merged scope differ."
+          }
+        ]
+      },
+      "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": 37,
+      "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": "a127d89a8356226894a044e9dcf6cf8e43584bd4103201d8fbcbb2c70d3cc048",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4675,
+        "estimated_input_tokens": 1169,
+        "estimated_eval_tokens": 2594
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4675,
+        "estimated_input_tokens": 1169,
+        "estimated_eval_tokens": 2594
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two PRs are unrelated feature additions in different model families: TimesFM covariate forecasting vs. DINOv3 image classification support.",
+        "confidence": 0.99,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "pull_request:41224 is the more focused representative: it targets one concrete model capability (DINOv3 image classification) with a narrow implementation scope.",
+        "best_issue_reason": null,
+        "best_pr_reason": "pull_request:41224 is the cleanest standalone PR in this cluster because it is narrowly scoped and directly tied to a single code path, unlike the broader TimesFM covariate work.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:40755",
+            "right": "pull_request:41224",
+            "accept": false,
+            "reason": "Different underlying changes and code paths: TimesFM forecasting with covariates vs. DINOv3 image classification support. They do not appear 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": 37,
+      "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": 37,
+      "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": "b8d030bf284ad626a8b8ef032485fbc5c8f33e0c93323d5ff8d38f9cc525f9ef",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 10436,
+        "estimated_input_tokens": 2609,
+        "estimated_eval_tokens": 5474
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 10436,
+        "estimated_input_tokens": 2609,
+        "estimated_eval_tokens": 5474
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are three separate model-addition PRs that share repository plumbing and docs scaffolding, but they target different models and codepaths. The overlap is superficial, not a duplicate change set.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:41797",
+            "right": "pull_request:45186",
+            "accept": false,
+            "reason": "Different models (DeepSeek OCR vs Isaac) with distinct configs, model code, processors, and tests; shared auto-registration/docs files are template overlap only."
+          },
+          {
+            "left": "pull_request:45186",
+            "right": "pull_request:45630",
+            "accept": false,
+            "reason": "Different models (Isaac vs Kimi2-6) and different implementation details; they share broad model-addition plumbing but 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": "pull_request",
+      "component_index": 6,
+      "component_count": 37,
+      "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": "d4c6892c3cc063345aa024f3fc96e819d9441121a294d9fca2410b673630ae57",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5934,
+        "estimated_input_tokens": 1484,
+        "estimated_eval_tokens": 3224
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5934,
+        "estimated_input_tokens": 1484,
+        "estimated_eval_tokens": 3224
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open workflow-maintenance PRs that both bump GitHub Actions dependencies, but they target different action families and different concrete compatibility concerns.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #42975 is the better representative only because it is broader and more clearly centered on the Node 24-compatible Actions refresh across many workflows; PR #42976 is a narrower Docker action version bump.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #42975 is the strongest cluster anchor: it updates multiple core workflow actions (checkout/cache/upload-artifact) across many files. PR #42976 is still substantial, but it specifically bumps docker/build-push-action and related workflow versions, so it is a separate change stream.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:42975",
+            "right": "pull_request:42976",
+            "accept": false,
+            "reason": "Do not merge: both are workflow dependency upgrades, but one is about general GitHub Actions Node 24 compatibility while the other is about docker/build-push-action and related version bumps. They touch overlapping files, 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": "pull_request",
+      "component_index": 7,
+      "component_count": 37,
+      "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-42978-5",
+      "nodes": [
+        "pull_request:42978",
+        "pull_request:43448",
+        "pull_request:43451",
+        "pull_request:45490",
+        "pull_request:46562"
+      ],
+      "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:45490|pull_request:46562"
+      ],
+      "prepared_review_unit_hash": "c727ec421c4e6cc014d54dab4377afae0489ec6f45071d399c117ecd765eea43",
+      "component_budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 7,
+        "serialized_chars": 17573,
+        "estimated_input_tokens": 4394,
+        "estimated_eval_tokens": 9044
+      },
+      "budget": {
+        "node_count": 5,
+        "item_count": 5,
+        "soft_pair_count": 7,
+        "serialized_chars": 17573,
+        "estimated_input_tokens": 4394,
+        "estimated_eval_tokens": 9044
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "All items are distinct model-addition PRs with only superficial overlap in docs/auto-registration boilerplate. They do not describe the same underlying change or fix, so none should be merged as duplicates.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:43448",
+            "accept": false,
+            "reason": "Different models (ViT NEPA vs Molmo) with separate model, processor, and test files; shared docs/auto registration is boilerplate, not the same code-path change."
+          },
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:43451",
+            "accept": false,
+            "reason": "ViT NEPA and Molmo2 are unrelated model additions; the overlap is only in generic transformers registration/docs, not a single mergeable fix."
+          },
+          {
+            "left": "pull_request:42978",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "CTSM and ViT NEPA are different model implementations and tests; no shared bug or identical change to merge together."
+          },
+          {
+            "left": "pull_request:43448",
+            "right": "pull_request:43451",
+            "accept": false,
+            "reason": "Molmo and Molmo2 are related by naming but are distinct model families with different files and functionality; not the same concrete change."
+          },
+          {
+            "left": "pull_request:43448",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "Molmo and CTSM are unrelated model-addition PRs; shared auto-map/docs edits are too broad to indicate duplication."
+          },
+          {
+            "left": "pull_request:43451",
+            "right": "pull_request:45490",
+            "accept": false,
+            "reason": "Molmo2 and CTSM touch different model codepaths and introduce different architectures; no evidence of a single underlying fix."
+          },
+          {
+            "left": "pull_request:45490",
+            "right": "pull_request:46562",
+            "accept": false,
+            "reason": "CTSM and Step3p7 are separate model additions with different configs, processors, and tests; similar repo plumbing does not make them duplicates."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 8,
+      "component_count": 37,
+      "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-43291-8",
+      "nodes": [
+        "pull_request:43291",
+        "pull_request:43498",
+        "pull_request:43665",
+        "pull_request:44815",
+        "pull_request:45114",
+        "pull_request:46380",
+        "pull_request:46539",
+        "pull_request:46579"
+      ],
+      "soft_pairs": [
+        "pull_request:43291|pull_request:43665",
+        "pull_request:43291|pull_request:44815",
+        "pull_request:43291|pull_request:46380",
+        "pull_request:43498|pull_request:46579",
+        "pull_request:43665|pull_request:46539",
+        "pull_request:43665|pull_request:44815",
+        "pull_request:43665|pull_request:46380",
+        "pull_request:43665|pull_request:46579",
+        "pull_request:44815|pull_request:46539",
+        "pull_request:44815|pull_request:46380",
+        "pull_request:44815|pull_request:46579",
+        "pull_request:45114|pull_request:46579",
+        "pull_request:46380|pull_request:46539",
+        "pull_request:46380|pull_request:46579"
+      ],
+      "prepared_review_unit_hash": "589b9f7ab521f675f5156cac7accf6ca484225e9c95acdc59ea032aca2de374b",
+      "component_budget": {
+        "node_count": 8,
+        "item_count": 8,
+        "soft_pair_count": 14,
+        "serialized_chars": 23156,
+        "estimated_input_tokens": 5789,
+        "estimated_eval_tokens": 11834
+      },
+      "budget": {
+        "node_count": 8,
+        "item_count": 8,
+        "soft_pair_count": 14,
+        "serialized_chars": 23156,
+        "estimated_input_tokens": 5789,
+        "estimated_eval_tokens": 11834
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These PRs are mostly unrelated fixes with generic titles or superficial overlap in tests/docs. I don\u2019t see a shared concrete bug or change set that would justify treating any pair as duplicates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": null,
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:43665",
+            "accept": false,
+            "reason": "Both touch test-related files, but one is a Whisper tokenizer fix and the other is unrelated CLIP/ViT test edits; different code paths and no shared bug."
+          },
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:44815",
+            "accept": false,
+            "reason": "Whisper decoding/tests are unrelated to FP8 dequant/loading logic; overlap is only broad 'fix' similarity."
+          },
+          {
+            "left": "pull_request:43291",
+            "right": "pull_request:46380",
+            "accept": false,
+            "reason": "Whisper tokenization/timestamp handling and generation watermarking validation are separate issues in different subsystems."
+          },
+          {
+            "left": "pull_request:43498",
+            "right": "pull_request:46579",
+            "accept": false,
+            "reason": "One adds a backward-compat alias for tie_weights; the other is a broad typo cleanup across many files. No same underlying bug."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:46539",
+            "accept": false,
+            "reason": "Model test assertion edits are unrelated to the Phimoe integration test/CI memory workaround."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:44815",
+            "accept": false,
+            "reason": "Test failures in CLIP/ViT are unrelated to the dequant fix in core loading and FP8 integration."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:46380",
+            "accept": false,
+            "reason": "These touch different areas: vision model tests vs generation config/Whisper changes; no common concrete fix."
+          },
+          {
+            "left": "pull_request:43665",
+            "right": "pull_request:46579",
+            "accept": false,
+            "reason": "Generic test-file edits do not match a code-typo sweep across many source files."
+          },
+          {
+            "left": "pull_request:44815",
+            "right": "pull_request:46539",
+            "accept": false,
+            "reason": "FP8/dequant changes are unrelated to Phimoe integration testing and CI environment memory limits."
+          },
+          {
+            "left": "pull_request:44815",
+            "right": "pull_request:46380",
+            "accept": false,
+            "reason": "Both involve generation-related files, but one is a dequant/loading fix and the other is a validation/Whisper integration fix; not the same bug."
+          },
+          {
+            "left": "pull_request:44815",
+            "right": "pull_request:46579",
+            "accept": false,
+            "reason": "Shared file names are incidental; FP8 dequant fixes and typo cleanups address different changes."
+          },
+          {
+            "left": "pull_request:45114",
+            "right": "pull_request:46579",
+            "accept": false,
+            "reason": "Documentation doctest cleanup is not the same as code typo fixes across implementation files."
+          },
+          {
+            "left": "pull_request:46380",
+            "right": "pull_request:46539",
+            "accept": false,
+            "reason": "Generation/Whisper validation fixes are unrelated to the Phimoe integration test CI memory cap change."
+          },
+          {
+            "left": "pull_request:46380",
+            "right": "pull_request:46579",
+            "accept": false,
+            "reason": "One fixes generation config validation and Whisper behavior; the other is a broad typo cleanup, so they are not mergeable 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": 9,
+      "component_count": 37,
+      "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": "dbc629f0b0bad3a827a6db245d9d9e9633ed2684b8b13700ff09dd0a4179e514",
+      "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": "Two open '[don't merge]' pull requests with only superficial overlap on a Vit test file; they target different code paths and are not duplicates.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No canonical PR: 43488 is a packaging/test-dependency/version-bump check, while 43532 is a broader workflow/testing-utility change; they do not represent the same underlying fix.",
+        "best_issue_reason": null,
+        "best_pr_reason": "If one must be chosen as the more representative cluster member, 43532 has the richer, broader change set, but it still should not be treated as a duplicate of 43488.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43488",
+            "right": "pull_request:43532",
+            "accept": false,
+            "reason": "Reject: these are different synthetic maintenance changes. 43488 edits setup/dependency/version files and a Vit test, while 43532 changes GitHub workflows plus testing/notification utilities. Shared test-file touch is incidental, not the same bug or fix."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 10,
+      "component_count": 37,
+      "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": "4a9f2f6b9bcccf3ebeff2e0c5b832acb050c6a2fc1c77cff4954d0390f79af34",
+      "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 related but distinct feature PRs: one adds a Promptable Concept Segmentation pipeline, the other adds a Promptable Visual Segmentation pipeline. They share some scaffolding, but they target different tasks, model integrations, and test surfaces, so they should not be treated as duplicates.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No single canonical PR: the two PRs implement different pipelines with different underlying functionality and model coverage rather than the same fix/change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "Neither PR is a global best representative for a duplicate cluster, because they are separate feature additions instead of alternative implementations of the same code-path problem.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43612",
+            "right": "pull_request:43613",
+            "accept": false,
+            "reason": "Different features: concept segmentation vs visual segmentation. The diffs show different pipeline modules, docs pages, and model/test integrations (SAM3-centric versus SAM/SAM2/EdgeTam), so they are not the same underlying 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": 11,
+      "component_count": 37,
+      "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": "6f659977cc638e23fc61e39d66304fbfb0e843a989462bc90fb4d5435d14f327",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4293,
+        "estimated_input_tokens": 1074,
+        "estimated_eval_tokens": 2404
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4293,
+        "estimated_input_tokens": 1074,
+        "estimated_eval_tokens": 2404
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open PRs in `trainer.py` add separate trainer extension hooks: one for gradient-accumulation loss scaling and one for overriding signature-based column filtering. They touch the same file but solve different problems.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #43651 is the more representative change because it introduces a concrete new trainer behavior around loss scaling in the training step, which is a clearer standalone capability than the helper refactor in #43663.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #43651 is the best standalone PR in this cluster: it addresses a specific training-path behavior change and is easy to understand as a distinct user-facing fix/feature.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43651",
+            "right": "pull_request:43663",
+            "accept": false,
+            "reason": "They are different changes in the same file: #43651 adds `_loss_is_scaled_for_ga` to control gradient accumulation loss normalization, while #43663 adds `_get_signature_columns` to customize column filtering. They do not appear to fix the same underlying bug or be 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": 12,
+      "component_count": 37,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-43656-2",
+      "nodes": [
+        "pull_request:43656",
+        "pull_request:43842"
+      ],
+      "soft_pairs": [
+        "pull_request:43656|pull_request:43842"
+      ],
+      "prepared_review_unit_hash": "d3a697ec81c71df5b72e767cb08408e0f8e36f426d83c0b44167b7b179390f17",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 3663,
+        "estimated_input_tokens": 916,
+        "estimated_eval_tokens": 2088
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 3663,
+        "estimated_input_tokens": 916,
+        "estimated_eval_tokens": 2088
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open PRs touch the same CLI file and the same TypeAdapter area, but they address different NameError scenarios: one is a broader annotation/runtime typing fix, the other is specifically for missing pydantic. They are related, not duplicates.",
+        "confidence": 0.89,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "pull_request:43656 is the better canonical representative because it is the broader CLI typing fix with an explicit issue link; pull_request:43842 is a separate, narrower pydantic-optional import/annotation guard.",
+        "best_issue_reason": null,
+        "best_pr_reason": "pull_request:43656 is the strongest single representative of the cluster: it fixes a concrete CLI NameError path and makes the typing/runtime behavior coherent. pull_request:43842 targets a different failure mode and should remain separate.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43656",
+            "right": "pull_request:43842",
+            "accept": false,
+            "reason": "Reject: both PRs touch src/transformers/cli/serve.py and TypeAdapter, but they fix different bugs. 43656 addresses a broader annotation/runtime NameError in the CLI, while 43842 only guards the TypeAdapter type import when pydantic is absent. They are not the same underlying change 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": 13,
+      "component_count": 37,
+      "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": 37,
+      "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:44019|pull_request:44085",
+        "pull_request:44086|pull_request:44722",
+        "pull_request:44018|pull_request:44019",
+        "pull_request:44019|pull_request:44071",
+        "pull_request:44068|pull_request:44116",
+        "pull_request:44018|pull_request:44116",
+        "pull_request:44068|pull_request:44722",
+        "pull_request:44071|pull_request:44722",
+        "pull_request:44018|pull_request:44085",
+        "pull_request:44018|pull_request:44086",
+        "pull_request:44068|pull_request:44086",
+        "pull_request:44071|pull_request:44085",
+        "pull_request:44059|pull_request:44086",
+        "pull_request:44085|pull_request:44086",
+        "pull_request:44098|pull_request:44101",
+        "pull_request:44018|pull_request:44071",
+        "pull_request:44059|pull_request:44722",
+        "pull_request:44017|pull_request:44161",
+        "pull_request:44017|pull_request:44722",
+        "pull_request:44017|pull_request:44073",
+        "pull_request:44068|pull_request:44071",
+        "pull_request:43997|pull_request:44073",
+        "pull_request:44071|pull_request:44086",
+        "pull_request:44017|pull_request:44071",
+        "pull_request:44073|pull_request:44101",
+        "pull_request:44073|pull_request:44161",
+        "pull_request:44056|pull_request:44101",
+        "pull_request:44059|pull_request:44071",
+        "pull_request:44018|pull_request:44059",
+        "pull_request:44059|pull_request:44068",
+        "pull_request:44101|pull_request:44161",
+        "pull_request:43997|pull_request:44017",
+        "pull_request:43997|pull_request:44056",
+        "pull_request:44071|pull_request:44076",
+        "pull_request:44074|pull_request:44101",
+        "pull_request:44056|pull_request:44161",
+        "pull_request:43997|pull_request:44018",
+        "pull_request:43997|pull_request:44161",
+        "pull_request:44074|pull_request:44161",
+        "pull_request:44018|pull_request:44161",
+        "pull_request:43997|pull_request:44071",
+        "pull_request:44010|pull_request:44071",
+        "pull_request:44010|pull_request:44073",
+        "pull_request:44018|pull_request:44129",
+        "pull_request:43997|pull_request:44074",
+        "pull_request:43997|pull_request:44129",
+        "pull_request:44010|pull_request:44018",
+        "pull_request:44071|pull_request:44129",
+        "pull_request:43996|pull_request:44161",
+        "pull_request:43996|pull_request:44017",
+        "pull_request:44010|pull_request:44044",
+        "pull_request:44044|pull_request:44074",
+        "pull_request:44066|pull_request:44722",
+        "pull_request:44085|pull_request:44722",
+        "pull_request:44085|pull_request:44129",
+        "pull_request:43996|pull_request:44007",
+        "pull_request:43996|pull_request:44129",
+        "pull_request:44044|pull_request:44129",
+        "pull_request:44044|pull_request:44154"
+      ],
+      "prepared_review_unit_hash": "a47f17aec74d9ec27f72c99c6785b5090cd477b1195e63656bcbc1fde62b7e3c",
+      "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": 75093,
+        "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 set of model-specific output-tracing decorator refactors under the same tracking theme (#43979), not true duplicates across architectures. The only strong duplicate candidates are the GPT-J/CodeGen variants and the GPT-Neo variants; the rest are related but distinct model migrations.",
+        "confidence": 0.77,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #44007 is the best representative of the cluster: it explicitly targets #43979, has a coherent standardized output-tracing migration, and spans multiple vision model files rather than a narrow one-off tweak.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #44007 is the strongest single representative of the broader output-tracing refactor work: explicit tracking issue, broad cross-model implementation, and a clear decorator-based migration path.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "Different model families and different code paths; sharing #43979 is not enough to make them duplicates."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "Both are output-tracing refactors, but for different architectures (CVT/FNet vs DeBERTa) and not the same underlying change."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44085",
+            "accept": true,
+            "reason": "Both are GPT-J output-tracing refactors in the same model file and appear to be iterative versions of the same change."
+          },
+          {
+            "left": "pull_request:44007",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "Different model families and implementation details; they are separate output-tracing migrations, not one fix."
+          },
+          {
+            "left": "pull_request:44072",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "EfficientNet vs GPT-J/CodeGen are unrelated code paths despite both following the same refactor pattern."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "Different models and different files; same initiative, but not the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:44013",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "MobileNetV2 and DeBERTa are unrelated model implementations, so these are not duplicates."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-J/CodeGen output tracing is unrelated to MGP-STR output tracing; different concrete code paths."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-J/CodeGen and MPT are different architectures; this is the same refactor theme, not the same change."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44068",
+            "accept": true,
+            "reason": "Both are GPT-Neo output-tracing refactors touching the same model file and look like alternate/updated versions of the same work."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "GPT-J/CodeGen vs GPT-Neo are different models and code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44019",
+            "accept": false,
+            "reason": "GPT-Neo and ResNet are unrelated; the similar wording is only the shared standardization pattern."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "ResNet output tracing and GPT-J/CodeGen output tracing are separate model-specific migrations."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "ResNet vs GPT-J are different concrete implementations, so they are not duplicates."
+          },
+          {
+            "left": "pull_request:44086",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "MGP-STR and GPT-J/CodeGen are unrelated model code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44019",
+            "accept": false,
+            "reason": "GPT-Neo and ResNet are different architectures; no shared underlying bug/change."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "ResNet and MPT are separate model refactors, not the same fix."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44116",
+            "accept": false,
+            "reason": "GPT-Neo and Flaubert/XLM output tracing are different model-specific changes."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44116",
+            "accept": false,
+            "reason": "GPT-Neo and Flaubert are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "GPT-Neo is unrelated to GPT-J/CodeGen despite the shared refactor theme."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "MPT and MGP-STR are different architectures and separate changes."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J are different model implementations."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo vs MGP-STR are not the same code-path problem."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo and MGP-STR are unrelated model migrations."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-Neo and MPT are different models with separate output-tracing implementations."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-2 and MGP-STR are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-J output tracing and MGP-STR output tracing are different model-specific changes."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-2 and MPT are different architectures and not duplicates."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44059",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-2 are distinct model refactors, even if both standardize output handling."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44017",
+            "accept": false,
+            "reason": "RegNet and SegFormer are different model families; same issue target does not make them duplicates."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "MPT and ImageGPT are separate model code paths."
+          },
+          {
+            "left": "pull_request:44074",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "TextNet and Flaubert/XLM are unrelated implementations."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "MPNet and Flaubert/XLM are different model-specific migrations."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "MPNet and LongT5 are separate architectures and not the same underlying change."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "GPT-2 and GPT-J/CodeGen are different models and code paths."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "SegFormer and LongT5 are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "RegNet and GPT-Neo are different model families."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44056",
+            "accept": false,
+            "reason": "RegNet and MPNet are separate output-tracing migrations."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "SqueezeBERT and MPT are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44074",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "TextNet and LongT5 are distinct models with separate changes."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "MPNet and GPT-Neo are different architectures."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "SqueezeBERT and GPT-Neo are unrelated model migrations."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "MPT and SpeechT5 are different code paths."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "MPT and MGP-STR are different models; not the same bug/change."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "CVT/FNet and LongT5 are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44017",
+            "accept": false,
+            "reason": "CVT/FNet and SegFormer are different model-specific refactors."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "SqueezeBERT and DeBERTa are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "DeBERTa and TextNet are separate model migrations."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44722",
+            "accept": true,
+            "reason": "Both target the same GPT-J/CodeGen output-tracing refactor and 44722 looks like a later, broader revision of the same change."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44722",
+            "accept": true,
+            "reason": "Both modify GPT-J output tracing in the same file; 44722 is a natural superset/revision of the same work."
+          },
+          {
+            "left": "pull_request:44085",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "GPT-J and SpeechT5 are unrelated models."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44007",
+            "accept": false,
+            "reason": "CVT/FNet and ResNet/RegNet/RT-DETR are different model migrations, not one duplicate change."
+          },
+          {
+            "left": "pull_request:43996",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "CVT/FNet and SpeechT5 are unrelated output-tracing changes."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "DeBERTa and SpeechT5 are different architectures and code paths."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44154",
+            "accept": false,
+            "reason": "DeBERTa and VITS are separate model-specific changes."
+          },
+          {
+            "left": "pull_request:44066",
+            "right": "pull_request:44072",
+            "accept": false,
+            "reason": "GPT-J/CodeGen and EfficientNet are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:44072",
+            "right": "pull_request:44722",
+            "accept": false,
+            "reason": "EfficientNet and GPT-J/CodeGen are different code paths."
+          },
+          {
+            "left": "pull_request:44013",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "MobileNetV2 and DeBERTa are unrelated models."
+          },
+          {
+            "left": "pull_request:44098",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "ViLT and Flaubert/XLM are separate model-specific migrations."
+          },
+          {
+            "left": "pull_request:44098",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "ViLT and VisualBERT are related visually, but they are different implementations and not the same concrete change."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "MPNet and Flaubert/XLM are different models and not duplicates."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-2 and MGP-STR are unrelated output-tracing changes."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "RegNet and TextNet are different model implementations."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "RegNet and SpeechT5 are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "GPT-Neo and LongT5 are distinct architectures."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "RegNet and GPT-Neo are different models."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "MPT and SpeechT5 are not the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44068",
+            "accept": false,
+            "reason": "GPT-2 and GPT-Neo are different model families."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "VisualBERT and Flaubert/XLM are separate model migrations."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44161",
+            "accept": false,
+            "reason": "VisualBERT and LongT5 are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "MPNet and Flaubert/XLM are different model-specific refactors."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-2 and MPT are unrelated models."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44059",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-2 are different architectures."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44066",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J/CodeGen are different code paths."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44085",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-J are different model implementations."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo and MGP-STR are unrelated changes."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-Neo and MGP-STR are separate model migrations."
+          },
+          {
+            "left": "pull_request:44068",
+            "right": "pull_request:44071",
+            "accept": false,
+            "reason": "GPT-Neo and MPT are different architectures."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "GPT-2 and TextNet are unrelated code paths."
+          },
+          {
+            "left": "pull_request:44019",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "ResNet and Flaubert/XLM are different model families."
+          },
+          {
+            "left": "pull_request:44059",
+            "right": "pull_request:44086",
+            "accept": false,
+            "reason": "GPT-2 and MGP-STR are distinct model-specific refactors."
+          },
+          {
+            "left": "pull_request:44098",
+            "right": "pull_request:44101",
+            "accept": false,
+            "reason": "ViLT and Flaubert/XLM are not the same underlying change."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "RegNet and GPT-Neo are unrelated model implementations."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "MPT and ImageGPT are different code paths."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44018",
+            "accept": false,
+            "reason": "SqueezeBERT and GPT-Neo are unrelated."
+          },
+          {
+            "left": "pull_request:44071",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "MPT and SpeechT5 are separate model migrations."
+          },
+          {
+            "left": "pull_request:44017",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "SegFormer and VisualBERT are different model-specific changes."
+          },
+          {
+            "left": "pull_request:44018",
+            "right": "pull_request:44059",
+            "accept": false,
+            "reason": "GPT-Neo and GPT-2 are different output-tracing implementations."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "SqueezeBERT and VisualBERT are unrelated model families."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "RegNet and TextNet are distinct model code paths."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44044",
+            "accept": false,
+            "reason": "SqueezeBERT and DeBERTa are unrelated."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "DeBERTa and TextNet are different architectures."
+          },
+          {
+            "left": "pull_request:44044",
+            "right": "pull_request:44154",
+            "accept": false,
+            "reason": "DeBERTa and VITS are separate model-specific migrations."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: it correctly treats most PRs as architecture-specific output-tracing refactors under #43979, and it only elevates the GPT-J/CodeGen and GPT-Neo pairs as likely duplicates/revisions. The rejection calls are conservative, and the accepted pairs are plausibly mergeable revisions of the same model-specific bugfix."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 14,
+      "component_count": 37,
+      "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": "03f80aa1483d98ee382c6f68bd4070f4cc10996d29505197da02bee326ccf972",
+      "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": 17124,
+        "estimated_input_tokens": 4281,
+        "estimated_eval_tokens": 8818
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": true,
+      "analyst_result": {
+        "summary": "These PRs share a broad theme of migrating model forward passes to standardized output tracing, but they are separate model-specific refactors rather than duplicates of one underlying bug or change.",
+        "confidence": 0.93,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #44129 is the most well-scoped and explicit representative of the theme because it has a linked issue and a concrete SpeechT5 output-tracing refactor, but it is still model-specific rather than a duplicate of the others.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #44129 is the strongest single representative of the cluster\u2019s topic, with an explicit issue target and a focused code change; however, the cluster does not contain a single shared PR-worthy change across all items.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44024",
+            "right": "pull_request:44076",
+            "accept": false,
+            "reason": "Both are model-specific output-tracing migrations, but FocalNet and ImageGPT are different code paths and architectures, so they are not the same concrete change."
+          },
+          {
+            "left": "pull_request:44073",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "VisualBert and TextNet are separate models; despite similar decorator refactors, these are distinct per-model implementations, not one mergeable PR."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44073",
+            "accept": false,
+            "reason": "MPNet and VisualBert touch different model stacks and output-capture wiring, so they are related only by pattern, not by the same underlying fix."
+          },
+          {
+            "left": "pull_request:44056",
+            "right": "pull_request:44074",
+            "accept": false,
+            "reason": "These are separate model migrations to standardized output capture, not the same code-path problem or a single PR that could plausibly be merged together."
+          },
+          {
+            "left": "pull_request:43997",
+            "right": "pull_request:44010",
+            "accept": false,
+            "reason": "RegNet and SqueezeBert are unrelated model implementations; the shared decorator migration is too broad to treat as a duplicate."
+          },
+          {
+            "left": "pull_request:44010",
+            "right": "pull_request:44129",
+            "accept": false,
+            "reason": "Although both refactor output collection, SqueezeBert and SpeechT5 are distinct model-specific changes and should not be merged as one PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 15,
+      "component_count": 37,
+      "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": "6bcf3a2530400554388b5ade249c8422d76e6a577d1c6136b0f777ec9ac1fa39",
+      "component_budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 40,
+        "serialized_chars": 40172,
+        "estimated_input_tokens": 10043,
+        "estimated_eval_tokens": 20342
+      },
+      "budget": {
+        "node_count": 13,
+        "item_count": 13,
+        "soft_pair_count": 40,
+        "serialized_chars": 40172,
+        "estimated_input_tokens": 10043,
+        "estimated_eval_tokens": 20342
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are separate model-specific output-tracing refactors, not duplicate fixes; they share the same tooling pattern but not the same code path or mergeable change.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 44025 is the best representative of the cluster: it applies the output-tracing refactor across two closely related depth_anything variants and clearly shows the new can_return_tuple pattern.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 44025 is the strongest single stand-in for the group because it is a clean, substantive example of the refactor family without being limited to a single tiny code tweak.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Different model families (timm_backbone vs upernet); same refactor style only."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44000",
+            "accept": false,
+            "reason": "Different architectures and code paths; both are output-tracing cleanups, not the same change."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Model-specific refactors in unrelated models; not one underlying bug or fix."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "timm_backbone and mamba are unrelated code paths; shared theme is too broad."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Different models and implementations; cannot plausibly merge as one PR."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Different model families; same output-capturing pattern only."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "MobileNetV1 and UperNet touch different forward paths and model outputs."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "Different models and API surfaces; not the same concrete fix."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44026",
+            "accept": false,
+            "reason": "MobileNetV1 refactor is unrelated to vision_encoder_decoder's wrapper logic."
+          },
+          {
+            "left": "pull_request:43999",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "Different multimodal wrappers and code paths; not duplicates."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44001",
+            "accept": false,
+            "reason": "Both are refactors, but for different model implementations and forward signatures."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "Different model subsystems; same broad cleanup theme only."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "vision_text_dual_encoder and falcon_mamba do not share the same underlying change."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "Different model families and forward paths; not mergeable into one PR."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44002",
+            "accept": false,
+            "reason": "UnivNet and UperNet are unrelated model code paths."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "Same refactor pattern, but different model-specific behavior."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44025",
+            "accept": false,
+            "reason": "Different models and output contracts; not the same fix."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Different architectures and output plumbing; too broad to treat as duplicate."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44003",
+            "accept": false,
+            "reason": "UperNet vs mamba are distinct code paths with no shared concrete bug."
+          },
+          {
+            "left": "pull_request:44003",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Different model families and implementations; only the refactor motif matches."
+          },
+          {
+            "left": "pull_request:44003",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "mamba and codegen changes are unrelated forward-path edits."
+          },
+          {
+            "left": "pull_request:43998",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Different models and distinct output handling; not the same underlying issue."
+          },
+          {
+            "left": "pull_request:44001",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "UnivNet and CodeGen are separate code paths; same cleanup theme only."
+          },
+          {
+            "left": "pull_request:44002",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "UperNet and CodeGen are unrelated; cannot be merged as one bug fix."
+          },
+          {
+            "left": "pull_request:44000",
+            "right": "pull_request:44004",
+            "accept": false,
+            "reason": "Different model types and APIs; shared title is not enough for duplication."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44026",
+            "accept": false,
+            "reason": "Depth Anything and Vision-Encoder-Decoder are different model wrappers and fixes."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "Different encoder-decoder stacks; same refactor pattern only."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Depth Anything and SuperPoint are unrelated model paths."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Different model architectures; not the same concrete change."
+          },
+          {
+            "left": "pull_request:44025",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Depth Anything and DPR do different things and touch different forward logic."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Vision encoder-decoder and RWKV are unrelated implementations."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44027",
+            "accept": false,
+            "reason": "Vision and speech encoder-decoder wrappers are separate code paths."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Different models and different output-tracing behavior; not duplicates."
+          },
+          {
+            "left": "pull_request:44026",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Distinct model families; same broad cleanup is insufficient."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44028",
+            "accept": false,
+            "reason": "Speech encoder-decoder and SuperPoint are unrelated model code paths."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "Different architectures and forward signatures; not the same fix."
+          },
+          {
+            "left": "pull_request:44027",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Speech encoder-decoder vs DPR are separate changes with no shared code path."
+          },
+          {
+            "left": "pull_request:44028",
+            "right": "pull_request:44029",
+            "accept": false,
+            "reason": "SuperPoint and RWKV are unrelated; only the refactor theme overlaps."
+          },
+          {
+            "left": "pull_request:44028",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "Different model families and behavior; not one underlying bug."
+          },
+          {
+            "left": "pull_request:44029",
+            "right": "pull_request:44030",
+            "accept": false,
+            "reason": "RWKV and DPR are distinct implementations; same output-tracing label 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": 16,
+      "component_count": 37,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-44662-2",
+      "nodes": [
+        "pull_request:44662",
+        "pull_request:45401"
+      ],
+      "soft_pairs": [
+        "pull_request:44662|pull_request:45401"
+      ],
+      "prepared_review_unit_hash": "6b32b9128918ddf7d9cb847d10f46607950449b9c83aa54de04f35b6fc2b9417",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6551,
+        "estimated_input_tokens": 1638,
+        "estimated_eval_tokens": 3532
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 6551,
+        "estimated_input_tokens": 1638,
+        "estimated_eval_tokens": 3532
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are two separate model-integration PRs: one adds PenguinVL, the other adds Voxtral TTS. They share framework registration and docs plumbing, but they target different models and code paths, so they are not duplicates.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45401 is the more canonical representative of its own change because it has an explicit issue target (#45027) and a tightly scoped Voxtral TTS implementation.",
+        "best_issue_reason": null,
+        "best_pr_reason": "If one PR must be picked as the better representative, 45401 is more clearly anchored and narrowly scoped; 44662 is a separate PenguinVL feature, not the same change.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44662",
+            "right": "pull_request:45401",
+            "accept": false,
+            "reason": "Different model additions with different implementation targets (PenguinVL vs Voxtral TTS). Shared auto-registration/docs files are boilerplate and do not indicate the same underlying bug or feature."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 17,
+      "component_count": 37,
+      "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": "a511a5642e3e0221792a75ac56dcffb5d3668018a395d428bbd33304dbc83ae1",
+      "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": "The two PRs are related around module/kernel fusion, but they are not the same underlying change: #44979 introduces a general Module Fusion API, while #45363 adds a specific KernelConfig-based n-to-1 kernel fusion integration and example. This is a relationship/dependency, not a duplicate.",
+        "confidence": 0.88,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "#44979 is the foundational API PR: it introduces the generic module fusion abstraction itself, while #45363 appears to build on that idea with a specific kernel-fusion integration.",
+        "best_issue_reason": null,
+        "best_pr_reason": "#45363 is the more concrete end-to-end user-facing feature PR, with integration points and an example, so it is the stronger representative of the applied fusion workflow.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:44979",
+            "right": "pull_request:45363",
+            "accept": false,
+            "reason": "Different changes and codepaths: one adds a generic fusion API module, the other wires KernelConfig into model loading/integrations for a specific n-to-1 fusion use case. They are related, but not the same bug or mergeable duplicate PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 18,
+      "component_count": 37,
+      "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": "adb29ae74101b09d97da167fc060cc09ce7c23b18e00e9445b0f1ea1a5b10d2b",
+      "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 implement the same Mistral regex fix in the same code path; 45317 is the more complete version because it includes tests.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45317 is the better canonical representative: it fixes the same AttributeError in `_patch_mistral_regex` and also adds test coverage, making it the more complete change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45317 is the best overall artifact because it contains the same fix as 45086 plus validation in tests, so it is more suitable as the cluster representative.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45086",
+            "right": "pull_request:45317",
+            "accept": true,
+            "reason": "Same concrete bug fix in `src/transformers/tokenization_utils_tokenizers.py` for `_patch_mistral_regex`; the second PR is the same patch with added tests, so they are merge-equivalent."
+          }
+        ]
+      },
+      "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": 37,
+      "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-45115-4",
+      "nodes": [
+        "pull_request:45115",
+        "pull_request:45569",
+        "pull_request:46332",
+        "pull_request:46565"
+      ],
+      "soft_pairs": [
+        "pull_request:45115|pull_request:45569",
+        "pull_request:45569|pull_request:46565",
+        "pull_request:46332|pull_request:46565"
+      ],
+      "prepared_review_unit_hash": "f5074be12dac8269bc9d8afdfa31b3875aa37baff9d3bfbced568153e124b56c",
+      "component_budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 3,
+        "serialized_chars": 13472,
+        "estimated_input_tokens": 3368,
+        "estimated_eval_tokens": 6992
+      },
+      "budget": {
+        "node_count": 4,
+        "item_count": 4,
+        "soft_pair_count": 3,
+        "serialized_chars": 13472,
+        "estimated_input_tokens": 3368,
+        "estimated_eval_tokens": 6992
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This is a mixed cluster: two NemotronH architecture/model-split PRs and two ASR streaming PRs. The overlaps are mostly broad subsystem similarity (auto mapping, docs, model scaffolding), not the same underlying change.",
+        "confidence": 0.86,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 46565 is the newest and broadest ASR-streaming implementation in the set, so it is the most reasonable single representative, though the cluster is not a true duplicate group.",
+        "best_issue_reason": null,
+        "best_pr_reason": "46565 is the strongest standalone representative because it has the clearest end-to-end feature scope and the most direct overlap with one other PR in the cluster.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45115",
+            "right": "pull_request:45569",
+            "accept": false,
+            "reason": "Both concern NemotronH, but one is a checkpoint/key-mapping refactor and the other is a larger model-split/backward-compat implementation; related area, not the same concrete change."
+          },
+          {
+            "left": "pull_request:45569",
+            "right": "pull_request:46565",
+            "accept": false,
+            "reason": "They touch different model families entirely: NemotronH vs Nemotron ASR. Shared auto/docs files are too broad to treat as the same fix."
+          },
+          {
+            "left": "pull_request:46332",
+            "right": "pull_request:46565",
+            "accept": false,
+            "reason": "Both are ASR streaming PRs, but they target different model versions (Nemotron ASR vs Nemotron 3.5 ASR) with separate model files and support code, so they are not the same underlying change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 20,
+      "component_count": 37,
+      "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": "3dd5c3885e559939a4f8932a99d2070ba877476223cb5b80a587256e4d5ad801",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5466,
+        "estimated_input_tokens": 1367,
+        "estimated_eval_tokens": 2990
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5466,
+        "estimated_input_tokens": 1367,
+        "estimated_eval_tokens": 2990
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are related feature PRs for the same heterogeneous-model initiative, but they are not duplicates: one adds per-layer config support, while the other extends that work into model/cache/masking integration and broader heterogeneous modeling behavior.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45332 is the broader umbrella change and includes the downstream modeling/cache integration built on top of heterogeneous configuration support.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45332 is the best representative for the cluster because it spans the full heterogeneous-model feature set rather than just the config layer.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45332",
+            "right": "pull_request:45333",
+            "accept": false,
+            "reason": "Related feature work, but not the same concrete change: 45333 is focused on introducing per-layer configuration primitives, while 45332 builds on that to add heterogeneous modeling and cache/masking integration. They could be sequenced, not merged as one PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 21,
+      "component_count": 37,
+      "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": "acf5393e3c86341c3874e329fbd5bdb37600cd232763b738006c09b12526f8ab",
+      "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 the same underlying change: one is a broad repo-wide type-checking sweep, the other adds typing helpers and fixes in core internals.",
+        "confidence": 0.9,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No canonical PR: 45415 is a wide type-checking cleanup across many files, while 45425 introduces typing protocol/alias helpers and a smaller set of modeling changes; they solve different problems.",
+        "best_issue_reason": null,
+        "best_pr_reason": "45425 is the more coherent, self-contained typing-infrastructure change, but it is not a duplicate of the broader type-checking sweep in 45415.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45415",
+            "right": "pull_request:45425",
+            "accept": false,
+            "reason": "Overlapping type-related edits and shared files are incidental; 45415 is a broad type-checking pass, while 45425 adds specific typing abstractions and targeted fixes, so they are not mergeable 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": 22,
+      "component_count": 37,
+      "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-45513-2",
+      "nodes": [
+        "pull_request:45513",
+        "pull_request:46084"
+      ],
+      "soft_pairs": [
+        "pull_request:45513|pull_request:46084"
+      ],
+      "prepared_review_unit_hash": "10bd2eb9b92c2c559d7b934ba8c4a2f7f98288795caeb38d12d9f80290f1fb81",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4903,
+        "estimated_input_tokens": 1226,
+        "estimated_eval_tokens": 2708
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4903,
+        "estimated_input_tokens": 1226,
+        "estimated_eval_tokens": 2708
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are both PRs about cached forward paths with sequence length > 1, but they target different model families and different concrete bugs. They should not be deduplicated.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #46084 is the better canonical PR for the cluster because it has an explicit issue target, a focused fix in one model, and the change is tightly scoped to a concrete cache-handling bug.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #46084 is the strongest standalone PR: it directly fixes a specific use_cache/seq_len>1 failure mode, references a tracking issue, and has a narrow, testable code change.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45513",
+            "right": "pull_request:46084",
+            "accept": false,
+            "reason": "Different underlying problems in different architectures (Qwen3.5/OLMo hybrid linear attention vs Mamba2 state caching). They share a broad theme of cached multi-token forward bugs, but not the same code path or the same concrete fix."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 23,
+      "component_count": 37,
+      "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": "3730dcf06d58f19c49b830c89974bdd7b8423660362b289e4a6fde764838646a",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4508,
+        "estimated_input_tokens": 1127,
+        "estimated_eval_tokens": 2510
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4508,
+        "estimated_input_tokens": 1127,
+        "estimated_eval_tokens": 2510
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "The two PRs are both about expert parallelism, but they fix different bugs in different loading paths: one is about preserving distributed_config/EP state during Accelerate + ZeRO-3 loading, the other is about preventing expert weights from being overwritten during FSDP2/rank-0 broadcast. They are related, but not duplicates.",
+        "confidence": 0.95,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45662 is the better representative for this cluster because it targets a concrete EP weight-loading corruption path with a direct fix in model loading and related MoE/tensor-parallel logic.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45662 is the strongest standalone change: it addresses a specific, user-visible correctness bug and the diff shows a direct fix in the affected code path rather than only plumbing state through configuration.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45548",
+            "right": "pull_request:45662",
+            "accept": false,
+            "reason": "Both mention EP/loading, but #45548 fixes serialization and EP detection for Accelerate/ZeRO-3, while #45662 fixes expert parameters being overwritten during FSDP2 broadcast. Different underlying bugs and not mergeable 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": 24,
+      "component_count": 37,
+      "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": "ac2e9d880f243c401e87dc90d31cda7b1d6d7a00acfd22096e790345335f5641",
+      "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 address the same xdist-related captured_info collision problem around `testing_utils.py`; PR 45645 is the broader end-to-end fix, while 45639 is a narrower subset.",
+        "confidence": 0.9,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 45645 is the best canonical PR because it contains the core xdist-safe output-file fix plus the CI/workflow changes needed to consume multiple `captured_info*.txt` files, making it the more complete representation of the bug fix.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 45645 is the strongest representative of the cluster: it fixes the same concrete collision bug and also updates the reporting path that reads the artifacts, so it is more comprehensive than 45639.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45639",
+            "right": "pull_request:45645",
+            "accept": true,
+            "reason": "Accept: both target the same captured_info/xdist collision in the testing debug logs; 45645 extends the same fix with CI artifact handling, so they are the same underlying change family."
+          }
+        ]
+      },
+      "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": 37,
+      "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-45670-3",
+      "nodes": [
+        "pull_request:45670",
+        "pull_request:46320",
+        "pull_request:46401"
+      ],
+      "soft_pairs": [
+        "pull_request:45670|pull_request:46320",
+        "pull_request:45670|pull_request:46401",
+        "pull_request:46320|pull_request:46401"
+      ],
+      "prepared_review_unit_hash": "93b17b824c62ebd199d4eef9c8feb6540a860498328ff704c2cace5aeeda6c58",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 7525,
+        "estimated_input_tokens": 1882,
+        "estimated_eval_tokens": 4020
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 7525,
+        "estimated_input_tokens": 1882,
+        "estimated_eval_tokens": 4020
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These PRs are not duplicates. One is an AutoModel mapping nit for GLM-ASR, while the other two add attention-implementation dispatch support to different model families (RecurrentGemma vs BLIP-2/InstructBLIP/InstructBLIPVideo). They share a broad theme but not the same concrete code path or mergeable change.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 46401 is the strongest representative of the attn_implementation work because it is the broadest concrete implementation across multiple related BLIP-family models, with corresponding model and test changes. It is still not a duplicate of the others, just the most substantial of the three.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 46401 is the best single representative if one had to pick a canonical PR from this set, because it contains the most complete implementation and test coverage for a specific feature area. PR 46320 is similar in theme but targets a different model family; PR 45670 is unrelated.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45670",
+            "right": "pull_request:46320",
+            "accept": false,
+            "reason": "Different changes: AutoModel registry addition for glmasr vs RecurrentGemma attn_implementation dispatch. No shared concrete bug or mergeable code path."
+          },
+          {
+            "left": "pull_request:45670",
+            "right": "pull_request:46401",
+            "accept": false,
+            "reason": "Different scope and subsystem: a model mapping tweak vs BLIP-family attention dispatch support. Same broad repo area only, not the same fix."
+          },
+          {
+            "left": "pull_request:46320",
+            "right": "pull_request:46401",
+            "accept": false,
+            "reason": "Both add attn_implementation support, but for different model families and distinct files/tests. They are similar feature work, not the same underlying change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 26,
+      "component_count": 37,
+      "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-45720-2",
+      "nodes": [
+        "pull_request:45720",
+        "pull_request:46259"
+      ],
+      "soft_pairs": [
+        "pull_request:45720|pull_request:46259"
+      ],
+      "prepared_review_unit_hash": "5b50eedc7805034b983560b5f86f3ce89227faff1ae76451d45ef7f629ebb8bf",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 7248,
+        "estimated_input_tokens": 1812,
+        "estimated_eval_tokens": 3880
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 7248,
+        "estimated_input_tokens": 1812,
+        "estimated_eval_tokens": 3880
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These two pull requests are not duplicates: they target different linter rules and different code-path changes. PR 45720 adds `@can_return_tuple` and removes explicit `return_dict` handling in several model `forward()` methods, while PR 46259 changes weight initialization logic for `_init_weights` across many model classes. The overlap in filenames is just broad model coverage, not the same fix.",
+        "confidence": 0.96,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "If a representative PR is needed, 46259 is the better canonical choice because its scope is clearer and more self-contained: it implements TRF018 via `_init_weights` refactoring across model implementations. 45720 is also a legitimate standalone PR, but it is only a partial implementation of TRF003 and has a different code change pattern.",
+        "best_issue_reason": null,
+        "best_pr_reason": "46259 is the better overall representative of the cluster because its change intent is explicit in the title and diff, and it affects a consistent code path across models. However, it should not be merged with 45720; they fix different problems.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45720",
+            "right": "pull_request:46259",
+            "accept": false,
+            "reason": "Different underlying changes: 45720 modifies forward-return behavior with `can_return_tuple`/`return_dict`, while 46259 rewrites `_init_weights` initialization logic for a different linter rule. Shared model filenames do not indicate the same bug or mergeable PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 27,
+      "component_count": 37,
+      "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": 28,
+      "component_count": 37,
+      "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-45932-2",
+      "nodes": [
+        "pull_request:45932",
+        "pull_request:46100"
+      ],
+      "soft_pairs": [
+        "pull_request:45932|pull_request:46100"
+      ],
+      "prepared_review_unit_hash": "99b0a9d2b637e281d581e833ce2fe9a3547155c49c2cbf3ee8a5ccdb70f7aba8",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5083,
+        "estimated_input_tokens": 1271,
+        "estimated_eval_tokens": 2798
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5083,
+        "estimated_input_tokens": 1271,
+        "estimated_eval_tokens": 2798
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are two unrelated open pull requests that only look similar at a high level: one is a code/CI change across model files and workflow secrets, while the other is a ViT test migration. They do not appear to fix the same concrete problem or belong in one merged PR.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR #45932 is the better canonical PR because it is the broader, more substantive code change and has clear review activity; #46100 is a narrower test-only migration and does not represent the same change stream.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #45932 is the best representative of this cluster because it touches the actual implementation/workflow surface and is more central than the test migration PR.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:45932",
+            "right": "pull_request:46100",
+            "accept": false,
+            "reason": "Reject: the PRs address different things. #45932 changes model code and CI workflow secrets; #46100 adds/migrates ViT tests. Shared broad area (transformers/ViT) is not enough to treat them as the same underlying fix."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 29,
+      "component_count": 37,
+      "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-46073-3",
+      "nodes": [
+        "pull_request:46073",
+        "pull_request:46154",
+        "pull_request:46196"
+      ],
+      "soft_pairs": [
+        "pull_request:46073|pull_request:46154",
+        "pull_request:46073|pull_request:46196",
+        "pull_request:46154|pull_request:46196"
+      ],
+      "prepared_review_unit_hash": "915790d9ac68beec66a591ea0c343c6e6a1e3ce6bbfe4917078156c1945e3d62",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 8398,
+        "estimated_input_tokens": 2100,
+        "estimated_eval_tokens": 4456
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 3,
+        "serialized_chars": 8398,
+        "estimated_input_tokens": 2100,
+        "estimated_eval_tokens": 4456
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are three separate model-specific PRs enabling SDPA/attn_implementation dispatch in different architectures. They share the same feature theme, but they are not the same concrete change and should not be deduplicated.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "Use PR #46196 as the representative because it is the broadest multi-model SDPA dispatch effort, but it is not a duplicate of the CTRL or T5 PRs.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR #46196 is the best cluster representative: it covers the largest set of affected models and includes the same SDPA-dispatch feature family as the others.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46073",
+            "right": "pull_request:46154",
+            "accept": false,
+            "reason": "Both add attention-dispatch support, but they target different model code paths (CTRL vs T5). They are related feature work, not the same fix and would not be merged into one PR."
+          },
+          {
+            "left": "pull_request:46073",
+            "right": "pull_request:46196",
+            "accept": false,
+            "reason": "CTRL and wav2vec2-family SDPA dispatch are separate model-specific implementations with different files and execution paths. Same general capability, not the same underlying change."
+          },
+          {
+            "left": "pull_request:46154",
+            "right": "pull_request:46196",
+            "accept": false,
+            "reason": "T5 and wav2vec2-family dispatch support affect different architectures and code paths. They share the SDPA theme but are not duplicate PRs."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 30,
+      "component_count": 37,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46126-2",
+      "nodes": [
+        "pull_request:46126",
+        "pull_request:46269"
+      ],
+      "soft_pairs": [
+        "pull_request:46126|pull_request:46269"
+      ],
+      "prepared_review_unit_hash": "ba72bee9426b79b545b3014aa8501a4715f73c0057c470c720e26de0757c6ee4",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5572,
+        "estimated_input_tokens": 1393,
+        "estimated_eval_tokens": 3042
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5572,
+        "estimated_input_tokens": 1393,
+        "estimated_eval_tokens": 3042
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs live in the distributed-training area, but they are not the same change: #46126 is a focused EP/Trainer integration feature, while #46269 is a broad distributed branch base touching many subsystems. The overlap in files is not enough to treat them as duplicates.",
+        "confidence": 0.9,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "#46126 is the more focused, concrete change with a specific code-path goal (EP routing/trainer integration). #46269 is a broad base branch spanning many distributed components and is too diffuse to serve as the canonical duplicate target.",
+        "best_issue_reason": null,
+        "best_pr_reason": "#46126 is the best representative of a single, coherent PR because it targets one feature area and has a clear implementation surface. #46269 is too large and cross-cutting to be a good canonical example.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46126",
+            "right": "pull_request:46269",
+            "accept": false,
+            "reason": "Shared distributed filenames do not indicate the same underlying bug/change. One PR adds EP/Trainer integration, the other is a much broader distributed branch base; they would not plausibly be merged as one PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 31,
+      "component_count": 37,
+      "review_unit_index": 1,
+      "review_unit_count": 1,
+      "status": "reviewed",
+      "reason": null,
+      "source": "llm",
+      "cache_hit": false,
+      "model": "gpt-5.4-mini?service_tier=flex",
+      "cluster_id": "cluster-46138-9",
+      "nodes": [
+        "pull_request:46138",
+        "pull_request:46432",
+        "pull_request:46501",
+        "pull_request:46502",
+        "pull_request:46503",
+        "pull_request:46504",
+        "pull_request:46554",
+        "pull_request:46558",
+        "pull_request:46563"
+      ],
+      "soft_pairs": [
+        "pull_request:46138|pull_request:46554",
+        "pull_request:46138|pull_request:46558",
+        "pull_request:46138|pull_request:46563",
+        "pull_request:46138|pull_request:46504",
+        "pull_request:46138|pull_request:46432",
+        "pull_request:46432|pull_request:46503",
+        "pull_request:46432|pull_request:46501",
+        "pull_request:46432|pull_request:46502",
+        "pull_request:46432|pull_request:46554",
+        "pull_request:46432|pull_request:46558",
+        "pull_request:46501|pull_request:46503",
+        "pull_request:46501|pull_request:46502",
+        "pull_request:46501|pull_request:46504",
+        "pull_request:46501|pull_request:46554",
+        "pull_request:46502|pull_request:46503",
+        "pull_request:46502|pull_request:46504",
+        "pull_request:46502|pull_request:46554",
+        "pull_request:46503|pull_request:46504",
+        "pull_request:46503|pull_request:46554",
+        "pull_request:46504|pull_request:46554",
+        "pull_request:46504|pull_request:46558",
+        "pull_request:46554|pull_request:46558",
+        "pull_request:46554|pull_request:46563",
+        "pull_request:46558|pull_request:46563",
+        "pull_request:46504|pull_request:46563",
+        "pull_request:46432|pull_request:46563"
+      ],
+      "prepared_review_unit_hash": "3b2d1962ea007ba4571cf50fc9dac02842627cd26d45c0ac44f982aa42a9df89",
+      "component_budget": {
+        "node_count": 9,
+        "item_count": 9,
+        "soft_pair_count": 26,
+        "serialized_chars": 23365,
+        "estimated_input_tokens": 5842,
+        "estimated_eval_tokens": 11940
+      },
+      "budget": {
+        "node_count": 9,
+        "item_count": 9,
+        "soft_pair_count": 26,
+        "serialized_chars": 23365,
+        "estimated_input_tokens": 5842,
+        "estimated_eval_tokens": 11940
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "This cluster is mostly unrelated workflow-only PRs. Even the three pr-ci-security-gate changes touch different concerns (security checks, permissions/approval flow, and comments), so none should be merged as duplicates.",
+        "confidence": 0.91,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "No single canonical PR stands out: the cluster spans several independent workflow edits. If forced to pick a representative, 46554 is the most substantive pr-ci-security-gate change, but it is not a duplicate of the others.",
+        "best_issue_reason": null,
+        "best_pr_reason": "46554 is the best representative only because it has the clearest, self-contained security-gate logic change; however, the surrounding PRs are separate workflow tweaks rather than one underlying fix.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Different workflow files and different changes: self-comment CI message plumbing vs pr-ci-security-gate logic. Not the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46558",
+            "accept": false,
+            "reason": "Unrelated workflow updates in different files; one changes CI comment input handling, the other adjusts PR CI security-gate permissions/approval flow."
+          },
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Different file and purpose: self-comment CI invocation fix vs a comment-only reusable-workflow note in pr-ci-security-gate."
+          },
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46504",
+            "accept": false,
+            "reason": "Build CI docker image trigger handling is unrelated to self-comment CI stdin handling; not the same code-path problem."
+          },
+          {
+            "left": "pull_request:46138",
+            "right": "pull_request:46432",
+            "accept": false,
+            "reason": "Different workflows and different concerns: self-comment CI messaging vs ssh-runner hardening/secret handling."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46503",
+            "accept": false,
+            "reason": "Both are workflow edits, but one is SSH runner hardening and the other adds workflow_dispatch to benchmark_v2; not the same fix."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46501",
+            "accept": false,
+            "reason": "Different workflow files and different changes: SSH runner secret handling vs anti-slop permission placement."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46502",
+            "accept": false,
+            "reason": "SSH runner hardening and assign-reviewers checkout ref selection solve unrelated problems."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Security/hardening theme is too broad; ssh-runner changes are about remote-run safety, while 46554 changes PR CI security-gate behavior."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46558",
+            "accept": false,
+            "reason": "Different workflows and different fixes: SSH runner environment sanitization vs PR CI security-gate permissions and gating."
+          },
+          {
+            "left": "pull_request:46501",
+            "right": "pull_request:46503",
+            "accept": false,
+            "reason": "Anti-slop permission scoping and benchmark_v2 dispatch support are unrelated workflow changes."
+          },
+          {
+            "left": "pull_request:46501",
+            "right": "pull_request:46502",
+            "accept": false,
+            "reason": "Different workflows and different purposes: anti-slop permissions vs assign-reviewers checkout base ref."
+          },
+          {
+            "left": "pull_request:46501",
+            "right": "pull_request:46504",
+            "accept": false,
+            "reason": "Anti-slop workflow permission change is unrelated to build-ci-docker-images trigger parsing."
+          },
+          {
+            "left": "pull_request:46501",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Same broad CI area, but not the same bug/change: anti-slop permissions vs PR CI security-gate logic."
+          },
+          {
+            "left": "pull_request:46502",
+            "right": "pull_request:46503",
+            "accept": false,
+            "reason": "Assign-reviewers base-ref checkout and benchmark_v2 workflow dispatch are unrelated."
+          },
+          {
+            "left": "pull_request:46502",
+            "right": "pull_request:46504",
+            "accept": false,
+            "reason": "Different workflows and separate fixes: reviewer assignment checkout behavior vs build-ci-docker-images trigger detection."
+          },
+          {
+            "left": "pull_request:46502",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "No shared concrete issue: reviewer assignment checkout ref vs PR CI security-gate checks."
+          },
+          {
+            "left": "pull_request:46503",
+            "right": "pull_request:46504",
+            "accept": false,
+            "reason": "Benchmark_v2 dispatch support and build-ci-docker-images commit-message handling are separate workflow changes."
+          },
+          {
+            "left": "pull_request:46503",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Benchmark execution triggers are unrelated to PR CI security-gate logic."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46554",
+            "accept": false,
+            "reason": "Different workflows and different code paths: build-ci-docker-images commit-message parsing vs PR CI security-gate checks."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46558",
+            "accept": false,
+            "reason": "Build-image tag trigger parsing is unrelated to PR CI security-gate permissions/approval changes."
+          },
+          {
+            "left": "pull_request:46554",
+            "right": "pull_request:46558",
+            "accept": false,
+            "reason": "Same file, but different fixes: 46554 adds checkout/diff-based security checks, while 46558 changes permissions and approval gating. Not the same underlying bug/change."
+          },
+          {
+            "left": "pull_request:46554",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "46554 changes the security-check logic; 46563 only adds explanatory comments. These are not duplicate fixes."
+          },
+          {
+            "left": "pull_request:46558",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Both touch the same workflow file, but one alters permissions/approval flow and the other adds comments only; not a duplicate code-path fix."
+          },
+          {
+            "left": "pull_request:46504",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "Build CI docker image workflow update is unrelated to pr-ci-security-gate comment-only changes."
+          },
+          {
+            "left": "pull_request:46432",
+            "right": "pull_request:46563",
+            "accept": false,
+            "reason": "SSH runner hardening is unrelated to pr-ci-security-gate comments; different workflow and different problem."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 32,
+      "component_count": 37,
+      "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-46152-2",
+      "nodes": [
+        "pull_request:46152",
+        "pull_request:46207"
+      ],
+      "soft_pairs": [
+        "pull_request:46152|pull_request:46207"
+      ],
+      "prepared_review_unit_hash": "d3d2394f36e8bdee3dc84fe328af0f3d75c217bdf76da2c39dd47d9929fb5e2f",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4276,
+        "estimated_input_tokens": 1069,
+        "estimated_eval_tokens": 2394
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4276,
+        "estimated_input_tokens": 1069,
+        "estimated_eval_tokens": 2394
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These PRs are thematically similar because both replace brittle file-open source inspection, but they touch different modules and different code paths, so they are not duplicates.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": null,
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 46207 is the better representative of the cluster theme: it makes a concrete source-inspection fix in _can_set_attn_implementation and includes tests.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46152",
+            "right": "pull_request:46207",
+            "accept": false,
+            "reason": "Different fixes in different modules: one changes import_utils resource lookup, the other changes modeling_utils source inspection for attention implementation. Same broad idea, not the same underlying bug or mergeable change."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 33,
+      "component_count": 37,
+      "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-46173-3",
+      "nodes": [
+        "pull_request:46173",
+        "pull_request:46436",
+        "pull_request:46474"
+      ],
+      "soft_pairs": [
+        "pull_request:46436|pull_request:46474",
+        "pull_request:46173|pull_request:46474"
+      ],
+      "prepared_review_unit_hash": "c963914ae25d9de77ab1282269240ca8a06d8ae74812b6346043918c4dceffe2",
+      "component_budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 5921,
+        "estimated_input_tokens": 1481,
+        "estimated_eval_tokens": 3218
+      },
+      "budget": {
+        "node_count": 3,
+        "item_count": 3,
+        "soft_pair_count": 2,
+        "serialized_chars": 5921,
+        "estimated_input_tokens": 1481,
+        "estimated_eval_tokens": 3218
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "PR 46173 and PR 46474 are duplicates: both make the same StaticLayer.get_seq_length type-hint change in cache_utils.py for issue #45987. PR 46436 is unrelated, fixing static-cache max length handling in generation/utils.py for a different issue.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 46173 is the clean canonical choice: it is the earliest exact-match fix for StaticLayer.get_seq_length on the same file and issue target (#45987).",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 46173 best represents the duplicate cluster because it matches PR 46474 almost exactly and is directly scoped to the type-hint fix, unlike PR 46436 which addresses a different cache-length bug.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46436",
+            "right": "pull_request:46474",
+            "accept": false,
+            "reason": "Different bugs and code paths: 46436 changes static cache length preparation in generation/utils.py, while 46474 only adjusts StaticLayer.get_seq_length annotations in cache_utils.py."
+          },
+          {
+            "left": "pull_request:46173",
+            "right": "pull_request:46474",
+            "accept": true,
+            "reason": "Same underlying change: both PRs make the same StaticLayer.get_seq_length return-type update in cache_utils.py and share the same issue target (#45987)."
+          }
+        ]
+      },
+      "evaluator_result": {
+        "accept": true,
+        "feedback": "The summary is grounded in the packet: 46173 and 46474 have the same file, same explicit issue target, and identical diff for the StaticLayer.get_seq_length type-hint change. The 46436 vs 46474 verdict is also conservative, since they touch different files and target different issues. No evidence is overstated."
+      },
+      "evaluator_used": true,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": true,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 34,
+      "component_count": 37,
+      "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-46236-2",
+      "nodes": [
+        "pull_request:46236",
+        "pull_request:46621"
+      ],
+      "soft_pairs": [
+        "pull_request:46236|pull_request:46621"
+      ],
+      "prepared_review_unit_hash": "1275997b08282b88072e1cf0eaf5289d1798efd657d385bf53d4923c4131137f",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5922,
+        "estimated_input_tokens": 1481,
+        "estimated_eval_tokens": 3218
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5922,
+        "estimated_input_tokens": 1481,
+        "estimated_eval_tokens": 3218
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are two unrelated documentation translation PRs: one adds Bengali docs for a large set of pages, the other adds a Chinese translation for a single kernels page. They do not describe the same underlying change and should not be deduplicated.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 46236 is the better canonical PR for this cluster because it is the larger, issue-linked localization change and more representative of the translation work here.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 46236 is the strongest standalone PR in the set due to its explicit issue target, broader scope, and clearer project context; PR 46621 is a separate single-page translation.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46236",
+            "right": "pull_request:46621",
+            "accept": false,
+            "reason": "Different languages and different documentation content: Bengali translation batch vs. Chinese kernels page. No shared files, targets, or concrete code/doc change to 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": 35,
+      "component_count": 37,
+      "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-46310-2",
+      "nodes": [
+        "pull_request:46310",
+        "pull_request:46546"
+      ],
+      "soft_pairs": [
+        "pull_request:46310|pull_request:46546"
+      ],
+      "prepared_review_unit_hash": "d1d49e2ea32c41cdfc2b5524b8ad388dc948a5c313086b79419f8cbe98ff694f",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4850,
+        "estimated_input_tokens": 1213,
+        "estimated_eval_tokens": 2682
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 4850,
+        "estimated_input_tokens": 1213,
+        "estimated_eval_tokens": 2682
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Both PRs fix the same off-by-one bug in the Moonshine streaming sliding-window mask: the right-side boundary uses `<` instead of `<=`. PR 46546 is the broader superset, adding the same Moonshine fix plus Gemma4 and tests; PR 46310 is the narrower Moonshine-only patch.",
+        "confidence": 0.97,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 46546 is the better canonical representative because it contains the Moonshine boundary fix, related Gemma4 adjustment, and a test update, making it the more complete change set.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 46546 is the strongest overall fit: it addresses the same concrete off-by-one boundary bug as 46310, but with broader coverage and validation; 46310 is a subset of the same fix.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46310",
+            "right": "pull_request:46546",
+            "accept": true,
+            "reason": "Same underlying off-by-one bug in the Moonshine streaming right-window mask; the code change is identical on the shared files, and 46546 includes the broader follow-up/test coverage."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 36,
+      "component_count": 37,
+      "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-46376-2",
+      "nodes": [
+        "pull_request:46376",
+        "pull_request:46417"
+      ],
+      "soft_pairs": [
+        "pull_request:46376|pull_request:46417"
+      ],
+      "prepared_review_unit_hash": "b3ef4fb8a2aa6a0b77df8c82e0818c539736a1246e9a4b7cb174e22d9f36bf16",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5939,
+        "estimated_input_tokens": 1485,
+        "estimated_eval_tokens": 3226
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5939,
+        "estimated_input_tokens": 1485,
+        "estimated_eval_tokens": 3226
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "These are not duplicates: one PR adds multimodal support to continuous batching infrastructure, while the other adds the HunYuan VL model, its docs, and tests. The soft similarity is superficial and not enough to merge them.",
+        "confidence": 0.98,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 46417 is the better representative of the cluster because it is a more complete, self-contained feature addition with model code, docs, and tests; PR 46376 is a draft infrastructure change and is unrelated in scope.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 46417 is the strongest standalone PR: it has a full implementation surface and clearer deliverables, whereas PR 46376 is draft-stage and addresses a different subsystem.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46376",
+            "right": "pull_request:46417",
+            "accept": false,
+            "reason": "Different underlying changes: continuous batching multimodal infrastructure vs. HunYuan VL model integration. No shared files or concrete code-path overlap, so this is not a duplicate."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    },
+    {
+      "label": "pull_request",
+      "component_index": 37,
+      "component_count": 37,
+      "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-46603-2",
+      "nodes": [
+        "pull_request:46603",
+        "pull_request:46604"
+      ],
+      "soft_pairs": [
+        "pull_request:46603|pull_request:46604"
+      ],
+      "prepared_review_unit_hash": "9d09d98b83e178507f3ee0af98294f159f6a76bdc1d280850edea00f4bb5eaaa",
+      "component_budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5495,
+        "estimated_input_tokens": 1374,
+        "estimated_eval_tokens": 3004
+      },
+      "budget": {
+        "node_count": 2,
+        "item_count": 2,
+        "soft_pair_count": 1,
+        "serialized_chars": 5495,
+        "estimated_input_tokens": 1374,
+        "estimated_eval_tokens": 3004
+      },
+      "overflow_policy": "truncate_then_skip",
+      "trimmed": false,
+      "aggressively_trimmed": false,
+      "split": false,
+      "analyst_result": {
+        "summary": "Two open Mistral PRs overlap only at the package/file level; they target different changes. PR 46603 is a refactor that moves `MistralConverter` into `integrations/mistral/`, while PR 46604 adds tekken tokenizer detection/conversion/save utilities. They should not be merged as duplicates.",
+        "confidence": 0.94,
+        "canonical_issue_reason": null,
+        "canonical_pr_reason": "PR 46604 is the better canonical PR because it delivers the concrete tokenizer feature work; PR 46603 is a narrower code move/refactor that is not the same change.",
+        "best_issue_reason": null,
+        "best_pr_reason": "PR 46604 best represents the cluster since it contains the substantive Mistral tokenizer behavior change, whereas PR 46603 is mainly structural packaging cleanup.",
+        "soft_edge_verdicts": [
+          {
+            "left": "pull_request:46603",
+            "right": "pull_request:46604",
+            "accept": false,
+            "reason": "Different underlying changes: one relocates `MistralConverter`, the other implements tekken tokenizer detection/conversion/save utilities. Shared Mistral files and subsystem are not enough to treat them as the same PR."
+          }
+        ]
+      },
+      "evaluator_result": null,
+      "evaluator_used": false,
+      "retried": false,
+      "accepted_nontrivial_soft_edge": false,
+      "error_kind": null,
+      "error_message": null
+    }
+  ]
+}
diff --git a/snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/manifest.json b/snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..e304ba913cd8f321ca2925f63e1f0faa8475e031
--- /dev/null
+++ b/snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/manifest.json
@@ -0,0 +1,14 @@
+{
+  "analysis_id": "hybrid-model-20260613t180123z",
+  "artifacts": {
+    "hybrid": "snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.json",
+    "hybrid_reviews": "snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.llm-reviews.json"
+  },
+  "channel": "canonical",
+  "model": null,
+  "published_at": "2026-06-13T18:06:05Z",
+  "repo": "huggingface/transformers",
+  "schema_version": 1,
+  "snapshot_id": "20260613T180123Z",
+  "variant": "hybrid"
+}
diff --git a/snapshots/20260613T180123Z/manifest.json b/snapshots/20260613T180123Z/manifest.json
index 4261823ea7512cb07da889f7523148879c3ed30c..a1218fb44a41c48d9fb9ffa280a0e822ce9f821d 100644
--- a/snapshots/20260613T180123Z/manifest.json
+++ b/snapshots/20260613T180123Z/manifest.json
@@ -32,6 +32,24 @@
     "timeline_events": 138
   },
   "extracted_at": "2026-06-13T18:01:23Z",
+  "published_analysis": {
+    "canonical_analysis_id": "hybrid-model-20260613t180123z",
+    "runs": {
+      "hybrid-model-20260613t180123z": {
+        "analysis_id": "hybrid-model-20260613t180123z",
+        "artifacts": {
+          "hybrid": "snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.json",
+          "hybrid_reviews": "snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/analysis-report-hybrid.llm-reviews.json"
+        },
+        "channel": "canonical",
+        "manifest_path": "snapshots/20260613T180123Z/analysis-runs/hybrid-model-20260613t180123z/manifest.json",
+        "model": null,
+        "published_at": "2026-06-13T18:06:05Z",
+        "variant": "hybrid"
+      }
+    },
+    "schema_version": 1
+  },
   "repo": "huggingface/transformers",
   "snapshot_id": "20260613T180123Z",
   "watermark": {