Publish Qwen3.8-27B distribution-fidelity artifact

Inferact-Qwen3.8-27B-NVFP4/compliance.json

@@ -1,8 +1,9 @@
 {
   "attribution": null,
   "candidate": "/media/fmodels2/Inferact/Qwen3.8-27B-NVFP4",
-  "candidate_weights_sha256": null,
+  "candidate_weights_sha256": "83cf20bf984d554a10ed908da7866ece8f248f9b6f45fb806154c5c6791f3697",
   "comparability_key": {
+    "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
     "context_length": 2048,
     "driver": "580.173.02",
     "gpu_names": [
@@ -12,19 +13,22 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 248044,
-    "laws_version": 9,
+    "kv_cache_dtype": "bfloat16",
+    "laws_version": 15,
     "model_runner_v2": false,
+    "numerics_digest": "251a9225b37415b91fc58d975b0670fc4b481e6f5a816b663573c42e161438f8",
     "reference_config_sha256": "191e0af232104ed8b65258cf3fb2b842e288008baca7633c11b82a1ac7203aab",
     "rows": 768,
     "score_from": 0,
     "stride": 2048,
+    "substituted_parameters": [],
     "suite_id": "qwen3.8-27b-fidelity-1024x2048-v1",
     "tensor_parallel_size": 1,
     "token_sha256": "9c935708bbffbf45d5baa2931fb4af7e7b22df1e041aa0b6f4ca44d3315e543b",
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T12:56:04.271735+00:00",
+  "evaluated_at": "2026-09-11T20:24:26.982179+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -52,13 +56,13 @@
       "title": "Real vocabulary"
     },
     {
-      "detail": "all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc",
+      "detail": "all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3",
       "law": 5,
       "status": "pass",
       "title": "Manifest binding"
     },
     {
-      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s)",
+      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s)",
       "law": 6,
       "status": "pass",
       "title": "Provenance"
@@ -70,13 +74,13 @@
       "title": "Storage integrity"
     },
     {
-      "detail": "trunk 0.13959233, deployed 0.13875295, delta -0.0008393859011140425",
+      "detail": "trunk 0.13975674, deployed 0.13893424, delta -0.0008225048750587216",
       "law": 8,
       "status": "pass",
       "title": "Head transparency"
     },
     {
-      "detail": "mean 0.13875295, median 0.04445935, max 29.81742477, 4 depth buckets",
+      "detail": "mean 0.13893424, median 0.04438159, max 28.27250290, 4 depth buckets",
       "law": 9,
       "status": "pass",
       "title": "Tail and depth disclosure"
@@ -103,38 +107,37 @@
       "detail": "reference declares no experts",
       "law": 14,
       "status": "not_applicable",
-      "title": "Component attribution"
+      "title": "Routed-model intervention"
     },
     {
-      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.32244942, strongest worked_math_reasoning at 0.07214179, spread 4.5x",
+      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.32376633, strongest worked_math_reasoning at 0.07241381, spread 4.5x",
       "law": 15,
       "status": "pass",
       "title": "Domain disclosure"
     },
     {
-      "approval": {
-        "approver": "Andy Kitzke",
-        "justification": "These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger.",
-        "timestamp": "2026-09-02T05:15:00Z"
-      },
-      "detail": "the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else",
+      "detail": "scored weights 83cf20bf984d554a as inspected",
       "law": 16,
-      "status": "override",
+      "status": "pass",
       "title": "Candidate weight binding"
     },
     {
-      "detail": "1 override(s), each fully attributed",
+      "detail": "every scored layer used the checkpoint's own quantization parameters",
+      "law": 17,
+      "status": "pass",
+      "title": "Substitution disclosure"
+    },
+    {
+      "detail": "no overrides claimed",
       "law": 13,
       "status": "pass",
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 9,
-  "mean_kld": 0.1387529468264112,
+  "laws_version": 15,
+  "mean_kld": 0.13893423537794009,
   "nondeterminism_floor": 0.0,
-  "overridden_laws": [
-    16
-  ],
+  "overridden_laws": [],
   "partition": "analysis",
   "program": "Local Inference Lab \u2014 Distribution Fidelity",
   "ranking_floor": null,
@@ -147,16 +150,16 @@
     "overall": {
       "deployed": {
         "contexts": 768,
-        "max_kld": 29.817424774169922,
-        "mean_kld": 0.1387529468264112,
-        "mean_ref_top1_prob": 0.6415440866166368,
-        "median_context_kld": 0.0948157276574599,
-        "median_context_p99": 0.760625958442688,
-        "p90_context_kld": 0.27379871622917323,
+        "max_kld": 28.272502899169922,
+        "mean_kld": 0.13893423537794009,
+        "mean_ref_top1_prob": 0.6415473983687051,
+        "median_context_kld": 0.09546938309119762,
+        "median_context_p99": 0.7544186115264893,
+        "p90_context_kld": 0.2722677825694095,
         "positions": 1572096,
-        "top1_agreement": 0.8597254875020355,
+        "top1_agreement": 0.859637070509689,
         "worst_context_id": 454,
-        "worst_context_kld": 1.9809056674647718
+        "worst_context_kld": 2.0724223042466337
       }
     },
     "primary": "deployed",
@@ -165,201 +168,201 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 29.817424774169922,
-            "mean_kld": 0.3224494154204839,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.2704242110574551,
-            "median_context_p99": 4.772159099578857,
-            "p90_context_kld": 0.6499366568065644,
+            "max_kld": 28.272502899169922,
+            "mean_kld": 0.3237663331062429,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.26837684090210884,
+            "median_context_p99": 4.754233360290527,
+            "p90_context_kld": 0.6971572786024132,
             "positions": 196512,
-            "top1_agreement": 0.8537646555935515,
+            "top1_agreement": 0.8548078488845465,
             "worst_context_id": 454,
-            "worst_context_kld": 1.9809056674647718
+            "worst_context_kld": 2.0724223042466337
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 2.323910394666347
+        "relative_to_run": 2.3303567491878994
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.021528244018555,
-            "mean_kld": 0.22299356229430725,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.15146792872386075,
-            "median_context_p99": 1.1896569728851318,
-            "p90_context_kld": 0.38827515649388716,
+            "max_kld": 8.725299835205078,
+            "mean_kld": 0.22251604001072897,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.15263639992821296,
+            "median_context_p99": 1.196325421333313,
+            "p90_context_kld": 0.387976897349159,
             "positions": 147384,
-            "top1_agreement": 0.79339684090539,
+            "top1_agreement": 0.7934511208815068,
             "worst_context_id": 905,
-            "worst_context_kld": 0.7735271078252318
+            "worst_context_kld": 0.7599954233707221
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.6071266765475363
+        "relative_to_run": 1.6015925765555403
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 10.653387069702148,
-            "mean_kld": 0.1370184207671435,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.12206410714064975,
-            "median_context_p99": 1.0921473503112793,
-            "p90_context_kld": 0.18849303508813178,
+            "max_kld": 13.785171508789062,
+            "mean_kld": 0.13692233186327699,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.12210456756146866,
+            "median_context_p99": 1.0520226955413818,
+            "p90_context_kld": 0.18501191849037915,
             "positions": 196512,
-            "top1_agreement": 0.8466862074580688,
+            "top1_agreement": 0.8468439586386582,
             "worst_context_id": 6,
-            "worst_context_kld": 0.5272718036607824
+            "worst_context_kld": 0.5222000123857532
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.9874991767818977
+        "relative_to_run": 0.9855190226570855
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 9.7470703125,
-            "mean_kld": 0.11774063579793023,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.11220843445138266,
-            "median_context_p99": 0.8903030753135681,
-            "p90_context_kld": 0.17592728118249437,
+            "max_kld": 9.485733032226562,
+            "mean_kld": 0.11789006132413633,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.1104174708454115,
+            "median_context_p99": 0.8959710001945496,
+            "p90_context_kld": 0.17768971710652043,
             "positions": 73692,
-            "top1_agreement": 0.8540275742278673,
-            "worst_context_id": 953,
-            "worst_context_kld": 0.19032780062111215
+            "top1_agreement": 0.8525213048906258,
+            "worst_context_id": 959,
+            "worst_context_kld": 0.185708233196314
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.848563136790394
+        "relative_to_run": 0.848531400510769
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.67590045928955,
-            "mean_kld": 0.11673723786276567,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.10286570023584779,
-            "median_context_p99": 0.7754952311515808,
-            "p90_context_kld": 0.21367443152790838,
+            "max_kld": 8.996603012084961,
+            "mean_kld": 0.11715024080914344,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.10216221687613633,
+            "median_context_p99": 0.7410627007484436,
+            "p90_context_kld": 0.21365635485502046,
             "positions": 147384,
-            "top1_agreement": 0.8564294631710362,
+            "top1_agreement": 0.8566669380665473,
             "worst_context_id": 275,
-            "worst_context_kld": 0.28412494747417477
+            "worst_context_kld": 0.2925384386270357
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.8413315935467044
+        "relative_to_run": 0.8432064313771327
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 9.671066284179688,
-            "mean_kld": 0.11331059360869399,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.09982347202097942,
-            "median_context_p99": 0.5975978970527649,
-            "p90_context_kld": 0.13298753666949875,
+            "max_kld": 10.53090763092041,
+            "mean_kld": 0.11299148107033156,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.10009519240273125,
+            "median_context_p99": 0.5999166369438171,
+            "p90_context_kld": 0.1339625543235338,
             "positions": 147384,
-            "top1_agreement": 0.8377571513868534,
+            "top1_agreement": 0.83713293166151,
             "worst_context_id": 443,
-            "worst_context_kld": 0.4831157392859844
+            "worst_context_kld": 0.47158011349891993
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.8166355828856937
+        "relative_to_run": 0.8132731343211631
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 11.367209434509277,
-            "mean_kld": 0.08946067960442923,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.07114210014218113,
-            "median_context_p99": 0.7475235462188721,
-            "p90_context_kld": 0.1406945040734226,
+            "max_kld": 16.9556884765625,
+            "mean_kld": 0.09007301542454976,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.07053174211201081,
+            "median_context_p99": 0.7423832416534424,
+            "p90_context_kld": 0.14488682669665146,
             "positions": 196512,
-            "top1_agreement": 0.9138220566682951,
-            "worst_context_id": 582,
-            "worst_context_kld": 0.49920624666515057
+            "top1_agreement": 0.9138882103891874,
+            "worst_context_id": 667,
+            "worst_context_kld": 0.4984387221219649
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.6447479613989766
+        "relative_to_run": 0.6483140399450568
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 10.58599853515625,
-            "mean_kld": 0.07836417196189645,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.06484397643575615,
-            "median_context_p99": 0.9599510431289673,
-            "p90_context_kld": 0.12365192357983938,
+            "max_kld": 9.473292350769043,
+            "mean_kld": 0.07786292927453185,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.06419700892962933,
+            "median_context_p99": 0.9900352954864502,
+            "p90_context_kld": 0.12212695387331837,
             "positions": 73692,
-            "top1_agreement": 0.8043749660750149,
+            "top1_agreement": 0.8038050263257884,
             "worst_context_id": 985,
-            "worst_context_kld": 0.26076892411661884
+            "worst_context_kld": 0.2582430189493058
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.5647748300433217
+        "relative_to_run": 0.5604301133030519
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 9.688684463500977,
-            "mean_kld": 0.07563291716412826,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.07650770511032035,
-            "median_context_p99": 0.5419639348983765,
-            "p90_context_kld": 0.09455487405505722,
+            "max_kld": 17.140230178833008,
+            "mean_kld": 0.07539770427345319,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.07524519797729494,
+            "median_context_p99": 0.5095043778419495,
+            "p90_context_kld": 0.0943307534521116,
             "positions": 196512,
-            "top1_agreement": 0.8805620013027194,
+            "top1_agreement": 0.8794119443087445,
             "worst_context_id": 246,
-            "worst_context_kld": 0.1169571107310768
+            "worst_context_kld": 0.11872880759564826
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.5450905288429648
+        "relative_to_run": 0.5426862865610501
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.191375732421875,
-            "mean_kld": 0.07214179342084462,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.06596337549559317,
-            "median_context_p99": 0.4456246495246887,
-            "p90_context_kld": 0.08905786780333834,
+            "max_kld": 15.153870582580566,
+            "mean_kld": 0.07241380546384427,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.06545848173091594,
+            "median_context_p99": 0.4294142723083496,
+            "p90_context_kld": 0.09084108125371913,
             "positions": 196512,
-            "top1_agreement": 0.8953804347826086,
+            "top1_agreement": 0.8955839846930468,
             "worst_context_id": 825,
-            "worst_context_kld": 0.21149815280834836
+            "worst_context_kld": 0.2163537839192812
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.5199298110122202
+        "relative_to_run": 0.521209227278348
       }
     ]
   },

Inferact-Qwen3.8-27B-NVFP4/inspect.json

@@ -1,31 +1,37 @@
 {
+  "model": "/media/fmodels2/Inferact/Qwen3.8-27B-NVFP4",
+  "inspect_version": 5,
+  "weights_sha256": "83cf20bf984d554a10ed908da7866ece8f248f9b6f45fb806154c5c6791f3697",
+  "weights_bytes": 26381249312,
+  "quant_method": "modelopt",
+  "declared": {},
+  "detected_scheme": "nvfp4",
+  "detected_block": 16,
+  "quant_algorithm": "modelopt",
+  "scheme_mix": null,
+  "kv_cache_scheme": null,
   "coverage": {
-    "attention": {
-      "quantized": 64,
-      "weights": 68
-    },
-    "dense_mlp": {
-      "quantized": 192,
-      "weights": 195
+    "router": {
+      "weights": 0,
+      "quantized": 0
     },
     "experts": {
-      "quantized": 0,
-      "weights": 0
-    },
-    "router": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
     },
     "shared_expert": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
+    },
+    "attention": {
+      "weights": 68,
+      "quantized": 64
+    },
+    "dense_mlp": {
+      "weights": 195,
+      "quantized": 192
     }
   },
-  "declared": {},
-  "detected_block": null,
-  "detected_scheme": "fp8_per_tensor",
-  "model": "/media/fmodels2/Inferact/Qwen3.8-27B-NVFP4",
-  "quant_method": "modelopt",
-  "weights_bytes": 26381249312,
-  "weights_bytes_source": "hub"
+  "unloadable_reason": null,
+  "quantized_names_sha256": "eccbe240ef5caa18a896c8a311d4a9b4257c68423f0d8f9ae1c982d53797e657"
 }

Inferact-Qwen3.8-27B-NVFP4/report.md

@@ -1,12 +1,12 @@
 # Qwen3.8-27B / Inferact-Qwen3.8-27B-NVFP4: distribution fidelity
 
-**Mean KLD(reference || candidate) = 0.13875295** over 1572096 scored positions.
+**Mean KLD(reference || candidate) = 0.13893424** over 1572096 scored positions.
 
 | Mean | Median | p90 | p99 | Max | Top-1 agreement |
 |---|---|---|---|---|---|
-| 0.13875295 | 0.04445935 | 0.24611130 | 1.76721573 | 29.81742477 | 85.9725% |
+| 0.13893424 | 0.04438159 | 0.24611974 | 1.76796739 | 28.27250290 | 85.9637% |
 
-Reverse direction, KLD(candidate || reference): 0.15276902.
+Reverse direction, KLD(candidate || reference): 0.15297289.
 
 ## Identity
 
@@ -15,10 +15,10 @@ Reverse direction, KLD(candidate || reference): 0.15276902.
 | Reference checkpoint | /media/fmodels2/Qwen/Qwen3.8-27B |
 | Reference config SHA-256 | 191e0af23210 |
 | Candidate checkpoint | /media/fmodels2/Inferact/Qwen3.8-27B-NVFP4 |
-| Candidate weights SHA-256 | unbound (Law 16 gap) |
+| Candidate weights SHA-256 | 83cf20bf984d554a |
 | Suite | qwen3.8-27b-fidelity-1024x2048-v1 |
 | Suite token SHA-256 | 9c935708bbffbf45 |
-| Capture manifest SHA-256 | 304794b4e35b326d |
+| Capture manifest SHA-256 | dbfacc9cbb6d3e06 |
 | Tokenizer | None |
 | Scored vocabulary | 248044 |
 | Declared vocabulary | 248320 |
@@ -27,32 +27,48 @@ Reverse direction, KLD(candidate || reference): 0.15276902.
 | Model runner | V1 |
 | Tensor parallel | 1 |
 | Eager enforced | True |
+| KV cache | bfloat16 |
 | Prefix caching | False |
 | max_num_seqs | 1 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45 |
+| vLLM commit | 60071d1ab732 |
+| vLLM dirty digest | e3b0c44298fc1c14 |
+| Numerics digest | 251a9225b37415b9 |
+| Compiled extensions | f2fbc7537b0f01f6 |
+| FlashInfer | 0.6.17 |
 | torch | 2.13.0+cu132 |
 | Driver | 580.173.02 |
 | GPUs | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
-| Laws version | 9 |
+| Laws version | 15 |
 | Partition | analysis |
 
+The commit says when this number was taken. The numerics digest says what took it: a hash of the runtime and the scorer, which moves only when code that can change a logit changes. Comparability is bounded by the digest, so a result stays current across commits that cannot reach a number, and two results carrying the same digest were computed by the same code whatever their commits say.
+
+## Expert kernels: declared against built
+
+| Property | Value |
+|---|---|
+| Declared for its experts | `4-bit float` |
+| Expert implementation built | n/a |
+| Expert kernel built | n/a |
+| Expert layers carrying an activation scale | n/a |
+
 ## Fidelity by domain
 
-The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
+The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `deployed` is QxQ (natural student routing). `bxq` is the teacher-ID counterfactual on the same weights. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
 
 | Domain | Contexts | Reference top-1 | Mean KLD | x run | Worst context |
 |---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.32244942 | 2.32 | 1.98090567 |
-| Chinese across several content types | 72 | 55.6% | 0.22299356 | 1.61 | 0.77352711 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.13701842 | 0.99 | 0.52727180 |
-| Other multilingual content | 36 | 64.5% | 0.11774064 | 0.85 | 0.19032780 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.11673724 | 0.84 | 0.28412495 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.11331059 | 0.82 | 0.48311574 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.08946068 | 0.64 | 0.49920625 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.07836417 | 0.56 | 0.26076892 |
-| Scientific and technical exposition | 96 | 60.3% | 0.07563292 | 0.55 | 0.11695711 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.07214179 | 0.52 | 0.21149815 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.32376633 | 2.33 | 2.07242230 |
+| Chinese across several content types | 72 | 55.6% | 0.22251604 | 1.60 | 0.75999542 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.13692233 | 0.99 | 0.52220001 |
+| Other multilingual content | 36 | 64.5% | 0.11789006 | 0.85 | 0.18570823 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.11715024 | 0.84 | 0.29253844 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.11299148 | 0.81 | 0.47158011 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.09007302 | 0.65 | 0.49843872 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.07786293 | 0.56 | 0.25824302 |
+| Scientific and technical exposition | 96 | 60.3% | 0.07539770 | 0.54 | 0.11872881 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.07241381 | 0.52 | 0.21635378 |
 
 **Natural dialogue, instruction following, and assistance** is this candidate's weakest domain and **Worked mathematics, science, and formal reasoning** its strongest, a spread of 4.5x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
 
@@ -60,9 +76,9 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Component | Value |
 |---|---|
-| Trunk (candidate hidden states, reference head) | 0.13959233 |
-| Deployed (candidate's own head) | 0.13875295 |
-| Head-associated delta (not additive) | -0.00083939 |
+| Trunk (candidate hidden states, reference head) | 0.13975674 |
+| Deployed (candidate's own head) | 0.13893424 |
+| Head-associated delta (not additive) | -0.00082250 |
 | Reference head | {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]} |
 | Candidate head | {'runtime': {'state': 'quantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedLinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedLinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'quantized'}]}, 'state': 'quantized', 'static': {'ignored': True, 'lm_head_dtypes': {'lm_head.weight': 'BF16', 'model.language_model.embed_tokens.weight': 'BF16'}, 'lm_head_keys': ['lm_head.weight'], 'output_weight_keys': ['lm_head.weight'], 'packed_keys': [], 'quant_method': 'modelopt', 'state': 'unquantized', 'tie_word_embeddings': False}} |
 
@@ -70,26 +86,26 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Position range | Positions | Mean KLD |
 |---|---|---|
-| 0–511 | 393216 | 0.12720580 |
-| 512–1023 | 393216 | 0.13076163 |
-| 1024–1535 | 393216 | 0.14287953 |
-| 1536–2046 | 392448 | 0.15419499 |
+| 0–511 | 393216 | 0.12704732 |
+| 512–1023 | 393216 | 0.13100820 |
+| 1024–1535 | 393216 | 0.14367785 |
+| 1536–2046 | 392448 | 0.15403306 |
 
 ## Error by reference confidence
 
 | Reference top-1 probability | Positions | Share | Mean KLD |
 |---|---|---|---|
-| [0.00, 0.25) | 240569 | 15.3% | 0.18162765 |
-| [0.25, 0.50) | 346390 | 22.0% | 0.19529914 |
-| [0.50, 0.75) | 273384 | 17.4% | 0.19340607 |
-| [0.75, 0.95) | 249939 | 15.9% | 0.14813560 |
-| [0.95, 1.00) | 461814 | 29.4% | 0.03657384 |
+| [0.00, 0.25) | 240522 | 15.3% | 0.18171434 |
+| [0.25, 0.50) | 346508 | 22.0% | 0.19638587 |
+| [0.50, 0.75) | 273204 | 17.4% | 0.19209288 |
+| [0.75, 0.95) | 250041 | 15.9% | 0.14889459 |
+| [0.95, 1.00) | 461821 | 29.4% | 0.03670705 |
 
 ## Top-K set agreement
 
 | K=1 | K=2 | K=3 | K=4 | K=5 |
 |---|---|---|---|---|
-| 85.9038% | 59.4774% | 35.6480% | 19.1110% | 9.6474% |
+| 85.8959% | 59.4529% | 35.5793% | 19.0534% | 9.6315% |
 
 ## Law compliance
 
@@ -101,22 +117,19 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | 2 | Determinism | PASS | eager enforced; prefix caching off, max_num_seqs=1 |
 | 3 | Frozen input | PASS | token hash matches suite qwen3.8-27b-fidelity-1024x2048-v1 [analysis]: 9c935708bbffbf45 |
 | 4 | Real vocabulary | PASS | scored 248044 real tokens of 248320 declared (276 padding rows) |
-| 5 | Manifest binding | PASS | all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc |
-| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s) |
+| 5 | Manifest binding | PASS | all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3 |
+| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s) |
 | 7 | Storage integrity | PASS | hidden storage with bitwise-exact replay |
-| 8 | Head transparency | PASS | trunk 0.13959233, deployed 0.13875295, delta -0.0008393859011140425 |
-| 9 | Tail and depth disclosure | PASS | mean 0.13875295, median 0.04445935, max 29.81742477, 4 depth buckets |
+| 8 | Head transparency | PASS | trunk 0.13975674, deployed 0.13893424, delta -0.0008225048750587216 |
+| 9 | Tail and depth disclosure | PASS | mean 0.13893424, median 0.04438159, max 28.27250290, 4 depth buckets |
 | 10 | Comparability | PASS | comparability key fully resolved |
 | 11 | Freeze before qualification | NOT_APPLICABLE | partition is 'analysis' |
 | 12 | Reusable reference | PASS | suite, reference, head, and checksums present; published reference matches the scored capture on every bound field |
-| 14 | Component attribution | NOT_APPLICABLE | reference declares no experts |
-| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.32244942, strongest worked_math_reasoning at 0.07214179, spread 4.5x |
-| 16 | Candidate weight binding | OVERRIDE | the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else |
-| 13 | Recorded deviation | PASS | 1 override(s), each fully attributed |
-
-### Recorded deviations
-
-- **Law 16 (Candidate weight binding)** overridden by Andy Kitzke at 2026-09-02T05:15:00Z. Justification: These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger. Underlying finding: the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else.
+| 14 | Routed-model intervention | NOT_APPLICABLE | reference declares no experts |
+| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.32376633, strongest worked_math_reasoning at 0.07241381, spread 4.5x |
+| 16 | Candidate weight binding | PASS | scored weights 83cf20bf984d554a as inspected |
+| 17 | Substitution disclosure | PASS | every scored layer used the checkpoint's own quantization parameters |
+| 13 | Recorded deviation | PASS | no overrides claimed |
 
 ## Environment
 
@@ -126,7 +139,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | Platform | Linux-6.8.0-137-generic-x86_64-with-glibc2.39 |
 | Python | 3.12.3 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45fbe0b0205e70893bcd28fa928a88 |
+| vLLM commit | 60071d1ab73229321712254b1350dcaaac1c125a |
 | torch | 2.13.0+cu132 |
 | torch CUDA runtime | 13.2 |
 | cuDNN | 9.20.0 (92000) |
@@ -134,7 +147,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | CUDA arch list | sm_75, sm_80, sm_86, sm_90, sm_100, sm_120 |
 | nvcc | Cuda compilation tools, release 13.0, V13.0.88 |
 | gcc | gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0 |
-| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.8) 2.39 |
+| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.9) 2.39 |
 | NVIDIA driver | 580.173.02 |
 | float32 matmul precision | highest |
 | TF32 (matmul / cuDNN) | False / True |
@@ -157,8 +170,18 @@ Credential values are never published. Set but redacted: `HF_TOKEN`.
 
 | Variable | Value |
 |---|---|
+| `CUBLAS_WORKSPACE_CONFIG` | `:4096:8` |
 | `CUDA_HOME` | `/usr/local/cuda-13.0` |
 | `HF_TOKEN` | `<redacted>` |
+| `NCCL_DETERMINISTIC` | `1` |
+| `PYTORCH_NVML_BASED_CUDA_CHECK` | `1` |
+| `TORCHINDUCTOR_CACHE_DIR` | `/tmp/torchinductor_phaedawg` |
+| `TORCHINDUCTOR_COMPILE_THREADS` | `1` |
+| `TRITON_CACHE_AUTOTUNING` | `1` |
+| `TRITON_PTXAS_BLACKWELL_PATH` | `/usr/local/cuda-13.0/bin/ptxas` |
+| `VLLM_BATCH_INVARIANT` | `1` |
+| `VLLM_MARLIN_USE_ATOMIC_ADD` | `0` |
+| `VLLM_MOE_USE_DEEP_GEMM` | `0` |
 
 ## Files in this artifact
 
@@ -166,26 +189,26 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 
 | Path | Size | What it is |
 |---|---|---|
-| `Inferact-Qwen3.8-27B-NVFP4/report.md` | 12.36 KiB | This document. |
-| `Inferact-Qwen3.8-27B-NVFP4/report.json` | 252.83 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
-| `Inferact-Qwen3.8-27B-NVFP4/manifest.json` | 85.43 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
-| `Inferact-Qwen3.8-27B-NVFP4/compliance.json` | 12.99 KiB | The law-by-law receipt, including the comparability key. |
-| `baselines/self-kld.json` | 4.74 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
+| `Inferact-Qwen3.8-27B-NVFP4/report.md` | 13.20 KiB | This document. |
+| `Inferact-Qwen3.8-27B-NVFP4/report.json` | 255.33 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
+| `Inferact-Qwen3.8-27B-NVFP4/manifest.json` | 86.99 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
+| `Inferact-Qwen3.8-27B-NVFP4/compliance.json` | 12.25 KiB | The law-by-law receipt, including the comparability key. |
+| `baselines/self-kld.json` | 7.06 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
 | `suite/suite-manifest.json` | 382.68 KiB | The frozen evaluation input's identity: token hashes per context and per partition, sources, strata, and the analysis/qualification split. |
 | `suite/tokens` | 15.49 MiB in 1024 files | The token IDs themselves. These are the evaluation input, not a description of it; retokenizing source text does not reproduce them. |
 | `suite/sources.json` | 737.42 KiB | Per-context provenance: dataset, revision, licence, source unit, and the deterministic token offset chosen within the document. |
 | `suite/validation/capability-overlap.json` | 1.48 KiB | The benchmark-contamination scan and every document it blocked. |
-| `reference/manifest.json` | 85.43 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
+| `reference/manifest.json` | 86.99 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
 | `reference` | 19.25 GiB in 771 files | The reusable reference distributions. Pass this directory as `--reference-logits` to score a new candidate against the same reference without loading the reference checkpoint. |
 | `reference/lm_head.safetensors` | 2.37 GiB | The reference language-model head, which turns the stored hidden states back into reference logits. |
-| `environment/runtime.json` | 2.26 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
+| `environment/runtime.json` | 5.42 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
 | `environment/summary.md` | 1.18 KiB | The same provenance as prose, plus an index of every captured file. |
 | `environment/toolchain-nvcc.txt` | 243 B | `nvcc --version` verbatim; `toolchain-gcc.txt` and `toolchain-ldd.txt` sit beside it. |
 | `environment/gpu-smi-query.txt` | 54.73 KiB | `nvidia-smi -q` verbatim: ECC state, persistence mode, clocks, and throttle reasons, any of which can move a bitwise result. |
-| `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
-| `environment/models` | 7.22 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
-| `checksums.txt` | 182.99 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
+| `environment/pip-freeze.txt` | 4.47 KiB | Every installed package version in the scoring environment. |
+| `environment/models` | 19.46 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
+| `checksums.txt` | 183.22 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
+| `LAWS.md` | 40.20 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

Inferact-Qwen3.8-27B-NVFP4/strata.json

@@ -8,381 +8,381 @@
         "cells": {
           "deployed": {
             "contexts": 33,
-            "max_kld": 8.021528244018555,
-            "mean_kld": 0.3389235920971226,
-            "mean_ref_top1_prob": 0.6040606719329803,
-            "median_context_kld": 0.285559165801828,
-            "median_context_p99": 2.089630126953125,
-            "p90_context_kld": 0.6934425062474486,
+            "max_kld": 8.725299835205078,
+            "mean_kld": 0.33867063919541485,
+            "mean_ref_top1_prob": 0.6040861921682756,
+            "median_context_kld": 0.28253108540139005,
+            "median_context_p99": 1.908400297164917,
+            "p90_context_kld": 0.712074033997245,
             "positions": 67551,
-            "top1_agreement": 0.7561694127400038,
+            "top1_agreement": 0.75849358262646,
             "worst_context_id": 905,
-            "worst_context_kld": 0.7735271078252318
+            "worst_context_kld": 0.7599954233707221
           }
         },
         "key": "wikisource_zh",
         "label": "wikisource_zh",
-        "relative_to_run": 2.442640677903134
+        "relative_to_run": 2.4376327279891505
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 29.817424774169922,
-            "mean_kld": 0.3224494154204839,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.2704242110574551,
-            "median_context_p99": 4.772159099578857,
-            "p90_context_kld": 0.6499366568065644,
+            "max_kld": 28.272502899169922,
+            "mean_kld": 0.3237663331062429,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.26837684090210884,
+            "median_context_p99": 4.754233360290527,
+            "p90_context_kld": 0.6971572786024132,
             "positions": 196512,
-            "top1_agreement": 0.8537646555935515,
+            "top1_agreement": 0.8548078488845465,
             "worst_context_id": 454,
-            "worst_context_kld": 1.9809056674647718
+            "worst_context_kld": 2.0724223042466337
           }
         },
         "key": "wildchat",
         "label": "wildchat",
-        "relative_to_run": 2.323910394666347
+        "relative_to_run": 2.3303567491878994
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 9.7470703125,
-            "mean_kld": 0.1616994939671555,
-            "mean_ref_top1_prob": 0.7357366218042676,
-            "median_context_kld": 0.17592728118249437,
-            "median_context_p99": 2.627157688140869,
-            "p90_context_kld": 0.19032780062111215,
+            "max_kld": 9.485733032226562,
+            "mean_kld": 0.16177371106019364,
+            "mean_ref_top1_prob": 0.7358200961573914,
+            "median_context_kld": 0.17768971710652043,
+            "median_context_p99": 2.8594753742218018,
+            "p90_context_kld": 0.185708233196314,
             "positions": 14329,
-            "top1_agreement": 0.8487682322562635,
-            "worst_context_id": 953,
-            "worst_context_kld": 0.19032780062111215
+            "top1_agreement": 0.8498150603670878,
+            "worst_context_id": 959,
+            "worst_context_kld": 0.185708233196314
           }
         },
         "key": "wikipedia_de",
         "label": "wikipedia_de",
-        "relative_to_run": 1.165377007592148
+        "relative_to_run": 1.1643905522647011
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 8.67590045928955,
-            "mean_kld": 0.15381675863389638,
-            "mean_ref_top1_prob": 0.7232628548554616,
-            "median_context_kld": 0.1946865052886737,
-            "median_context_p99": 2.3454291820526123,
-            "p90_context_kld": 0.24028980404604577,
+            "max_kld": 8.996603012084961,
+            "mean_kld": 0.15446324511839943,
+            "mean_ref_top1_prob": 0.7233191001139795,
+            "median_context_kld": 0.1947756225639116,
+            "median_context_p99": 2.3370704650878906,
+            "p90_context_kld": 0.2414717674721937,
             "positions": 47081,
-            "top1_agreement": 0.8701599371296277,
+            "top1_agreement": 0.8723264161763769,
             "worst_context_id": 275,
-            "worst_context_kld": 0.28412494747417477
+            "worst_context_kld": 0.2925384386270357
           }
         },
         "key": "regulations",
         "label": "regulations",
-        "relative_to_run": 1.108565707266246
+        "relative_to_run": 1.1117723770402312
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 10.653387069702148,
-            "mean_kld": 0.1370184207671435,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.12206410714064975,
-            "median_context_p99": 1.0921473503112793,
-            "p90_context_kld": 0.18849303508813178,
+            "max_kld": 13.785171508789062,
+            "mean_kld": 0.13692233186327699,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.12210456756146866,
+            "median_context_p99": 1.0520226955413818,
+            "p90_context_kld": 0.18501191849037915,
             "positions": 196512,
-            "top1_agreement": 0.8466862074580688,
+            "top1_agreement": 0.8468439586386582,
             "worst_context_id": 6,
-            "worst_context_kld": 0.5272718036607824
+            "worst_context_kld": 0.5222000123857532
           }
         },
         "key": "wikipedia_en",
         "label": "wikipedia_en",
-        "relative_to_run": 0.9874991767818977
+        "relative_to_run": 0.9855190226570855
       },
       {
         "cells": {
           "deployed": {
             "contexts": 39,
-            "max_kld": 7.907846927642822,
-            "mean_kld": 0.12489892169192499,
-            "mean_ref_top1_prob": 0.515358929042611,
-            "median_context_kld": 0.12155689460254131,
-            "median_context_p99": 0.8025867938995361,
-            "p90_context_kld": 0.16630316278016485,
+            "max_kld": 6.134963035583496,
+            "mean_kld": 0.12423137916214863,
+            "mean_ref_top1_prob": 0.5153760832971095,
+            "median_context_kld": 0.12106671125027292,
+            "median_context_p99": 0.7975048422813416,
+            "p90_context_kld": 0.16468408417279906,
             "positions": 79833,
-            "top1_agreement": 0.8248969724299475,
+            "top1_agreement": 0.8230305763280849,
             "worst_context_id": 880,
-            "worst_context_kld": 0.19174844904134652
+            "worst_context_kld": 0.185623894076553
           }
         },
         "key": "wikipedia_zh",
         "label": "wikipedia_zh",
-        "relative_to_run": 0.9001532907851068
+        "relative_to_run": 0.8941739868809472
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 4.0597004890441895,
-            "mean_kld": 0.11922249983613585,
-            "mean_ref_top1_prob": 0.5702914711488575,
-            "median_context_kld": 0.11584520546788203,
-            "median_context_p99": 0.6899316906929016,
-            "p90_context_kld": 0.1401531873133934,
+            "max_kld": 3.892676591873169,
+            "mean_kld": 0.12022743159239367,
+            "mean_ref_top1_prob": 0.5701150598950927,
+            "median_context_kld": 0.11619470036624585,
+            "median_context_p99": 0.7556959986686707,
+            "p90_context_kld": 0.14259579128472266,
             "positions": 14329,
-            "top1_agreement": 0.8346709470304976,
+            "top1_agreement": 0.8362760834670947,
             "worst_context_id": 968,
-            "worst_context_kld": 0.1401531873133934
+            "worst_context_kld": 0.14259579128472266
           }
         },
         "key": "wikipedia_ja",
         "label": "wikipedia_ja",
-        "relative_to_run": 0.8592430111433295
+        "relative_to_run": 0.8653549736344058
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 9.671066284179688,
-            "mean_kld": 0.11331059360869399,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.09982347202097942,
-            "median_context_p99": 0.5975978970527649,
-            "p90_context_kld": 0.13298753666949875,
+            "max_kld": 10.53090763092041,
+            "mean_kld": 0.11299148107033156,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.10009519240273125,
+            "median_context_p99": 0.5999166369438171,
+            "p90_context_kld": 0.1339625543235338,
             "positions": 147384,
-            "top1_agreement": 0.8377571513868534,
+            "top1_agreement": 0.83713293166151,
             "worst_context_id": 443,
-            "worst_context_kld": 0.4831157392859844
+            "worst_context_kld": 0.47158011349891993
           }
         },
         "key": "public_domain_books",
         "label": "public_domain_books",
-        "relative_to_run": 0.8166355828856937
+        "relative_to_run": 0.8132731343211631
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 8.629904747009277,
-            "mean_kld": 0.11186995757553385,
-            "mean_ref_top1_prob": 0.591765751203999,
-            "median_context_kld": 0.11220843445138266,
-            "median_context_p99": 0.9299942255020142,
-            "p90_context_kld": 0.1281363217103897,
+            "max_kld": 7.5889129638671875,
+            "mean_kld": 0.11225478144981937,
+            "mean_ref_top1_prob": 0.591761624645417,
+            "median_context_kld": 0.10861866027083132,
+            "median_context_p99": 0.9978964328765869,
+            "p90_context_kld": 0.13055007740284663,
             "positions": 14329,
-            "top1_agreement": 0.8475120385232745,
+            "top1_agreement": 0.8422778979691534,
             "worst_context_id": 936,
-            "worst_context_kld": 0.1281363217103897
+            "worst_context_kld": 0.13055007740284663
           }
         },
         "key": "wikipedia_es",
         "label": "wikipedia_es",
-        "relative_to_run": 0.806252840997246
+        "relative_to_run": 0.8079706282937024
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 6.607489585876465,
-            "mean_kld": 0.11185847155608147,
-            "mean_ref_top1_prob": 0.6892386900239509,
-            "median_context_kld": 0.11088604138522333,
-            "median_context_p99": 0.8480619192123413,
-            "p90_context_kld": 0.1397956107460902,
+            "max_kld": 5.6331024169921875,
+            "mean_kld": 0.1109637072734158,
+            "mean_ref_top1_prob": 0.689239984531695,
+            "median_context_kld": 0.1104174708454115,
+            "median_context_p99": 0.99306720495224,
+            "p90_context_kld": 0.13728585527420786,
             "positions": 12282,
-            "top1_agreement": 0.866389838788471,
+            "top1_agreement": 0.8654942191825435,
             "worst_context_id": 958,
-            "worst_context_kld": 0.1397956107460902
+            "worst_context_kld": 0.13728585527420786
           }
         },
         "key": "wikipedia_cs",
         "label": "wikipedia_cs",
-        "relative_to_run": 0.8061700606331883
+        "relative_to_run": 0.7986779282411092
       },
       {
         "cells": {
           "deployed": {
             "contexts": 26,
-            "max_kld": 7.503324031829834,
-            "mean_kld": 0.1080877859518247,
-            "mean_ref_top1_prob": 0.5182113024080622,
-            "median_context_kld": 0.10286570023584779,
-            "median_context_p99": 0.7569081783294678,
-            "p90_context_kld": 0.15311613346491215,
+            "max_kld": 8.907048225402832,
+            "mean_kld": 0.10832642963679921,
+            "mean_ref_top1_prob": 0.5182496321630543,
+            "median_context_kld": 0.10216221687613633,
+            "median_context_p99": 0.7380863428115845,
+            "p90_context_kld": 0.156248984352194,
             "positions": 53222,
-            "top1_agreement": 0.845515012588779,
+            "top1_agreement": 0.8456653263687949,
             "worst_context_id": 312,
-            "worst_context_kld": 0.21855730135295334
+            "worst_context_kld": 0.21618519222618846
           }
         },
         "key": "public_domain_review",
         "label": "public_domain_review",
-        "relative_to_run": 0.7789945253346541
+        "relative_to_run": 0.7796957268460214
       },
       {
         "cells": {
           "deployed": {
             "contexts": 44,
-            "max_kld": 11.367209434509277,
-            "mean_kld": 0.09297266299875274,
-            "mean_ref_top1_prob": 0.720394649593742,
-            "median_context_kld": 0.07943809610231228,
-            "median_context_p99": 0.7475235462188721,
-            "p90_context_kld": 0.1406945040734226,
+            "max_kld": 11.36757755279541,
+            "mean_kld": 0.0934585211427134,
+            "mean_ref_top1_prob": 0.7203910827522294,
+            "median_context_kld": 0.07820657085516014,
+            "median_context_p99": 0.7423832416534424,
+            "p90_context_kld": 0.14488682669665146,
             "positions": 90068,
-            "top1_agreement": 0.897388639694453,
+            "top1_agreement": 0.8981991384287428,
             "worst_context_id": 582,
-            "worst_context_kld": 0.49920624666515057
+            "worst_context_kld": 0.49227887184721075
           }
         },
         "key": "stackv2",
         "label": "stackv2",
-        "relative_to_run": 0.6700590158641277
+        "relative_to_run": 0.6726817251952336
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 5.699172496795654,
-            "mean_kld": 0.09243902333424184,
-            "mean_ref_top1_prob": 0.6660615117323249,
-            "median_context_kld": 0.09578638517320973,
-            "median_context_p99": 0.7872180938720703,
-            "p90_context_kld": 0.11733642057703256,
+            "max_kld": 4.931934833526611,
+            "mean_kld": 0.0925989252299416,
+            "mean_ref_top1_prob": 0.6660535411394803,
+            "median_context_kld": 0.09808316004676107,
+            "median_context_p99": 0.8350603580474854,
+            "p90_context_kld": 0.1181526341377527,
             "positions": 12282,
-            "top1_agreement": 0.8737176355642403,
+            "top1_agreement": 0.8696466373554795,
             "worst_context_id": 945,
-            "worst_context_kld": 0.11733642057703256
+            "worst_context_kld": 0.1181526341377527
           }
         },
         "key": "wikipedia_ru",
         "label": "wikipedia_ru",
-        "relative_to_run": 0.6662130459102173
+        "relative_to_run": 0.666494654669143
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 6.9558820724487305,
-            "mean_kld": 0.08943535838226391,
-            "mean_ref_top1_prob": 0.5379292386352562,
-            "median_context_kld": 0.08905932993476644,
-            "median_context_p99": 0.5505937337875366,
-            "p90_context_kld": 0.11123378901814941,
+            "max_kld": 6.965224742889404,
+            "mean_kld": 0.0898119795642766,
+            "mean_ref_top1_prob": 0.537879873360319,
+            "median_context_kld": 0.0876688278041105,
+            "median_context_p99": 0.5209121704101562,
+            "p90_context_kld": 0.11554724911424986,
             "positions": 47081,
-            "top1_agreement": 0.8550370637836919,
+            "top1_agreement": 0.8534440644846116,
             "worst_context_id": 343,
-            "worst_context_kld": 0.14255780810163965
+            "worst_context_kld": 0.14500321168035077
           }
         },
         "key": "open_news",
         "label": "open_news",
-        "relative_to_run": 0.6445654699799152
+        "relative_to_run": 0.6464351951839863
       },
       {
         "cells": {
           "deployed": {
             "contexts": 3,
-            "max_kld": 4.1011199951171875,
-            "mean_kld": 0.0877780865772576,
-            "mean_ref_top1_prob": 0.6023603953518782,
-            "median_context_kld": 0.0831387116085892,
-            "median_context_p99": 0.548545777797699,
-            "p90_context_kld": 0.09831847005756625,
+            "max_kld": 5.12807035446167,
+            "mean_kld": 0.08762498131063912,
+            "mean_ref_top1_prob": 0.6023746847155801,
+            "median_context_kld": 0.08312882094316615,
+            "median_context_p99": 0.5453658699989319,
+            "p90_context_kld": 0.09901612002324117,
             "positions": 6141,
-            "top1_agreement": 0.8625631004722358,
+            "top1_agreement": 0.8604461814036802,
             "worst_context_id": 967,
-            "worst_context_kld": 0.09831847005756625
+            "worst_context_kld": 0.09901612002324117
           }
         },
         "key": "wikipedia_fr",
         "label": "wikipedia_fr",
-        "relative_to_run": 0.6326214223548966
+        "relative_to_run": 0.6306939471921704
       },
       {
         "cells": {
           "deployed": {
             "contexts": 52,
-            "max_kld": 11.20549201965332,
-            "mean_kld": 0.08648900134769394,
-            "mean_ref_top1_prob": 0.8582575125835092,
-            "median_context_kld": 0.05967764827901316,
-            "median_context_p99": 0.7507213354110718,
-            "p90_context_kld": 0.1339402104231225,
+            "max_kld": 16.9556884765625,
+            "mean_kld": 0.0872083567399498,
+            "mean_ref_top1_prob": 0.8582941084526002,
+            "median_context_kld": 0.06078220694907813,
+            "median_context_p99": 0.7528872489929199,
+            "p90_context_kld": 0.1405896036534089,
             "positions": 106444,
-            "top1_agreement": 0.9277272556461613,
+            "top1_agreement": 0.9271635789711021,
             "worst_context_id": 667,
-            "worst_context_kld": 0.48678355155357317
+            "worst_context_kld": 0.4984387221219649
           }
         },
         "key": "github_code",
         "label": "github_code",
-        "relative_to_run": 0.6233309153130795
+        "relative_to_run": 0.6276952293487534
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 10.58599853515625,
-            "mean_kld": 0.07836417196189645,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.06484397643575615,
-            "median_context_p99": 0.9599510431289673,
-            "p90_context_kld": 0.12365192357983938,
+            "max_kld": 9.473292350769043,
+            "mean_kld": 0.07786292927453185,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.06419700892962933,
+            "median_context_p99": 0.9900352954864502,
+            "p90_context_kld": 0.12212695387331837,
             "positions": 73692,
-            "top1_agreement": 0.8043749660750149,
+            "top1_agreement": 0.8038050263257884,
             "worst_context_id": 985,
-            "worst_context_kld": 0.26076892411661884
+            "worst_context_kld": 0.2582430189493058
           }
         },
         "key": "starcoder_structured",
         "label": "starcoder_structured",
-        "relative_to_run": 0.5647748300433217
+        "relative_to_run": 0.5604301133030519
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 9.688684463500977,
-            "mean_kld": 0.07563291716412826,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.07650770511032035,
-            "median_context_p99": 0.5419639348983765,
-            "p90_context_kld": 0.09455487405505722,
+            "max_kld": 17.140230178833008,
+            "mean_kld": 0.07539770427345319,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.07524519797729494,
+            "median_context_p99": 0.5095043778419495,
+            "p90_context_kld": 0.0943307534521116,
             "positions": 196512,
-            "top1_agreement": 0.8805620013027194,
+            "top1_agreement": 0.8794119443087445,
             "worst_context_id": 246,
-            "worst_context_kld": 0.1169571107310768
+            "worst_context_kld": 0.11872880759564826
           }
         },
         "key": "scientific_papers",
         "label": "scientific_papers",
-        "relative_to_run": 0.5450905288429648
+        "relative_to_run": 0.5426862865610501
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.191375732421875,
-            "mean_kld": 0.07214179342084462,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.06596337549559317,
-            "median_context_p99": 0.4456246495246887,
-            "p90_context_kld": 0.08905786780333834,
+            "max_kld": 15.153870582580566,
+            "mean_kld": 0.07241380546384427,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.06545848173091594,
+            "median_context_p99": 0.4294142723083496,
+            "p90_context_kld": 0.09084108125371913,
             "positions": 196512,
-            "top1_agreement": 0.8953804347826086,
+            "top1_agreement": 0.8955839846930468,
             "worst_context_id": 825,
-            "worst_context_kld": 0.21149815280834836
+            "worst_context_kld": 0.2163537839192812
           }
         },
         "key": "libretexts",
         "label": "libretexts",
-        "relative_to_run": 0.5199298110122202
+        "relative_to_run": 0.521209227278348
       }
     ],
     "stratum": [
@@ -390,217 +390,217 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 29.817424774169922,
-            "mean_kld": 0.3224494154204839,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.2704242110574551,
-            "median_context_p99": 4.772159099578857,
-            "p90_context_kld": 0.6499366568065644,
+            "max_kld": 28.272502899169922,
+            "mean_kld": 0.3237663331062429,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.26837684090210884,
+            "median_context_p99": 4.754233360290527,
+            "p90_context_kld": 0.6971572786024132,
             "positions": 196512,
-            "top1_agreement": 0.8537646555935515,
+            "top1_agreement": 0.8548078488845465,
             "worst_context_id": 454,
-            "worst_context_kld": 1.9809056674647718
+            "worst_context_kld": 2.0724223042466337
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 2.323910394666347
+        "relative_to_run": 2.3303567491878994
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.021528244018555,
-            "mean_kld": 0.22299356229430725,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.15146792872386075,
-            "median_context_p99": 1.1896569728851318,
-            "p90_context_kld": 0.38827515649388716,
+            "max_kld": 8.725299835205078,
+            "mean_kld": 0.22251604001072897,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.15263639992821296,
+            "median_context_p99": 1.196325421333313,
+            "p90_context_kld": 0.387976897349159,
             "positions": 147384,
-            "top1_agreement": 0.79339684090539,
+            "top1_agreement": 0.7934511208815068,
             "worst_context_id": 905,
-            "worst_context_kld": 0.7735271078252318
+            "worst_context_kld": 0.7599954233707221
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.6071266765475363
+        "relative_to_run": 1.6015925765555403
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 10.653387069702148,
-            "mean_kld": 0.1370184207671435,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.12206410714064975,
-            "median_context_p99": 1.0921473503112793,
-            "p90_context_kld": 0.18849303508813178,
+            "max_kld": 13.785171508789062,
+            "mean_kld": 0.13692233186327699,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.12210456756146866,
+            "median_context_p99": 1.0520226955413818,
+            "p90_context_kld": 0.18501191849037915,
             "positions": 196512,
-            "top1_agreement": 0.8466862074580688,
+            "top1_agreement": 0.8468439586386582,
             "worst_context_id": 6,
-            "worst_context_kld": 0.5272718036607824
+            "worst_context_kld": 0.5222000123857532
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.9874991767818977
+        "relative_to_run": 0.9855190226570855
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 9.7470703125,
-            "mean_kld": 0.11774063579793023,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.11220843445138266,
-            "median_context_p99": 0.8903030753135681,
-            "p90_context_kld": 0.17592728118249437,
+            "max_kld": 9.485733032226562,
+            "mean_kld": 0.11789006132413633,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.1104174708454115,
+            "median_context_p99": 0.8959710001945496,
+            "p90_context_kld": 0.17768971710652043,
             "positions": 73692,
-            "top1_agreement": 0.8540275742278673,
-            "worst_context_id": 953,
-            "worst_context_kld": 0.19032780062111215
+            "top1_agreement": 0.8525213048906258,
+            "worst_context_id": 959,
+            "worst_context_kld": 0.185708233196314
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.848563136790394
+        "relative_to_run": 0.848531400510769
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.67590045928955,
-            "mean_kld": 0.11673723786276567,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.10286570023584779,
-            "median_context_p99": 0.7754952311515808,
-            "p90_context_kld": 0.21367443152790838,
+            "max_kld": 8.996603012084961,
+            "mean_kld": 0.11715024080914344,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.10216221687613633,
+            "median_context_p99": 0.7410627007484436,
+            "p90_context_kld": 0.21365635485502046,
             "positions": 147384,
-            "top1_agreement": 0.8564294631710362,
+            "top1_agreement": 0.8566669380665473,
             "worst_context_id": 275,
-            "worst_context_kld": 0.28412494747417477
+            "worst_context_kld": 0.2925384386270357
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.8413315935467044
+        "relative_to_run": 0.8432064313771327
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 9.671066284179688,
-            "mean_kld": 0.11331059360869399,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.09982347202097942,
-            "median_context_p99": 0.5975978970527649,
-            "p90_context_kld": 0.13298753666949875,
+            "max_kld": 10.53090763092041,
+            "mean_kld": 0.11299148107033156,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.10009519240273125,
+            "median_context_p99": 0.5999166369438171,
+            "p90_context_kld": 0.1339625543235338,
             "positions": 147384,
-            "top1_agreement": 0.8377571513868534,
+            "top1_agreement": 0.83713293166151,
             "worst_context_id": 443,
-            "worst_context_kld": 0.4831157392859844
+            "worst_context_kld": 0.47158011349891993
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.8166355828856937
+        "relative_to_run": 0.8132731343211631
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 11.367209434509277,
-            "mean_kld": 0.08946067960442923,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.07114210014218113,
-            "median_context_p99": 0.7475235462188721,
-            "p90_context_kld": 0.1406945040734226,
+            "max_kld": 16.9556884765625,
+            "mean_kld": 0.09007301542454976,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.07053174211201081,
+            "median_context_p99": 0.7423832416534424,
+            "p90_context_kld": 0.14488682669665146,
             "positions": 196512,
-            "top1_agreement": 0.9138220566682951,
-            "worst_context_id": 582,
-            "worst_context_kld": 0.49920624666515057
+            "top1_agreement": 0.9138882103891874,
+            "worst_context_id": 667,
+            "worst_context_kld": 0.4984387221219649
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.6447479613989766
+        "relative_to_run": 0.6483140399450568
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 10.58599853515625,
-            "mean_kld": 0.07836417196189645,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.06484397643575615,
-            "median_context_p99": 0.9599510431289673,
-            "p90_context_kld": 0.12365192357983938,
+            "max_kld": 9.473292350769043,
+            "mean_kld": 0.07786292927453185,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.06419700892962933,
+            "median_context_p99": 0.9900352954864502,
+            "p90_context_kld": 0.12212695387331837,
             "positions": 73692,
-            "top1_agreement": 0.8043749660750149,
+            "top1_agreement": 0.8038050263257884,
             "worst_context_id": 985,
-            "worst_context_kld": 0.26076892411661884
+            "worst_context_kld": 0.2582430189493058
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.5647748300433217
+        "relative_to_run": 0.5604301133030519
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 9.688684463500977,
-            "mean_kld": 0.07563291716412826,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.07650770511032035,
-            "median_context_p99": 0.5419639348983765,
-            "p90_context_kld": 0.09455487405505722,
+            "max_kld": 17.140230178833008,
+            "mean_kld": 0.07539770427345319,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.07524519797729494,
+            "median_context_p99": 0.5095043778419495,
+            "p90_context_kld": 0.0943307534521116,
             "positions": 196512,
-            "top1_agreement": 0.8805620013027194,
+            "top1_agreement": 0.8794119443087445,
             "worst_context_id": 246,
-            "worst_context_kld": 0.1169571107310768
+            "worst_context_kld": 0.11872880759564826
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.5450905288429648
+        "relative_to_run": 0.5426862865610501
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.191375732421875,
-            "mean_kld": 0.07214179342084462,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.06596337549559317,
-            "median_context_p99": 0.4456246495246887,
-            "p90_context_kld": 0.08905786780333834,
+            "max_kld": 15.153870582580566,
+            "mean_kld": 0.07241380546384427,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.06545848173091594,
+            "median_context_p99": 0.4294142723083496,
+            "p90_context_kld": 0.09084108125371913,
             "positions": 196512,
-            "top1_agreement": 0.8953804347826086,
+            "top1_agreement": 0.8955839846930468,
             "worst_context_id": 825,
-            "worst_context_kld": 0.21149815280834836
+            "worst_context_kld": 0.2163537839192812
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.5199298110122202
+        "relative_to_run": 0.521209227278348
       }
     ]
   },
   "overall": {
     "deployed": {
       "contexts": 768,
-      "max_kld": 29.817424774169922,
-      "mean_kld": 0.1387529468264112,
-      "mean_ref_top1_prob": 0.6415440866166368,
-      "median_context_kld": 0.0948157276574599,
-      "median_context_p99": 0.760625958442688,
-      "p90_context_kld": 0.27379871622917323,
+      "max_kld": 28.272502899169922,
+      "mean_kld": 0.13893423537794009,
+      "mean_ref_top1_prob": 0.6415473983687051,
+      "median_context_kld": 0.09546938309119762,
+      "median_context_p99": 0.7544186115264893,
+      "p90_context_kld": 0.2722677825694095,
       "positions": 1572096,
-      "top1_agreement": 0.8597254875020355,
+      "top1_agreement": 0.859637070509689,
       "worst_context_id": 454,
-      "worst_context_kld": 1.9809056674647718
+      "worst_context_kld": 2.0724223042466337
     }
   },
   "primary": "deployed"

Inferact-Qwen3.8-27B-NVFP4/strata.md

@@ -1,52 +1,52 @@
 # Fidelity by domain - Qwen3.8-27B / Inferact-Qwen3.8-27B-NVFP4
 
-768 contexts, 1572096 scored positions, mean 0.13875295, reference top-1 64.2%, top-1 agreement 85.9725%.
+768 contexts, 1572096 scored positions, mean 0.13893424, reference top-1 64.2%, top-1 agreement 85.9637%.
 
-`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
+`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. The `deployed` cell is QxQ; `bxq` is the teacher-ID counterfactual on the same student weights. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
 
 ## Ranked by domain
 
 | Domain | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.32244942 | 2.32 | 0.27042421 | 0.64993666 | 1.980906 |
-| Chinese across several content types | 72 | 55.6% | 0.22299356 | 1.61 | 0.15146793 | 0.38827516 | 0.773527 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.13701842 | 0.99 | 0.12206411 | 0.18849304 | 0.527272 |
-| Other multilingual content | 36 | 64.5% | 0.11774064 | 0.85 | 0.11220843 | 0.17592728 | 0.190328 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.11673724 | 0.84 | 0.10286570 | 0.21367443 | 0.284125 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.11331059 | 0.82 | 0.09982347 | 0.13298754 | 0.483116 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.08946068 | 0.64 | 0.07114210 | 0.14069450 | 0.499206 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.07836417 | 0.56 | 0.06484398 | 0.12365192 | 0.260769 |
-| Scientific and technical exposition | 96 | 60.3% | 0.07563292 | 0.55 | 0.07650771 | 0.09455487 | 0.116957 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.07214179 | 0.52 | 0.06596338 | 0.08905787 | 0.211498 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.32376633 | 2.33 | 0.26837684 | 0.69715728 | 2.072422 |
+| Chinese across several content types | 72 | 55.6% | 0.22251604 | 1.60 | 0.15263640 | 0.38797690 | 0.759995 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.13692233 | 0.99 | 0.12210457 | 0.18501192 | 0.522200 |
+| Other multilingual content | 36 | 64.5% | 0.11789006 | 0.85 | 0.11041747 | 0.17768972 | 0.185708 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.11715024 | 0.84 | 0.10216222 | 0.21365635 | 0.292538 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.11299148 | 0.81 | 0.10009519 | 0.13396255 | 0.471580 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.09007302 | 0.65 | 0.07053174 | 0.14488683 | 0.498439 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.07786293 | 0.56 | 0.06419701 | 0.12212695 | 0.258243 |
+| Scientific and technical exposition | 96 | 60.3% | 0.07539770 | 0.54 | 0.07524520 | 0.09433075 | 0.118729 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.07241381 | 0.52 | 0.06545848 | 0.09084108 | 0.216354 |
 
 ## Ranked by source dataset
 
 | Source | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| wikisource_zh | 33 | 60.4% | 0.33892359 | 2.44 | 0.28555917 | 0.69344251 | 0.773527 |
-| wildchat | 96 | 73.5% | 0.32244942 | 2.32 | 0.27042421 | 0.64993666 | 1.980906 |
-| wikipedia_de | 7 | 73.6% | 0.16169949 | 1.17 | 0.17592728 | 0.19032780 | 0.190328 |
-| regulations | 23 | 72.3% | 0.15381676 | 1.11 | 0.19468651 | 0.24028980 | 0.284125 |
-| wikipedia_en | 96 | 59.5% | 0.13701842 | 0.99 | 0.12206411 | 0.18849304 | 0.527272 |
-| wikipedia_zh | 39 | 51.5% | 0.12489892 | 0.90 | 0.12155689 | 0.16630316 | 0.191748 |
-| wikipedia_ja | 7 | 57.0% | 0.11922250 | 0.86 | 0.11584521 | 0.14015319 | 0.140153 |
-| public_domain_books | 72 | 50.6% | 0.11331059 | 0.82 | 0.09982347 | 0.13298754 | 0.483116 |
-| wikipedia_es | 7 | 59.2% | 0.11186996 | 0.81 | 0.11220843 | 0.12813632 | 0.128136 |
-| wikipedia_cs | 6 | 68.9% | 0.11185847 | 0.81 | 0.11088604 | 0.13979561 | 0.139796 |
-| public_domain_review | 26 | 51.8% | 0.10808779 | 0.78 | 0.10286570 | 0.15311613 | 0.218557 |
-| stackv2 | 44 | 72.0% | 0.09297266 | 0.67 | 0.07943810 | 0.14069450 | 0.499206 |
-| wikipedia_ru | 6 | 66.6% | 0.09243902 | 0.67 | 0.09578639 | 0.11733642 | 0.117336 |
-| open_news | 23 | 53.8% | 0.08943536 | 0.64 | 0.08905933 | 0.11123379 | 0.142558 |
-| wikipedia_fr | 3 | 60.2% | 0.08777809 | 0.63 | 0.08313871 | 0.09831847 | 0.098318 |
-| github_code | 52 | 85.8% | 0.08648900 | 0.62 | 0.05967765 | 0.13394021 | 0.486784 |
-| starcoder_structured | 36 | 70.7% | 0.07836417 | 0.56 | 0.06484398 | 0.12365192 | 0.260769 |
-| scientific_papers | 96 | 60.3% | 0.07563292 | 0.55 | 0.07650771 | 0.09455487 | 0.116957 |
-| libretexts | 96 | 65.8% | 0.07214179 | 0.52 | 0.06596338 | 0.08905787 | 0.211498 |
+| wikisource_zh | 33 | 60.4% | 0.33867064 | 2.44 | 0.28253109 | 0.71207403 | 0.759995 |
+| wildchat | 96 | 73.5% | 0.32376633 | 2.33 | 0.26837684 | 0.69715728 | 2.072422 |
+| wikipedia_de | 7 | 73.6% | 0.16177371 | 1.16 | 0.17768972 | 0.18570823 | 0.185708 |
+| regulations | 23 | 72.3% | 0.15446325 | 1.11 | 0.19477562 | 0.24147177 | 0.292538 |
+| wikipedia_en | 96 | 59.5% | 0.13692233 | 0.99 | 0.12210457 | 0.18501192 | 0.522200 |
+| wikipedia_zh | 39 | 51.5% | 0.12423138 | 0.89 | 0.12106671 | 0.16468408 | 0.185624 |
+| wikipedia_ja | 7 | 57.0% | 0.12022743 | 0.87 | 0.11619470 | 0.14259579 | 0.142596 |
+| public_domain_books | 72 | 50.6% | 0.11299148 | 0.81 | 0.10009519 | 0.13396255 | 0.471580 |
+| wikipedia_es | 7 | 59.2% | 0.11225478 | 0.81 | 0.10861866 | 0.13055008 | 0.130550 |
+| wikipedia_cs | 6 | 68.9% | 0.11096371 | 0.80 | 0.11041747 | 0.13728586 | 0.137286 |
+| public_domain_review | 26 | 51.8% | 0.10832643 | 0.78 | 0.10216222 | 0.15624898 | 0.216185 |
+| stackv2 | 44 | 72.0% | 0.09345852 | 0.67 | 0.07820657 | 0.14488683 | 0.492279 |
+| wikipedia_ru | 6 | 66.6% | 0.09259893 | 0.67 | 0.09808316 | 0.11815263 | 0.118153 |
+| open_news | 23 | 53.8% | 0.08981198 | 0.65 | 0.08766883 | 0.11554725 | 0.145003 |
+| wikipedia_fr | 3 | 60.2% | 0.08762498 | 0.63 | 0.08312882 | 0.09901612 | 0.099016 |
+| github_code | 52 | 85.8% | 0.08720836 | 0.63 | 0.06078221 | 0.14058960 | 0.498439 |
+| starcoder_structured | 36 | 70.7% | 0.07786293 | 0.56 | 0.06419701 | 0.12212695 | 0.258243 |
+| scientific_papers | 96 | 60.3% | 0.07539770 | 0.54 | 0.07524520 | 0.09433075 | 0.118729 |
+| libretexts | 96 | 65.8% | 0.07241381 | 0.52 | 0.06545848 | 0.09084108 | 0.216354 |
 
 ## Reading
 
-- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.32244942, 2.32x the run mean of 0.13875295 over 96 context(s).
-- Strongest domain: **Worked mathematics, science, and formal reasoning** at 0.07214179, 0.52x the run mean. The spread across domains is 4.5x.
+- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.32376633, 2.33x the run mean of 0.13893424 over 96 context(s).
+- Strongest domain: **Worked mathematics, science, and formal reasoning** at 0.07241381, 0.52x the run mean. The spread across domains is 4.5x.
 - This is a strong finding: the reference was *more* certain on Natural dialogue, instruction following, and assistance (top-1 73.5% against 64.2% overall) and the candidate still diverged most there, so predictability does not explain it.
-- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 1.980906 against a median of 0.27042421 (context 454). Read the documents before treating the domain as weak.
+- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 2.072422 against a median of 0.26837684 (context 454). Read the documents before treating the domain as weak.
 

LAWS.md

@@ -1,6 +1,6 @@
 # Local Inference Lab — Distribution Fidelity Laws
 
-**Laws version:** 8
+**Laws version:** 15
 **Status:** draft, pending coordination with `local-inference-lab` on the
 publication namespace and suite format.
 
@@ -75,6 +75,83 @@ disagreement with the field, the position, and both values named. A version-8
 receipt from a single-worker run becomes a version-9 receipt by reassembling it;
 a multi-worker candidate that version 8 refused must be scored.
 
+Version 10 replaces Law 14's synthetic component-cell interpretation with a
+paired routed-model intervention. QxQ is the unchanged quantized candidate under
+its natural routing. BxQ is the same candidate with the BF16 teacher's expert IDs
+forced while the student computes its own gating weights for those experts.
+Earlier `expert_cell`, `router_cell`, and `composite_cell` results remain useful
+synthetic QDQ diagnostics, but none measured BxQ or QxQ. Routed candidates must
+therefore be rescored for BxQ; an existing deployed report may supply QxQ only
+after the paired-run controls and bindings are established.
+
+Version 11 replaces Law 14's fixed bitwise-style natural-control threshold with
+a measured repeatability envelope. The paired protocol runs two natural and two
+forced-natural samples, retains the fixed numerical floor for repeatable
+kernels, and otherwise permits no more than twice the larger observed
+within-path span. Natural expert IDs may differ across the two samples when
+arithmetic drift propagates into later routers, so their selection flip rate is
+recorded. Both forced-natural controls still replay the first QxQ sample's exact
+IDs.
+The paired routed-score and BxQ protocol versions are now 3. Version-10 paired
+reports must be rescored rather than relabeled.
+
+Version 12 restores Law 14's publication requirement to exact repeated QxQ and
+BxQ. Version 11's repeatability envelope remains diagnostic evidence of kernel
+drift, but it does not authorize a drifting canonical score. Publication
+requires two natural QxQ samples, two forced-natural controls, and two BxQ
+samples to agree exactly in routes and per-position KLD. Uncertified or
+nondeterministic backends fail. Version-11 paired reports must be rescored
+rather than relabeled. The paired routed-score and BxQ protocol versions are
+now 4.
+
+Version 13 adds Law 17. Versions 1 through 12 checked at length that a kernel
+computed repeatably and never asked whether it computed what the checkpoint
+exported, so a batch-invariant, fully certified run could substitute a
+quantization parameter and publish the result under the checkpoint's name. A
+version-12 report does not record what it substituted and cannot be relabeled as
+one that substituted nothing; the record comes from the loaded model during
+scoring. Version-12 paired reports must be rescored. The paired routed-score and
+BxQ protocol versions are now 5.
+
+Version 14 adds no law. It puts `kv_cache_dtype` into the comparability key
+(Law 10), the bound capture fields (Law 5), and the baseline agreement Law 1
+already required of the model runner, because scoring had left it at
+`auto`, under which vLLM resolves the KV cache dtype from a scheme declared in
+the candidate's own config. One candidate therefore ran its attention through an
+8-bit KV cache while the candidates it was ranked against ran unquantized, and
+nothing in versions 1 through 13 could see the difference. The KV cache is now
+pinned unquantized and never inherits a checkpoint's declaration. A version-13
+report cannot be relabeled: it does not record which cache it used, and one that
+used a quantized cache measured something else. Version-13 reports must be
+rescored, and their captures retaken, since the capture is bound to the cache it
+was taken under.
+
+Version 15 adds no law and changes what two of them bind to.
+
+Law 10's comparability key drops `vllm_commit` and `vllm_dirty_digest` for
+`numerics_digest`, a hash of every `.py` under `vllm/` together with the scorer.
+Binding to the commit meant a documentation edit, a campaign config, or a fix to
+the orchestration invalidated every published number and cost GPU-days to
+reproduce values that were already correct. The commit is still recorded and still
+published under Law 6, where it says when a number was taken; currency asks
+whether the code that computed it would compute it again, and the digest answers
+that. `compiled_extensions_sha256` continues to cover the built kernels, so a
+rebuild still invalidates. A capture is bound the same way: it is retaken when the
+digest or the kernels move, not when the commit does.
+
+Law 1 no longer requires the baseline and the candidate to cache in the *same*
+unquantized dtype, only that each is unquantized. Version 14 held a literal
+`bfloat16`, which refused three checkpoints published as float16 outright:
+FlashAttention will not take a float16 query against a bfloat16 key. Each side now
+caches in its own checkpoint's compute dtype, both dtypes are recorded, and the
+candidate's stays in the comparability key, so a difference is disclosed and never
+silently ranked across.
+
+A version-14 report has no `numerics_digest` and cannot be relabeled into one; a
+report that records no digest cannot be shown to have been computed by the code
+running now. Version-14 reports must be rescored. This is the last rescore either
+change forces: after it, a commit that cannot reach a number costs nothing.
+
 These laws govern every distribution-fidelity measurement this program
 publishes. They are not guidance. The pipeline refuses to produce or upload an
 artifact that violates one, and the only way past a refusal is a recorded,
@@ -140,7 +217,10 @@ override, ranking two candidates requires repeated candidate capture as well.
 **Required.** Eager execution enforced. No autotuned kernel selection, no
 inference-time JIT kernel selection, no CUDA graphs, prefix caching disabled,
 and a fixed `max_num_seqs`. Tensor-parallel size is recorded, and reference and
-candidate are scored under identical settings.
+candidate are scored under identical settings, with one exception: each caches in
+its own checkpoint's compute dtype, because a float16 checkpoint cannot read a
+bfloat16 key. Both must be unquantized, both are recorded, and the candidate's is
+in the comparability key.
 
 **Why.** Timing-based autotuners and JIT kernel choice make the arithmetic a
 function of machine load, which silently converts run-to-run noise into apparent
@@ -189,8 +269,8 @@ is strictly less than or equal to the checkpoint's declared `vocab_size`.
 
 **Required.** A reference capture binds itself to the tokenizer identity, token
 hash, context length, row count, `score_from`, scored vocabulary size, tensor
-parallel size, eager mode, and runtime identity. Scoring against a capture whose
-manifest disagrees with the live configuration aborts.
+parallel size, eager mode, KV cache dtype, and runtime identity. Scoring against
+a capture whose manifest disagrees with the live configuration aborts.
 
 **Why.** Reusing a capture across configurations is the easiest way to publish a
 number that compares two different things. The binding must fail closed, because
@@ -315,9 +395,14 @@ number is presented as a capability, accuracy, or general quality claim.
 depth distribution, the vocabulary width, and the runtime. Cross-harness
 comparison is the single easiest way to publish a confident falsehood.
 
-**Check.** Each result records its suite ID, geometry, laws version, and runtime
-manifest hash. The leaderboard groups strictly by that tuple and refuses to place
-rows from differing tuples in one ranking. The suite identity is read from what the
+**Check.** Each result records its suite ID, geometry, laws version, KV cache
+dtype, and runtime manifest hash. The leaderboard groups strictly by that tuple
+and refuses to place rows from differing tuples in one ranking. One field of the
+key is deliberately not a section boundary: a substituted quantization parameter
+describes the candidate rather than how the number was taken, so such a row is
+marked and footnoted in place. Splitting it out would leave a reader comparing
+exports of one model unable to see the two side by side, and the substitution
+remains in the receipt's key, which is what binds it (Law 17). The suite identity is read from what the
 scoring run recorded, never from a suite manifest supplied to the audit, or a run
 that tokenized at run time reports a complete key by borrowing the identity of a
 suite it never opened.
@@ -391,107 +476,78 @@ indistinguishable from a bug.
 
 **Override.** None. This law has no exceptions.
 
-## Law 14 — Component attribution on routed models
-
-**Required.** For a checkpoint that routes tokens to experts, a fidelity number
-does not publish as a single mean. The artifact carries, measured on the same
-tokens against the same reference and at the deployed checkpoint's own scheme and
-granularity:
-
-1. the **expert cell** — the reference with only its expert weights rounded
-   through the deployed scheme;
-2. the **measured routing divergence** — the candidate's own expert selections
-   compared against the reference's over the same frozen tokens;
-3. the **router weight cell** — the reference with only its router rounded, or
-   `not_applicable` with inspection evidence when the deployed checkpoint leaves
-   the router unquantized;
-4. the **deployed cell** — the candidate as it ships.
-
-**Why.** Routing cost saturates, and a saturating term cannot rank anything. A
-perturbation changes an expert selection only when it crosses a near-tie, so the
-count of changed selections is governed by how many near-ties the model has, not
-by how large the perturbation was. Measured on one MoE checkpoint, NVFP4 and
-MXFP8 expert weights differed by a factor of 36 (0.1065 against 0.0030), while the
-deployed means differed by a factor of 1.2 (0.2534 against 0.2119), because a
-routing term near 0.20 dominated both. A reader given only the deployed means
-would conclude the two formats are nearly equivalent. They are not.
-
-**Routing is not router precision.** The mechanism is the activations arriving at
-the router, not the router's own weights. A checkpoint that ships its router in
-BF16 has a bit-identical router and still selects different experts, because
-quantized attention and quantized experts upstream have already moved the
-residual stream. It follows that the router weight cell is not the routing term
-and cannot stand in for it: on such a checkpoint that cell is exactly zero while
-routing divergence is not. The cell is retained only as the evidence that router
-weight precision is a red herring.
-
-It also follows that no quantize-dequantize cell is routing-free. Rounding any
-weight moves the residual stream, so every router downstream of it sees different
-inputs. An artifact must not describe a cell as holding routing fixed; each cell
-reports the share of its own selections that changed, which is what makes the
-claim checkable rather than assumed.
-
-**Measurement, not emulation.** The routing term is measured by recording the
-reference's selected experts per token and per layer, then comparing the
-candidate's selections over the same tokens. The artifact reports the share of
-(token, layer) selections that changed, the share of scored positions where any
-layer rerouted, the mean divergence at positions where routing held against
-positions where it changed, and how divergence grows with the number of rerouted
-layers.
-
-**Floor.** The floor is the excess the mean carries because rerouted positions
-diverge more than positions whose selection survived: the flipped fraction times
-the difference of the two means. It is a floor in the sense Law 1's repeat spread
-is a floor. Two candidates whose deployed means differ by less than it are not
-ranked by that difference, and an artifact that ranks them ranks them on the
-expert cell and says so in the same sentence as the claim.
-
-**A floor that is not a number.** The excess is undefined when one of its two
-populations is empty, and the two cases are opposites. If no position rerouted,
-routing cost nothing and the floor is zero. If every position rerouted, no
-held-routing population remains and the floor is unmeasurable — not small, and
-no rescore produces one, because the perturbation is large enough that routing
-divergence is the whole picture. Both satisfy this law when the run measured
-routing and the artifact states which case it is; what the law refuses is a
-routing measurement that never ran, and an artifact that omits the routing term
-so that saturation reads as zero cost. A saturated candidate publishes with its
-deployed mean marked unranked.
-
-**Naming.** Cells are named for the component that carries the error, never as
-`B×Q` or `Q×B`. That notation does not say which factor is the router, and two
-readers will order it two ways and mean opposite things by the same symbol. A cell
-is `expert_cell`, `router_cell`, or `composite_cell`, and a published artifact
-spells out what each one isolates.
-
-**Weight rounding is not the deployment.** A QDQ cell rounds weights and runs on
-BF16 kernels. A deployed checkpoint may also quantize activations and does use
-quantized kernels, so the artifact reports the deployed mean minus the composite
-cell. On one FP8 MoE checkpoint that term was 40% of the deployed mean, which no
-arrangement of weight-only cells can see. A component decomposition that omits it
-attributes a deployment to weight precision alone and understates it.
-
-**Ladder.** A campaign on a routed reference also scores the expert weights at
-each scheme on its ladder, not only at the deployed one. Ladder rungs are
-component-wide by construction: every expert weight is rounded, so the rungs
-differ only in format. The expert cell that decomposes the deployed mean matches
-per tensor and may therefore round a subset; it is not the deployed ladder rung.
-The cost of one more cell is a QDQ pass and a scoring run against a capture that
-already exists; the cost of not having it is a comparison nobody can make later
-without redoing the campaign.
-
-**Check.** For a reference whose capture manifest records declared experts, the
-compliance receipt requires an expert cell and a router weight cell, each naming
-the variant checkpoint and its QDQ manifest, and each carrying the same partition,
-token digest, and reference config digest as the deployed cell. A cell measured on
-other tokens is not a decomposition of this number and is rejected as one. The
-receipt also requires a measured routing term from the run itself; a routed
-candidate scored without one fails, and an unquantized router is not accepted as
-a reason to omit it.
-
-**Override.** Permitted under Law 13 for a dense checkpoint misdetected as routed,
-and for an exploratory result that is never published as a ranking. Not permitted
-for a published comparison between two quantization schemes, which is the case the
-law exists for.
+## Law 14 — Routed-model QxQ/BxQ intervention
+
+**Required.** A reference that routes tokens to experts publishes a paired
+intervention over the identical frozen tokens:
+
+1. **QxQ** (`qxq_cell`) is the unchanged quantized candidate running normally,
+   with its natural expert IDs and its own gating weights.
+2. **BxQ** (`bxq_cell`) is that same unchanged quantized candidate with the BF16
+   teacher's ordered logical expert IDs forced at every routed layer. The student
+   computes the gating weights for those forced experts from its own router
+   logits, using its native scoring, bias, normalization, and scaling rules.
+
+The first axis is the source of the expert IDs: `B` means BF16 teacher IDs and
+`Q` means the quantized student's natural IDs. The second `Q` is the candidate,
+which is identical in both runs. It does not mean "quantized experts" in a
+synthetic checkpoint, and the first axis does not mean router-weight precision.
+
+**Binding.** Both cells carry `mean_kld`, their supporting report path, partition,
+token SHA-256, reference-config SHA-256, and `candidate_weights_sha256`. Every
+binding equals the deployed report, and QxQ's mean equals the deployed natural
+mean. BxQ additionally carries a 64-hex `routing_trace_sha256`,
+`routing_mode: teacher_ids_student_weights`, and the supported
+`protocol_version`. The trace binds the forced IDs to the same teacher, tokens,
+layer order, and routing geometry used by both scores.
+
+**Backend control.** Exact ID replay is a capability, not an assumption.
+`backend_evidence` names the active backend or kernel, and `replay_supported` is
+true only when that path injects logical teacher IDs before placement mapping and
+dispatch. If replay uses a different kernel path, the candidate is first run with
+that path and no override. The control protocol runs two deployed-natural, two
+forced-natural, and two BxQ samples. Publication requires exact agreement:
+identical natural expert IDs across the two QxQ samples, identical per-position
+KLD across QxQ repeats, forced-natural controls, and BxQ repeats, all within
+the fixed numerical floor. Observed spans and route-flip rates are stored as a
+diagnostic addendum and never authorize a drifting canonical score. There is no
+backend-name allowlist or blanket tolerance. `natural_control_parity.passed` is
+true only when those exact-repeat conditions hold. An uncertified or
+nondeterministic backend fails. The candidate weight digest must remain unchanged
+and `candidate_weights_unchanged` must be true.
+Forced-natural controls and BxQ remain bound to the first QxQ sample's exact
+ordered IDs. QxQ always uses the model's own router.
+
+**Natural routing divergence.** The QxQ run also measures the student's natural
+expert IDs against the teacher trace. The artifact reports selection flip rate,
+position flip rate, conditional KLD where routing held and flipped, per-layer
+rates, and the existing routing-excess state. Conditioning after the run is not
+BxQ: it observes the subset where routes happened to agree, while BxQ forces the
+teacher IDs over the full suite.
+
+**Paired delta.** `routing_intervention_delta = QxQ mean KLD - BxQ mean KLD`.
+It may be positive or negative. It is a paired intervention result, not an
+additive decomposition and not a claim that routing alone contributed that much.
+
+**Synthetic QDQ diagnostics remain synthetic.** `expert_cell`, `router_cell`,
+`composite_cell`, and ladder rungs round selected BF16 weights, run on BF16
+kernels, and route naturally. They may still diagnose weight rounding and report
+their own route flips, but they are never labeled BxQ or QxQ. In particular,
+`expert_cell` is not BxQ because it changes the checkpoint and does not force
+teacher IDs.
+
+**Check.** For a manifest declaring experts, compliance requires both paired
+cells and all bindings above; exact QxQ/deployed equivalence; a complete trace
+digest; protocol, certified backend, and passing exact-repeat control evidence;
+the exact paired delta; measured natural routing divergence; and QxQ plus BxQ
+domain records. A manifest declaring no experts is
+`not_applicable`. Missing replay support, stale or unbound traces, failed exact
+repeat, an uncertified backend, and missing BxQ are failures. They are never
+converted into an override or a synthetic substitute.
+
+**Override.** Not permitted. Dense models satisfy this law as
+`not_applicable`; a routed model without supported replay and exact repeated
+QxQ/BxQ does not.
 
 ## Law 15 — Domain disclosure
 
@@ -558,12 +614,11 @@ artifact attributing one vendor's numbers to another's work. That is the worst
 error this program can make, because it is invisible: the number is real, the
 receipt is honest, and the name on it is wrong.
 
-**Digest, not a file list.** The bond is over tensor names, dtypes, and shapes,
-plus each shard's name and size, read from safetensors headers. It costs one
-header per shard rather than a pass over hundreds of gigabytes, and it cannot be
-preserved by a repack: a checkpoint that quantized one more layer, or stored a
-scale at a different width, is a different checkpoint under this digest. Two
-directories that share it hold the same weights, whatever their repos are called.
+**Content digest, not a file list.** The bond hashes each shard name and every
+byte of every safetensors shard, including its header. A checkpoint with the same
+tensor names, dtypes, shapes, and file sizes but different values therefore
+cannot preserve the digest. Two directories that share it hold byte-identical
+weight shards, whatever their repos are called.
 
 **The same mean from different weights is refused, not explained.** Distinct
 quantizations of one reference do not land on an identical mean to eighteen
@@ -572,8 +627,9 @@ one report was scored against the other's checkpoint, and nothing in the
 artifacts says which. Both results are withdrawn and rescored. The converse —
 two candidates sharing a digest — is not an error in the measurement but a fact
 about the upstream repositories: one is a verbatim re-upload of the other. It
-publishes once, and the artifact names the re-upload as such rather than
-presenting it as an independent quantization.
+publishes once only when both reports agree; conflicting scores from identical
+weights are refused. The artifact names an agreeing re-upload as such rather
+than presenting it as an independent quantization.
 
 **Check.** The compliance receipt requires `student_weights_sha256` on the report
 and an inspection of the published candidate carrying the same digest.
@@ -586,6 +642,78 @@ without a rescore. The deviation is printed beside the number, which then states
 that its weights are unbound. Not permitted for a new measurement, and never
 permitted for a refused identical mean.
 
+## Law 17 — Numerical substitution disclosure
+
+**Required.** When the kernel that scored a candidate did not use a quantization
+parameter as the checkpoint exported it, the result names that parameter, states
+what was put in its place, reports how many of the scored layers it reached, and
+bounds how far the replacement had to stretch. The substitution enters the
+candidate's comparability key, so a substituted result never ranks against one
+measured on its own parameters. A report that carries no substitution field at all
+fails: silence is not the same claim as "none". An empty field is a pass only when
+the run actually walked the loaded kernels and found nothing replaced. A run with
+no kernel inspection to walk, such as a dense candidate with no routed experts, is
+not applicable rather than clean; an empty field must never be read as a clean bill
+of health for parameters nobody looked at.
+
+**Why.** A kernel can be perfectly deterministic, pass every certification this
+program runs, and still not be computing what the checkpoint describes. Two
+substitutions this program has actually measured: the NVFP4 emulation experts
+replace each expert's activation scale with one scalar for the whole layer
+(measured once at a factor of 150 on a real export, and no longer the scoring
+path), and a checkpoint that omitted per-expert keys has those slots filled from
+the layer maximum so CUTLASS does not run on uninitialized memory. Both produce
+a real, bitwise-repeatable mean that answers a slightly different question than
+the checkpoint posed.
+
+**Disclosure, not withdrawal.** This is the law's central choice. A substituted
+result is published. Withdrawal is reserved for a result that cannot be
+interpreted at all, such as one bound to a reference capture the family no longer
+publishes; it is not a penalty for a number that came out badly, and a fidelity
+index that quietly drops the checkpoints its kernels handle worst is not an index
+of checkpoints but of kernel coverage. The failure mode here is never the
+measurement, it is an unqualified label on it. So the remedy is a label.
+
+**The spread is disclosed but does not bound comparability.** Only the identity of
+the substituted parameter enters the comparability key. Two candidates the kernel
+substituted the same way remain rankable against each other — which is the entire
+purpose of keeping them — while the spread tells a reader how much to discount the
+absolute value. A wide spread does not make a result less comparable to its peers;
+it makes the whole group further from the deployment all of them describe.
+
+**Per-cell states.** Each cell of a routed result reports one of three states.
+`measured` used the checkpoint's own parameters. `substituted` holds a real
+repeatable number obtained under a named replacement. `unavailable` holds no
+number, with the reason it could not be taken. The three are not collapsible: an
+absent cell and a substituted one are both unlike a plain measurement and nothing
+like each other. Where a substitution reaches both cells of a routed pair it
+partly cancels in `QxQ − BxQ`, but never exactly — clipping is nonlinear and the
+two runs route to different experts — so the delta is the sounder of the three
+numbers without being clean.
+
+**Check.** Compliance requires `quantization_substitutions` on the report. Each
+entry states its parameter, kind, affected and scored layer counts, and worst
+spread; a missing field, a malformed entry, or one that omits its spread fails.
+The check then confirms the comparability key still binds the substituted
+parameters. Scoring records the field from the loaded model, per tensor-parallel
+worker, reporting the worst case across ranks.
+
+**What is inspected so far, stated as a limit.** Routed expert layers are
+inspected; dense linear layers are not. This is a real gap and it fails closed:
+a dense family scored today reports no substitution field and Law 17 refuses it,
+which is the correct outcome — the law says nothing about that checkpoint until
+someone measures it. Recording an empty list instead would assert that nothing was
+substituted, and for NVFP4 that assertion would be false: the dense path collapses
+the input scale across each fused parallel layer exactly as the expert path does,
+under vLLM's own warning that it will likely reduce accuracy. Closing the gap
+needs more than a new inspector, because that path deletes `input_scale` once it
+has taken the maximum. The evidence does not survive loading, so it has to be
+captured while the substitution happens rather than observed afterwards.
+
+**Override.** Not permitted. An approval could only assert that an undisclosed
+substitution is acceptable, which is the one thing this law exists to refuse.
+A result whose substitution cannot be determined is unavailable, not approved.
+
 ## Numbering
 
 Laws are append-only. A published receipt cites its laws by number, so renumbering

QXQ.md

@@ -0,0 +1,705 @@
+# QxQ and BxQ — Routed-Model Distribution Fidelity
+
+**Applies to:** laws version 14, Laws 14 and 17.
+**Read first:** [`LAWS.md`](LAWS.md) for the laws, [`README.md`](README.md) for the
+harness.
+
+This document explains what the QxQ and BxQ cells measure, why measuring them at
+all requires bitwise-exact repeat from every kernel on the path, which MoE
+backends have earned the right to produce a published number, and how the
+pipeline refuses the ones that have not.
+
+## 1. Why one number is not enough for a routed model
+
+For a dense model, quantization error is weight rounding. One KLD against the
+BF16 reference describes it, and the only question left is how it distributes
+across text.
+
+A Mixture-of-Experts model breaks that. Two unrelated things happen at once, and
+a single mean hides both:
+
+1. The experts round their weights, exactly as a dense layer would.
+2. The router sends tokens to *different experts* than the reference would,
+   because the activations arriving at the router are themselves quantized.
+
+These have opposite consequences. Rounding is a fidelity tax that shrinks
+predictably with bit width. Routing divergence means the model is evaluating a
+different function — a different subnetwork per token — and it does not shrink
+predictably with anything. A candidate whose loss is mostly rounding and a
+candidate whose loss is mostly rerouting can post identical means and behave
+differently in deployment. Ranking them on that shared mean is the error Law 14
+exists to prevent.
+
+## 2. The two cells
+
+Both cells score the same candidate against the same BF16 teacher on the same
+frozen token suite, and differ only in who chooses the experts.
+
+**QxQ** is the candidate exactly as deployed: quantized weights, and its own
+router choosing its own experts. This is the number that describes what a user
+would actually run, and it is what routed candidates are ranked on.
+
+**BxQ** replays the same tokens with the teacher's expert IDs forced into the
+student, while the student keeps its own gating *weights*. Forcing the IDs
+removes the routing disagreement and leaves the rounding, so BxQ is the fidelity
+the candidate would have if its router agreed with the reference.
+
+The keeping of student gating weights is deliberate and easy to get wrong. If
+BxQ also took the teacher's gate values it would stop being a measurement of the
+student's experts and become a partial reconstruction of the teacher. Only the
+selection is intervened on; the weighting stays the candidate's own.
+
+**QxQ − BxQ** is therefore the routing contribution, reported as the paired
+intervention delta. The **natural route flip rate** reports how often the two
+disagree at all, measured over `(token, layer)` choices on the natural QxQ run —
+not on the intervened one, where by construction there is nothing to count.
+
+A worked example from the gemma-4-26B-A4B-it family: `Intel/gemma-4-26B-A4B-it-int4-AutoRound`
+scores QxQ 1.17816288 and BxQ 0.90308200, a delta of +0.27508088, and flips
+93.07% of natural routing choices. It sits mid-pack on deployed fidelity, but the
+decomposition says something no mean could: nearly all of its loss is the router,
+not the rounding. Its neighbours in the same group carry deltas near +0.15 at
+comparable flip rates.
+
+## 3. Why the subtraction demands bitwise-exact repeat
+
+QxQ − BxQ is a difference between two separate forward passes. If scoring the
+same tokens twice does not reproduce the same logits bit for bit, then a delta of
++0.15 is indistinguishable from kernel noise, and the decomposition is
+storytelling.
+
+So the pipeline does not assume reproducibility, it measures it. Every routed
+candidate is scored, replayed, and the two results compared; the published
+`repeat_delta` is that difference and it is required to be exactly `0.0`. A
+candidate whose backend cannot deliver that is marked uncertified for exact
+repeat and cannot be published as a Law 14 measurement.
+
+This is the entire reason the effort behind this document was determinism work
+rather than metric work. The metric is a subtraction; the difficulty is earning
+the right to subtract.
+
+## 4. How MoE kernels break exact repeat
+
+Batch invariance is the property that a token's output does not depend on which
+other tokens share its batch, nor on their order. Most MoE kernels violate it,
+and not by accident: grouping tokens by expert and reducing partial sums in
+whatever order the grouping produced is the fast way to do it. That is a correct
+optimization for serving, where nobody compares two runs, and it is fatal here.
+
+Four fixes were required before routed models could be scored at all.
+
+**Marlin token ordering and reduction.** The Marlin MoE path was ported to a
+canonical token order with a full-K reduction, which is what makes
+`MarlinExperts` and `BatchedMarlinExperts` batch invariant. This carries the bulk
+of the field: every int4 AWQ, GPTQ, and W4A16 routed candidate scores here.
+
+**Qwen GDN attention.** Certified only on its NVIDIA CUDA, non-speculative,
+per-sequence path, with FlashInfer GDN context parallelism disabled. Without this
+the Qwen3.6 family could not run under `VLLM_BATCH_INVARIANT` at all.
+
+**Dense NVFP4 linear layers.** Weight-only W4A16 NVFP4 linear layers use
+deterministic emulation, because dense Marlin is not batch invariant.
+
+**Expert parallelism is never certified.** An EP path reduces across ranks in
+completion order. No allowlist entry overrides this.
+
+## 5. Certification fails closed
+
+A kernel's `_supports_batch_invariance()` is a claim about itself. It is not
+evidence. `CutlassExpertsFp4` declared True and still produced NaN on W4A4
+checkpoints; the NaN was uninitialized per-expert activation scales, not the
+kernel, but the lesson stands: a self-declaration is not a probe.
+
+`inspect_model_moe_backends` in `vllm/v1/sample/kld.py` therefore grants
+`certified_for_exact_repeat` only when all three of these hold:
+
+1. The expert class is named in `_EXACT_REPEAT_CERTIFIED_EXPERTS`, an allowlist
+   earned by passing an exact-repeat probe on real suite content. The report
+   carries this separately as `exact_repeat_probed`.
+2. The kernel also self-declares batch invariance.
+3. The run is not expert-parallel.
+
+The allowlist currently holds `MarlinExperts`, `BatchedMarlinExperts`,
+`TritonExperts`, `Nvfp4QuantizationEmulationTritonExperts`, and
+`CutlassExpertsFp4`. CUTLASS earned its place after the SM120 bitwise
+permutation test (24/24, `atol=0`) and after the W4A4 NaN was shown to be a
+loader gap, not a kernel defect. The run's own zero-tolerance exact-repeat
+control remains the binding gate.
+
+The default is refusal. An unprobed backend — including one that arrives with a
+future vLLM bump — is uncertified until somebody probes it, and an uncertified
+backend is reported as uncertified rather than published as a number. This
+matters more than it sounds: the failure mode it replaces is a kernel quietly
+producing garbage that the harness dutifully writes down as a fidelity result.
+
+DeepGEMM, FlashInfer MoE, AITER, XPU, CPU, and every expert-parallel path remain
+uncertified.
+
+**Certification and backend selection now meet.** With the MoE backend pin removed
+(§7) the choice belongs to vLLM's oracle, and the oracle ranks by expected
+performance, not by whether a kernel is certified here. `AVAILABLE_BACKENDS` in
+`vllm/model_executor/layers/fused_moe/oracle/nvfp4.py` puts `FLASHINFER_TRTLLM`
+first and nothing in that path consults `VLLM_BATCH_INVARIANT`, so nothing
+*guarantees* a certified kernel. The refusal is upstream of the cost, at least:
+scoring reads certification off the loaded model and raises before any forward pass,
+so an uncertified choice costs a model load rather than a scored pass.
+
+Measured, on ten gemma4-26b-a4b candidates at one row each: the oracle chose
+`VLLM_CUTLASS` for all four NVFP4 MoE exports, `TRITON` for the FP8 export, and
+`MARLIN` for the four INT4 ones. Every one is in the allowlist and every control
+repeated at exactly 0.000e+00. FlashInfer was first in the list each time and its
+`is_supported_config` declined each time, for reasons logged only at debug level.
+So the outcome is good and it is a reading rather than a promise: run the `smoke`
+stage over a routed family before committing a campaign to it, and read the backend
+each candidate names.
+
+**A kernel is verified on the next run, not pinned.** The oracle chooses per layer
+from the checkpoint, the device, the installed FlashInfer and its own source, and
+§9's comparability key binds all four — the last of them because the oracle is a
+`.py` file under `vllm/` and therefore inside the numerics digest. Under one binding
+the choice is a function, so a second run of it owes the same answer. Scoring holds
+it to that: `_prior_kernel_identity` reads the previous report for the candidate,
+and when its binding matches the live one, the expert class and kernel read back
+after the student load must match what that report recorded. A mismatch is refused
+before any forward pass, because a published number that a rerun would not
+reproduce is worse than no number. The comparison is narrowed to `quant_method`,
+`kernel` and `experts`: tensor and expert parallelism are recorded on the report but
+a rescore is entitled to move them. Where the binding differs the expectation is
+released, since new numerics or newly built kernels are allowed a new choice — that
+is a new number, not a broken one. This is the safe half of pinning. A pin had to
+predict the kernel before the load and could refuse a checkpoint outright (§7); a
+comparison happens after the load and can only refuse a contradiction.
+
+## 6. Case study: uninitialized scales, not a broken kernel
+
+`CutlassExpertsFp4` self-declares batch invariance. Under
+`VLLM_BATCH_INVARIANT=1`, on two W4A4 NVFP4 checkpoints, it took finite inputs
+to NaN. The scorer refused a non-finite KLD and the engine died. The first
+reading was that the kernel's claim was false. That was wrong.
+
+The kernel is bitwise batch-invariant on this hardware. The CUDA grouped GEMM
+asserts at compile time that it uses `PersistentTileSchedulerSm100Group` "for
+batch invariance", and `tests/v1/determinism/test_cutlass_batch_invariance.py`
+passed 24/24 on SM120 at `atol=0, rtol=0` across both activations, both expert
+counts (40, 64), both top-k values, and all three shape cases.
+
+The NaN was uninitialized memory. NVFP4 allocated per-expert activation scales
+with `torch.empty`. Checkpoints that omitted `input_scale` keys for some experts
+in some layers — `bg-digitalservices/Gemma-4-26B-A4B-it-NVFP4` (14 layers short)
+and `Neural-ICE/Gemma-4-26B-A4B-it-NVFP4` (16 layers short) — left those slots
+holding whatever was on the GPU. CUTLASS consumes one scale per expert
+(`a1_gscale` / `a2_gscale` of length `e`); a NaN slot damages that expert. The
+unsloth export, with every slot written, never went NaN.
+
+Loading does not catch this. Strict all-parameters-loaded tracking is off by
+default for quantized models, and any module with `process_weights_after_loading`
+has every parameter force-marked as loaded, so a partially filled tensor is
+invisible to it by construction.
+
+Localizing it took two purpose-built tools.
+
+`scripts/scan_checkpoint_nonfinite.py` reads a checkpoint's tensors and counts
+non-finite values, zero scales, and per-row magnitude outliers in scale tensors.
+`torch.isfinite` has no CPU kernel for several float8 dtypes, so `nonfinite_mask`
+tests FP8 bit patterns directly — `(bits & 0x7F) == 0x7F` for `e4m3fn`, and
+`>= 0x7C` for `e5m2`. The scan came back clean, which established that the NaN
+was generated at runtime rather than baked into the weights, and moved the
+investigation off the checkpoint.
+
+`scripts/nan_first_module_probe.py` installs forward hooks and reports the first
+module whose output goes non-finite from finite inputs, together with the first
+row index at which it happens. Two capabilities were decisive. `--context-file`
+replays a real suite context instead of a synthetic prompt, because synthetic
+prompts are repetitive and never route diversely enough to trigger the bug.
+`--prompt-logprobs` computes logits at every prompt position, which is what the
+scoring harness does and what a generation-only probe does not.
+
+With both, the failure is exactly reproducible: `moe.experts` at layer 0, row 814
+of `context-0002`, and layer 14, row 170 of `context-0004`.
+
+The `--moe-backend` override then isolated the failure to the native path on
+the two incomplete checkpoints, not to batch invariance itself:
+
+| `VLLM_BATCH_INVARIANT` | Backend | Result |
+| --- | --- | --- |
+| 0 | `FLASHINFER_CUTLASS` (auto) | pass |
+| 1 | `VLLM_CUTLASS` (auto) | **NaN** on incomplete exports |
+| 1 | `marlin` (forced) | pass (drops activation scales) |
+| 1 | `emulation` (forced) | pass (collapses per-expert scales) |
+
+The fail-closed allowlist was the right reaction to a NaN that looked like a
+kernel defect. Once the scales were the cause, keeping CUTLASS off the list
+was the thing standing between native BxQ and a published number.
+
+## 7. W4A4 NVFP4 and the kernel the loader actually builds
+
+A W4A16 export and a W4A4 export of the same model differ only in whether
+activations are also quantized, and that difference decides which kernels can
+score the checkpoint honestly. A repository name says nothing reliable about it,
+so `_declared_expert_activation_quant` in
+`examples/offline_inference/score_mode_kld.py` reads the checkpoint: for every
+`quantization_config.config_groups` entry that targets experts, it reports what
+that group declares for `input_activations` — `4-bit float`, `8-bit float`, or
+`unquantized`. Scoped to expert groups, because a group covering only attention
+says nothing about the kernels the MoE will run, and scanning every group is what
+made a checkpoint with W4A4 attention and FP8 experts look like a W4A4 MoE.
+
+It reports a list, not a verdict, because a checkpoint can declare more than one
+width inside its own MoE. And it is a reading of the checkpoint, not a prediction
+about the run: what the loader builds is a separate fact, read back after load.
+
+**The scorer no longer pins a MoE backend, and the removal was not a
+simplification.** It used to pin `moe_backend="cutlass"` for anything it judged
+W4A4, and two real checkpoints showed that a pre-load pin is being asked two
+questions it cannot answer. `nvidia/gemma-4-26B-A4B-it-NVFP4` declares
+`num_bits: 4, type: float` weights *and* activations for `mlp.experts`; the
+modelopt `MIXED_PRECISION` loader built those experts weight-only, and the pinned
+W4A4 CUTLASS kernel refused a configuration ending `(symmetric)xNone` — the pin
+failed the candidate over a kernel the checkpoint was never going to get.
+`unsloth/gemma-4-26B-A4B-it-NVFP4` declares FP8 experts in its last eight layers
+and W4A4 in the rest, which no single global pin can serve at all. Both failures
+were ours, not the checkpoints'.
+
+So the choice is left to the oracle, which answers per layer and knows what it
+built, and `inspect_model_moe_backends` reads the answer back off the loaded
+model. That is the measurement this project prefers to a declaration, and it is
+the same discipline as `assert_unquantized_kv_cache` in §9.
+
+The kernels themselves still differ, and which one ran still decides what a number
+means. vLLM CUTLASS is the only native W4A4 MoE path that keeps a per-expert
+activation-scale vector. FlashInfer collapses every expert to one scalar via
+`amax_for_moe_activation_quant(...).repeat(num_experts)` — the same defect as
+emulation. Marlin drops activation scales and scores W4A4 as W4A16.
+
+Unwritten slots are now a NaN sentinel, filled from the maximum of the present
+per-expert scales before CUTLASS fuses them into the weight alphas. Too large
+wastes quantization range; too small overflows e4m3. A layer with no finite
+positive slot is refused rather than invented. The fill is recorded on the
+layer at fill time as `uncalibrated_experts_filled_from_layer_max` and Law 17
+discloses it. A complete export scored on CUTLASS records an empty substitution
+list, because the kernel used the checkpoint's own scales. A complete export is
+not automatically such a run: with the pin gone, unsloth's mixed-precision MoE
+gets whichever kernel the loader builds for it, and if that kernel collapses the
+scales the substitution is disclosed and then priced.
+
+Both halves of the model are walked, because a dense NVFP4 projection can omit a
+shard's scale exactly as an expert can. `inspect_model_moe_backends` covers the
+routed experts and `inspect_model_nvfp4_dense_scales` covers everything else,
+reported separately so the denominators stay meaningful — a fill on 3 of 200
+dense layers is a different claim than 3 of 30 experts. The fill records a scan
+marker on every layer it visits, filled or not, because counting NVFP4 layers
+after load is guesswork: the dense paths delete or overwrite the very parameter
+they were named for. Without the dense walk a dense W4A4 candidate reported no
+substitution, and Law 17 read that silence as a clean bill of health rather than
+as an absence of evidence, which is why it now returns `NOT_APPLICABLE` when
+nothing was inspected at all.
+
+| Path | Per-expert activation scales | Exact repeat |
+| --- | --- | --- |
+| vLLM CUTLASS FP4 | honoured (per-expert vector) | certified on SM120 |
+| FlashInfer FP4 | collapsed to one scalar for the layer | uncertified |
+| Marlin | dropped entirely (scores W4A16) | certified |
+| Emulation | collapsed to a layer maximum | certified |
+
+**A collapse is priced, not just disclosed.** Law 17 states that a run used one
+layer-wide scalar where the checkpoint exported a scale per expert, and that left
+a reader to guess whether the substitution was worth a decimal place or the
+ranking. When a report discloses `per_expert_collapsed_to_layer_scalar`, the
+campaign scores that candidate once more with `--moe-backend cutlass`, which keeps
+the per-expert scales, and records the difference as `per_expert_collapse_cost`.
+The one-pager prints both numbers and the delta.
+
+Conditional by construction: a candidate that collapsed nothing pays nothing,
+which is every dense candidate and every routed one whose experts kept their own
+scales. A candidate that did collapse pays one student load and one scoring pass,
+because the teacher capture does not depend on which kernel the experts run and is
+shared with the run that just finished. If the non-collapsing kernel will not load
+the checkpoint on this hardware, the cost is recorded as unpriced with the reason;
+withdrawing a measured, compliant, deployed number because a second run vLLM never
+has to perform did not work would be the wrong trade.
+
+The deployed number does not change. Which kernel the oracle builds for a
+checkpoint on this hardware is a fact about deploying it, and this index reports
+what deployment does. The counterfactual is priced beside it, never in place of it,
+and the two are not comparable to each other's leaderboard groups because the
+substitution is part of the comparability key.
+
+**A named backend belongs on the student only.** `--moe-backend` reaches
+`student_kwargs`, not the shared `llm_kwargs`. Applied to the latter it propagates
+to the unquantized BF16 teacher, which has no such scheme, and reference engine
+initialization fails.
+
+A weight-only W4A16 NVFP4 result still measures the quantization scheme rather than
+a native FP4 kernel's own rounding, because dense Marlin is not batch invariant
+and those linear layers use deterministic emulation. That limit is stated on the
+published card.
+
+### History: the emulation collapse we published through
+
+Before the loader gap was understood, W4A4 scoring was pinned to emulation.
+The emulation branch of `convert_to_nvfp4_moe_kernel_format` collapses the
+per-expert activation scales to one scalar per layer —
+`a13_scale = 1.0 / a13_scale.max()` — and vLLM's own comment says taking the
+largest global scale "likely results in overflowing the FP8 range for other
+experts." Since `a2_scale` holds `1 / w2_input_global_scale`, `.max()` applies
+the *smallest* per-expert scale. Measured on
+`unsloth/gemma-4-26B-A4B-it-NVFP4`:
+
+| Parameter | Slots | Distinct values | Worst spread | Collapsed |
+| --- | --- | --- | --- | --- |
+| `w13_input_global_scale` | 256 | 1 | 1.0x | no |
+| `w2_input_global_scale` | 128 | 86–99 | 150.5x | yes, all 30 layers |
+
+Four NVFP4 exports carrying byte-identical QDQ diagnostics (0.66652247 and
+1.43642041) scored QxQ between 1.16299506 and 1.82545539 on that path, and that
+0.6 nat spread was written up as living "entirely in the activation scheme and
+kernel path." That reading was unsupported. The spread was substantially an
+artifact of the measurement. It stays here because it is the exact shape of
+mistake this document exists to prevent: a real, reproducible, bitwise-exact
+number that is nonetheless measuring the harness rather than the checkpoint.
+
+Law 17 still exists for the remaining real substitutions — an uncalibrated
+expert filled from the layer maximum is one — so a substituted result ranks
+only against candidates measured the same way, and is never withdrawn for
+having a disclosed fill.
+
+### How far to discount a filled result
+
+The disclosure states a spread. It does not say what that spread costs, and the
+first native-CUTLASS campaign answered the question by accident.
+
+The two filled candidates landed on top of each other:
+
+| Candidate | Layers filled | Worst spread | QxQ | BxQ |
+| --- | --- | --- | --- | --- |
+| `Neural-ICE/Gemma-4-26B-A4B-it-NVFP4` | 15 of 30 | 230.6x | 1.82281421 | 1.58512109 |
+| `bg-digitalservices/Gemma-4-26B-A4B-it-NVFP4` | 12 of 30 | 240.3x | 1.82307292 | 1.59361302 |
+
+0.00026 nats apart on QxQ, from different publishers, with different fill
+footprints, and with one quantizing its LM head while the other does not. They
+are not the same checkpoint: distinct `student_weights_sha256`, and distinct
+per-position KLD digests in `kld_evidence`, so the two runs did not produce
+bitwise-identical output either.
+
+Two independent quantizations do not agree to four decimal places on their own.
+The reading that fits is that once a tensor's per-expert scales are replaced by
+one layer maximum on half the layers, the fill sets the number and the
+checkpoint's own choices stop being visible in it. On the same suite the clean
+NVFP4 exports separate normally — 1.13846019 for unsloth against 1.77968754 for
+RedHatAI — so the collapse is not the suite failing to discriminate.
+
+This is n=2 and therefore a strong hint rather than a proof; a fill-direction
+sweep on one checkpoint would settle it, and has not been run. Treat a filled
+QxQ as an upper bound on that family of exports rather than a measurement of the
+particular one. The leaderboard says so on the row: a filled candidate is ranked
+beside the clean ones, because it was measured on the same suite, geometry, and
+runtime, and carries a `†` naming what was substituted. It is not exiled into a
+section of its own, which would hide the comparison a reader came for while the
+substitution stays bound in its comparability key either way.
+It is also the argument against ever promoting the fill out of the harness: see
+§13.
+
+## 8. Checkpoint defects the pipeline had to fix, not tolerate
+
+Determinism gets you a repeatable number. It does not get you a *correct* one if
+the checkpoint is being loaded wrong, and two classes of loading bug were found
+by scoring rather than by tests.
+
+**AutoRound int4 geometry.** AutoRound exports use group sizes that do not divide
+the input dimension, producing a partial final group. The packed-row and
+scale-group arithmetic — `scales_size`, `num_groups`, `size_k`, and the encoded
+symmetric zero used to pad — was wrong at exactly those boundaries, in both the
+Marlin and AutoGPTQ paths. A read-only `safetensors` header probe established the
+real geometry before any code changed; assumptions about row counts had already
+cost several wrong fixes. Regression tests cover the padded reduction allocation,
+the encoded-zero padding, single-rank partial groups, and the dense path's packed
+rows with a partial scale group.
+
+**A quantized router with nowhere to land.** The AutoRound Gemma-4 export
+quantizes the router projection. vLLM's `Gemma4Router` hardcodes its `GateLinear`
+as unquantized BF16, so the checkpoint's `qweight`, `qzeros`, and `scales` had no
+destination, no loader complained, and `router.proj.weight` was left as
+uninitialized memory. The router emitted NaN from finite inputs and KLD went
+non-finite.
+
+Note how this presents: identical symptom to §6, entirely different cause. The
+module probe is what separated them, naming `layers.0.router.proj` rather than
+`moe.experts`. The fix dequantizes those tensors into BF16 at load time in
+`Gemma4Model.load_weights`, with tests in `tests/kernels/moe/test_gemma4router.py`
+covering the transposed packed values and the refusal of an indivisible group
+count. That checkpoint now scores 1.17816288 and passes all seventeen laws.
+
+## 9. The runtime the numbers are bound to
+
+Scoring pins `VLLM_BATCH_INVARIANT=1`, disables DeepGEMM
+(`VLLM_MOE_USE_DEEP_GEMM=0`) and FlashInfer autotune, sets `NCCL_DETERMINISTIC=1`
+and `CUBLAS_WORKSPACE_CONFIG=:4096:8`, enforces eager execution, disables prefix
+caching, holds `max_num_seqs=1`, and refuses a quantized KV cache.
+
+None of that is optional and none of it is a performance setting. Each one closes
+a path by which two runs of the same tokens could diverge.
+
+The KV cache closes the widest such path found so far, and it was open for
+the whole campaign. Left at `auto`, vLLM resolves the KV cache dtype from a
+scheme declared in the candidate's own config, so
+`unsloth/gemma-4-26B-A4B-it-NVFP4` — which declares an 8-bit float KV cache —
+had its attention read and written through a quantized cache while every
+candidate it was ranked against used an unquantized one. That is a difference in
+how the measurement was taken, not a property of the checkpoint being measured.
+The cache is never quantized: scoring holds 4096 tokens, so there is no memory
+pressure that quantizing it could relieve, and nothing to weigh against the loss
+of comparability.
+
+Unquantized is not one dtype. It was a literal `bfloat16` at first, which refused
+three AWQ checkpoints published as float16: FlashAttention will not take a float16
+query against a bfloat16 key, so the run died in `mha_varlen_fwd` rather than
+scoring. `unquantized_kv_cache_dtype` in `vllm/v1/sample/kld.py` reads each
+checkpoint's own `torch_dtype` and caches in that, so a float16 model caches in
+float16 and a bfloat16 model in bfloat16. Both are unquantized, which is the
+property the policy is about; the resolved dtype is recorded on the report and
+carried in the comparability key, so a reader is told which one a number was taken
+under and two candidates cached differently are not silently ranked together.
+`assert_unquantized_kv_cache` refuses anything outside
+`UNQUANTIZED_KV_CACHE_DTYPES`, and still refuses `auto`.
+
+The rescore that followed measured what it had cost, and doubled as the cleanest
+determinism evidence in this document. unsloth is the only one of the nine
+candidates whose inspection reported a declared KV cache scheme, and it is the
+only one whose number moved: 1.16570700 to 1.13846019, so its declared FP8 cache
+had been costing it 0.02724681 nats, or 2.3% of its KLD, and enough to place it
+above `gemma-4-26B-A4B-it-W4A16` at 1.15926068 when it belongs below. Five of the
+remaining eight are quoted at their pre-pin values elsewhere in this document —
+0.69415039, 1.17816288 with its BxQ and delta, 1.77968754, 1.82281421, and
+1.82307292 — and every one came back identical to eight decimal places on a
+different `vllm_commit`. So the defect was worth a rescore and was not worth a
+panic, the blast radius was exactly the set the mechanism predicted, and the
+harness reproduced everything outside it. `assert_unquantized_kv_cache` reads the dtype the engine actually
+resolved and refuses both a quantized value and `auto`, because the failure being
+prevented is precisely a value nobody checked.
+
+**A number that moves on a rescore is not a number that was wrong.** The Qwen
+families were first scored before the NVFP4 uncalibrated-scale fill, before the
+Marlin determinism work, and before MoE batch invariance; rescored under all three,
+some of those values moved by as much as 6%. That is the correct amount of movement
+for a runtime that gained a fill where it had been consuming uninitialized scales
+and a batch-invariant expert path where it had not had one. Each value was a
+faithful measurement of the runtime that produced it, which is why the runtime is
+in the comparability key and why a legacy number is never quietly ranked beside a
+current one. Read a movement of this size as the runtime changing, and look for the
+change; the alarming case is a number that moves when nothing that computes it did,
+which is what the digest below is for.
+
+Because that runtime *is* part of the result, its identity is bound into the
+comparability key: `numerics_digest`,
+`compiled_extensions_sha256`, `torch`, `driver`, `gpu_names`, and
+`kv_cache_dtype`, alongside the suite and geometry. Two candidates are ranked
+against each other only when all of it matches. `kv_cache_dtype` is in that list
+because of the unsloth case above: the key's one job is to bound a ranking to
+runs that ran alike, and it had nothing to say about a candidate whose attention
+ran at a different precision than its neighbours'.
+
+The commit used to sit in that list, and the consequence was that **any commit
+invalidated every published number.** A documentation paragraph, a campaign
+config, a new script: the next scoring run read a different `vllm_commit`,
+declared 45 compliant reports stale, and spent GPU-days reproducing numbers that
+were already right — and, being a rescore rather than a refusal, it did so
+silently. That was over-refusal dressed as rigour. An index of quantization
+fidelity is under continuous development by construction, so a currency test that
+cannot tell a docs edit from a kernel change makes the index unmaintainable.
+
+What bounds a result is whether the code that computed it would compute it again.
+`numerics_digest` in `vllm/v1/sample/kld.py` hashes every `.py` under `vllm/`
+together with the scorer, and `compiled_extensions_sha256` covers the built
+kernels, so between them they answer that question directly. Kernel sources are
+not hashed, because a source edit cannot move a number until it is rebuilt and the
+rebuild changes the extension digest. The digest is deliberately coarse — a
+comment in `vllm/` moves it — because deciding which edits inside the runtime are
+numerically inert is exactly the judgement a currency test must not be trusted
+with, and the cost of that coarseness is a rescore rather than a wrong number.
+
+The commit is still recorded on every report and printed on every one-pager. It
+says *when* a number was taken, which is provenance under Law 6, and provenance is
+not a currency test. A result reads: at commit `abc123`, under numerics digest
+`def456`, this KLD was measured. The commit may have moved a hundred times since;
+the number stands until the digest moves. Harness fixes are still best batched,
+now because a rescore wave costs GPU-hours rather than because the alternative is
+a stale library.
+
+## 10. Gates that stop a correct measurement from being published wrong
+
+Each of the following was found by a law refusing to publish, at the end of a
+multi-hour campaign, rather than by a test. They are recorded here because the
+failure mode is characteristic: the measurement is fine, and the metadata binding
+it to a suite, a geometry, or a reference capture is not.
+
+**Suite-driven geometry.** `_expected_geometry` derives the expected row count and
+context length from the suite manifest and the active partition, not from the
+dataset-driven `rows` and `context_length` on the config. Those config fields
+describe a dataset-driven run only; reading them for a suite-driven one compared
+768 real rows against a config default of 1024 and failed every candidate.
+
+**A field nobody wrote.** `context_length` was absent from every report, so a gate
+comparing it always failed. The writer in `score_mode_kld.py` now records it.
+Existing reports were backfilled from their bound capture manifests after
+verifying the recorded `capture_manifest_sha256`, refusing any mismatch.
+
+**Captures that moved underneath a report.** Assembly publishes the reference
+capture beside the report that cites it. If the capture is rebuilt after a report
+scores, the pair cites a manifest nothing hashes to, and Laws 5 and 14 refuse it —
+after the whole campaign. `_score_report_is_current` now treats a changed capture
+manifest as staleness, where a rescore is cheap.
+
+**Currency must not depend on processing order.** The gate above compares against
+whatever capture is on disk at that instant, and silently passes when the file is
+absent. That is order-dependent: early candidates in a run match the capture their
+predecessor left behind and are skipped, then a later rescore replaces it and
+strands them. `_candidate_complete` therefore delegates to the same
+`_score_report_is_current` the scorer uses, which compares each report's recorded
+commit against the live runtime and cannot be defeated by ordering.
+
+The last one has a worked example. In the gemma-4-26B-A4B-it family, one candidate
+was skipped early in a run on a stale `ec11b8…` capture, while a later rescore
+rebuilt the capture as `d1f03dea…` on a new commit. Nine candidates published on
+the new commit; the tenth kept a report from `fdc0b57e` bound to a capture the
+family no longer published. It failed Law 12 alone and ranked in a comparability
+group of one — every other law passed, because they read the candidate's own
+manifest. It was withdrawn via `excluded_candidates`, which records the repo,
+revision, and reason in `excluded-candidates.json` beside the family, rather than
+being dropped from the config silently.
+
+## 11. Reproducing and extending
+
+Certify a new MoE backend before trusting it. In order:
+
+```bash
+# 1. Is the NaN in the weights, or generated at runtime?
+python scripts/scan_checkpoint_nonfinite.py --model <checkpoint> --stats
+
+# 2. Which module goes non-finite first, on real content, at which row?
+VLLM_BATCH_INVARIANT=1 python scripts/nan_first_module_probe.py \
+    --model <checkpoint> --context-file <suite>/contexts/context-0002.json \
+    --prompt-logprobs
+
+# 3. Is it the kernel? Re-run pinning each backend in turn.
+VLLM_BATCH_INVARIANT=1 python scripts/nan_first_module_probe.py \
+    --model <checkpoint> --context-file <suite>/contexts/context-0002.json \
+    --prompt-logprobs --moe-backend marlin
+
+# 4. One candidate end to end: finite KLD, and repeat exactly 0.
+python fidelity/campaign.py smoke --config <campaign>.json \
+    --only-candidate <name>
+
+# 5. The family.
+python fidelity/campaign.py all --config <campaign>.json
+```
+
+A backend that passes the probe on real content, at full context, with prompt
+logprobs on, may be added to `_EXACT_REPEAT_CERTIFIED_EXPERTS`. Nothing else
+qualifies it, and self-declaration never does.
+
+Note that `--only-candidate` applies to `smoke` only. To rescore a single
+candidate of an assembled family, delete its report from
+`<work>/reports/<tag>.json`; the completeness gate then rescores exactly that one
+and skips the rest. A rescore the campaign decides on for itself moves the old
+report to `<work>/prior/<tag>.json` and holds the rebuild to the kernel that report
+read back. Deleting the report by hand skips that, which is the way to accept a
+kernel change deliberately rather than argue with the check.
+
+## 12. What these numbers do not say
+
+**They are not comparable outside their group.** Not against numbers from another
+suite, geometry, runtime, or laws version, and not against any published
+elsewhere. The comparability key is printed with every leaderboard group for
+exactly this reason.
+
+**A W4A4 NVFP4 MoE result is not guaranteed to be a native CUTLASS number.** It
+used to be, by a pin, and the pin failed two checkpoints it should have scored
+(§7). A result is now a number from whichever expert kernel the loader built for
+that checkpoint on this hardware, which is the kernel deploying it would get. Read
+it off the one-pager's declared-against-built table rather than inferring it from
+the scheme label: a checkpoint declaring 4-bit activations for its experts may have
+been built weight-only, and one built on a kernel that collapses per-expert
+activation scales carries a disclosed substitution and a priced cost for it. A
+W4A16 dense NVFP4 result still measures the scheme rather than a native FP4 kernel,
+because dense Marlin is not batch invariant.
+
+**A declared quantization is not a performed one.** The checkpoint's
+`quantization_config` is a statement by whoever exported it, and the kernels vLLM
+builds are a separate fact. `nvidia/gemma-4-26B-A4B-it-NVFP4` declares W4A4
+experts and was built weight-only; its number is honest about what ran and says
+nothing about what a W4A4 kernel would have scored.
+
+**The QDQ ladder is diagnostic, never a candidate.** Those cells round weights on
+synthetic BF16 checkpoints and route naturally. They are not QxQ or BxQ, they are
+not rankable against deployed candidates, and the published tables separate them.
+
+**A scheme label names the narrowest group, not the whole model.** A checkpoint
+may quantize attention at one width and its experts at another, and may declare a
+KV cache scheme. `unsloth/gemma-4-26B-A4B-it-NVFP4` is `format:
+"mixed-precision"`: FP8 W8A8 attention, NVFP4 W4A4 experts and dense MLP, and a
+declared FP8 KV cache. Its QxQ of 1.13846019 against 1.77968754 for a complete
+all-`Linear` NVFP4 export is therefore mostly the 8-bit attention, not a better
+NVFP4 export, and it lands between the all-FP8 candidate at 0.69415039 and the
+all-NVFP4 ones exactly where a hybrid should. The declared KV cache is a separate
+matter and is no longer in that number: §9 holds an unquantized cache, which is
+worth 0.02724681 of the 0.64 separating it from the complete export, so the
+attention width still carries the result. This is not a comparability failure — the key deliberately excludes
+the candidate's scheme, because ranking schemes against one reference is the
+point — but it was a labelling one until `scheme_mix` and `kv_cache_scheme` were
+added to the inspection. Component coverage cannot substitute: it counts how many
+weights are quantized, not at what width, so a hybrid and a uniform export both
+read `all`.
+
+**A flip rate is not an error rate.** Two experts disagreeing on a token is not
+by itself a wrong answer; the delta is what quantifies the cost. AutoRound's
+93.07% flip rate with a +0.275 delta is the point — high disagreement, bounded
+consequence.
+
+**Certification is about repeatability, not accuracy.** `exact_repeat: certified`
+says two runs agree bit for bit. It says nothing about whether the kernel computes
+the right thing, which is what the zero baseline (Law 1) and the reference binding
+(Laws 12 and 16) are for.
+
+## 13. The fill is a harness policy, not a vLLM fix
+
+Half of the loader change here — allocating consumed NVFP4 scales as a NaN
+sentinel instead of `torch.empty`, so an unwritten slot is detectable rather
+than arbitrary — is not ours to contribute. Three open upstream PRs already do
+it on the same files by the same mechanism: #54444 on the ModelOpt linear
+methods and fused experts, #45320 on the ModelOpt per-expert scales, #52501 on
+the linear per-block `weight_scale`. A fourth, #55073, is actively reworking the
+same compressed-tensors and ModelOpt scale code.
+
+Where we differ is the policy after detection, and the difference is deliberate
+on both sides. All three upstream PRs **reject**: they raise at load time naming
+the parameter and the affected experts. #45320 states the position outright —
+"this remains fail-fast only, it does not guess missing calibration statistics
+or add an imputation policy." Filling from the layer maximum is exactly the
+imputation policy they declined.
+
+They are right for a serving engine, and the evidence for that is in §7's
+convergence table rather than in any argument from principle. A user served a
+filled checkpoint gets a model whose numerics are substantially set by an
+invented scale, with nothing on the surface to say so. Refusing to load is the
+better failure.
+
+The harness can do what the engine should not, because it discloses. Law 17
+records the fill on the report, the substituted parameters enter the
+comparability key, and a filled candidate ranks only against others measured the
+same way. That is the whole justification, and it does not transfer to a serving
+path that has no comparability key to put anything in.
+
+Two consequences follow.
+
+**Those two candidates will stop loading on stock vLLM.** When any of the
+rejecting PRs lands, `Neural-ICE/Gemma-4-26B-A4B-it-NVFP4` and
+`bg-digitalservices/Gemma-4-26B-A4B-it-NVFP4` will refuse at load. Their
+published numbers stay valid for what they are and become unreproducible without
+this fill, so the fill has to be maintained as a standing, disclosed divergence
+rather than treated as a fix awaiting merge.
+
+**The contribution worth making is evidence, not code.** #45320, #54444, and
+#55073 all report that no end-to-end evaluation was run: no Blackwell hardware,
+or no affected checkpoint, or both. This harness has SM120, two affected
+checkpoints, and measured numbers for what the missing scales cost. That is the
+gap in those PRs, and it is not a competing patch.

Qwen3.8-27B-AWQ-INT4/compliance.json

@@ -1,8 +1,9 @@
 {
   "attribution": null,
   "candidate": "/media/fmodels2/cyankiwi/Qwen3.8-27B-AWQ-INT4",
-  "candidate_weights_sha256": null,
+  "candidate_weights_sha256": "7efc0eee36683b616eea4d1147b381e07bf0f1a359abc150aa43f74126a38d63",
   "comparability_key": {
+    "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
     "context_length": 2048,
     "driver": "580.173.02",
     "gpu_names": [
@@ -12,19 +13,22 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 248044,
-    "laws_version": 9,
+    "kv_cache_dtype": "bfloat16",
+    "laws_version": 15,
     "model_runner_v2": false,
+    "numerics_digest": "251a9225b37415b91fc58d975b0670fc4b481e6f5a816b663573c42e161438f8",
     "reference_config_sha256": "191e0af232104ed8b65258cf3fb2b842e288008baca7633c11b82a1ac7203aab",
     "rows": 768,
     "score_from": 0,
     "stride": 2048,
+    "substituted_parameters": [],
     "suite_id": "qwen3.8-27b-fidelity-1024x2048-v1",
     "tensor_parallel_size": 1,
     "token_sha256": "9c935708bbffbf45d5baa2931fb4af7e7b22df1e041aa0b6f4ca44d3315e543b",
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T12:56:04.551912+00:00",
+  "evaluated_at": "2026-09-11T20:24:35.093765+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -52,13 +56,13 @@
       "title": "Real vocabulary"
     },
     {
-      "detail": "all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc",
+      "detail": "all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3",
       "law": 5,
       "status": "pass",
       "title": "Manifest binding"
     },
     {
-      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s)",
+      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s)",
       "law": 6,
       "status": "pass",
       "title": "Provenance"
@@ -70,13 +74,13 @@
       "title": "Storage integrity"
     },
     {
-      "detail": "trunk 0.03110211, deployed 0.03110211, delta -8.21671395159762e-09",
+      "detail": "trunk 0.03116103, deployed 0.03116102, delta -8.224659543698554e-09",
       "law": 8,
       "status": "pass",
       "title": "Head transparency"
     },
     {
-      "detail": "mean 0.03110211, median 0.00686589, max 28.04218483, 4 depth buckets",
+      "detail": "mean 0.03116102, median 0.00686925, max 25.42261505, 4 depth buckets",
       "law": 9,
       "status": "pass",
       "title": "Tail and depth disclosure"
@@ -103,38 +107,37 @@
       "detail": "reference declares no experts",
       "law": 14,
       "status": "not_applicable",
-      "title": "Component attribution"
+      "title": "Routed-model intervention"
     },
     {
-      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.09667677, strongest scientific_technical at 0.01271958, spread 7.6x",
+      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.09701459, strongest scientific_technical at 0.01271533, spread 7.6x",
       "law": 15,
       "status": "pass",
       "title": "Domain disclosure"
     },
     {
-      "approval": {
-        "approver": "Andy Kitzke",
-        "justification": "These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger.",
-        "timestamp": "2026-09-02T05:15:00Z"
-      },
-      "detail": "the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else",
+      "detail": "scored weights 7efc0eee36683b61 as inspected",
       "law": 16,
-      "status": "override",
+      "status": "pass",
       "title": "Candidate weight binding"
     },
     {
-      "detail": "1 override(s), each fully attributed",
+      "detail": "every scored layer used the checkpoint's own quantization parameters",
+      "law": 17,
+      "status": "pass",
+      "title": "Substitution disclosure"
+    },
+    {
+      "detail": "no overrides claimed",
       "law": 13,
       "status": "pass",
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 9,
-  "mean_kld": 0.031102106439063696,
+  "laws_version": 15,
+  "mean_kld": 0.031161018371537674,
   "nondeterminism_floor": 0.0,
-  "overridden_laws": [
-    16
-  ],
+  "overridden_laws": [],
   "partition": "analysis",
   "program": "Local Inference Lab \u2014 Distribution Fidelity",
   "ranking_floor": null,
@@ -147,16 +150,16 @@
     "overall": {
       "deployed": {
         "contexts": 768,
-        "max_kld": 28.042184829711914,
-        "mean_kld": 0.031102106439063696,
-        "mean_ref_top1_prob": 0.6415440866166368,
-        "median_context_kld": 0.016576553745145464,
-        "median_context_p99": 0.13887158036231995,
-        "p90_context_kld": 0.0612287023900835,
+        "max_kld": 25.42261505126953,
+        "mean_kld": 0.031161018371537674,
+        "mean_ref_top1_prob": 0.6415473983687051,
+        "median_context_kld": 0.016547730692520052,
+        "median_context_p99": 0.14009475708007812,
+        "p90_context_kld": 0.06096497974345977,
         "positions": 1572096,
-        "top1_agreement": 0.9387569206969549,
+        "top1_agreement": 0.9386646871437877,
         "worst_context_id": 454,
-        "worst_context_kld": 0.7412628312695106
+        "worst_context_kld": 0.7633895164572437
       }
     },
     "primary": "deployed",
@@ -165,201 +168,201 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 28.042184829711914,
-            "mean_kld": 0.09667677342740076,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.06543842591323477,
-            "median_context_p99": 1.2771141529083252,
-            "p90_context_kld": 0.21150114342673582,
+            "max_kld": 25.42261505126953,
+            "mean_kld": 0.09701458859996956,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.06514060327445419,
+            "median_context_p99": 1.2380735874176025,
+            "p90_context_kld": 0.2367520687257336,
             "positions": 196512,
-            "top1_agreement": 0.9359275769418661,
+            "top1_agreement": 0.9359581094284318,
             "worst_context_id": 454,
-            "worst_context_kld": 0.7412628312695106
+            "worst_context_kld": 0.7633895164572437
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 3.108367390382808
+        "relative_to_run": 3.1133317737966557
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.694039344787598,
-            "mean_kld": 0.045539586959129484,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.028735511319572565,
-            "median_context_p99": 0.21844643354415894,
-            "p90_context_kld": 0.08768965578529091,
+            "max_kld": 7.141540050506592,
+            "mean_kld": 0.04554699851371887,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.028760450221885604,
+            "median_context_p99": 0.22136780619621277,
+            "p90_context_kld": 0.08768560320948965,
             "positions": 147384,
-            "top1_agreement": 0.9074865657059111,
+            "top1_agreement": 0.9072898007924877,
             "worst_context_id": 893,
-            "worst_context_kld": 0.1958269979242176
+            "worst_context_kld": 0.19577244735318067
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.4641962289066237
+        "relative_to_run": 1.461665917674927
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 5.444983005523682,
-            "mean_kld": 0.0287845435718118,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.02293439457839732,
-            "median_context_p99": 0.22981658577919006,
-            "p90_context_kld": 0.04189508559087464,
+            "max_kld": 5.054895877838135,
+            "mean_kld": 0.02878963941388972,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.023144045316762365,
+            "median_context_p99": 0.23143187165260315,
+            "p90_context_kld": 0.04200163334689099,
             "positions": 196512,
-            "top1_agreement": 0.9313884139390979,
+            "top1_agreement": 0.9312357515062694,
             "worst_context_id": 65,
-            "worst_context_kld": 0.14457800293637935
+            "worst_context_kld": 0.14448823853579743
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.9254853406217824
+        "relative_to_run": 0.923899182967205
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.748706817626953,
-            "mean_kld": 0.02181122798031287,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.016825883979479615,
-            "median_context_p99": 0.14032632112503052,
-            "p90_context_kld": 0.044644123456438076,
+            "max_kld": 7.8731608390808105,
+            "mean_kld": 0.0218276826276566,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.016630140947741442,
+            "median_context_p99": 0.1429976224899292,
+            "p90_context_kld": 0.04534338454113899,
             "positions": 147384,
-            "top1_agreement": 0.9390775118058948,
+            "top1_agreement": 0.9387925419312816,
             "worst_context_id": 275,
-            "worst_context_kld": 0.06380190983164141
+            "worst_context_kld": 0.06302050823430973
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.7012781601479685
+        "relative_to_run": 0.7004804004606571
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 6.197922706604004,
-            "mean_kld": 0.020862326095255422,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.01816635310248603,
-            "median_context_p99": 0.14613811671733856,
-            "p90_context_kld": 0.03870429632334617,
+            "max_kld": 6.146731376647949,
+            "mean_kld": 0.02090740290564167,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.017961478379677633,
+            "median_context_p99": 0.15575742721557617,
+            "p90_context_kld": 0.03910626657243996,
             "positions": 73692,
-            "top1_agreement": 0.9396949465342235,
+            "top1_agreement": 0.9397085165282527,
             "worst_context_id": 953,
-            "worst_context_kld": 0.04634572315564929
+            "worst_context_kld": 0.0471025725329215
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.6707689119426558
+        "relative_to_run": 0.6709473566094487
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.178435802459717,
-            "mean_kld": 0.020749322754068735,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.017448991748261974,
-            "median_context_p99": 0.10535755008459091,
-            "p90_context_kld": 0.024722712384587904,
+            "max_kld": 5.839554786682129,
+            "mean_kld": 0.020770561889322404,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.01741421428303659,
+            "median_context_p99": 0.10827115178108215,
+            "p90_context_kld": 0.0246539960080243,
             "positions": 147384,
-            "top1_agreement": 0.930820170439125,
+            "top1_agreement": 0.9303791456331759,
             "worst_context_id": 443,
-            "worst_context_kld": 0.13945793604077322
+            "worst_context_kld": 0.13891611533058892
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.667135610082279
+        "relative_to_run": 0.6665559399141504
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 5.842103004455566,
-            "mean_kld": 0.017170366302393433,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.011055941157220542,
-            "median_context_p99": 0.1271955668926239,
-            "p90_context_kld": 0.031480838601946864,
+            "max_kld": 13.393780708312988,
+            "mean_kld": 0.017258125689529703,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.011162757498727624,
+            "median_context_p99": 0.13722017407417297,
+            "p90_context_kld": 0.03180781198777742,
             "positions": 196512,
-            "top1_agreement": 0.9632948624002605,
+            "top1_agreement": 0.9635849210226347,
             "worst_context_id": 667,
-            "worst_context_kld": 0.1262930956785631
+            "worst_context_kld": 0.12542341486746875
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.5520644184031136
+        "relative_to_run": 0.5538370243153925
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 6.695211887359619,
-            "mean_kld": 0.015954869541938466,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.011525465745476264,
-            "median_context_p99": 0.16844584047794342,
-            "p90_context_kld": 0.030351257627919313,
+            "max_kld": 7.090137481689453,
+            "mean_kld": 0.015983325490243932,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.011221187727829521,
+            "median_context_p99": 0.1766842156648636,
+            "p90_context_kld": 0.02967326012736375,
             "positions": 73692,
-            "top1_agreement": 0.8993784942734625,
+            "top1_agreement": 0.8989578244585572,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.09074432302239947
+            "worst_context_kld": 0.09236369531223719
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.5129835682736726
+        "relative_to_run": 0.5129269300403553
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 23.66070556640625,
-            "mean_kld": 0.01358403856720618,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.010927746679615437,
-            "median_context_p99": 0.07393613457679749,
-            "p90_context_kld": 0.020596134704460638,
+            "max_kld": 17.859786987304688,
+            "mean_kld": 0.01356750565584152,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.010833755501125017,
+            "median_context_p99": 0.07359921932220459,
+            "p90_context_kld": 0.020571213900340293,
             "positions": 196512,
-            "top1_agreement": 0.9565624491125224,
+            "top1_agreement": 0.9566845790587852,
             "worst_context_id": 825,
-            "worst_context_kld": 0.04992035635412224
+            "worst_context_kld": 0.05003944289278355
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.4367562240139744
+        "relative_to_run": 0.4353999440606863
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 10.160826683044434,
-            "mean_kld": 0.012719578009616381,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.012256373709563865,
-            "median_context_p99": 0.09276100993156433,
-            "p90_context_kld": 0.016576553745145464,
+            "max_kld": 10.008549690246582,
+            "mean_kld": 0.012715332191590384,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.012210063005714826,
+            "median_context_p99": 0.09308900684118271,
+            "p90_context_kld": 0.016547730692520052,
             "positions": 196512,
-            "top1_agreement": 0.9501913369158117,
+            "top1_agreement": 0.9500081419964175,
             "worst_context_id": 246,
-            "worst_context_kld": 0.023991239175797346
+            "worst_context_kld": 0.02355286353559141
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.40896194714454503
+        "relative_to_run": 0.4080525238290834
       }
     ]
   },

Qwen3.8-27B-AWQ-INT4/inspect.json

@@ -1,31 +1,37 @@
 {
+  "model": "/media/fmodels2/cyankiwi/Qwen3.8-27B-AWQ-INT4",
+  "inspect_version": 5,
+  "weights_sha256": "7efc0eee36683b616eea4d1147b381e07bf0f1a359abc150aa43f74126a38d63",
+  "weights_bytes": 21018000928,
+  "quant_method": "compressed-tensors",
+  "declared": {},
+  "detected_scheme": "int4_g32_asym",
+  "detected_block": 32,
+  "quant_algorithm": "round_to_nearest",
+  "scheme_mix": null,
+  "kv_cache_scheme": null,
   "coverage": {
-    "attention": {
-      "quantized": 64,
-      "weights": 4
-    },
-    "dense_mlp": {
-      "quantized": 192,
-      "weights": 3
+    "router": {
+      "weights": 0,
+      "quantized": 0
     },
     "experts": {
-      "quantized": 0,
-      "weights": 0
-    },
-    "router": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
     },
     "shared_expert": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
+    },
+    "attention": {
+      "weights": 68,
+      "quantized": 64
+    },
+    "dense_mlp": {
+      "weights": 195,
+      "quantized": 192
     }
   },
-  "declared": {},
-  "detected_block": null,
-  "detected_scheme": null,
-  "model": "/media/fmodels2/cyankiwi/Qwen3.8-27B-AWQ-INT4",
-  "quant_method": "compressed-tensors",
-  "weights_bytes": 21018000928,
-  "weights_bytes_source": "hub"
+  "unloadable_reason": null,
+  "quantized_names_sha256": "eccbe240ef5caa18a896c8a311d4a9b4257c68423f0d8f9ae1c982d53797e657"
 }

Qwen3.8-27B-AWQ-INT4/report.md

@@ -1,12 +1,12 @@
 # Qwen3.8-27B / Qwen3.8-27B-AWQ-INT4: distribution fidelity
 
-**Mean KLD(reference || candidate) = 0.03110211** over 1572096 scored positions.
+**Mean KLD(reference || candidate) = 0.03116102** over 1572096 scored positions.
 
 | Mean | Median | p90 | p99 | Max | Top-1 agreement |
 |---|---|---|---|---|---|
-| 0.03110211 | 0.00686589 | 0.04071126 | 0.37582302 | 28.04218483 | 93.8757% |
+| 0.03116102 | 0.00686925 | 0.04070211 | 0.37613413 | 25.42261505 | 93.8665% |
 
-Reverse direction, KLD(candidate || reference): 0.03041559.
+Reverse direction, KLD(candidate || reference): 0.03049716.
 
 ## Identity
 
@@ -15,10 +15,10 @@ Reverse direction, KLD(candidate || reference): 0.03041559.
 | Reference checkpoint | /media/fmodels2/Qwen/Qwen3.8-27B |
 | Reference config SHA-256 | 191e0af23210 |
 | Candidate checkpoint | /media/fmodels2/cyankiwi/Qwen3.8-27B-AWQ-INT4 |
-| Candidate weights SHA-256 | unbound (Law 16 gap) |
+| Candidate weights SHA-256 | 7efc0eee36683b61 |
 | Suite | qwen3.8-27b-fidelity-1024x2048-v1 |
 | Suite token SHA-256 | 9c935708bbffbf45 |
-| Capture manifest SHA-256 | 304794b4e35b326d |
+| Capture manifest SHA-256 | dbfacc9cbb6d3e06 |
 | Tokenizer | None |
 | Scored vocabulary | 248044 |
 | Declared vocabulary | 248320 |
@@ -27,32 +27,48 @@ Reverse direction, KLD(candidate || reference): 0.03041559.
 | Model runner | V1 |
 | Tensor parallel | 1 |
 | Eager enforced | True |
+| KV cache | bfloat16 |
 | Prefix caching | False |
 | max_num_seqs | 1 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45 |
+| vLLM commit | 60071d1ab732 |
+| vLLM dirty digest | e3b0c44298fc1c14 |
+| Numerics digest | 251a9225b37415b9 |
+| Compiled extensions | f2fbc7537b0f01f6 |
+| FlashInfer | 0.6.17 |
 | torch | 2.13.0+cu132 |
 | Driver | 580.173.02 |
 | GPUs | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
-| Laws version | 9 |
+| Laws version | 15 |
 | Partition | analysis |
 
+The commit says when this number was taken. The numerics digest says what took it: a hash of the runtime and the scorer, which moves only when code that can change a logit changes. Comparability is bounded by the digest, so a result stays current across commits that cannot reach a number, and two results carrying the same digest were computed by the same code whatever their commits say.
+
+## Expert kernels: declared against built
+
+| Property | Value |
+|---|---|
+| Declared for its experts | `unquantized` |
+| Expert implementation built | n/a |
+| Expert kernel built | n/a |
+| Expert layers carrying an activation scale | n/a |
+
 ## Fidelity by domain
 
-The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
+The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `deployed` is QxQ (natural student routing). `bxq` is the teacher-ID counterfactual on the same weights. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
 
 | Domain | Contexts | Reference top-1 | Mean KLD | x run | Worst context |
 |---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.09667677 | 3.11 | 0.74126283 |
-| Chinese across several content types | 72 | 55.6% | 0.04553959 | 1.46 | 0.19582700 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.02878454 | 0.93 | 0.14457800 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.02181123 | 0.70 | 0.06380191 |
-| Other multilingual content | 36 | 64.5% | 0.02086233 | 0.67 | 0.04634572 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.02074932 | 0.67 | 0.13945794 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.01717037 | 0.55 | 0.12629310 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.01595487 | 0.51 | 0.09074432 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.01358404 | 0.44 | 0.04992036 |
-| Scientific and technical exposition | 96 | 60.3% | 0.01271958 | 0.41 | 0.02399124 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.09701459 | 3.11 | 0.76338952 |
+| Chinese across several content types | 72 | 55.6% | 0.04554700 | 1.46 | 0.19577245 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.02878964 | 0.92 | 0.14448824 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.02182768 | 0.70 | 0.06302051 |
+| Other multilingual content | 36 | 64.5% | 0.02090740 | 0.67 | 0.04710257 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.02077056 | 0.67 | 0.13891612 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.01725813 | 0.55 | 0.12542341 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.01598333 | 0.51 | 0.09236370 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.01356751 | 0.44 | 0.05003944 |
+| Scientific and technical exposition | 96 | 60.3% | 0.01271533 | 0.41 | 0.02355286 |
 
 **Natural dialogue, instruction following, and assistance** is this candidate's weakest domain and **Scientific and technical exposition** its strongest, a spread of 7.6x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
 
@@ -60,8 +76,8 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Component | Value |
 |---|---|
-| Trunk (candidate hidden states, reference head) | 0.03110211 |
-| Deployed (candidate's own head) | 0.03110211 |
+| Trunk (candidate hidden states, reference head) | 0.03116103 |
+| Deployed (candidate's own head) | 0.03116102 |
 | Head-associated delta (not additive) | -0.00000001 |
 | Reference head | {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]} |
 | Candidate head | {'runtime': {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]}, 'state': 'unquantized', 'static': {'ignored': True, 'lm_head_dtypes': {'lm_head.weight': 'BF16', 'model.language_model.embed_tokens.weight': 'BF16'}, 'lm_head_keys': ['lm_head.weight'], 'output_weight_keys': ['lm_head.weight'], 'packed_keys': [], 'quant_method': 'compressed-tensors', 'state': 'unquantized', 'tie_word_embeddings': False}} |
@@ -70,26 +86,26 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Position range | Positions | Mean KLD |
 |---|---|---|
-| 0–511 | 393216 | 0.02493864 |
-| 512–1023 | 393216 | 0.02846178 |
-| 1024–1535 | 393216 | 0.03394437 |
-| 1536–2046 | 392448 | 0.03707531 |
+| 0–511 | 393216 | 0.02499198 |
+| 512–1023 | 393216 | 0.02859100 |
+| 1024–1535 | 393216 | 0.03378435 |
+| 1536–2046 | 392448 | 0.03728870 |
 
 ## Error by reference confidence
 
 | Reference top-1 probability | Positions | Share | Mean KLD |
 |---|---|---|---|
-| [0.00, 0.25) | 240569 | 15.3% | 0.03377519 |
-| [0.25, 0.50) | 346390 | 22.0% | 0.04387957 |
-| [0.50, 0.75) | 273384 | 17.4% | 0.04991703 |
-| [0.75, 0.95) | 249939 | 15.9% | 0.03557873 |
-| [0.95, 1.00) | 461814 | 29.4% | 0.00656489 |
+| [0.00, 0.25) | 240522 | 15.3% | 0.03373890 |
+| [0.25, 0.50) | 346508 | 22.0% | 0.04475667 |
+| [0.50, 0.75) | 273204 | 17.4% | 0.04847260 |
+| [0.75, 0.95) | 250041 | 15.9% | 0.03574825 |
+| [0.95, 1.00) | 461821 | 29.4% | 0.00689268 |
 
 ## Top-K set agreement
 
 | K=1 | K=2 | K=3 | K=4 | K=5 |
 |---|---|---|---|---|
-| 93.6654% | 79.7744% | 62.9561% | 46.7627% | 33.4554% |
+| 93.6671% | 79.7419% | 62.9510% | 46.7486% | 33.4278% |
 
 ## Law compliance
 
@@ -101,22 +117,19 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | 2 | Determinism | PASS | eager enforced; prefix caching off, max_num_seqs=1 |
 | 3 | Frozen input | PASS | token hash matches suite qwen3.8-27b-fidelity-1024x2048-v1 [analysis]: 9c935708bbffbf45 |
 | 4 | Real vocabulary | PASS | scored 248044 real tokens of 248320 declared (276 padding rows) |
-| 5 | Manifest binding | PASS | all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc |
-| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s) |
+| 5 | Manifest binding | PASS | all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3 |
+| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s) |
 | 7 | Storage integrity | PASS | hidden storage with bitwise-exact replay |
-| 8 | Head transparency | PASS | trunk 0.03110211, deployed 0.03110211, delta -8.21671395159762e-09 |
-| 9 | Tail and depth disclosure | PASS | mean 0.03110211, median 0.00686589, max 28.04218483, 4 depth buckets |
+| 8 | Head transparency | PASS | trunk 0.03116103, deployed 0.03116102, delta -8.224659543698554e-09 |
+| 9 | Tail and depth disclosure | PASS | mean 0.03116102, median 0.00686925, max 25.42261505, 4 depth buckets |
 | 10 | Comparability | PASS | comparability key fully resolved |
 | 11 | Freeze before qualification | NOT_APPLICABLE | partition is 'analysis' |
 | 12 | Reusable reference | PASS | suite, reference, head, and checksums present; published reference matches the scored capture on every bound field |
-| 14 | Component attribution | NOT_APPLICABLE | reference declares no experts |
-| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.09667677, strongest scientific_technical at 0.01271958, spread 7.6x |
-| 16 | Candidate weight binding | OVERRIDE | the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else |
-| 13 | Recorded deviation | PASS | 1 override(s), each fully attributed |
-
-### Recorded deviations
-
-- **Law 16 (Candidate weight binding)** overridden by Andy Kitzke at 2026-09-02T05:15:00Z. Justification: These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger. Underlying finding: the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else.
+| 14 | Routed-model intervention | NOT_APPLICABLE | reference declares no experts |
+| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.09701459, strongest scientific_technical at 0.01271533, spread 7.6x |
+| 16 | Candidate weight binding | PASS | scored weights 7efc0eee36683b61 as inspected |
+| 17 | Substitution disclosure | PASS | every scored layer used the checkpoint's own quantization parameters |
+| 13 | Recorded deviation | PASS | no overrides claimed |
 
 ## Environment
 
@@ -126,7 +139,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | Platform | Linux-6.8.0-137-generic-x86_64-with-glibc2.39 |
 | Python | 3.12.3 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45fbe0b0205e70893bcd28fa928a88 |
+| vLLM commit | 60071d1ab73229321712254b1350dcaaac1c125a |
 | torch | 2.13.0+cu132 |
 | torch CUDA runtime | 13.2 |
 | cuDNN | 9.20.0 (92000) |
@@ -134,7 +147,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | CUDA arch list | sm_75, sm_80, sm_86, sm_90, sm_100, sm_120 |
 | nvcc | Cuda compilation tools, release 13.0, V13.0.88 |
 | gcc | gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0 |
-| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.8) 2.39 |
+| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.9) 2.39 |
 | NVIDIA driver | 580.173.02 |
 | float32 matmul precision | highest |
 | TF32 (matmul / cuDNN) | False / True |
@@ -157,8 +170,18 @@ Credential values are never published. Set but redacted: `HF_TOKEN`.
 
 | Variable | Value |
 |---|---|
+| `CUBLAS_WORKSPACE_CONFIG` | `:4096:8` |
 | `CUDA_HOME` | `/usr/local/cuda-13.0` |
 | `HF_TOKEN` | `<redacted>` |
+| `NCCL_DETERMINISTIC` | `1` |
+| `PYTORCH_NVML_BASED_CUDA_CHECK` | `1` |
+| `TORCHINDUCTOR_CACHE_DIR` | `/tmp/torchinductor_phaedawg` |
+| `TORCHINDUCTOR_COMPILE_THREADS` | `1` |
+| `TRITON_CACHE_AUTOTUNING` | `1` |
+| `TRITON_PTXAS_BLACKWELL_PATH` | `/usr/local/cuda-13.0/bin/ptxas` |
+| `VLLM_BATCH_INVARIANT` | `1` |
+| `VLLM_MARLIN_USE_ATOMIC_ADD` | `0` |
+| `VLLM_MOE_USE_DEEP_GEMM` | `0` |
 
 ## Files in this artifact
 
@@ -166,26 +189,26 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 
 | Path | Size | What it is |
 |---|---|---|
-| `Qwen3.8-27B-AWQ-INT4/report.md` | 12.34 KiB | This document. |
-| `Qwen3.8-27B-AWQ-INT4/report.json` | 254.91 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
-| `Qwen3.8-27B-AWQ-INT4/manifest.json` | 85.43 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
-| `Qwen3.8-27B-AWQ-INT4/compliance.json` | 13.04 KiB | The law-by-law receipt, including the comparability key. |
-| `baselines/self-kld.json` | 4.74 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
+| `Qwen3.8-27B-AWQ-INT4/report.md` | 13.18 KiB | This document. |
+| `Qwen3.8-27B-AWQ-INT4/report.json` | 257.43 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
+| `Qwen3.8-27B-AWQ-INT4/manifest.json` | 86.99 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
+| `Qwen3.8-27B-AWQ-INT4/compliance.json` | 12.30 KiB | The law-by-law receipt, including the comparability key. |
+| `baselines/self-kld.json` | 7.06 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
 | `suite/suite-manifest.json` | 382.68 KiB | The frozen evaluation input's identity: token hashes per context and per partition, sources, strata, and the analysis/qualification split. |
 | `suite/tokens` | 15.49 MiB in 1024 files | The token IDs themselves. These are the evaluation input, not a description of it; retokenizing source text does not reproduce them. |
 | `suite/sources.json` | 737.42 KiB | Per-context provenance: dataset, revision, licence, source unit, and the deterministic token offset chosen within the document. |
 | `suite/validation/capability-overlap.json` | 1.48 KiB | The benchmark-contamination scan and every document it blocked. |
-| `reference/manifest.json` | 85.43 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
+| `reference/manifest.json` | 86.99 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
 | `reference` | 19.25 GiB in 771 files | The reusable reference distributions. Pass this directory as `--reference-logits` to score a new candidate against the same reference without loading the reference checkpoint. |
 | `reference/lm_head.safetensors` | 2.37 GiB | The reference language-model head, which turns the stored hidden states back into reference logits. |
-| `environment/runtime.json` | 2.26 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
+| `environment/runtime.json` | 5.42 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
 | `environment/summary.md` | 1.18 KiB | The same provenance as prose, plus an index of every captured file. |
 | `environment/toolchain-nvcc.txt` | 243 B | `nvcc --version` verbatim; `toolchain-gcc.txt` and `toolchain-ldd.txt` sit beside it. |
 | `environment/gpu-smi-query.txt` | 54.73 KiB | `nvidia-smi -q` verbatim: ECC state, persistence mode, clocks, and throttle reasons, any of which can move a bitwise result. |
-| `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
-| `environment/models` | 7.22 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
-| `checksums.txt` | 182.99 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
+| `environment/pip-freeze.txt` | 4.47 KiB | Every installed package version in the scoring environment. |
+| `environment/models` | 19.46 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
+| `checksums.txt` | 183.22 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
+| `LAWS.md` | 40.20 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

Qwen3.8-27B-AWQ-INT4/strata.json

@@ -8,381 +8,381 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 28.042184829711914,
-            "mean_kld": 0.09667677342740076,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.06543842591323477,
-            "median_context_p99": 1.2771141529083252,
-            "p90_context_kld": 0.21150114342673582,
+            "max_kld": 25.42261505126953,
+            "mean_kld": 0.09701458859996956,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.06514060327445419,
+            "median_context_p99": 1.2380735874176025,
+            "p90_context_kld": 0.2367520687257336,
             "positions": 196512,
-            "top1_agreement": 0.9359275769418661,
+            "top1_agreement": 0.9359581094284318,
             "worst_context_id": 454,
-            "worst_context_kld": 0.7412628312695106
+            "worst_context_kld": 0.7633895164572437
           }
         },
         "key": "wildchat",
         "label": "wildchat",
-        "relative_to_run": 3.108367390382808
+        "relative_to_run": 3.1133317737966557
       },
       {
         "cells": {
           "deployed": {
             "contexts": 33,
-            "max_kld": 7.694039344787598,
-            "mean_kld": 0.07473911160197923,
-            "mean_ref_top1_prob": 0.6040606719329803,
-            "median_context_kld": 0.05613039184561264,
-            "median_context_p99": 0.4171757102012634,
-            "p90_context_kld": 0.17763363437416263,
+            "max_kld": 7.141540050506592,
+            "mean_kld": 0.07477739616771742,
+            "mean_ref_top1_prob": 0.6040861921682756,
+            "median_context_kld": 0.056872185852744755,
+            "median_context_p99": 0.40756842494010925,
+            "p90_context_kld": 0.17603352375213194,
             "positions": 67551,
-            "top1_agreement": 0.8850202069547453,
+            "top1_agreement": 0.8846501162084943,
             "worst_context_id": 893,
-            "worst_context_kld": 0.1958269979242176
+            "worst_context_kld": 0.19577244735318067
           }
         },
         "key": "wikisource_zh",
         "label": "wikisource_zh",
-        "relative_to_run": 2.403024108621409
+        "relative_to_run": 2.3997096396573077
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 4.922831058502197,
-            "mean_kld": 0.035849603841147874,
-            "mean_ref_top1_prob": 0.7357366218042676,
-            "median_context_kld": 0.03870429632334617,
-            "median_context_p99": 0.6272720098495483,
-            "p90_context_kld": 0.04634572315564929,
+            "max_kld": 5.16567325592041,
+            "mean_kld": 0.036092572653876616,
+            "mean_ref_top1_prob": 0.7358200961573914,
+            "median_context_kld": 0.03910626657243996,
+            "median_context_p99": 0.646285355091095,
+            "p90_context_kld": 0.0471025725329215,
             "positions": 14329,
-            "top1_agreement": 0.9298625165747785,
+            "top1_agreement": 0.9294437853304487,
             "worst_context_id": 953,
-            "worst_context_kld": 0.04634572315564929
+            "worst_context_kld": 0.0471025725329215
           }
         },
         "key": "wikipedia_de",
         "label": "wikipedia_de",
-        "relative_to_run": 1.1526423109439754
+        "relative_to_run": 1.158260369527698
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 5.373136520385742,
-            "mean_kld": 0.031211160784369987,
-            "mean_ref_top1_prob": 0.7232628548554616,
-            "median_context_kld": 0.03553082132454644,
-            "median_context_p99": 0.4050675630569458,
-            "p90_context_kld": 0.060331689381902705,
+            "max_kld": 5.600569248199463,
+            "mean_kld": 0.031223567048228287,
+            "mean_ref_top1_prob": 0.7233191001139795,
+            "median_context_kld": 0.03542479524833575,
+            "median_context_p99": 0.3749263882637024,
+            "p90_context_kld": 0.060297125465259956,
             "positions": 47081,
-            "top1_agreement": 0.9426307852424545,
+            "top1_agreement": 0.9418236655975871,
             "worst_context_id": 275,
-            "worst_context_kld": 0.06380190983164141
+            "worst_context_kld": 0.06302050823430973
           }
         },
         "key": "regulations",
         "label": "regulations",
-        "relative_to_run": 1.0035063331006198
+        "relative_to_run": 1.002007273188085
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 5.444983005523682,
-            "mean_kld": 0.0287845435718118,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.02293439457839732,
-            "median_context_p99": 0.22981658577919006,
-            "p90_context_kld": 0.04189508559087464,
+            "max_kld": 5.054895877838135,
+            "mean_kld": 0.02878963941388972,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.023144045316762365,
+            "median_context_p99": 0.23143187165260315,
+            "p90_context_kld": 0.04200163334689099,
             "positions": 196512,
-            "top1_agreement": 0.9313884139390979,
+            "top1_agreement": 0.9312357515062694,
             "worst_context_id": 65,
-            "worst_context_kld": 0.14457800293637935
+            "worst_context_kld": 0.14448823853579743
           }
         },
         "key": "wikipedia_en",
         "label": "wikipedia_en",
-        "relative_to_run": 0.9254853406217824
+        "relative_to_run": 0.923899182967205
       },
       {
         "cells": {
           "deployed": {
             "contexts": 39,
-            "max_kld": 4.944775104522705,
-            "mean_kld": 0.020832296876718157,
-            "mean_ref_top1_prob": 0.515358929042611,
-            "median_context_kld": 0.01954589701205417,
-            "median_context_p99": 0.14377470314502716,
-            "p90_context_kld": 0.030699087382658682,
+            "max_kld": 4.685449123382568,
+            "mean_kld": 0.020813585114181635,
+            "mean_ref_top1_prob": 0.5153760832971095,
+            "median_context_kld": 0.01960896569649871,
+            "median_context_p99": 0.15238429605960846,
+            "p90_context_kld": 0.030369273721799726,
             "positions": 79833,
-            "top1_agreement": 0.9264965615722821,
+            "top1_agreement": 0.9264464569789436,
             "worst_context_id": 880,
-            "worst_context_kld": 0.03422051701032839
+            "worst_context_kld": 0.034152012582933926
           }
         },
         "key": "wikipedia_zh",
         "label": "wikipedia_zh",
-        "relative_to_run": 0.6698034076094974
+        "relative_to_run": 0.6679366144590662
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.178435802459717,
-            "mean_kld": 0.020749322754068735,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.017448991748261974,
-            "median_context_p99": 0.10535755008459091,
-            "p90_context_kld": 0.024722712384587904,
+            "max_kld": 5.839554786682129,
+            "mean_kld": 0.020770561889322404,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.01741421428303659,
+            "median_context_p99": 0.10827115178108215,
+            "p90_context_kld": 0.0246539960080243,
             "positions": 147384,
-            "top1_agreement": 0.930820170439125,
+            "top1_agreement": 0.9303791456331759,
             "worst_context_id": 443,
-            "worst_context_kld": 0.13945793604077322
+            "worst_context_kld": 0.13891611533058892
           }
         },
         "key": "public_domain_books",
         "label": "public_domain_books",
-        "relative_to_run": 0.667135610082279
+        "relative_to_run": 0.6665559399141504
       },
       {
         "cells": {
           "deployed": {
             "contexts": 26,
-            "max_kld": 7.748706817626953,
-            "mean_kld": 0.02022364180898306,
-            "mean_ref_top1_prob": 0.5182113024080622,
-            "median_context_kld": 0.01705912196093182,
-            "median_context_p99": 0.14032632112503052,
-            "p90_context_kld": 0.033483738416120845,
+            "max_kld": 7.8731608390808105,
+            "mean_kld": 0.020213784227122354,
+            "mean_ref_top1_prob": 0.5182496321630543,
+            "median_context_kld": 0.016992442443620782,
+            "median_context_p99": 0.1429976224899292,
+            "p90_context_kld": 0.033787311125783294,
             "positions": 53222,
-            "top1_agreement": 0.9344256134681147,
+            "top1_agreement": 0.9343880350231107,
             "worst_context_id": 312,
-            "worst_context_kld": 0.05213722910501061
+            "worst_context_kld": 0.052073676788751425
           }
         },
         "key": "public_domain_review",
         "label": "public_domain_review",
-        "relative_to_run": 0.6502338305801219
+        "relative_to_run": 0.6486881778416308
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 1.2518802881240845,
-            "mean_kld": 0.01883299186653763,
-            "mean_ref_top1_prob": 0.5702914711488575,
-            "median_context_kld": 0.01816635310248603,
-            "median_context_p99": 0.1268729716539383,
-            "p90_context_kld": 0.023229672955958933,
+            "max_kld": 1.2985306978225708,
+            "mean_kld": 0.018780109997211636,
+            "mean_ref_top1_prob": 0.5701150598950927,
+            "median_context_kld": 0.01790622692756403,
+            "median_context_p99": 0.12759386003017426,
+            "p90_context_kld": 0.023116627858478397,
             "positions": 14329,
-            "top1_agreement": 0.9350268685881779,
+            "top1_agreement": 0.9346779258845698,
             "worst_context_id": 968,
-            "worst_context_kld": 0.023229672955958933
+            "worst_context_kld": 0.023116627858478397
           }
         },
         "key": "wikipedia_ja",
         "label": "wikipedia_ja",
-        "relative_to_run": 0.6055214267701086
+        "relative_to_run": 0.6026795970944678
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 2.443483352661133,
-            "mean_kld": 0.018556687548381266,
-            "mean_ref_top1_prob": 0.591765751203999,
-            "median_context_kld": 0.018242773135174754,
-            "median_context_p99": 0.18581736087799072,
-            "p90_context_kld": 0.022747585317861696,
+            "max_kld": 2.3282008171081543,
+            "mean_kld": 0.018525593784418537,
+            "mean_ref_top1_prob": 0.591761624645417,
+            "median_context_kld": 0.018175120030234318,
+            "median_context_p99": 0.18835358321666718,
+            "p90_context_kld": 0.02282773946679107,
             "positions": 14329,
-            "top1_agreement": 0.9364226394026101,
+            "top1_agreement": 0.9375392560541559,
             "worst_context_id": 966,
-            "worst_context_kld": 0.022747585317861696
+            "worst_context_kld": 0.02282773946679107
           }
         },
         "key": "wikipedia_es",
         "label": "wikipedia_es",
-        "relative_to_run": 0.5966376452584702
+        "relative_to_run": 0.594511821261263
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 6.197922706604004,
-            "mean_kld": 0.018394448322257166,
-            "mean_ref_top1_prob": 0.6892386900239509,
-            "median_context_kld": 0.019485870194125742,
-            "median_context_p99": 0.14441317319869995,
-            "p90_context_kld": 0.02192389090120193,
+            "max_kld": 6.146731376647949,
+            "mean_kld": 0.018496405910725683,
+            "mean_ref_top1_prob": 0.689239984531695,
+            "median_context_kld": 0.019470748973926016,
+            "median_context_p99": 0.15875497460365295,
+            "p90_context_kld": 0.02206932290149989,
             "positions": 12282,
-            "top1_agreement": 0.9468327633935841,
+            "top1_agreement": 0.9460185637518319,
             "worst_context_id": 958,
-            "worst_context_kld": 0.02192389090120193
+            "worst_context_kld": 0.02206932290149989
           }
         },
         "key": "wikipedia_cs",
         "label": "wikipedia_cs",
-        "relative_to_run": 0.5914213031936019
+        "relative_to_run": 0.5935751421917652
       },
       {
         "cells": {
           "deployed": {
             "contexts": 52,
-            "max_kld": 5.504080772399902,
-            "mean_kld": 0.017945425675690362,
-            "mean_ref_top1_prob": 0.8582575125835092,
-            "median_context_kld": 0.010655520236978994,
-            "median_context_p99": 0.14233162999153137,
-            "p90_context_kld": 0.03161436646252504,
+            "max_kld": 4.879175662994385,
+            "mean_kld": 0.017993323901892008,
+            "mean_ref_top1_prob": 0.8582941084526002,
+            "median_context_kld": 0.010814381290457163,
+            "median_context_p99": 0.1424388885498047,
+            "p90_context_kld": 0.03237368735157185,
             "positions": 106444,
-            "top1_agreement": 0.9672128067340573,
+            "top1_agreement": 0.9676449588516027,
             "worst_context_id": 667,
-            "worst_context_kld": 0.1262930956785631
+            "worst_context_kld": 0.12542341486746875
           }
         },
         "key": "github_code",
         "label": "github_code",
-        "relative_to_run": 0.5769842538109002
+        "relative_to_run": 0.5774305475949086
       },
       {
         "cells": {
           "deployed": {
             "contexts": 44,
-            "max_kld": 5.842103004455566,
-            "mean_kld": 0.016254387043042516,
-            "mean_ref_top1_prob": 0.720394649593742,
-            "median_context_kld": 0.011693944644111916,
-            "median_context_p99": 0.1271955668926239,
-            "p90_context_kld": 0.02565605743086988,
+            "max_kld": 13.393780708312988,
+            "mean_kld": 0.016389255074919708,
+            "mean_ref_top1_prob": 0.7203910827522294,
+            "median_context_kld": 0.011781157981404973,
+            "median_context_p99": 0.13722017407417297,
+            "p90_context_kld": 0.029811882521120738,
             "positions": 90068,
-            "top1_agreement": 0.958664564551228,
+            "top1_agreement": 0.9587866944974908,
             "worst_context_id": 582,
-            "worst_context_kld": 0.10557822110925151
+            "worst_context_kld": 0.10451388947053483
           }
         },
         "key": "stackv2",
         "label": "stackv2",
-        "relative_to_run": 0.5226137038302747
+        "relative_to_run": 0.5259537695305098
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 6.695211887359619,
-            "mean_kld": 0.015954869541938466,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.011525465745476264,
-            "median_context_p99": 0.16844584047794342,
-            "p90_context_kld": 0.030351257627919313,
+            "max_kld": 7.090137481689453,
+            "mean_kld": 0.015983325490243932,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.011221187727829521,
+            "median_context_p99": 0.1766842156648636,
+            "p90_context_kld": 0.02967326012736375,
             "positions": 73692,
-            "top1_agreement": 0.8993784942734625,
+            "top1_agreement": 0.8989578244585572,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.09074432302239947
+            "worst_context_kld": 0.09236369531223719
           }
         },
         "key": "starcoder_structured",
         "label": "starcoder_structured",
-        "relative_to_run": 0.5129835682736726
+        "relative_to_run": 0.5129269300403553
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 2.0435266494750977,
-            "mean_kld": 0.01451226913676392,
-            "mean_ref_top1_prob": 0.6660615117323249,
-            "median_context_kld": 0.015073299627354781,
-            "median_context_p99": 0.13546854257583618,
-            "p90_context_kld": 0.018706599086283214,
+            "max_kld": 1.8480932712554932,
+            "mean_kld": 0.01452934850799381,
+            "mean_ref_top1_prob": 0.6660535411394803,
+            "median_context_kld": 0.015385738591614252,
+            "median_context_p99": 0.13763126730918884,
+            "p90_context_kld": 0.018730638469995477,
             "positions": 12282,
-            "top1_agreement": 0.9486240026054389,
+            "top1_agreement": 0.9487054225696141,
             "worst_context_id": 945,
-            "worst_context_kld": 0.018706599086283214
+            "worst_context_kld": 0.018730638469995477
           }
         },
         "key": "wikipedia_ru",
         "label": "wikipedia_ru",
-        "relative_to_run": 0.4666008447111725
+        "relative_to_run": 0.46626680600608506
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 2.7419142723083496,
-            "mean_kld": 0.014205957804715537,
-            "mean_ref_top1_prob": 0.5379292386352562,
-            "median_context_kld": 0.014389406218182936,
-            "median_context_p99": 0.10889367014169693,
-            "p90_context_kld": 0.019335491066219627,
+            "max_kld": 2.9592385292053223,
+            "mean_kld": 0.01425620509464537,
+            "mean_ref_top1_prob": 0.537879873360319,
+            "median_context_kld": 0.014296799834929014,
+            "median_context_p99": 0.10684597492218018,
+            "p90_context_kld": 0.019538807512203465,
             "positions": 47081,
-            "top1_agreement": 0.9407829060555213,
+            "top1_agreement": 0.9407404260742125,
             "worst_context_id": 343,
-            "worst_context_kld": 0.02154162281618613
+            "worst_context_kld": 0.022069511261969762
           }
         },
         "key": "open_news",
         "label": "open_news",
-        "relative_to_run": 0.4567522727937522
+        "relative_to_run": 0.45750125765038924
       },
       {
         "cells": {
           "deployed": {
             "contexts": 3,
-            "max_kld": 1.4447282552719116,
-            "mean_kld": 0.013642817294200454,
-            "mean_ref_top1_prob": 0.6023603953518782,
-            "median_context_kld": 0.013129884177378225,
-            "median_context_p99": 0.1089385598897934,
-            "p90_context_kld": 0.016000769355042722,
+            "max_kld": 1.622402310371399,
+            "mean_kld": 0.013574681014078556,
+            "mean_ref_top1_prob": 0.6023746847155801,
+            "median_context_kld": 0.013048632467336398,
+            "median_context_p99": 0.10719217360019684,
+            "p90_context_kld": 0.015888242927918444,
             "positions": 6141,
-            "top1_agreement": 0.9490311024263149,
+            "top1_agreement": 0.949845302068067,
             "worst_context_id": 967,
-            "worst_context_kld": 0.016000769355042722
+            "worst_context_kld": 0.015888242927918444
           }
         },
         "key": "wikipedia_fr",
         "label": "wikipedia_fr",
-        "relative_to_run": 0.43864608723302795
+        "relative_to_run": 0.4356302111896832
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 23.66070556640625,
-            "mean_kld": 0.01358403856720618,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.010927746679615437,
-            "median_context_p99": 0.07393613457679749,
-            "p90_context_kld": 0.020596134704460638,
+            "max_kld": 17.859786987304688,
+            "mean_kld": 0.01356750565584152,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.010833755501125017,
+            "median_context_p99": 0.07359921932220459,
+            "p90_context_kld": 0.020571213900340293,
             "positions": 196512,
-            "top1_agreement": 0.9565624491125224,
+            "top1_agreement": 0.9566845790587852,
             "worst_context_id": 825,
-            "worst_context_kld": 0.04992035635412224
+            "worst_context_kld": 0.05003944289278355
           }
         },
         "key": "libretexts",
         "label": "libretexts",
-        "relative_to_run": 0.4367562240139744
+        "relative_to_run": 0.4353999440606863
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 10.160826683044434,
-            "mean_kld": 0.012719578009616381,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.012256373709563865,
-            "median_context_p99": 0.09276100993156433,
-            "p90_context_kld": 0.016576553745145464,
+            "max_kld": 10.008549690246582,
+            "mean_kld": 0.012715332191590384,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.012210063005714826,
+            "median_context_p99": 0.09308900684118271,
+            "p90_context_kld": 0.016547730692520052,
             "positions": 196512,
-            "top1_agreement": 0.9501913369158117,
+            "top1_agreement": 0.9500081419964175,
             "worst_context_id": 246,
-            "worst_context_kld": 0.023991239175797346
+            "worst_context_kld": 0.02355286353559141
           }
         },
         "key": "scientific_papers",
         "label": "scientific_papers",
-        "relative_to_run": 0.40896194714454503
+        "relative_to_run": 0.4080525238290834
       }
     ],
     "stratum": [
@@ -390,217 +390,217 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 28.042184829711914,
-            "mean_kld": 0.09667677342740076,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.06543842591323477,
-            "median_context_p99": 1.2771141529083252,
-            "p90_context_kld": 0.21150114342673582,
+            "max_kld": 25.42261505126953,
+            "mean_kld": 0.09701458859996956,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.06514060327445419,
+            "median_context_p99": 1.2380735874176025,
+            "p90_context_kld": 0.2367520687257336,
             "positions": 196512,
-            "top1_agreement": 0.9359275769418661,
+            "top1_agreement": 0.9359581094284318,
             "worst_context_id": 454,
-            "worst_context_kld": 0.7412628312695106
+            "worst_context_kld": 0.7633895164572437
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 3.108367390382808
+        "relative_to_run": 3.1133317737966557
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.694039344787598,
-            "mean_kld": 0.045539586959129484,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.028735511319572565,
-            "median_context_p99": 0.21844643354415894,
-            "p90_context_kld": 0.08768965578529091,
+            "max_kld": 7.141540050506592,
+            "mean_kld": 0.04554699851371887,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.028760450221885604,
+            "median_context_p99": 0.22136780619621277,
+            "p90_context_kld": 0.08768560320948965,
             "positions": 147384,
-            "top1_agreement": 0.9074865657059111,
+            "top1_agreement": 0.9072898007924877,
             "worst_context_id": 893,
-            "worst_context_kld": 0.1958269979242176
+            "worst_context_kld": 0.19577244735318067
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.4641962289066237
+        "relative_to_run": 1.461665917674927
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 5.444983005523682,
-            "mean_kld": 0.0287845435718118,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.02293439457839732,
-            "median_context_p99": 0.22981658577919006,
-            "p90_context_kld": 0.04189508559087464,
+            "max_kld": 5.054895877838135,
+            "mean_kld": 0.02878963941388972,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.023144045316762365,
+            "median_context_p99": 0.23143187165260315,
+            "p90_context_kld": 0.04200163334689099,
             "positions": 196512,
-            "top1_agreement": 0.9313884139390979,
+            "top1_agreement": 0.9312357515062694,
             "worst_context_id": 65,
-            "worst_context_kld": 0.14457800293637935
+            "worst_context_kld": 0.14448823853579743
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.9254853406217824
+        "relative_to_run": 0.923899182967205
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.748706817626953,
-            "mean_kld": 0.02181122798031287,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.016825883979479615,
-            "median_context_p99": 0.14032632112503052,
-            "p90_context_kld": 0.044644123456438076,
+            "max_kld": 7.8731608390808105,
+            "mean_kld": 0.0218276826276566,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.016630140947741442,
+            "median_context_p99": 0.1429976224899292,
+            "p90_context_kld": 0.04534338454113899,
             "positions": 147384,
-            "top1_agreement": 0.9390775118058948,
+            "top1_agreement": 0.9387925419312816,
             "worst_context_id": 275,
-            "worst_context_kld": 0.06380190983164141
+            "worst_context_kld": 0.06302050823430973
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.7012781601479685
+        "relative_to_run": 0.7004804004606571
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 6.197922706604004,
-            "mean_kld": 0.020862326095255422,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.01816635310248603,
-            "median_context_p99": 0.14613811671733856,
-            "p90_context_kld": 0.03870429632334617,
+            "max_kld": 6.146731376647949,
+            "mean_kld": 0.02090740290564167,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.017961478379677633,
+            "median_context_p99": 0.15575742721557617,
+            "p90_context_kld": 0.03910626657243996,
             "positions": 73692,
-            "top1_agreement": 0.9396949465342235,
+            "top1_agreement": 0.9397085165282527,
             "worst_context_id": 953,
-            "worst_context_kld": 0.04634572315564929
+            "worst_context_kld": 0.0471025725329215
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.6707689119426558
+        "relative_to_run": 0.6709473566094487
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.178435802459717,
-            "mean_kld": 0.020749322754068735,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.017448991748261974,
-            "median_context_p99": 0.10535755008459091,
-            "p90_context_kld": 0.024722712384587904,
+            "max_kld": 5.839554786682129,
+            "mean_kld": 0.020770561889322404,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.01741421428303659,
+            "median_context_p99": 0.10827115178108215,
+            "p90_context_kld": 0.0246539960080243,
             "positions": 147384,
-            "top1_agreement": 0.930820170439125,
+            "top1_agreement": 0.9303791456331759,
             "worst_context_id": 443,
-            "worst_context_kld": 0.13945793604077322
+            "worst_context_kld": 0.13891611533058892
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.667135610082279
+        "relative_to_run": 0.6665559399141504
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 5.842103004455566,
-            "mean_kld": 0.017170366302393433,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.011055941157220542,
-            "median_context_p99": 0.1271955668926239,
-            "p90_context_kld": 0.031480838601946864,
+            "max_kld": 13.393780708312988,
+            "mean_kld": 0.017258125689529703,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.011162757498727624,
+            "median_context_p99": 0.13722017407417297,
+            "p90_context_kld": 0.03180781198777742,
             "positions": 196512,
-            "top1_agreement": 0.9632948624002605,
+            "top1_agreement": 0.9635849210226347,
             "worst_context_id": 667,
-            "worst_context_kld": 0.1262930956785631
+            "worst_context_kld": 0.12542341486746875
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.5520644184031136
+        "relative_to_run": 0.5538370243153925
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 6.695211887359619,
-            "mean_kld": 0.015954869541938466,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.011525465745476264,
-            "median_context_p99": 0.16844584047794342,
-            "p90_context_kld": 0.030351257627919313,
+            "max_kld": 7.090137481689453,
+            "mean_kld": 0.015983325490243932,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.011221187727829521,
+            "median_context_p99": 0.1766842156648636,
+            "p90_context_kld": 0.02967326012736375,
             "positions": 73692,
-            "top1_agreement": 0.8993784942734625,
+            "top1_agreement": 0.8989578244585572,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.09074432302239947
+            "worst_context_kld": 0.09236369531223719
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.5129835682736726
+        "relative_to_run": 0.5129269300403553
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 23.66070556640625,
-            "mean_kld": 0.01358403856720618,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.010927746679615437,
-            "median_context_p99": 0.07393613457679749,
-            "p90_context_kld": 0.020596134704460638,
+            "max_kld": 17.859786987304688,
+            "mean_kld": 0.01356750565584152,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.010833755501125017,
+            "median_context_p99": 0.07359921932220459,
+            "p90_context_kld": 0.020571213900340293,
             "positions": 196512,
-            "top1_agreement": 0.9565624491125224,
+            "top1_agreement": 0.9566845790587852,
             "worst_context_id": 825,
-            "worst_context_kld": 0.04992035635412224
+            "worst_context_kld": 0.05003944289278355
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.4367562240139744
+        "relative_to_run": 0.4353999440606863
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 10.160826683044434,
-            "mean_kld": 0.012719578009616381,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.012256373709563865,
-            "median_context_p99": 0.09276100993156433,
-            "p90_context_kld": 0.016576553745145464,
+            "max_kld": 10.008549690246582,
+            "mean_kld": 0.012715332191590384,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.012210063005714826,
+            "median_context_p99": 0.09308900684118271,
+            "p90_context_kld": 0.016547730692520052,
             "positions": 196512,
-            "top1_agreement": 0.9501913369158117,
+            "top1_agreement": 0.9500081419964175,
             "worst_context_id": 246,
-            "worst_context_kld": 0.023991239175797346
+            "worst_context_kld": 0.02355286353559141
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.40896194714454503
+        "relative_to_run": 0.4080525238290834
       }
     ]
   },
   "overall": {
     "deployed": {
       "contexts": 768,
-      "max_kld": 28.042184829711914,
-      "mean_kld": 0.031102106439063696,
-      "mean_ref_top1_prob": 0.6415440866166368,
-      "median_context_kld": 0.016576553745145464,
-      "median_context_p99": 0.13887158036231995,
-      "p90_context_kld": 0.0612287023900835,
+      "max_kld": 25.42261505126953,
+      "mean_kld": 0.031161018371537674,
+      "mean_ref_top1_prob": 0.6415473983687051,
+      "median_context_kld": 0.016547730692520052,
+      "median_context_p99": 0.14009475708007812,
+      "p90_context_kld": 0.06096497974345977,
       "positions": 1572096,
-      "top1_agreement": 0.9387569206969549,
+      "top1_agreement": 0.9386646871437877,
       "worst_context_id": 454,
-      "worst_context_kld": 0.7412628312695106
+      "worst_context_kld": 0.7633895164572437
     }
   },
   "primary": "deployed"

Qwen3.8-27B-AWQ-INT4/strata.md

@@ -1,52 +1,52 @@
 # Fidelity by domain - Qwen3.8-27B / Qwen3.8-27B-AWQ-INT4
 
-768 contexts, 1572096 scored positions, mean 0.03110211, reference top-1 64.2%, top-1 agreement 93.8757%.
+768 contexts, 1572096 scored positions, mean 0.03116102, reference top-1 64.2%, top-1 agreement 93.8665%.
 
-`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
+`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. The `deployed` cell is QxQ; `bxq` is the teacher-ID counterfactual on the same student weights. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
 
 ## Ranked by domain
 
 | Domain | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.09667677 | 3.11 | 0.06543843 | 0.21150114 | 0.741263 |
-| Chinese across several content types | 72 | 55.6% | 0.04553959 | 1.46 | 0.02873551 | 0.08768966 | 0.195827 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.02878454 | 0.93 | 0.02293439 | 0.04189509 | 0.144578 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.02181123 | 0.70 | 0.01682588 | 0.04464412 | 0.063802 |
-| Other multilingual content | 36 | 64.5% | 0.02086233 | 0.67 | 0.01816635 | 0.03870430 | 0.046346 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.02074932 | 0.67 | 0.01744899 | 0.02472271 | 0.139458 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.01717037 | 0.55 | 0.01105594 | 0.03148084 | 0.126293 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.01595487 | 0.51 | 0.01152547 | 0.03035126 | 0.090744 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.01358404 | 0.44 | 0.01092775 | 0.02059613 | 0.049920 |
-| Scientific and technical exposition | 96 | 60.3% | 0.01271958 | 0.41 | 0.01225637 | 0.01657655 | 0.023991 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.09701459 | 3.11 | 0.06514060 | 0.23675207 | 0.763390 |
+| Chinese across several content types | 72 | 55.6% | 0.04554700 | 1.46 | 0.02876045 | 0.08768560 | 0.195772 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.02878964 | 0.92 | 0.02314405 | 0.04200163 | 0.144488 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.02182768 | 0.70 | 0.01663014 | 0.04534338 | 0.063021 |
+| Other multilingual content | 36 | 64.5% | 0.02090740 | 0.67 | 0.01796148 | 0.03910627 | 0.047103 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.02077056 | 0.67 | 0.01741421 | 0.02465400 | 0.138916 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.01725813 | 0.55 | 0.01116276 | 0.03180781 | 0.125423 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.01598333 | 0.51 | 0.01122119 | 0.02967326 | 0.092364 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.01356751 | 0.44 | 0.01083376 | 0.02057121 | 0.050039 |
+| Scientific and technical exposition | 96 | 60.3% | 0.01271533 | 0.41 | 0.01221006 | 0.01654773 | 0.023553 |
 
 ## Ranked by source dataset
 
 | Source | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| wildchat | 96 | 73.5% | 0.09667677 | 3.11 | 0.06543843 | 0.21150114 | 0.741263 |
-| wikisource_zh | 33 | 60.4% | 0.07473911 | 2.40 | 0.05613039 | 0.17763363 | 0.195827 |
-| wikipedia_de | 7 | 73.6% | 0.03584960 | 1.15 | 0.03870430 | 0.04634572 | 0.046346 |
-| regulations | 23 | 72.3% | 0.03121116 | 1.00 | 0.03553082 | 0.06033169 | 0.063802 |
-| wikipedia_en | 96 | 59.5% | 0.02878454 | 0.93 | 0.02293439 | 0.04189509 | 0.144578 |
-| wikipedia_zh | 39 | 51.5% | 0.02083230 | 0.67 | 0.01954590 | 0.03069909 | 0.034221 |
-| public_domain_books | 72 | 50.6% | 0.02074932 | 0.67 | 0.01744899 | 0.02472271 | 0.139458 |
-| public_domain_review | 26 | 51.8% | 0.02022364 | 0.65 | 0.01705912 | 0.03348374 | 0.052137 |
-| wikipedia_ja | 7 | 57.0% | 0.01883299 | 0.61 | 0.01816635 | 0.02322967 | 0.023230 |
-| wikipedia_es | 7 | 59.2% | 0.01855669 | 0.60 | 0.01824277 | 0.02274759 | 0.022748 |
-| wikipedia_cs | 6 | 68.9% | 0.01839445 | 0.59 | 0.01948587 | 0.02192389 | 0.021924 |
-| github_code | 52 | 85.8% | 0.01794543 | 0.58 | 0.01065552 | 0.03161437 | 0.126293 |
-| stackv2 | 44 | 72.0% | 0.01625439 | 0.52 | 0.01169394 | 0.02565606 | 0.105578 |
-| starcoder_structured | 36 | 70.7% | 0.01595487 | 0.51 | 0.01152547 | 0.03035126 | 0.090744 |
-| wikipedia_ru | 6 | 66.6% | 0.01451227 | 0.47 | 0.01507330 | 0.01870660 | 0.018707 |
-| open_news | 23 | 53.8% | 0.01420596 | 0.46 | 0.01438941 | 0.01933549 | 0.021542 |
-| wikipedia_fr | 3 | 60.2% | 0.01364282 | 0.44 | 0.01312988 | 0.01600077 | 0.016001 |
-| libretexts | 96 | 65.8% | 0.01358404 | 0.44 | 0.01092775 | 0.02059613 | 0.049920 |
-| scientific_papers | 96 | 60.3% | 0.01271958 | 0.41 | 0.01225637 | 0.01657655 | 0.023991 |
+| wildchat | 96 | 73.5% | 0.09701459 | 3.11 | 0.06514060 | 0.23675207 | 0.763390 |
+| wikisource_zh | 33 | 60.4% | 0.07477740 | 2.40 | 0.05687219 | 0.17603352 | 0.195772 |
+| wikipedia_de | 7 | 73.6% | 0.03609257 | 1.16 | 0.03910627 | 0.04710257 | 0.047103 |
+| regulations | 23 | 72.3% | 0.03122357 | 1.00 | 0.03542480 | 0.06029713 | 0.063021 |
+| wikipedia_en | 96 | 59.5% | 0.02878964 | 0.92 | 0.02314405 | 0.04200163 | 0.144488 |
+| wikipedia_zh | 39 | 51.5% | 0.02081359 | 0.67 | 0.01960897 | 0.03036927 | 0.034152 |
+| public_domain_books | 72 | 50.6% | 0.02077056 | 0.67 | 0.01741421 | 0.02465400 | 0.138916 |
+| public_domain_review | 26 | 51.8% | 0.02021378 | 0.65 | 0.01699244 | 0.03378731 | 0.052074 |
+| wikipedia_ja | 7 | 57.0% | 0.01878011 | 0.60 | 0.01790623 | 0.02311663 | 0.023117 |
+| wikipedia_es | 7 | 59.2% | 0.01852559 | 0.59 | 0.01817512 | 0.02282774 | 0.022828 |
+| wikipedia_cs | 6 | 68.9% | 0.01849641 | 0.59 | 0.01947075 | 0.02206932 | 0.022069 |
+| github_code | 52 | 85.8% | 0.01799332 | 0.58 | 0.01081438 | 0.03237369 | 0.125423 |
+| stackv2 | 44 | 72.0% | 0.01638926 | 0.53 | 0.01178116 | 0.02981188 | 0.104514 |
+| starcoder_structured | 36 | 70.7% | 0.01598333 | 0.51 | 0.01122119 | 0.02967326 | 0.092364 |
+| wikipedia_ru | 6 | 66.6% | 0.01452935 | 0.47 | 0.01538574 | 0.01873064 | 0.018731 |
+| open_news | 23 | 53.8% | 0.01425621 | 0.46 | 0.01429680 | 0.01953881 | 0.022070 |
+| wikipedia_fr | 3 | 60.2% | 0.01357468 | 0.44 | 0.01304863 | 0.01588824 | 0.015888 |
+| libretexts | 96 | 65.8% | 0.01356751 | 0.44 | 0.01083376 | 0.02057121 | 0.050039 |
+| scientific_papers | 96 | 60.3% | 0.01271533 | 0.41 | 0.01221006 | 0.01654773 | 0.023553 |
 
 ## Reading
 
-- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.09667677, 3.11x the run mean of 0.03110211 over 96 context(s).
-- Strongest domain: **Scientific and technical exposition** at 0.01271958, 0.41x the run mean. The spread across domains is 7.6x.
+- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.09701459, 3.11x the run mean of 0.03116102 over 96 context(s).
+- Strongest domain: **Scientific and technical exposition** at 0.01271533, 0.41x the run mean. The spread across domains is 7.6x.
 - This is a strong finding: the reference was *more* certain on Natural dialogue, instruction following, and assistance (top-1 73.5% against 64.2% overall) and the candidate still diverged most there, so predictability does not explain it.
-- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 0.741263 against a median of 0.06543843 (context 454). Read the documents before treating the domain as weak.
+- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 0.763390 against a median of 0.06514060 (context 454). Read the documents before treating the domain as weak.
 

Qwen3.8-27B-FP8/compliance.json

@@ -1,8 +1,9 @@
 {
   "attribution": null,
   "candidate": "/media/fmodels2/Qwen/Qwen3.8-27B-FP8",
-  "candidate_weights_sha256": null,
+  "candidate_weights_sha256": "dffe578ba76dead963135c7e988e25acefdfc3cf9e19b8352653683b4edf199e",
   "comparability_key": {
+    "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
     "context_length": 2048,
     "driver": "580.173.02",
     "gpu_names": [
@@ -12,19 +13,22 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 248044,
-    "laws_version": 9,
+    "kv_cache_dtype": "bfloat16",
+    "laws_version": 15,
     "model_runner_v2": false,
+    "numerics_digest": "251a9225b37415b91fc58d975b0670fc4b481e6f5a816b663573c42e161438f8",
     "reference_config_sha256": "191e0af232104ed8b65258cf3fb2b842e288008baca7633c11b82a1ac7203aab",
     "rows": 768,
     "score_from": 0,
     "stride": 2048,
+    "substituted_parameters": [],
     "suite_id": "qwen3.8-27b-fidelity-1024x2048-v1",
     "tensor_parallel_size": 1,
     "token_sha256": "9c935708bbffbf45d5baa2931fb4af7e7b22df1e041aa0b6f4ca44d3315e543b",
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T12:56:03.993488+00:00",
+  "evaluated_at": "2026-09-11T20:24:18.945645+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -52,13 +56,13 @@
       "title": "Real vocabulary"
     },
     {
-      "detail": "all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc",
+      "detail": "all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3",
       "law": 5,
       "status": "pass",
       "title": "Manifest binding"
     },
     {
-      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s)",
+      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s)",
       "law": 6,
       "status": "pass",
       "title": "Provenance"
@@ -70,13 +74,13 @@
       "title": "Storage integrity"
     },
     {
-      "detail": "trunk 0.01104734, deployed 0.01104734, delta 3.8109852079637463e-10",
+      "detail": "trunk 0.01097467, deployed 0.01097467, delta 4.072260306742237e-10",
       "law": 8,
       "status": "pass",
       "title": "Head transparency"
     },
     {
-      "detail": "mean 0.01104734, median 0.00189959, max 24.25029755, 4 depth buckets",
+      "detail": "mean 0.01097467, median 0.00190434, max 29.40929031, 4 depth buckets",
       "law": 9,
       "status": "pass",
       "title": "Tail and depth disclosure"
@@ -103,38 +107,37 @@
       "detail": "reference declares no experts",
       "law": 14,
       "status": "not_applicable",
-      "title": "Component attribution"
+      "title": "Routed-model intervention"
     },
     {
-      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.05388702, strongest scientific_technical at 0.00327748, spread 16.4x",
+      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.05343841, strongest scientific_technical at 0.00325771, spread 16.4x",
       "law": 15,
       "status": "pass",
       "title": "Domain disclosure"
     },
     {
-      "approval": {
-        "approver": "Andy Kitzke",
-        "justification": "These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger.",
-        "timestamp": "2026-09-02T05:15:00Z"
-      },
-      "detail": "the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else",
+      "detail": "scored weights dffe578ba76dead9 as inspected",
       "law": 16,
-      "status": "override",
+      "status": "pass",
       "title": "Candidate weight binding"
     },
     {
-      "detail": "1 override(s), each fully attributed",
+      "detail": "every scored layer used the checkpoint's own quantization parameters",
+      "law": 17,
+      "status": "pass",
+      "title": "Substitution disclosure"
+    },
+    {
+      "detail": "no overrides claimed",
       "law": 13,
       "status": "pass",
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 9,
-  "mean_kld": 0.01104733806059744,
+  "laws_version": 15,
+  "mean_kld": 0.010974670174294718,
   "nondeterminism_floor": 0.0,
-  "overridden_laws": [
-    16
-  ],
+  "overridden_laws": [],
   "partition": "analysis",
   "program": "Local Inference Lab \u2014 Distribution Fidelity",
   "ranking_floor": null,
@@ -147,16 +150,16 @@
     "overall": {
       "deployed": {
         "contexts": 768,
-        "max_kld": 24.25029754638672,
-        "mean_kld": 0.011047338060597442,
-        "mean_ref_top1_prob": 0.6415440866166368,
-        "median_context_kld": 0.003959233488364337,
-        "median_context_p99": 0.0307154580950737,
-        "p90_context_kld": 0.018751727010197644,
+        "max_kld": 29.409290313720703,
+        "mean_kld": 0.01097467017429472,
+        "mean_ref_top1_prob": 0.6415473983687051,
+        "median_context_kld": 0.003896204956611177,
+        "median_context_p99": 0.030610591173171997,
+        "p90_context_kld": 0.01781900445289071,
         "positions": 1572096,
-        "top1_agreement": 0.9670777102670575,
+        "top1_agreement": 0.9670738937062368,
         "worst_context_id": 454,
-        "worst_context_kld": 0.4895784940445628
+        "worst_context_kld": 0.45672114841615463
       }
     },
     "primary": "deployed",
@@ -165,201 +168,201 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 23.4235782623291,
-            "mean_kld": 0.05388702262702585,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.032507454946453476,
-            "median_context_p99": 0.5249273777008057,
-            "p90_context_kld": 0.1189590889010681,
+            "max_kld": 29.409290313720703,
+            "mean_kld": 0.05343841309902987,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.03457746948537548,
+            "median_context_p99": 0.48533299565315247,
+            "p90_context_kld": 0.13830447681183197,
             "positions": 196512,
-            "top1_agreement": 0.9609540384302231,
+            "top1_agreement": 0.9609489496824621,
             "worst_context_id": 454,
-            "worst_context_kld": 0.4895784940445628
+            "worst_context_kld": 0.45672114841615463
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 4.877828697867478
+        "relative_to_run": 4.86925003215088
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 1.8940415382385254,
-            "mean_kld": 0.008991718850348643,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.005429992133758688,
-            "median_context_p99": 0.04178592562675476,
-            "p90_context_kld": 0.01635930600748916,
+            "max_kld": 1.4653857946395874,
+            "mean_kld": 0.008966686634686733,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.005603489085350864,
+            "median_context_p99": 0.043331339955329895,
+            "p90_context_kld": 0.016244145148751315,
             "positions": 147384,
-            "top1_agreement": 0.9569220539542963,
+            "top1_agreement": 0.9570238289095153,
             "worst_context_id": 893,
-            "worst_context_kld": 0.03802517904088536
+            "worst_context_kld": 0.03694091126483796
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 0.8139262871315055
+        "relative_to_run": 0.8170347256256356
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 3.1001851558685303,
-            "mean_kld": 0.0063881460233417755,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.005259914155344338,
-            "median_context_p99": 0.05044623836874962,
-            "p90_context_kld": 0.009174476160926261,
+            "max_kld": 2.1070759296417236,
+            "mean_kld": 0.006326082165752138,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.005294499655798116,
+            "median_context_p99": 0.04691578447818756,
+            "p90_context_kld": 0.008994613024871681,
             "positions": 196512,
-            "top1_agreement": 0.9654779351897085,
-            "worst_context_id": 6,
-            "worst_context_kld": 0.03374026097547736
+            "top1_agreement": 0.9654321364598599,
+            "worst_context_id": 65,
+            "worst_context_kld": 0.03102557076205901
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.5782520629224144
+        "relative_to_run": 0.5764257207992749
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 1.6461397409439087,
-            "mean_kld": 0.0048897095708480606,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.004029185885625382,
-            "median_context_p99": 0.030439462512731552,
-            "p90_context_kld": 0.009268822767389277,
+            "max_kld": 1.8053836822509766,
+            "mean_kld": 0.004882264113287302,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.0041416290521338285,
+            "median_context_p99": 0.031151100993156433,
+            "p90_context_kld": 0.009558583566512019,
             "positions": 147384,
-            "top1_agreement": 0.9685176138522499,
+            "top1_agreement": 0.9686465287955274,
             "worst_context_id": 275,
-            "worst_context_kld": 0.014223638338661505
+            "worst_context_kld": 0.014266868176758726
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.44261427902602124
+        "relative_to_run": 0.44486659150110247
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 2.1852290630340576,
-            "mean_kld": 0.004737305064792011,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.0039352627612185285,
-            "median_context_p99": 0.03270211070775986,
-            "p90_context_kld": 0.009271094984799227,
+            "max_kld": 1.9099842309951782,
+            "mean_kld": 0.0047161988101167476,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.003908452587081799,
+            "median_context_p99": 0.029179947450757027,
+            "p90_context_kld": 0.008553707324987884,
             "positions": 73692,
-            "top1_agreement": 0.9693725234760897,
+            "top1_agreement": 0.9691011235955056,
             "worst_context_id": 953,
-            "worst_context_kld": 0.011588845220650184
+            "worst_context_kld": 0.01122486185099774
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.42881869268476214
+        "relative_to_run": 0.4297349018436293
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 1.4300388097763062,
-            "mean_kld": 0.004714006706887872,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.004113897056717312,
-            "median_context_p99": 0.024136444553732872,
-            "p90_context_kld": 0.00561418505282599,
+            "max_kld": 1.165907382965088,
+            "mean_kld": 0.004699737983564273,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.00391996123187938,
+            "median_context_p99": 0.024366630241274834,
+            "p90_context_kld": 0.005416538875010646,
             "positions": 147384,
-            "top1_agreement": 0.9647315855181023,
+            "top1_agreement": 0.964704445530044,
             "worst_context_id": 443,
-            "worst_context_kld": 0.027228787454472464
+            "worst_context_kld": 0.027905596454365553
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.4267097359590476
+        "relative_to_run": 0.42823500924630736
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 2.663503885269165,
-            "mean_kld": 0.004137738224923952,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.003228649903561243,
-            "median_context_p99": 0.03253652900457382,
-            "p90_context_kld": 0.005859838909925266,
+            "max_kld": 8.64619255065918,
+            "mean_kld": 0.004225324387170233,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.003200656097436244,
+            "median_context_p99": 0.03228462114930153,
+            "p90_context_kld": 0.00674611330592368,
             "positions": 196512,
-            "top1_agreement": 0.9797518726591761,
+            "top1_agreement": 0.9793040628562123,
             "worst_context_id": 667,
-            "worst_context_kld": 0.023852551947544348
+            "worst_context_kld": 0.026218505325394093
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.37454617594097434
+        "relative_to_run": 0.38500695875735247
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 3.2773475646972656,
-            "mean_kld": 0.003763443945706886,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.002558179151218321,
-            "median_context_p99": 0.035675883293151855,
-            "p90_context_kld": 0.006300030598680195,
+            "max_kld": 1.8485336303710938,
+            "mean_kld": 0.0037327059215536552,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.0027118550407615144,
+            "median_context_p99": 0.03762500360608101,
+            "p90_context_kld": 0.0068965462201931475,
             "positions": 73692,
-            "top1_agreement": 0.9357189382836671,
+            "top1_agreement": 0.9357460782717255,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.019168267112011875
+            "worst_context_kld": 0.01781900445289071
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.3406652285884115
+        "relative_to_run": 0.340120100401426
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 24.25029754638672,
-            "mean_kld": 0.0035539641212878384,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.002869848601152158,
-            "median_context_p99": 0.018992044031620026,
-            "p90_context_kld": 0.0045202462159064935,
+            "max_kld": 11.368499755859375,
+            "mean_kld": 0.0034699791707562626,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.002857173882379502,
+            "median_context_p99": 0.018883395940065384,
+            "p90_context_kld": 0.004444196419291071,
             "positions": 196512,
-            "top1_agreement": 0.9762915241817294,
+            "top1_agreement": 0.9766731802638007,
             "worst_context_id": 709,
-            "worst_context_kld": 0.027977719611953035
+            "worst_context_kld": 0.02107575118209227
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.3217032104741836
+        "relative_to_run": 0.3161807248552924
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 2.442533493041992,
-            "mean_kld": 0.003277476263199609,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.003168675314512905,
-            "median_context_p99": 0.022142626345157623,
-            "p90_context_kld": 0.0041321819357580175,
+            "max_kld": 2.281120538711548,
+            "mean_kld": 0.0032577067486191185,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.0031907973419261406,
+            "median_context_p99": 0.021607771515846252,
+            "p90_context_kld": 0.004144200098769572,
             "positions": 196512,
-            "top1_agreement": 0.9721085735222277,
-            "worst_context_id": 246,
-            "worst_context_kld": 0.005805297005183099
+            "top1_agreement": 0.9721340172610324,
+            "worst_context_id": 174,
+            "worst_context_kld": 0.005824702771456004
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.2966756557300793
+        "relative_to_run": 0.2968386928155199
       }
     ]
   },

Qwen3.8-27B-FP8/inspect.json

@@ -1,38 +1,44 @@
 {
-  "coverage": {
-    "attention": {
-      "quantized": 68,
-      "weights": 68
-    },
-    "dense_mlp": {
-      "quantized": 195,
-      "weights": 195
-    },
-    "experts": {
-      "quantized": 0,
-      "weights": 0
-    },
-    "router": {
-      "quantized": 0,
-      "weights": 0
-    },
-    "shared_expert": {
-      "quantized": 0,
-      "weights": 0
-    }
-  },
+  "model": "/media/fmodels2/Qwen/Qwen3.8-27B-FP8",
+  "inspect_version": 5,
+  "weights_sha256": "dffe578ba76dead963135c7e988e25acefdfc3cf9e19b8352653683b4edf199e",
+  "weights_bytes": 30866866928,
+  "quant_method": "fp8",
   "declared": {
-    "activation_scheme": "dynamic",
     "fmt": "e4m3",
+    "activation_scheme": "dynamic",
     "weight_block_size": [
       128,
       128
     ]
   },
-  "detected_block": 128,
   "detected_scheme": "fp8_block",
-  "model": "/media/fmodels2/Qwen/Qwen3.8-27B-FP8",
-  "quant_method": "fp8",
-  "weights_bytes": 30866866928,
-  "weights_bytes_source": "hub"
+  "detected_block": 128,
+  "quant_algorithm": null,
+  "scheme_mix": null,
+  "kv_cache_scheme": null,
+  "coverage": {
+    "router": {
+      "weights": 0,
+      "quantized": 0
+    },
+    "experts": {
+      "weights": 0,
+      "quantized": 0
+    },
+    "shared_expert": {
+      "weights": 0,
+      "quantized": 0
+    },
+    "attention": {
+      "weights": 68,
+      "quantized": 68
+    },
+    "dense_mlp": {
+      "weights": 195,
+      "quantized": 195
+    }
+  },
+  "unloadable_reason": null,
+  "quantized_names_sha256": "51488c3d2617a817296f4e55f15dfdc44668611ae32c7bcbffdaa57717090fce"
 }

Qwen3.8-27B-FP8/report.md

@@ -1,12 +1,12 @@
 # Qwen3.8-27B / Qwen3.8-27B-FP8: distribution fidelity
 
-**Mean KLD(reference || candidate) = 0.01104734** over 1572096 scored positions.
+**Mean KLD(reference || candidate) = 0.01097467** over 1572096 scored positions.
 
 | Mean | Median | p90 | p99 | Max | Top-1 agreement |
 |---|---|---|---|---|---|
-| 0.01104734 | 0.00189959 | 0.00940180 | 0.08646069 | 24.25029755 | 96.7078% |
+| 0.01097467 | 0.00190434 | 0.00940006 | 0.08594550 | 29.40929031 | 96.7074% |
 
-Reverse direction, KLD(candidate || reference): 0.01087530.
+Reverse direction, KLD(candidate || reference): 0.01099299.
 
 ## Identity
 
@@ -15,10 +15,10 @@ Reverse direction, KLD(candidate || reference): 0.01087530.
 | Reference checkpoint | /media/fmodels2/Qwen/Qwen3.8-27B |
 | Reference config SHA-256 | 191e0af23210 |
 | Candidate checkpoint | /media/fmodels2/Qwen/Qwen3.8-27B-FP8 |
-| Candidate weights SHA-256 | unbound (Law 16 gap) |
+| Candidate weights SHA-256 | dffe578ba76dead9 |
 | Suite | qwen3.8-27b-fidelity-1024x2048-v1 |
 | Suite token SHA-256 | 9c935708bbffbf45 |
-| Capture manifest SHA-256 | 304794b4e35b326d |
+| Capture manifest SHA-256 | dbfacc9cbb6d3e06 |
 | Tokenizer | None |
 | Scored vocabulary | 248044 |
 | Declared vocabulary | 248320 |
@@ -27,32 +27,39 @@ Reverse direction, KLD(candidate || reference): 0.01087530.
 | Model runner | V1 |
 | Tensor parallel | 1 |
 | Eager enforced | True |
+| KV cache | bfloat16 |
 | Prefix caching | False |
 | max_num_seqs | 1 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45 |
+| vLLM commit | 60071d1ab732 |
+| vLLM dirty digest | e3b0c44298fc1c14 |
+| Numerics digest | 251a9225b37415b9 |
+| Compiled extensions | f2fbc7537b0f01f6 |
+| FlashInfer | 0.6.17 |
 | torch | 2.13.0+cu132 |
 | Driver | 580.173.02 |
 | GPUs | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
-| Laws version | 9 |
+| Laws version | 15 |
 | Partition | analysis |
 
+The commit says when this number was taken. The numerics digest says what took it: a hash of the runtime and the scorer, which moves only when code that can change a logit changes. Comparability is bounded by the digest, so a result stays current across commits that cannot reach a number, and two results carrying the same digest were computed by the same code whatever their commits say.
+
 ## Fidelity by domain
 
-The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
+The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `deployed` is QxQ (natural student routing). `bxq` is the teacher-ID counterfactual on the same weights. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
 
 | Domain | Contexts | Reference top-1 | Mean KLD | x run | Worst context |
 |---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.05388702 | 4.88 | 0.48957849 |
-| Chinese across several content types | 72 | 55.6% | 0.00899172 | 0.81 | 0.03802518 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.00638815 | 0.58 | 0.03374026 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.00488971 | 0.44 | 0.01422364 |
-| Other multilingual content | 36 | 64.5% | 0.00473731 | 0.43 | 0.01158885 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.00471401 | 0.43 | 0.02722879 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.00413774 | 0.37 | 0.02385255 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.00376344 | 0.34 | 0.01916827 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.00355396 | 0.32 | 0.02797772 |
-| Scientific and technical exposition | 96 | 60.3% | 0.00327748 | 0.30 | 0.00580530 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.05343841 | 4.87 | 0.45672115 |
+| Chinese across several content types | 72 | 55.6% | 0.00896669 | 0.82 | 0.03694091 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.00632608 | 0.58 | 0.03102557 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.00488226 | 0.44 | 0.01426687 |
+| Other multilingual content | 36 | 64.5% | 0.00471620 | 0.43 | 0.01122486 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.00469974 | 0.43 | 0.02790560 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.00422532 | 0.39 | 0.02621851 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.00373271 | 0.34 | 0.01781900 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.00346998 | 0.32 | 0.02107575 |
+| Scientific and technical exposition | 96 | 60.3% | 0.00325771 | 0.30 | 0.00582470 |
 
 **Natural dialogue, instruction following, and assistance** is this candidate's weakest domain and **Scientific and technical exposition** its strongest, a spread of 16.4x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
 
@@ -60,8 +67,8 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Component | Value |
 |---|---|
-| Trunk (candidate hidden states, reference head) | 0.01104734 |
-| Deployed (candidate's own head) | 0.01104734 |
+| Trunk (candidate hidden states, reference head) | 0.01097467 |
+| Deployed (candidate's own head) | 0.01097467 |
 | Head-associated delta (not additive) | 0.00000000 |
 | Reference head | {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]} |
 | Candidate head | {'runtime': {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]}, 'state': 'unquantized', 'static': {'ignored': True, 'lm_head_dtypes': {'lm_head.weight': 'BF16', 'model.language_model.embed_tokens.weight': 'BF16'}, 'lm_head_keys': ['lm_head.weight'], 'output_weight_keys': ['lm_head.weight'], 'packed_keys': [], 'quant_method': 'fp8', 'state': 'unquantized', 'tie_word_embeddings': False}} |
@@ -70,26 +77,26 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Position range | Positions | Mean KLD |
 |---|---|---|
-| 0–511 | 393216 | 0.00649119 |
-| 512–1023 | 393216 | 0.00882887 |
-| 1024–1535 | 393216 | 0.01316371 |
-| 1536–2046 | 392448 | 0.01571470 |
+| 0–511 | 393216 | 0.00655353 |
+| 512–1023 | 393216 | 0.00893310 |
+| 1024–1535 | 393216 | 0.01307841 |
+| 1536–2046 | 392448 | 0.01534217 |
 
 ## Error by reference confidence
 
 | Reference top-1 probability | Positions | Share | Mean KLD |
 |---|---|---|---|
-| [0.00, 0.25) | 240569 | 15.3% | 0.00839035 |
-| [0.25, 0.50) | 346390 | 22.0% | 0.01436432 |
-| [0.50, 0.75) | 273384 | 17.4% | 0.01933117 |
-| [0.75, 0.95) | 249939 | 15.9% | 0.01405977 |
-| [0.95, 1.00) | 461814 | 29.4% | 0.00340925 |
+| [0.00, 0.25) | 240522 | 15.3% | 0.00832364 |
+| [0.25, 0.50) | 346508 | 22.0% | 0.01508628 |
+| [0.50, 0.75) | 273204 | 17.4% | 0.01818135 |
+| [0.75, 0.95) | 250041 | 15.9% | 0.01448346 |
+| [0.95, 1.00) | 461821 | 29.4% | 0.00310732 |
 
 ## Top-K set agreement
 
 | K=1 | K=2 | K=3 | K=4 | K=5 |
 |---|---|---|---|---|
-| 96.4861% | 88.4746% | 77.6329% | 65.6430% | 54.3723% |
+| 96.4821% | 88.4600% | 77.6550% | 65.7186% | 54.4059% |
 
 ## Law compliance
 
@@ -101,22 +108,19 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | 2 | Determinism | PASS | eager enforced; prefix caching off, max_num_seqs=1 |
 | 3 | Frozen input | PASS | token hash matches suite qwen3.8-27b-fidelity-1024x2048-v1 [analysis]: 9c935708bbffbf45 |
 | 4 | Real vocabulary | PASS | scored 248044 real tokens of 248320 declared (276 padding rows) |
-| 5 | Manifest binding | PASS | all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc |
-| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s) |
+| 5 | Manifest binding | PASS | all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3 |
+| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s) |
 | 7 | Storage integrity | PASS | hidden storage with bitwise-exact replay |
-| 8 | Head transparency | PASS | trunk 0.01104734, deployed 0.01104734, delta 3.8109852079637463e-10 |
-| 9 | Tail and depth disclosure | PASS | mean 0.01104734, median 0.00189959, max 24.25029755, 4 depth buckets |
+| 8 | Head transparency | PASS | trunk 0.01097467, deployed 0.01097467, delta 4.072260306742237e-10 |
+| 9 | Tail and depth disclosure | PASS | mean 0.01097467, median 0.00190434, max 29.40929031, 4 depth buckets |
 | 10 | Comparability | PASS | comparability key fully resolved |
 | 11 | Freeze before qualification | NOT_APPLICABLE | partition is 'analysis' |
 | 12 | Reusable reference | PASS | suite, reference, head, and checksums present; published reference matches the scored capture on every bound field |
-| 14 | Component attribution | NOT_APPLICABLE | reference declares no experts |
-| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.05388702, strongest scientific_technical at 0.00327748, spread 16.4x |
-| 16 | Candidate weight binding | OVERRIDE | the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else |
-| 13 | Recorded deviation | PASS | 1 override(s), each fully attributed |
-
-### Recorded deviations
-
-- **Law 16 (Candidate weight binding)** overridden by Andy Kitzke at 2026-09-02T05:15:00Z. Justification: These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger. Underlying finding: the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else.
+| 14 | Routed-model intervention | NOT_APPLICABLE | reference declares no experts |
+| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.05343841, strongest scientific_technical at 0.00325771, spread 16.4x |
+| 16 | Candidate weight binding | PASS | scored weights dffe578ba76dead9 as inspected |
+| 17 | Substitution disclosure | PASS | every scored layer used the checkpoint's own quantization parameters |
+| 13 | Recorded deviation | PASS | no overrides claimed |
 
 ## Environment
 
@@ -126,7 +130,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | Platform | Linux-6.8.0-137-generic-x86_64-with-glibc2.39 |
 | Python | 3.12.3 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45fbe0b0205e70893bcd28fa928a88 |
+| vLLM commit | 60071d1ab73229321712254b1350dcaaac1c125a |
 | torch | 2.13.0+cu132 |
 | torch CUDA runtime | 13.2 |
 | cuDNN | 9.20.0 (92000) |
@@ -134,7 +138,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | CUDA arch list | sm_75, sm_80, sm_86, sm_90, sm_100, sm_120 |
 | nvcc | Cuda compilation tools, release 13.0, V13.0.88 |
 | gcc | gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0 |
-| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.8) 2.39 |
+| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.9) 2.39 |
 | NVIDIA driver | 580.173.02 |
 | float32 matmul precision | highest |
 | TF32 (matmul / cuDNN) | False / True |
@@ -157,8 +161,18 @@ Credential values are never published. Set but redacted: `HF_TOKEN`.
 
 | Variable | Value |
 |---|---|
+| `CUBLAS_WORKSPACE_CONFIG` | `:4096:8` |
 | `CUDA_HOME` | `/usr/local/cuda-13.0` |
 | `HF_TOKEN` | `<redacted>` |
+| `NCCL_DETERMINISTIC` | `1` |
+| `PYTORCH_NVML_BASED_CUDA_CHECK` | `1` |
+| `TORCHINDUCTOR_CACHE_DIR` | `/tmp/torchinductor_phaedawg` |
+| `TORCHINDUCTOR_COMPILE_THREADS` | `1` |
+| `TRITON_CACHE_AUTOTUNING` | `1` |
+| `TRITON_PTXAS_BLACKWELL_PATH` | `/usr/local/cuda-13.0/bin/ptxas` |
+| `VLLM_BATCH_INVARIANT` | `1` |
+| `VLLM_MARLIN_USE_ATOMIC_ADD` | `0` |
+| `VLLM_MOE_USE_DEEP_GEMM` | `0` |
 
 ## Files in this artifact
 
@@ -166,26 +180,26 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 
 | Path | Size | What it is |
 |---|---|---|
-| `Qwen3.8-27B-FP8/report.md` | 12.30 KiB | This document. |
-| `Qwen3.8-27B-FP8/report.json` | 256.86 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
-| `Qwen3.8-27B-FP8/manifest.json` | 85.43 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
-| `Qwen3.8-27B-FP8/compliance.json` | 13.06 KiB | The law-by-law receipt, including the comparability key. |
-| `baselines/self-kld.json` | 4.74 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
+| `Qwen3.8-27B-FP8/report.md` | 13.13 KiB | This document. |
+| `Qwen3.8-27B-FP8/report.json` | 259.25 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
+| `Qwen3.8-27B-FP8/manifest.json` | 86.99 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
+| `Qwen3.8-27B-FP8/compliance.json` | 12.33 KiB | The law-by-law receipt, including the comparability key. |
+| `baselines/self-kld.json` | 7.06 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
 | `suite/suite-manifest.json` | 382.68 KiB | The frozen evaluation input's identity: token hashes per context and per partition, sources, strata, and the analysis/qualification split. |
 | `suite/tokens` | 15.49 MiB in 1024 files | The token IDs themselves. These are the evaluation input, not a description of it; retokenizing source text does not reproduce them. |
 | `suite/sources.json` | 737.42 KiB | Per-context provenance: dataset, revision, licence, source unit, and the deterministic token offset chosen within the document. |
 | `suite/validation/capability-overlap.json` | 1.48 KiB | The benchmark-contamination scan and every document it blocked. |
-| `reference/manifest.json` | 85.43 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
+| `reference/manifest.json` | 86.99 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
 | `reference` | 19.25 GiB in 771 files | The reusable reference distributions. Pass this directory as `--reference-logits` to score a new candidate against the same reference without loading the reference checkpoint. |
 | `reference/lm_head.safetensors` | 2.37 GiB | The reference language-model head, which turns the stored hidden states back into reference logits. |
-| `environment/runtime.json` | 2.26 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
+| `environment/runtime.json` | 5.42 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
 | `environment/summary.md` | 1.18 KiB | The same provenance as prose, plus an index of every captured file. |
 | `environment/toolchain-nvcc.txt` | 243 B | `nvcc --version` verbatim; `toolchain-gcc.txt` and `toolchain-ldd.txt` sit beside it. |
 | `environment/gpu-smi-query.txt` | 54.73 KiB | `nvidia-smi -q` verbatim: ECC state, persistence mode, clocks, and throttle reasons, any of which can move a bitwise result. |
-| `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
-| `environment/models` | 7.22 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
-| `checksums.txt` | 182.99 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
+| `environment/pip-freeze.txt` | 4.47 KiB | Every installed package version in the scoring environment. |
+| `environment/models` | 19.46 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
+| `checksums.txt` | 183.22 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
+| `LAWS.md` | 40.20 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

Qwen3.8-27B-FP8/strata.json

@@ -8,381 +8,381 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 23.4235782623291,
-            "mean_kld": 0.05388702262702585,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.032507454946453476,
-            "median_context_p99": 0.5249273777008057,
-            "p90_context_kld": 0.1189590889010681,
+            "max_kld": 29.409290313720703,
+            "mean_kld": 0.05343841309902987,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.03457746948537548,
+            "median_context_p99": 0.48533299565315247,
+            "p90_context_kld": 0.13830447681183197,
             "positions": 196512,
-            "top1_agreement": 0.9609540384302231,
+            "top1_agreement": 0.9609489496824621,
             "worst_context_id": 454,
-            "worst_context_kld": 0.4895784940445628
+            "worst_context_kld": 0.45672114841615463
           }
         },
         "key": "wildchat",
         "label": "wildchat",
-        "relative_to_run": 4.877828697867478
+        "relative_to_run": 4.86925003215088
       },
       {
         "cells": {
           "deployed": {
             "contexts": 33,
-            "max_kld": 1.8940415382385254,
-            "mean_kld": 0.014494251205789503,
-            "mean_ref_top1_prob": 0.6040606719329803,
-            "median_context_kld": 0.011235134472533293,
-            "median_context_p99": 0.08464735746383667,
-            "p90_context_kld": 0.03288934683322978,
+            "max_kld": 1.4653857946395874,
+            "mean_kld": 0.01450307849202896,
+            "mean_ref_top1_prob": 0.6040861921682756,
+            "median_context_kld": 0.010737033767105221,
+            "median_context_p99": 0.09213735163211823,
+            "p90_context_kld": 0.0319923099255563,
             "positions": 67551,
-            "top1_agreement": 0.9482465100442629,
+            "top1_agreement": 0.9481428846353126,
             "worst_context_id": 893,
-            "worst_context_kld": 0.03802517904088536
+            "worst_context_kld": 0.03694091126483796
           }
         },
         "key": "wikisource_zh",
         "label": "wikisource_zh",
-        "relative_to_run": 1.3120130049686967
+        "relative_to_run": 1.3215047251259184
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 2.1852290630340576,
-            "mean_kld": 0.008478143004339219,
-            "mean_ref_top1_prob": 0.7357366218042676,
-            "median_context_kld": 0.009271094984799227,
-            "median_context_p99": 0.12381438910961151,
-            "p90_context_kld": 0.011588845220650184,
+            "max_kld": 1.9099842309951782,
+            "mean_kld": 0.008550009321842978,
+            "mean_ref_top1_prob": 0.7358200961573914,
+            "median_context_kld": 0.008553707324987884,
+            "median_context_p99": 0.15561799705028534,
+            "p90_context_kld": 0.01122486185099774,
             "positions": 14329,
-            "top1_agreement": 0.96280270779538,
+            "top1_agreement": 0.9611277828180613,
             "worst_context_id": 953,
-            "worst_context_kld": 0.011588845220650184
+            "worst_context_kld": 0.01122486185099774
           }
         },
         "key": "wikipedia_de",
         "label": "wikipedia_de",
-        "relative_to_run": 0.767437635911426
+        "relative_to_run": 0.7790675424459796
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 1.6461397409439087,
-            "mean_kld": 0.006839122403048966,
-            "mean_ref_top1_prob": 0.7232628548554616,
-            "median_context_kld": 0.007780651702137065,
-            "median_context_p99": 0.0937204360961914,
-            "p90_context_kld": 0.011423241495751609,
+            "max_kld": 1.4340033531188965,
+            "mean_kld": 0.0067462983966682195,
+            "mean_ref_top1_prob": 0.7233191001139795,
+            "median_context_kld": 0.007960103791057228,
+            "median_context_p99": 0.10476938635110855,
+            "p90_context_kld": 0.011555328467113763,
             "positions": 47081,
-            "top1_agreement": 0.9701365731399079,
+            "top1_agreement": 0.9706675729062679,
             "worst_context_id": 275,
-            "worst_context_kld": 0.014223638338661505
+            "worst_context_kld": 0.014266868176758726
           }
         },
         "key": "regulations",
         "label": "regulations",
-        "relative_to_run": 0.6190742390189067
+        "relative_to_run": 0.6147153663414551
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 3.1001851558685303,
-            "mean_kld": 0.0063881460233417755,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.005259914155344338,
-            "median_context_p99": 0.05044623836874962,
-            "p90_context_kld": 0.009174476160926261,
+            "max_kld": 2.1070759296417236,
+            "mean_kld": 0.006326082165752138,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.005294499655798116,
+            "median_context_p99": 0.04691578447818756,
+            "p90_context_kld": 0.008994613024871681,
             "positions": 196512,
-            "top1_agreement": 0.9654779351897085,
-            "worst_context_id": 6,
-            "worst_context_kld": 0.03374026097547736
+            "top1_agreement": 0.9654321364598599,
+            "worst_context_id": 65,
+            "worst_context_kld": 0.03102557076205901
           }
         },
         "key": "wikipedia_en",
         "label": "wikipedia_en",
-        "relative_to_run": 0.5782520629224144
+        "relative_to_run": 0.5764257207992749
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 1.4300388097763062,
-            "mean_kld": 0.004714006706887872,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.004113897056717312,
-            "median_context_p99": 0.024136444553732872,
-            "p90_context_kld": 0.00561418505282599,
+            "max_kld": 1.165907382965088,
+            "mean_kld": 0.004699737983564273,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.00391996123187938,
+            "median_context_p99": 0.024366630241274834,
+            "p90_context_kld": 0.005416538875010646,
             "positions": 147384,
-            "top1_agreement": 0.9647315855181023,
+            "top1_agreement": 0.964704445530044,
             "worst_context_id": 443,
-            "worst_context_kld": 0.027228787454472464
+            "worst_context_kld": 0.027905596454365553
           }
         },
         "key": "public_domain_books",
         "label": "public_domain_books",
-        "relative_to_run": 0.4267097359590476
+        "relative_to_run": 0.42823500924630736
       },
       {
         "cells": {
           "deployed": {
             "contexts": 26,
-            "max_kld": 0.93686443567276,
-            "mean_kld": 0.004517050486458161,
-            "mean_ref_top1_prob": 0.5182113024080622,
-            "median_context_kld": 0.004118480529026551,
-            "median_context_p99": 0.030470337718725204,
-            "p90_context_kld": 0.0064679281424572565,
+            "max_kld": 1.8053836822509766,
+            "mean_kld": 0.0045295686855075,
+            "mean_ref_top1_prob": 0.5182496321630543,
+            "median_context_kld": 0.004108259074256645,
+            "median_context_p99": 0.02957768365740776,
+            "p90_context_kld": 0.006899708291828433,
             "positions": 53222,
-            "top1_agreement": 0.9665551839464883,
+            "top1_agreement": 0.9669497576190297,
             "worst_context_id": 312,
-            "worst_context_kld": 0.01186199497122189
+            "worst_context_kld": 0.01135261394611149
           }
         },
         "key": "public_domain_review",
         "label": "public_domain_review",
-        "relative_to_run": 0.408881348763023
+        "relative_to_run": 0.4127293680421325
       },
       {
         "cells": {
           "deployed": {
             "contexts": 39,
-            "max_kld": 1.0524587631225586,
-            "mean_kld": 0.004335729934206377,
-            "mean_ref_top1_prob": 0.515358929042611,
-            "median_context_kld": 0.004207335437167334,
-            "median_context_p99": 0.030693423002958298,
-            "p90_context_kld": 0.0059234729125623815,
+            "max_kld": 0.9606045484542847,
+            "mean_kld": 0.004282047370781772,
+            "mean_ref_top1_prob": 0.5153760832971095,
+            "median_context_kld": 0.004061216602046485,
+            "median_context_p99": 0.028410743921995163,
+            "p90_context_kld": 0.005994740834750328,
             "positions": 79833,
-            "top1_agreement": 0.9642628988012476,
+            "top1_agreement": 0.9645384740646099,
             "worst_context_id": 886,
-            "worst_context_kld": 0.007191435851410738
+            "worst_context_kld": 0.007109383381877114
           }
         },
         "key": "wikipedia_zh",
         "label": "wikipedia_zh",
-        "relative_to_run": 0.39246829511542075
+        "relative_to_run": 0.39017549527924245
       },
       {
         "cells": {
           "deployed": {
             "contexts": 52,
-            "max_kld": 2.0390231609344482,
-            "mean_kld": 0.004180312188396761,
-            "mean_ref_top1_prob": 0.8582575125835092,
-            "median_context_kld": 0.0031240299668609337,
-            "median_context_p99": 0.03311830013990402,
-            "p90_context_kld": 0.005859838909925266,
+            "max_kld": 6.946210861206055,
+            "mean_kld": 0.004274911986900791,
+            "mean_ref_top1_prob": 0.8582941084526002,
+            "median_context_kld": 0.0030968952218054235,
+            "median_context_p99": 0.03621552512049675,
+            "p90_context_kld": 0.00674611330592368,
             "positions": 106444,
-            "top1_agreement": 0.9817368757280823,
+            "top1_agreement": 0.9809383337717485,
             "worst_context_id": 667,
-            "worst_context_kld": 0.023852551947544348
+            "worst_context_kld": 0.026218505325394093
           }
         },
         "key": "github_code",
         "label": "github_code",
-        "relative_to_run": 0.37839995168670426
-      },
-      {
-        "cells": {
-          "deployed": {
-            "contexts": 7,
-            "max_kld": 0.26435965299606323,
-            "mean_kld": 0.0041055369373283355,
-            "mean_ref_top1_prob": 0.5702914711488575,
-            "median_context_kld": 0.0040482799366643365,
-            "median_context_p99": 0.027272282168269157,
-            "p90_context_kld": 0.004934028467450004,
-            "positions": 14329,
-            "top1_agreement": 0.9665712889943471,
-            "worst_context_id": 968,
-            "worst_context_kld": 0.004934028467450004
-          }
-        },
-        "key": "wikipedia_ja",
-        "label": "wikipedia_ja",
-        "relative_to_run": 0.37163133008227206
+        "relative_to_run": 0.38952532686710245
       },
       {
         "cells": {
           "deployed": {
             "contexts": 44,
-            "max_kld": 2.663503885269165,
-            "mean_kld": 0.004087423540819723,
-            "mean_ref_top1_prob": 0.720394649593742,
-            "median_context_kld": 0.0033318739900228,
-            "median_context_p99": 0.03225799277424812,
-            "p90_context_kld": 0.005771603222770657,
+            "max_kld": 8.64619255065918,
+            "mean_kld": 0.004166720860215938,
+            "mean_ref_top1_prob": 0.7203910827522294,
+            "median_context_kld": 0.0032651331572354555,
+            "median_context_p99": 0.030797146260738373,
+            "p90_context_kld": 0.00612854931655339,
             "positions": 90068,
-            "top1_agreement": 0.9774059599413776,
+            "top1_agreement": 0.9773726517742151,
             "worst_context_id": 582,
-            "worst_context_kld": 0.023764493762975004
+            "worst_context_kld": 0.023565068584076254
           }
         },
         "key": "stackv2",
         "label": "stackv2",
-        "relative_to_run": 0.36999171369602085
+        "relative_to_run": 0.3796670691731025
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 6,
-            "max_kld": 1.8690025806427002,
-            "mean_kld": 0.0040027544454333186,
-            "mean_ref_top1_prob": 0.6892386900239509,
-            "median_context_kld": 0.004155839709752397,
-            "median_context_p99": 0.03075869381427765,
-            "p90_context_kld": 0.005030124414311502,
-            "positions": 12282,
-            "top1_agreement": 0.973945611463931,
-            "worst_context_id": 958,
-            "worst_context_kld": 0.005030124414311502
+            "contexts": 7,
+            "max_kld": 0.3594842255115509,
+            "mean_kld": 0.00412385460898927,
+            "mean_ref_top1_prob": 0.5701150598950927,
+            "median_context_kld": 0.004194772862474528,
+            "median_context_p99": 0.028288114815950394,
+            "p90_context_kld": 0.004689954649223192,
+            "positions": 14329,
+            "top1_agreement": 0.9666410775350688,
+            "worst_context_id": 938,
+            "worst_context_kld": 0.004689954649223192
           }
         },
-        "key": "wikipedia_cs",
-        "label": "wikipedia_cs",
-        "relative_to_run": 0.3623275058187953
+        "key": "wikipedia_ja",
+        "label": "wikipedia_ja",
+        "relative_to_run": 0.37576114302262276
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 0.4819668233394623,
-            "mean_kld": 0.004000421899168063,
-            "mean_ref_top1_prob": 0.591765751203999,
-            "median_context_kld": 0.0039352627612185285,
-            "median_context_p99": 0.03636820986866951,
-            "p90_context_kld": 0.004933108867118309,
+            "max_kld": 0.8643730282783508,
+            "mean_kld": 0.0040610914246551125,
+            "mean_ref_top1_prob": 0.591761624645417,
+            "median_context_kld": 0.003908452587081799,
+            "median_context_p99": 0.03418981656432152,
+            "p90_context_kld": 0.004887819022356513,
             "positions": 14329,
-            "top1_agreement": 0.9682462139716659,
+            "top1_agreement": 0.9696419847860981,
             "worst_context_id": 936,
-            "worst_context_kld": 0.004933108867118309
+            "worst_context_kld": 0.004887819022356513
           }
         },
         "key": "wikipedia_es",
         "label": "wikipedia_es",
-        "relative_to_run": 0.36211636479527803
+        "relative_to_run": 0.37004222998584063
+      },
+      {
+        "cells": {
+          "deployed": {
+            "contexts": 6,
+            "max_kld": 0.8171505331993103,
+            "mean_kld": 0.003859297547196708,
+            "mean_ref_top1_prob": 0.689239984531695,
+            "median_context_kld": 0.004029460815379175,
+            "median_context_p99": 0.03684946149587631,
+            "p90_context_kld": 0.004699581690286988,
+            "positions": 12282,
+            "top1_agreement": 0.9728057319654779,
+            "worst_context_id": 958,
+            "worst_context_kld": 0.004699581690286988
+          }
+        },
+        "key": "wikipedia_cs",
+        "label": "wikipedia_cs",
+        "relative_to_run": 0.35165499153096175
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 3.2773475646972656,
-            "mean_kld": 0.003763443945706886,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.002558179151218321,
-            "median_context_p99": 0.035675883293151855,
-            "p90_context_kld": 0.006300030598680195,
+            "max_kld": 1.8485336303710938,
+            "mean_kld": 0.0037327059215536552,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.0027118550407615144,
+            "median_context_p99": 0.03762500360608101,
+            "p90_context_kld": 0.0068965462201931475,
             "positions": 73692,
-            "top1_agreement": 0.9357189382836671,
+            "top1_agreement": 0.9357460782717255,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.019168267112011875
+            "worst_context_kld": 0.01781900445289071
           }
         },
         "key": "starcoder_structured",
         "label": "starcoder_structured",
-        "relative_to_run": 0.3406652285884115
+        "relative_to_run": 0.340120100401426
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 24.25029754638672,
-            "mean_kld": 0.0035539641212878384,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.002869848601152158,
-            "median_context_p99": 0.018992044031620026,
-            "p90_context_kld": 0.0045202462159064935,
+            "max_kld": 11.368499755859375,
+            "mean_kld": 0.0034699791707562626,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.002857173882379502,
+            "median_context_p99": 0.018883395940065384,
+            "p90_context_kld": 0.004444196419291071,
             "positions": 196512,
-            "top1_agreement": 0.9762915241817294,
+            "top1_agreement": 0.9766731802638007,
             "worst_context_id": 709,
-            "worst_context_kld": 0.027977719611953035
+            "worst_context_kld": 0.02107575118209227
           }
         },
         "key": "libretexts",
         "label": "libretexts",
-        "relative_to_run": 0.3217032104741836
+        "relative_to_run": 0.3161807248552924
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 6,
-            "max_kld": 1.5419929027557373,
-            "mean_kld": 0.003466903276481711,
-            "mean_ref_top1_prob": 0.6660615117323249,
-            "median_context_kld": 0.0035648224349446134,
-            "median_context_p99": 0.03327962011098862,
-            "p90_context_kld": 0.004957788973098181,
-            "positions": 12282,
-            "top1_agreement": 0.973945611463931,
-            "worst_context_id": 945,
-            "worst_context_kld": 0.004957788973098181
+            "contexts": 23,
+            "max_kld": 0.6478263139724731,
+            "mean_kld": 0.003416929009135728,
+            "mean_ref_top1_prob": 0.537879873360319,
+            "median_context_kld": 0.003213981280852666,
+            "median_context_p99": 0.019698120653629303,
+            "p90_context_kld": 0.004322502291000816,
+            "positions": 47081,
+            "top1_agreement": 0.9685435738408276,
+            "worst_context_id": 343,
+            "worst_context_kld": 0.005585340689145614
           }
         },
-        "key": "wikipedia_ru",
-        "label": "wikipedia_ru",
-        "relative_to_run": 0.3138225025309147
+        "key": "open_news",
+        "label": "open_news",
+        "relative_to_run": 0.3113468518752378
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 23,
-            "max_kld": 0.5890254378318787,
-            "mean_kld": 0.0033615635296966057,
-            "mean_ref_top1_prob": 0.5379292386352562,
-            "median_context_kld": 0.0031743782651549855,
-            "median_context_p99": 0.022210456430912018,
-            "p90_context_kld": 0.004280879788646605,
-            "positions": 47081,
-            "top1_agreement": 0.9691170535884964,
-            "worst_context_id": 343,
-            "worst_context_kld": 0.005723521375894878
+            "contexts": 6,
+            "max_kld": 0.3777981996536255,
+            "mean_kld": 0.0033608578078174124,
+            "mean_ref_top1_prob": 0.6660535411394803,
+            "median_context_kld": 0.0034039176152432125,
+            "median_context_p99": 0.025507193058729172,
+            "p90_context_kld": 0.004574842338160697,
+            "positions": 12282,
+            "top1_agreement": 0.9745155512131575,
+            "worst_context_id": 945,
+            "worst_context_kld": 0.004574842338160697
           }
         },
-        "key": "open_news",
-        "label": "open_news",
-        "relative_to_run": 0.30428719672174237
+        "key": "wikipedia_ru",
+        "label": "wikipedia_ru",
+        "relative_to_run": 0.3062377050464203
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 2.442533493041992,
-            "mean_kld": 0.003277476263199609,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.003168675314512905,
-            "median_context_p99": 0.022142626345157623,
-            "p90_context_kld": 0.0041321819357580175,
+            "max_kld": 2.281120538711548,
+            "mean_kld": 0.0032577067486191185,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.0031907973419261406,
+            "median_context_p99": 0.021607771515846252,
+            "p90_context_kld": 0.004144200098769572,
             "positions": 196512,
-            "top1_agreement": 0.9721085735222277,
-            "worst_context_id": 246,
-            "worst_context_kld": 0.005805297005183099
+            "top1_agreement": 0.9721340172610324,
+            "worst_context_id": 174,
+            "worst_context_kld": 0.005824702771456004
           }
         },
         "key": "scientific_papers",
         "label": "scientific_papers",
-        "relative_to_run": 0.2966756557300793
+        "relative_to_run": 0.2968386928155199
       },
       {
         "cells": {
           "deployed": {
             "contexts": 3,
-            "max_kld": 0.705543577671051,
-            "mean_kld": 0.003212107705057633,
-            "mean_ref_top1_prob": 0.6023603953518782,
-            "median_context_kld": 0.0031552203512278053,
-            "median_context_p99": 0.021008508279919624,
-            "p90_context_kld": 0.0034167893172513076,
+            "max_kld": 0.32339420914649963,
+            "mean_kld": 0.003105845848568889,
+            "mean_ref_top1_prob": 0.6023746847155801,
+            "median_context_kld": 0.0030011382400890194,
+            "median_context_p99": 0.020637474954128265,
+            "p90_context_kld": 0.0035378718844978033,
             "positions": 6141,
-            "top1_agreement": 0.9755740107474352,
+            "top1_agreement": 0.973945611463931,
             "worst_context_id": 967,
-            "worst_context_kld": 0.0034167893172513076
+            "worst_context_kld": 0.0035378718844978033
           }
         },
         "key": "wikipedia_fr",
         "label": "wikipedia_fr",
-        "relative_to_run": 0.2907585236767817
+        "relative_to_run": 0.28300129290842074
       }
     ],
     "stratum": [
@@ -390,217 +390,217 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 23.4235782623291,
-            "mean_kld": 0.05388702262702585,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.032507454946453476,
-            "median_context_p99": 0.5249273777008057,
-            "p90_context_kld": 0.1189590889010681,
+            "max_kld": 29.409290313720703,
+            "mean_kld": 0.05343841309902987,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.03457746948537548,
+            "median_context_p99": 0.48533299565315247,
+            "p90_context_kld": 0.13830447681183197,
             "positions": 196512,
-            "top1_agreement": 0.9609540384302231,
+            "top1_agreement": 0.9609489496824621,
             "worst_context_id": 454,
-            "worst_context_kld": 0.4895784940445628
+            "worst_context_kld": 0.45672114841615463
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 4.877828697867478
+        "relative_to_run": 4.86925003215088
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 1.8940415382385254,
-            "mean_kld": 0.008991718850348643,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.005429992133758688,
-            "median_context_p99": 0.04178592562675476,
-            "p90_context_kld": 0.01635930600748916,
+            "max_kld": 1.4653857946395874,
+            "mean_kld": 0.008966686634686733,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.005603489085350864,
+            "median_context_p99": 0.043331339955329895,
+            "p90_context_kld": 0.016244145148751315,
             "positions": 147384,
-            "top1_agreement": 0.9569220539542963,
+            "top1_agreement": 0.9570238289095153,
             "worst_context_id": 893,
-            "worst_context_kld": 0.03802517904088536
+            "worst_context_kld": 0.03694091126483796
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 0.8139262871315055
+        "relative_to_run": 0.8170347256256356
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 3.1001851558685303,
-            "mean_kld": 0.0063881460233417755,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.005259914155344338,
-            "median_context_p99": 0.05044623836874962,
-            "p90_context_kld": 0.009174476160926261,
+            "max_kld": 2.1070759296417236,
+            "mean_kld": 0.006326082165752138,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.005294499655798116,
+            "median_context_p99": 0.04691578447818756,
+            "p90_context_kld": 0.008994613024871681,
             "positions": 196512,
-            "top1_agreement": 0.9654779351897085,
-            "worst_context_id": 6,
-            "worst_context_kld": 0.03374026097547736
+            "top1_agreement": 0.9654321364598599,
+            "worst_context_id": 65,
+            "worst_context_kld": 0.03102557076205901
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.5782520629224144
+        "relative_to_run": 0.5764257207992749
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 1.6461397409439087,
-            "mean_kld": 0.0048897095708480606,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.004029185885625382,
-            "median_context_p99": 0.030439462512731552,
-            "p90_context_kld": 0.009268822767389277,
+            "max_kld": 1.8053836822509766,
+            "mean_kld": 0.004882264113287302,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.0041416290521338285,
+            "median_context_p99": 0.031151100993156433,
+            "p90_context_kld": 0.009558583566512019,
             "positions": 147384,
-            "top1_agreement": 0.9685176138522499,
+            "top1_agreement": 0.9686465287955274,
             "worst_context_id": 275,
-            "worst_context_kld": 0.014223638338661505
+            "worst_context_kld": 0.014266868176758726
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.44261427902602124
+        "relative_to_run": 0.44486659150110247
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 2.1852290630340576,
-            "mean_kld": 0.004737305064792011,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.0039352627612185285,
-            "median_context_p99": 0.03270211070775986,
-            "p90_context_kld": 0.009271094984799227,
+            "max_kld": 1.9099842309951782,
+            "mean_kld": 0.0047161988101167476,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.003908452587081799,
+            "median_context_p99": 0.029179947450757027,
+            "p90_context_kld": 0.008553707324987884,
             "positions": 73692,
-            "top1_agreement": 0.9693725234760897,
+            "top1_agreement": 0.9691011235955056,
             "worst_context_id": 953,
-            "worst_context_kld": 0.011588845220650184
+            "worst_context_kld": 0.01122486185099774
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.42881869268476214
+        "relative_to_run": 0.4297349018436293
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 1.4300388097763062,
-            "mean_kld": 0.004714006706887872,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.004113897056717312,
-            "median_context_p99": 0.024136444553732872,
-            "p90_context_kld": 0.00561418505282599,
+            "max_kld": 1.165907382965088,
+            "mean_kld": 0.004699737983564273,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.00391996123187938,
+            "median_context_p99": 0.024366630241274834,
+            "p90_context_kld": 0.005416538875010646,
             "positions": 147384,
-            "top1_agreement": 0.9647315855181023,
+            "top1_agreement": 0.964704445530044,
             "worst_context_id": 443,
-            "worst_context_kld": 0.027228787454472464
+            "worst_context_kld": 0.027905596454365553
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.4267097359590476
+        "relative_to_run": 0.42823500924630736
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 2.663503885269165,
-            "mean_kld": 0.004137738224923952,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.003228649903561243,
-            "median_context_p99": 0.03253652900457382,
-            "p90_context_kld": 0.005859838909925266,
+            "max_kld": 8.64619255065918,
+            "mean_kld": 0.004225324387170233,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.003200656097436244,
+            "median_context_p99": 0.03228462114930153,
+            "p90_context_kld": 0.00674611330592368,
             "positions": 196512,
-            "top1_agreement": 0.9797518726591761,
+            "top1_agreement": 0.9793040628562123,
             "worst_context_id": 667,
-            "worst_context_kld": 0.023852551947544348
+            "worst_context_kld": 0.026218505325394093
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.37454617594097434
+        "relative_to_run": 0.38500695875735247
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 3.2773475646972656,
-            "mean_kld": 0.003763443945706886,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.002558179151218321,
-            "median_context_p99": 0.035675883293151855,
-            "p90_context_kld": 0.006300030598680195,
+            "max_kld": 1.8485336303710938,
+            "mean_kld": 0.0037327059215536552,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.0027118550407615144,
+            "median_context_p99": 0.03762500360608101,
+            "p90_context_kld": 0.0068965462201931475,
             "positions": 73692,
-            "top1_agreement": 0.9357189382836671,
+            "top1_agreement": 0.9357460782717255,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.019168267112011875
+            "worst_context_kld": 0.01781900445289071
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.3406652285884115
+        "relative_to_run": 0.340120100401426
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 24.25029754638672,
-            "mean_kld": 0.0035539641212878384,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.002869848601152158,
-            "median_context_p99": 0.018992044031620026,
-            "p90_context_kld": 0.0045202462159064935,
+            "max_kld": 11.368499755859375,
+            "mean_kld": 0.0034699791707562626,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.002857173882379502,
+            "median_context_p99": 0.018883395940065384,
+            "p90_context_kld": 0.004444196419291071,
             "positions": 196512,
-            "top1_agreement": 0.9762915241817294,
+            "top1_agreement": 0.9766731802638007,
             "worst_context_id": 709,
-            "worst_context_kld": 0.027977719611953035
+            "worst_context_kld": 0.02107575118209227
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.3217032104741836
+        "relative_to_run": 0.3161807248552924
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 2.442533493041992,
-            "mean_kld": 0.003277476263199609,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.003168675314512905,
-            "median_context_p99": 0.022142626345157623,
-            "p90_context_kld": 0.0041321819357580175,
+            "max_kld": 2.281120538711548,
+            "mean_kld": 0.0032577067486191185,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.0031907973419261406,
+            "median_context_p99": 0.021607771515846252,
+            "p90_context_kld": 0.004144200098769572,
             "positions": 196512,
-            "top1_agreement": 0.9721085735222277,
-            "worst_context_id": 246,
-            "worst_context_kld": 0.005805297005183099
+            "top1_agreement": 0.9721340172610324,
+            "worst_context_id": 174,
+            "worst_context_kld": 0.005824702771456004
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.2966756557300793
+        "relative_to_run": 0.2968386928155199
       }
     ]
   },
   "overall": {
     "deployed": {
       "contexts": 768,
-      "max_kld": 24.25029754638672,
-      "mean_kld": 0.011047338060597442,
-      "mean_ref_top1_prob": 0.6415440866166368,
-      "median_context_kld": 0.003959233488364337,
-      "median_context_p99": 0.0307154580950737,
-      "p90_context_kld": 0.018751727010197644,
+      "max_kld": 29.409290313720703,
+      "mean_kld": 0.01097467017429472,
+      "mean_ref_top1_prob": 0.6415473983687051,
+      "median_context_kld": 0.003896204956611177,
+      "median_context_p99": 0.030610591173171997,
+      "p90_context_kld": 0.01781900445289071,
       "positions": 1572096,
-      "top1_agreement": 0.9670777102670575,
+      "top1_agreement": 0.9670738937062368,
       "worst_context_id": 454,
-      "worst_context_kld": 0.4895784940445628
+      "worst_context_kld": 0.45672114841615463
     }
   },
   "primary": "deployed"

Qwen3.8-27B-FP8/strata.md

@@ -1,52 +1,52 @@
 # Fidelity by domain - Qwen3.8-27B / Qwen3.8-27B-FP8
 
-768 contexts, 1572096 scored positions, mean 0.01104734, reference top-1 64.2%, top-1 agreement 96.7078%.
+768 contexts, 1572096 scored positions, mean 0.01097467, reference top-1 64.2%, top-1 agreement 96.7074%.
 
-`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
+`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. The `deployed` cell is QxQ; `bxq` is the teacher-ID counterfactual on the same student weights. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
 
 ## Ranked by domain
 
 | Domain | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.05388702 | 4.88 | 0.03250745 | 0.11895909 | 0.489578 |
-| Chinese across several content types | 72 | 55.6% | 0.00899172 | 0.81 | 0.00542999 | 0.01635931 | 0.038025 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.00638815 | 0.58 | 0.00525991 | 0.00917448 | 0.033740 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.00488971 | 0.44 | 0.00402919 | 0.00926882 | 0.014224 |
-| Other multilingual content | 36 | 64.5% | 0.00473731 | 0.43 | 0.00393526 | 0.00927109 | 0.011589 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.00471401 | 0.43 | 0.00411390 | 0.00561419 | 0.027229 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.00413774 | 0.37 | 0.00322865 | 0.00585984 | 0.023853 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.00376344 | 0.34 | 0.00255818 | 0.00630003 | 0.019168 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.00355396 | 0.32 | 0.00286985 | 0.00452025 | 0.027978 |
-| Scientific and technical exposition | 96 | 60.3% | 0.00327748 | 0.30 | 0.00316868 | 0.00413218 | 0.005805 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.05343841 | 4.87 | 0.03457747 | 0.13830448 | 0.456721 |
+| Chinese across several content types | 72 | 55.6% | 0.00896669 | 0.82 | 0.00560349 | 0.01624415 | 0.036941 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.00632608 | 0.58 | 0.00529450 | 0.00899461 | 0.031026 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.00488226 | 0.44 | 0.00414163 | 0.00955858 | 0.014267 |
+| Other multilingual content | 36 | 64.5% | 0.00471620 | 0.43 | 0.00390845 | 0.00855371 | 0.011225 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.00469974 | 0.43 | 0.00391996 | 0.00541654 | 0.027906 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.00422532 | 0.39 | 0.00320066 | 0.00674611 | 0.026219 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.00373271 | 0.34 | 0.00271186 | 0.00689655 | 0.017819 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.00346998 | 0.32 | 0.00285717 | 0.00444420 | 0.021076 |
+| Scientific and technical exposition | 96 | 60.3% | 0.00325771 | 0.30 | 0.00319080 | 0.00414420 | 0.005825 |
 
 ## Ranked by source dataset
 
 | Source | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| wildchat | 96 | 73.5% | 0.05388702 | 4.88 | 0.03250745 | 0.11895909 | 0.489578 |
-| wikisource_zh | 33 | 60.4% | 0.01449425 | 1.31 | 0.01123513 | 0.03288935 | 0.038025 |
-| wikipedia_de | 7 | 73.6% | 0.00847814 | 0.77 | 0.00927109 | 0.01158885 | 0.011589 |
-| regulations | 23 | 72.3% | 0.00683912 | 0.62 | 0.00778065 | 0.01142324 | 0.014224 |
-| wikipedia_en | 96 | 59.5% | 0.00638815 | 0.58 | 0.00525991 | 0.00917448 | 0.033740 |
-| public_domain_books | 72 | 50.6% | 0.00471401 | 0.43 | 0.00411390 | 0.00561419 | 0.027229 |
-| public_domain_review | 26 | 51.8% | 0.00451705 | 0.41 | 0.00411848 | 0.00646793 | 0.011862 |
-| wikipedia_zh | 39 | 51.5% | 0.00433573 | 0.39 | 0.00420734 | 0.00592347 | 0.007191 |
-| github_code | 52 | 85.8% | 0.00418031 | 0.38 | 0.00312403 | 0.00585984 | 0.023853 |
-| wikipedia_ja | 7 | 57.0% | 0.00410554 | 0.37 | 0.00404828 | 0.00493403 | 0.004934 |
-| stackv2 | 44 | 72.0% | 0.00408742 | 0.37 | 0.00333187 | 0.00577160 | 0.023764 |
-| wikipedia_cs | 6 | 68.9% | 0.00400275 | 0.36 | 0.00415584 | 0.00503012 | 0.005030 |
-| wikipedia_es | 7 | 59.2% | 0.00400042 | 0.36 | 0.00393526 | 0.00493311 | 0.004933 |
-| starcoder_structured | 36 | 70.7% | 0.00376344 | 0.34 | 0.00255818 | 0.00630003 | 0.019168 |
-| libretexts | 96 | 65.8% | 0.00355396 | 0.32 | 0.00286985 | 0.00452025 | 0.027978 |
-| wikipedia_ru | 6 | 66.6% | 0.00346690 | 0.31 | 0.00356482 | 0.00495779 | 0.004958 |
-| open_news | 23 | 53.8% | 0.00336156 | 0.30 | 0.00317438 | 0.00428088 | 0.005724 |
-| scientific_papers | 96 | 60.3% | 0.00327748 | 0.30 | 0.00316868 | 0.00413218 | 0.005805 |
-| wikipedia_fr | 3 | 60.2% | 0.00321211 | 0.29 | 0.00315522 | 0.00341679 | 0.003417 |
+| wildchat | 96 | 73.5% | 0.05343841 | 4.87 | 0.03457747 | 0.13830448 | 0.456721 |
+| wikisource_zh | 33 | 60.4% | 0.01450308 | 1.32 | 0.01073703 | 0.03199231 | 0.036941 |
+| wikipedia_de | 7 | 73.6% | 0.00855001 | 0.78 | 0.00855371 | 0.01122486 | 0.011225 |
+| regulations | 23 | 72.3% | 0.00674630 | 0.61 | 0.00796010 | 0.01155533 | 0.014267 |
+| wikipedia_en | 96 | 59.5% | 0.00632608 | 0.58 | 0.00529450 | 0.00899461 | 0.031026 |
+| public_domain_books | 72 | 50.6% | 0.00469974 | 0.43 | 0.00391996 | 0.00541654 | 0.027906 |
+| public_domain_review | 26 | 51.8% | 0.00452957 | 0.41 | 0.00410826 | 0.00689971 | 0.011353 |
+| wikipedia_zh | 39 | 51.5% | 0.00428205 | 0.39 | 0.00406122 | 0.00599474 | 0.007109 |
+| github_code | 52 | 85.8% | 0.00427491 | 0.39 | 0.00309690 | 0.00674611 | 0.026219 |
+| stackv2 | 44 | 72.0% | 0.00416672 | 0.38 | 0.00326513 | 0.00612855 | 0.023565 |
+| wikipedia_ja | 7 | 57.0% | 0.00412385 | 0.38 | 0.00419477 | 0.00468995 | 0.004690 |
+| wikipedia_es | 7 | 59.2% | 0.00406109 | 0.37 | 0.00390845 | 0.00488782 | 0.004888 |
+| wikipedia_cs | 6 | 68.9% | 0.00385930 | 0.35 | 0.00402946 | 0.00469958 | 0.004700 |
+| starcoder_structured | 36 | 70.7% | 0.00373271 | 0.34 | 0.00271186 | 0.00689655 | 0.017819 |
+| libretexts | 96 | 65.8% | 0.00346998 | 0.32 | 0.00285717 | 0.00444420 | 0.021076 |
+| open_news | 23 | 53.8% | 0.00341693 | 0.31 | 0.00321398 | 0.00432250 | 0.005585 |
+| wikipedia_ru | 6 | 66.6% | 0.00336086 | 0.31 | 0.00340392 | 0.00457484 | 0.004575 |
+| scientific_papers | 96 | 60.3% | 0.00325771 | 0.30 | 0.00319080 | 0.00414420 | 0.005825 |
+| wikipedia_fr | 3 | 60.2% | 0.00310585 | 0.28 | 0.00300114 | 0.00353787 | 0.003538 |
 
 ## Reading
 
-- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.05388702, 4.88x the run mean of 0.01104734 over 96 context(s).
-- Strongest domain: **Scientific and technical exposition** at 0.00327748, 0.30x the run mean. The spread across domains is 16.4x.
+- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.05343841, 4.87x the run mean of 0.01097467 over 96 context(s).
+- Strongest domain: **Scientific and technical exposition** at 0.00325771, 0.30x the run mean. The spread across domains is 16.4x.
 - This is a strong finding: the reference was *more* certain on Natural dialogue, instruction following, and assistance (top-1 73.5% against 64.2% overall) and the candidate still diverged most there, so predictability does not explain it.
-- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 0.489578 against a median of 0.03250745 (context 454). Read the documents before treating the domain as weak.
+- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 0.456721 against a median of 0.03457747 (context 454). Read the documents before treating the domain as weak.
 

Qwen3.8-27B-INT4/compliance.json

@@ -1,8 +1,9 @@
 {
   "attribution": null,
   "candidate": "/media/fmodels2/RedHatAI/Qwen3.8-27B-INT4",
-  "candidate_weights_sha256": null,
+  "candidate_weights_sha256": "30328786e1d18b3932124b3d3c87c8282222e03bd6dced09f625fadb2f9aeba6",
   "comparability_key": {
+    "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
     "context_length": 2048,
     "driver": "580.173.02",
     "gpu_names": [
@@ -12,19 +13,22 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 248044,
-    "laws_version": 9,
+    "kv_cache_dtype": "bfloat16",
+    "laws_version": 15,
     "model_runner_v2": false,
+    "numerics_digest": "251a9225b37415b91fc58d975b0670fc4b481e6f5a816b663573c42e161438f8",
     "reference_config_sha256": "191e0af232104ed8b65258cf3fb2b842e288008baca7633c11b82a1ac7203aab",
     "rows": 768,
     "score_from": 0,
     "stride": 2048,
+    "substituted_parameters": [],
     "suite_id": "qwen3.8-27b-fidelity-1024x2048-v1",
     "tensor_parallel_size": 1,
     "token_sha256": "9c935708bbffbf45d5baa2931fb4af7e7b22df1e041aa0b6f4ca44d3315e543b",
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T12:56:04.735742+00:00",
+  "evaluated_at": "2026-09-11T20:24:40.478284+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -52,13 +56,13 @@
       "title": "Real vocabulary"
     },
     {
-      "detail": "all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc",
+      "detail": "all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3",
       "law": 5,
       "status": "pass",
       "title": "Manifest binding"
     },
     {
-      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s)",
+      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s)",
       "law": 6,
       "status": "pass",
       "title": "Provenance"
@@ -70,13 +74,13 @@
       "title": "Storage integrity"
     },
     {
-      "detail": "trunk 0.06297880, deployed 0.06297880, delta 4.952845258920924e-09",
+      "detail": "trunk 0.06290330, deployed 0.06290331, delta 5.024191465641259e-09",
       "law": 8,
       "status": "pass",
       "title": "Head transparency"
     },
     {
-      "detail": "mean 0.06297880, median 0.01324771, max 31.91850281, 4 depth buckets",
+      "detail": "mean 0.06290331, median 0.01324692, max 31.67959976, 4 depth buckets",
       "law": 9,
       "status": "pass",
       "title": "Tail and depth disclosure"
@@ -103,38 +107,37 @@
       "detail": "reference declares no experts",
       "law": 14,
       "status": "not_applicable",
-      "title": "Component attribution"
+      "title": "Routed-model intervention"
     },
     {
-      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.14460520, strongest scientific_technical at 0.02608180, spread 5.5x",
+      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.14393194, strongest scientific_technical at 0.02608979, spread 5.5x",
       "law": 15,
       "status": "pass",
       "title": "Domain disclosure"
     },
     {
-      "approval": {
-        "approver": "Andy Kitzke",
-        "justification": "These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger.",
-        "timestamp": "2026-09-02T05:15:00Z"
-      },
-      "detail": "the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else",
+      "detail": "scored weights 30328786e1d18b39 as inspected",
       "law": 16,
-      "status": "override",
+      "status": "pass",
       "title": "Candidate weight binding"
     },
     {
-      "detail": "1 override(s), each fully attributed",
+      "detail": "every scored layer used the checkpoint's own quantization parameters",
+      "law": 17,
+      "status": "pass",
+      "title": "Substitution disclosure"
+    },
+    {
+      "detail": "no overrides claimed",
       "law": 13,
       "status": "pass",
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 9,
-  "mean_kld": 0.06297880022919693,
+  "laws_version": 15,
+  "mean_kld": 0.06290330991231023,
   "nondeterminism_floor": 0.0,
-  "overridden_laws": [
-    16
-  ],
+  "overridden_laws": [],
   "partition": "analysis",
   "program": "Local Inference Lab \u2014 Distribution Fidelity",
   "ranking_floor": null,
@@ -147,16 +150,16 @@
     "overall": {
       "deployed": {
         "contexts": 768,
-        "max_kld": 31.918502807617188,
-        "mean_kld": 0.06297880022919694,
-        "mean_ref_top1_prob": 0.6415440866166368,
-        "median_context_kld": 0.03624156884367412,
-        "median_context_p99": 0.34093016386032104,
-        "p90_context_kld": 0.13468790532427413,
+        "max_kld": 31.67959976196289,
+        "mean_kld": 0.06290330991231023,
+        "mean_ref_top1_prob": 0.6415473983687051,
+        "median_context_kld": 0.03616175329293096,
+        "median_context_p99": 0.3476375341415405,
+        "p90_context_kld": 0.1364919523744591,
         "positions": 1572096,
-        "top1_agreement": 0.911584915933887,
+        "top1_agreement": 0.9116052709249308,
         "worst_context_id": 454,
-        "worst_context_kld": 1.0401913347215792
+        "worst_context_kld": 0.9887559796985864
       }
     },
     "primary": "deployed",
@@ -165,201 +168,201 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 31.918502807617188,
-            "mean_kld": 0.14460519754814047,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.12319854523112893,
-            "median_context_p99": 2.259077787399292,
-            "p90_context_kld": 0.2993194405564322,
+            "max_kld": 31.67959976196289,
+            "mean_kld": 0.14393193833588852,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.11954766931313557,
+            "median_context_p99": 2.314512252807617,
+            "p90_context_kld": 0.319737868304473,
             "positions": 196512,
-            "top1_agreement": 0.9142037127503664,
+            "top1_agreement": 0.9143818189219997,
             "worst_context_id": 454,
-            "worst_context_kld": 1.0401913347215792
+            "worst_context_kld": 0.9887559796985864
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 2.2960932412475774
+        "relative_to_run": 2.2881457038832376
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.908188343048096,
-            "mean_kld": 0.1345311533987298,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.0800067759461001,
-            "median_context_p99": 0.6115172505378723,
-            "p90_context_kld": 0.2779199934738752,
+            "max_kld": 7.039397239685059,
+            "mean_kld": 0.13454855501209217,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.0797083946501763,
+            "median_context_p99": 0.6599892377853394,
+            "p90_context_kld": 0.2783823526198344,
             "positions": 147384,
-            "top1_agreement": 0.845030668186506,
+            "top1_agreement": 0.8443250284969874,
             "worst_context_id": 893,
-            "worst_context_kld": 0.5674233458273176
+            "worst_context_kld": 0.5681165746871794
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 2.1361339515699638
+        "relative_to_run": 2.1389741684445274
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 8.464103698730469,
-            "mean_kld": 0.06222418345240118,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.04821683398172381,
-            "median_context_p99": 0.541857123374939,
-            "p90_context_kld": 0.09363392489135787,
+            "max_kld": 8.426396369934082,
+            "mean_kld": 0.062162872968970516,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.0483626568274087,
+            "median_context_p99": 0.5371996760368347,
+            "p90_context_kld": 0.09342251370086758,
             "positions": 196512,
-            "top1_agreement": 0.9018736769255822,
+            "top1_agreement": 0.9020823155837812,
             "worst_context_id": 6,
-            "worst_context_kld": 0.351286367924311
+            "worst_context_kld": 0.3535688326255857
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.988017923903766
+        "relative_to_run": 0.9882289668958293
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 9.018062591552734,
-            "mean_kld": 0.05436721019998957,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.0455443970797962,
-            "median_context_p99": 0.4305042028427124,
-            "p90_context_kld": 0.10590312001808456,
+            "max_kld": 8.899713516235352,
+            "mean_kld": 0.054424187539078905,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.04558276349529022,
+            "median_context_p99": 0.42817598581314087,
+            "p90_context_kld": 0.10664729455358758,
             "positions": 73692,
-            "top1_agreement": 0.9045758019866471,
+            "top1_agreement": 0.9036937523747489,
             "worst_context_id": 947,
-            "worst_context_kld": 0.1165199785020918
+            "worst_context_kld": 0.1190153221496802
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.8632620818772753
+        "relative_to_run": 0.8652038758365567
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.404207706451416,
-            "mean_kld": 0.04707976949886068,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.03371621264780603,
-            "median_context_p99": 0.2907443940639496,
-            "p90_context_kld": 0.11144659651404887,
+            "max_kld": 7.351467609405518,
+            "mean_kld": 0.04704711827297485,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.033557425424993736,
+            "median_context_p99": 0.29649659991264343,
+            "p90_context_kld": 0.11091345794084626,
             "positions": 147384,
-            "top1_agreement": 0.9157710470607393,
+            "top1_agreement": 0.9159406719861043,
             "worst_context_id": 317,
-            "worst_context_kld": 0.15066450098557546
+            "worst_context_kld": 0.15121817439717355
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.747549482167406
+        "relative_to_run": 0.7479275468740905
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.460888862609863,
-            "mean_kld": 0.044350334816430914,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.03592069204934934,
-            "median_context_p99": 0.24048706889152527,
-            "p90_context_kld": 0.0533476602926932,
+            "max_kld": 8.17362117767334,
+            "mean_kld": 0.04437409335639619,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.03576088007500376,
+            "median_context_p99": 0.23678484559059143,
+            "p90_context_kld": 0.05390548673075974,
             "positions": 147384,
-            "top1_agreement": 0.899609184171959,
+            "top1_agreement": 0.8995549041958422,
             "worst_context_id": 443,
-            "worst_context_kld": 0.31145168263573736
+            "worst_context_kld": 0.3110997550139541
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.7042105383879657
+        "relative_to_run": 0.7054333614281264
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 8.12700366973877,
-            "mean_kld": 0.040566097360712344,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.032986131219142754,
-            "median_context_p99": 0.4861080050468445,
-            "p90_context_kld": 0.06758088454436428,
+            "max_kld": 8.626458168029785,
+            "mean_kld": 0.04060035080074585,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.0334252776297643,
+            "median_context_p99": 0.5161235928535461,
+            "p90_context_kld": 0.06731926693285788,
             "positions": 73692,
-            "top1_agreement": 0.8673261683764859,
+            "top1_agreement": 0.8677468381913912,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.17541399166167326
+            "worst_context_kld": 0.17577993098848518
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.6441230574904779
+        "relative_to_run": 0.6454406112706055
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.63759994506836,
-            "mean_kld": 0.03917626801646627,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.023835338446136917,
-            "median_context_p99": 0.3008233308792114,
-            "p90_context_kld": 0.07485065233115942,
+            "max_kld": 12.795402526855469,
+            "mean_kld": 0.03929107550581172,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.02376802223446685,
+            "median_context_p99": 0.30457040667533875,
+            "p90_context_kld": 0.07510747058223882,
             "positions": 196512,
-            "top1_agreement": 0.9471991532323726,
+            "top1_agreement": 0.9473721706562449,
             "worst_context_id": 667,
-            "worst_context_kld": 0.30212707418198775
+            "worst_context_kld": 0.3038625150634299
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.6220548482011916
+        "relative_to_run": 0.6246265190271398
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 21.455047607421875,
-            "mean_kld": 0.026672020564325254,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.020993579961020265,
-            "median_context_p99": 0.16144904494285583,
-            "p90_context_kld": 0.040855219811793145,
+            "max_kld": 9.440866470336914,
+            "mean_kld": 0.02663927997817303,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.02102722273338084,
+            "median_context_p99": 0.1589260697364807,
+            "p90_context_kld": 0.04065772069359729,
             "positions": 196512,
-            "top1_agreement": 0.9386195245074093,
+            "top1_agreement": 0.9386958557238235,
             "worst_context_id": 825,
-            "worst_context_kld": 0.10989124282970031
+            "worst_context_kld": 0.11036421424426199
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.4235079180177859
+        "relative_to_run": 0.4234956795645455
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.329400062561035,
-            "mean_kld": 0.026081798631463017,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.024741105384693707,
-            "median_context_p99": 0.19944342970848083,
-            "p90_context_kld": 0.037098149314752825,
+            "max_kld": 12.36015510559082,
+            "mean_kld": 0.026089785651106444,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.024915956636511808,
+            "median_context_p99": 0.2050503045320511,
+            "p90_context_kld": 0.036752015789522266,
             "positions": 196512,
-            "top1_agreement": 0.9310118466047875,
+            "top1_agreement": 0.9311543315420941,
             "worst_context_id": 246,
-            "worst_context_kld": 0.060966793708638514
+            "worst_context_kld": 0.06105584639032175
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.41413616227276917
+        "relative_to_run": 0.4147601404040052
       }
     ]
   },

Qwen3.8-27B-INT4/inspect.json

@@ -1,31 +1,37 @@
 {
+  "model": "/media/fmodels2/RedHatAI/Qwen3.8-27B-INT4",
+  "inspect_version": 5,
+  "weights_sha256": "30328786e1d18b3932124b3d3c87c8282222e03bd6dced09f625fadb2f9aeba6",
+  "weights_bytes": 19452784112,
+  "quant_method": "compressed-tensors",
+  "declared": {},
+  "detected_scheme": "int4_g128_sym",
+  "detected_block": 128,
+  "quant_algorithm": "round_to_nearest",
+  "scheme_mix": null,
+  "kv_cache_scheme": null,
   "coverage": {
-    "attention": {
-      "quantized": 64,
-      "weights": 4
-    },
-    "dense_mlp": {
-      "quantized": 192,
-      "weights": 3
+    "router": {
+      "weights": 0,
+      "quantized": 0
     },
     "experts": {
-      "quantized": 0,
-      "weights": 0
-    },
-    "router": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
     },
     "shared_expert": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
+    },
+    "attention": {
+      "weights": 68,
+      "quantized": 64
+    },
+    "dense_mlp": {
+      "weights": 195,
+      "quantized": 192
     }
   },
-  "declared": {},
-  "detected_block": null,
-  "detected_scheme": null,
-  "model": "/media/fmodels2/RedHatAI/Qwen3.8-27B-INT4",
-  "quant_method": "compressed-tensors",
-  "weights_bytes": 19452784112,
-  "weights_bytes_source": "hub"
+  "unloadable_reason": null,
+  "quantized_names_sha256": "eccbe240ef5caa18a896c8a311d4a9b4257c68423f0d8f9ae1c982d53797e657"
 }

Qwen3.8-27B-INT4/report.md

@@ -1,12 +1,12 @@
 # Qwen3.8-27B / Qwen3.8-27B-INT4: distribution fidelity
 
-**Mean KLD(reference || candidate) = 0.06297880** over 1572096 scored positions.
+**Mean KLD(reference || candidate) = 0.06290331** over 1572096 scored positions.
 
 | Mean | Median | p90 | p99 | Max | Top-1 agreement |
 |---|---|---|---|---|---|
-| 0.06297880 | 0.01324771 | 0.09463928 | 0.92820790 | 31.91850281 | 91.1585% |
+| 0.06290331 | 0.01324692 | 0.09464278 | 0.92733533 | 31.67959976 | 91.1605% |
 
-Reverse direction, KLD(candidate || reference): 0.06984384.
+Reverse direction, KLD(candidate || reference): 0.06989908.
 
 ## Identity
 
@@ -15,10 +15,10 @@ Reverse direction, KLD(candidate || reference): 0.06984384.
 | Reference checkpoint | /media/fmodels2/Qwen/Qwen3.8-27B |
 | Reference config SHA-256 | 191e0af23210 |
 | Candidate checkpoint | /media/fmodels2/RedHatAI/Qwen3.8-27B-INT4 |
-| Candidate weights SHA-256 | unbound (Law 16 gap) |
+| Candidate weights SHA-256 | 30328786e1d18b39 |
 | Suite | qwen3.8-27b-fidelity-1024x2048-v1 |
 | Suite token SHA-256 | 9c935708bbffbf45 |
-| Capture manifest SHA-256 | 304794b4e35b326d |
+| Capture manifest SHA-256 | dbfacc9cbb6d3e06 |
 | Tokenizer | None |
 | Scored vocabulary | 248044 |
 | Declared vocabulary | 248320 |
@@ -27,32 +27,48 @@ Reverse direction, KLD(candidate || reference): 0.06984384.
 | Model runner | V1 |
 | Tensor parallel | 1 |
 | Eager enforced | True |
+| KV cache | bfloat16 |
 | Prefix caching | False |
 | max_num_seqs | 1 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45 |
+| vLLM commit | 60071d1ab732 |
+| vLLM dirty digest | e3b0c44298fc1c14 |
+| Numerics digest | 251a9225b37415b9 |
+| Compiled extensions | f2fbc7537b0f01f6 |
+| FlashInfer | 0.6.17 |
 | torch | 2.13.0+cu132 |
 | Driver | 580.173.02 |
 | GPUs | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
-| Laws version | 9 |
+| Laws version | 15 |
 | Partition | analysis |
 
+The commit says when this number was taken. The numerics digest says what took it: a hash of the runtime and the scorer, which moves only when code that can change a logit changes. Comparability is bounded by the digest, so a result stays current across commits that cannot reach a number, and two results carrying the same digest were computed by the same code whatever their commits say.
+
+## Expert kernels: declared against built
+
+| Property | Value |
+|---|---|
+| Declared for its experts | `unquantized` |
+| Expert implementation built | n/a |
+| Expert kernel built | n/a |
+| Expert layers carrying an activation scale | n/a |
+
 ## Fidelity by domain
 
-The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
+The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `deployed` is QxQ (natural student routing). `bxq` is the teacher-ID counterfactual on the same weights. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
 
 | Domain | Contexts | Reference top-1 | Mean KLD | x run | Worst context |
 |---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.14460520 | 2.30 | 1.04019133 |
-| Chinese across several content types | 72 | 55.6% | 0.13453115 | 2.14 | 0.56742335 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.06222418 | 0.99 | 0.35128637 |
-| Other multilingual content | 36 | 64.5% | 0.05436721 | 0.86 | 0.11651998 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.04707977 | 0.75 | 0.15066450 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.04435033 | 0.70 | 0.31145168 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.04056610 | 0.64 | 0.17541399 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.03917627 | 0.62 | 0.30212707 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.02667202 | 0.42 | 0.10989124 |
-| Scientific and technical exposition | 96 | 60.3% | 0.02608180 | 0.41 | 0.06096679 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.14393194 | 2.29 | 0.98875598 |
+| Chinese across several content types | 72 | 55.6% | 0.13454856 | 2.14 | 0.56811657 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.06216287 | 0.99 | 0.35356883 |
+| Other multilingual content | 36 | 64.5% | 0.05442419 | 0.87 | 0.11901532 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.04704712 | 0.75 | 0.15121817 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.04437409 | 0.71 | 0.31109976 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.04060035 | 0.65 | 0.17577993 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.03929108 | 0.62 | 0.30386252 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.02663928 | 0.42 | 0.11036421 |
+| Scientific and technical exposition | 96 | 60.3% | 0.02608979 | 0.41 | 0.06105585 |
 
 **Natural dialogue, instruction following, and assistance** is this candidate's weakest domain and **Scientific and technical exposition** its strongest, a spread of 5.5x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
 
@@ -60,9 +76,9 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Component | Value |
 |---|---|
-| Trunk (candidate hidden states, reference head) | 0.06297880 |
-| Deployed (candidate's own head) | 0.06297880 |
-| Head-associated delta (not additive) | 0.00000000 |
+| Trunk (candidate hidden states, reference head) | 0.06290330 |
+| Deployed (candidate's own head) | 0.06290331 |
+| Head-associated delta (not additive) | 0.00000001 |
 | Reference head | {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]} |
 | Candidate head | {'runtime': {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]}, 'state': 'unquantized', 'static': {'ignored': True, 'lm_head_dtypes': {'lm_head.weight': 'BF16', 'model.language_model.embed_tokens.weight': 'BF16'}, 'lm_head_keys': ['lm_head.weight'], 'output_weight_keys': ['lm_head.weight'], 'packed_keys': [], 'quant_method': 'compressed-tensors', 'state': 'unquantized', 'tie_word_embeddings': False}} |
 
@@ -70,26 +86,26 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Position range | Positions | Mean KLD |
 |---|---|---|
-| 0–511 | 393216 | 0.05613785 |
-| 512–1023 | 393216 | 0.05977902 |
-| 1024–1535 | 393216 | 0.06616904 |
-| 1536–2046 | 392448 | 0.06984270 |
+| 0–511 | 393216 | 0.05612833 |
+| 512–1023 | 393216 | 0.05990848 |
+| 1024–1535 | 393216 | 0.06592000 |
+| 1536–2046 | 392448 | 0.06966964 |
 
 ## Error by reference confidence
 
 | Reference top-1 probability | Positions | Share | Mean KLD |
 |---|---|---|---|
-| [0.00, 0.25) | 240569 | 15.3% | 0.07513872 |
-| [0.25, 0.50) | 346390 | 22.0% | 0.08400583 |
-| [0.50, 0.75) | 273384 | 17.4% | 0.09127543 |
-| [0.75, 0.95) | 249939 | 15.9% | 0.07164348 |
-| [0.95, 1.00) | 461814 | 29.4% | 0.01943240 |
+| [0.00, 0.25) | 240522 | 15.3% | 0.07494489 |
+| [0.25, 0.50) | 346508 | 22.0% | 0.08513541 |
+| [0.50, 0.75) | 273204 | 17.4% | 0.08957613 |
+| [0.75, 0.95) | 250041 | 15.9% | 0.07237178 |
+| [0.95, 1.00) | 461821 | 29.4% | 0.01904542 |
 
 ## Top-K set agreement
 
 | K=1 | K=2 | K=3 | K=4 | K=5 |
 |---|---|---|---|---|
-| 90.9672% | 72.3766% | 52.4421% | 35.3041% | 22.8329% |
+| 90.9625% | 72.3821% | 52.4397% | 35.2931% | 22.8305% |
 
 ## Law compliance
 
@@ -101,22 +117,19 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | 2 | Determinism | PASS | eager enforced; prefix caching off, max_num_seqs=1 |
 | 3 | Frozen input | PASS | token hash matches suite qwen3.8-27b-fidelity-1024x2048-v1 [analysis]: 9c935708bbffbf45 |
 | 4 | Real vocabulary | PASS | scored 248044 real tokens of 248320 declared (276 padding rows) |
-| 5 | Manifest binding | PASS | all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc |
-| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s) |
+| 5 | Manifest binding | PASS | all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3 |
+| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s) |
 | 7 | Storage integrity | PASS | hidden storage with bitwise-exact replay |
-| 8 | Head transparency | PASS | trunk 0.06297880, deployed 0.06297880, delta 4.952845258920924e-09 |
-| 9 | Tail and depth disclosure | PASS | mean 0.06297880, median 0.01324771, max 31.91850281, 4 depth buckets |
+| 8 | Head transparency | PASS | trunk 0.06290330, deployed 0.06290331, delta 5.024191465641259e-09 |
+| 9 | Tail and depth disclosure | PASS | mean 0.06290331, median 0.01324692, max 31.67959976, 4 depth buckets |
 | 10 | Comparability | PASS | comparability key fully resolved |
 | 11 | Freeze before qualification | NOT_APPLICABLE | partition is 'analysis' |
 | 12 | Reusable reference | PASS | suite, reference, head, and checksums present; published reference matches the scored capture on every bound field |
-| 14 | Component attribution | NOT_APPLICABLE | reference declares no experts |
-| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.14460520, strongest scientific_technical at 0.02608180, spread 5.5x |
-| 16 | Candidate weight binding | OVERRIDE | the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else |
-| 13 | Recorded deviation | PASS | 1 override(s), each fully attributed |
-
-### Recorded deviations
-
-- **Law 16 (Candidate weight binding)** overridden by Andy Kitzke at 2026-09-02T05:15:00Z. Justification: These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger. Underlying finding: the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else.
+| 14 | Routed-model intervention | NOT_APPLICABLE | reference declares no experts |
+| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.14393194, strongest scientific_technical at 0.02608979, spread 5.5x |
+| 16 | Candidate weight binding | PASS | scored weights 30328786e1d18b39 as inspected |
+| 17 | Substitution disclosure | PASS | every scored layer used the checkpoint's own quantization parameters |
+| 13 | Recorded deviation | PASS | no overrides claimed |
 
 ## Environment
 
@@ -126,7 +139,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | Platform | Linux-6.8.0-137-generic-x86_64-with-glibc2.39 |
 | Python | 3.12.3 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45fbe0b0205e70893bcd28fa928a88 |
+| vLLM commit | 60071d1ab73229321712254b1350dcaaac1c125a |
 | torch | 2.13.0+cu132 |
 | torch CUDA runtime | 13.2 |
 | cuDNN | 9.20.0 (92000) |
@@ -134,7 +147,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | CUDA arch list | sm_75, sm_80, sm_86, sm_90, sm_100, sm_120 |
 | nvcc | Cuda compilation tools, release 13.0, V13.0.88 |
 | gcc | gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0 |
-| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.8) 2.39 |
+| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.9) 2.39 |
 | NVIDIA driver | 580.173.02 |
 | float32 matmul precision | highest |
 | TF32 (matmul / cuDNN) | False / True |
@@ -157,8 +170,18 @@ Credential values are never published. Set but redacted: `HF_TOKEN`.
 
 | Variable | Value |
 |---|---|
+| `CUBLAS_WORKSPACE_CONFIG` | `:4096:8` |
 | `CUDA_HOME` | `/usr/local/cuda-13.0` |
 | `HF_TOKEN` | `<redacted>` |
+| `NCCL_DETERMINISTIC` | `1` |
+| `PYTORCH_NVML_BASED_CUDA_CHECK` | `1` |
+| `TORCHINDUCTOR_CACHE_DIR` | `/tmp/torchinductor_phaedawg` |
+| `TORCHINDUCTOR_COMPILE_THREADS` | `1` |
+| `TRITON_CACHE_AUTOTUNING` | `1` |
+| `TRITON_PTXAS_BLACKWELL_PATH` | `/usr/local/cuda-13.0/bin/ptxas` |
+| `VLLM_BATCH_INVARIANT` | `1` |
+| `VLLM_MARLIN_USE_ATOMIC_ADD` | `0` |
+| `VLLM_MOE_USE_DEEP_GEMM` | `0` |
 
 ## Files in this artifact
 
@@ -166,26 +189,26 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 
 | Path | Size | What it is |
 |---|---|---|
-| `Qwen3.8-27B-INT4/report.md` | 12.32 KiB | This document. |
-| `Qwen3.8-27B-INT4/report.json` | 254.05 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
-| `Qwen3.8-27B-INT4/manifest.json` | 85.43 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
-| `Qwen3.8-27B-INT4/compliance.json` | 13.00 KiB | The law-by-law receipt, including the comparability key. |
-| `baselines/self-kld.json` | 4.74 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
+| `Qwen3.8-27B-INT4/report.md` | 13.16 KiB | This document. |
+| `Qwen3.8-27B-INT4/report.json` | 256.51 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
+| `Qwen3.8-27B-INT4/manifest.json` | 86.99 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
+| `Qwen3.8-27B-INT4/compliance.json` | 12.27 KiB | The law-by-law receipt, including the comparability key. |
+| `baselines/self-kld.json` | 7.06 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
 | `suite/suite-manifest.json` | 382.68 KiB | The frozen evaluation input's identity: token hashes per context and per partition, sources, strata, and the analysis/qualification split. |
 | `suite/tokens` | 15.49 MiB in 1024 files | The token IDs themselves. These are the evaluation input, not a description of it; retokenizing source text does not reproduce them. |
 | `suite/sources.json` | 737.42 KiB | Per-context provenance: dataset, revision, licence, source unit, and the deterministic token offset chosen within the document. |
 | `suite/validation/capability-overlap.json` | 1.48 KiB | The benchmark-contamination scan and every document it blocked. |
-| `reference/manifest.json` | 85.43 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
+| `reference/manifest.json` | 86.99 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
 | `reference` | 19.25 GiB in 771 files | The reusable reference distributions. Pass this directory as `--reference-logits` to score a new candidate against the same reference without loading the reference checkpoint. |
 | `reference/lm_head.safetensors` | 2.37 GiB | The reference language-model head, which turns the stored hidden states back into reference logits. |
-| `environment/runtime.json` | 2.26 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
+| `environment/runtime.json` | 5.42 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
 | `environment/summary.md` | 1.18 KiB | The same provenance as prose, plus an index of every captured file. |
 | `environment/toolchain-nvcc.txt` | 243 B | `nvcc --version` verbatim; `toolchain-gcc.txt` and `toolchain-ldd.txt` sit beside it. |
 | `environment/gpu-smi-query.txt` | 54.73 KiB | `nvidia-smi -q` verbatim: ECC state, persistence mode, clocks, and throttle reasons, any of which can move a bitwise result. |
-| `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
-| `environment/models` | 7.22 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
-| `checksums.txt` | 182.99 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
+| `environment/pip-freeze.txt` | 4.47 KiB | Every installed package version in the scoring environment. |
+| `environment/models` | 19.46 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
+| `checksums.txt` | 183.22 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
+| `LAWS.md` | 40.20 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

Qwen3.8-27B-INT4/strata.json

@@ -8,381 +8,381 @@
         "cells": {
           "deployed": {
             "contexts": 33,
-            "max_kld": 6.908188343048096,
-            "mean_kld": 0.22439661797487742,
-            "mean_ref_top1_prob": 0.6040606719329803,
-            "median_context_kld": 0.17356615911634368,
-            "median_context_p99": 1.2664220333099365,
-            "p90_context_kld": 0.5323160650215791,
+            "max_kld": 7.039397239685059,
+            "mean_kld": 0.2244213433710845,
+            "mean_ref_top1_prob": 0.6040861921682756,
+            "median_context_kld": 0.17347105847250482,
+            "median_context_p99": 1.3065104484558105,
+            "p90_context_kld": 0.5332723756899984,
             "positions": 67551,
-            "top1_agreement": 0.8042219952332312,
+            "top1_agreement": 0.8038815117466803,
             "worst_context_id": 893,
-            "worst_context_kld": 0.5674233458273176
+            "worst_context_kld": 0.5681165746871794
           }
         },
         "key": "wikisource_zh",
         "label": "wikisource_zh",
-        "relative_to_run": 3.563050060627342
+        "relative_to_run": 3.567719149976956
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 31.918502807617188,
-            "mean_kld": 0.14460519754814047,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.12319854523112893,
-            "median_context_p99": 2.259077787399292,
-            "p90_context_kld": 0.2993194405564322,
+            "max_kld": 31.67959976196289,
+            "mean_kld": 0.14393193833588852,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.11954766931313557,
+            "median_context_p99": 2.314512252807617,
+            "p90_context_kld": 0.319737868304473,
             "positions": 196512,
-            "top1_agreement": 0.9142037127503664,
+            "top1_agreement": 0.9143818189219997,
             "worst_context_id": 454,
-            "worst_context_kld": 1.0401913347215792
+            "worst_context_kld": 0.9887559796985864
           }
         },
         "key": "wildchat",
         "label": "wildchat",
-        "relative_to_run": 2.2960932412475774
+        "relative_to_run": 2.2881457038832376
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 9.018062591552734,
-            "mean_kld": 0.09686725372101693,
-            "mean_ref_top1_prob": 0.7357366218042676,
-            "median_context_kld": 0.10590312001808456,
-            "median_context_p99": 1.9070539474487305,
-            "p90_context_kld": 0.1165199785020918,
+            "max_kld": 8.899713516235352,
+            "mean_kld": 0.09721810990547651,
+            "mean_ref_top1_prob": 0.7358200961573914,
+            "median_context_kld": 0.10664729455358758,
+            "median_context_p99": 1.952389121055603,
+            "p90_context_kld": 0.1190153221496802,
             "positions": 14329,
-            "top1_agreement": 0.8923162816665503,
+            "top1_agreement": 0.8901528369041803,
             "worst_context_id": 947,
-            "worst_context_kld": 0.1165199785020918
+            "worst_context_kld": 0.1190153221496802
           }
         },
         "key": "wikipedia_de",
         "label": "wikipedia_de",
-        "relative_to_run": 1.538093030805457
+        "relative_to_run": 1.5455166038321753
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 7.404207706451416,
-            "mean_kld": 0.07556401513269295,
-            "mean_ref_top1_prob": 0.7232628548554616,
-            "median_context_kld": 0.09591328063967895,
-            "median_context_p99": 1.3429927825927734,
-            "p90_context_kld": 0.1364198900844421,
+            "max_kld": 7.351467609405518,
+            "mean_kld": 0.07559704590519124,
+            "mean_ref_top1_prob": 0.7233191001139795,
+            "median_context_kld": 0.09568556294633192,
+            "median_context_p99": 1.3271344900131226,
+            "p90_context_kld": 0.1364784742438826,
             "positions": 47081,
-            "top1_agreement": 0.9151249973450012,
+            "top1_agreement": 0.91516747732631,
             "worst_context_id": 317,
-            "worst_context_kld": 0.15066450098557546
+            "worst_context_kld": 0.15121817439717355
           }
         },
         "key": "regulations",
         "label": "regulations",
-        "relative_to_run": 1.19983256044407
+        "relative_to_run": 1.2017975844288098
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 8.464103698730469,
-            "mean_kld": 0.06222418345240118,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.04821683398172381,
-            "median_context_p99": 0.541857123374939,
-            "p90_context_kld": 0.09363392489135787,
+            "max_kld": 8.426396369934082,
+            "mean_kld": 0.062162872968970516,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.0483626568274087,
+            "median_context_p99": 0.5371996760368347,
+            "p90_context_kld": 0.09342251370086758,
             "positions": 196512,
-            "top1_agreement": 0.9018736769255822,
+            "top1_agreement": 0.9020823155837812,
             "worst_context_id": 6,
-            "worst_context_kld": 0.351286367924311
+            "worst_context_kld": 0.3535688326255857
           }
         },
         "key": "wikipedia_en",
         "label": "wikipedia_en",
-        "relative_to_run": 0.988017923903766
+        "relative_to_run": 0.9882289668958293
       },
       {
         "cells": {
           "deployed": {
             "contexts": 39,
-            "max_kld": 6.461141586303711,
-            "mean_kld": 0.058491144911220275,
-            "mean_ref_top1_prob": 0.515358929042611,
-            "median_context_kld": 0.05639138502767969,
-            "median_context_p99": 0.41915103793144226,
-            "p90_context_kld": 0.08040768178868882,
+            "max_kld": 6.862327575683594,
+            "mean_kld": 0.05850234947756018,
+            "mean_ref_top1_prob": 0.5153760832971095,
+            "median_context_kld": 0.05616881779992842,
+            "median_context_p99": 0.4261801540851593,
+            "p90_context_kld": 0.08070244570411483,
             "positions": 79833,
-            "top1_agreement": 0.8795610837623539,
+            "top1_agreement": 0.8785464657472474,
             "worst_context_id": 886,
-            "worst_context_kld": 0.10124797409878483
+            "worst_context_kld": 0.10120748320986789
           }
         },
         "key": "wikipedia_zh",
         "label": "wikipedia_zh",
-        "relative_to_run": 0.9287433977521822
+        "relative_to_run": 0.9300361071478565
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 3.1750786304473877,
-            "mean_kld": 0.05241202866872408,
-            "mean_ref_top1_prob": 0.6892386900239509,
-            "median_context_kld": 0.049644270441966495,
-            "median_context_p99": 0.43684738874435425,
-            "p90_context_kld": 0.06319859898613685,
+            "max_kld": 3.189382791519165,
+            "mean_kld": 0.05242830122112039,
+            "mean_ref_top1_prob": 0.689239984531695,
+            "median_context_kld": 0.05001896781455315,
+            "median_context_p99": 0.4306614100933075,
+            "p90_context_kld": 0.06349262218184967,
             "positions": 12282,
             "top1_agreement": 0.905634261520925,
             "worst_context_id": 958,
-            "worst_context_kld": 0.06319859898613685
+            "worst_context_kld": 0.06349262218184967
           }
         },
         "key": "wikipedia_cs",
         "label": "wikipedia_cs",
-        "relative_to_run": 0.832217007595929
+        "relative_to_run": 0.8334744434626344
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 3.409838914871216,
-            "mean_kld": 0.04765670760024254,
-            "mean_ref_top1_prob": 0.5702914711488575,
-            "median_context_kld": 0.045956297929979316,
-            "median_context_p99": 0.32476675510406494,
-            "p90_context_kld": 0.05639627012088693,
+            "max_kld": 3.2061312198638916,
+            "mean_kld": 0.047665426183328016,
+            "mean_ref_top1_prob": 0.5701150598950927,
+            "median_context_kld": 0.04594255523616804,
+            "median_context_p99": 0.3387206792831421,
+            "p90_context_kld": 0.05671795958192163,
             "positions": 14329,
-            "top1_agreement": 0.8983878847093307,
+            "top1_agreement": 0.8987368274129388,
             "worst_context_id": 968,
-            "worst_context_kld": 0.05639627012088693
+            "worst_context_kld": 0.05671795958192163
           }
         },
         "key": "wikipedia_ja",
         "label": "wikipedia_ja",
-        "relative_to_run": 0.7567103124671613
+        "relative_to_run": 0.7577570441011062
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.460888862609863,
-            "mean_kld": 0.044350334816430914,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.03592069204934934,
-            "median_context_p99": 0.24048706889152527,
-            "p90_context_kld": 0.0533476602926932,
+            "max_kld": 8.17362117767334,
+            "mean_kld": 0.04437409335639619,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.03576088007500376,
+            "median_context_p99": 0.23678484559059143,
+            "p90_context_kld": 0.05390548673075974,
             "positions": 147384,
-            "top1_agreement": 0.899609184171959,
+            "top1_agreement": 0.8995549041958422,
             "worst_context_id": 443,
-            "worst_context_kld": 0.31145168263573736
+            "worst_context_kld": 0.3110997550139541
           }
         },
         "key": "public_domain_books",
         "label": "public_domain_books",
-        "relative_to_run": 0.7042105383879657
+        "relative_to_run": 0.7054333614281264
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 4.176137447357178,
-            "mean_kld": 0.04229308581186657,
-            "mean_ref_top1_prob": 0.591765751203999,
-            "median_context_kld": 0.043453887484235905,
-            "median_context_p99": 0.477895051240921,
-            "p90_context_kld": 0.056212005211155165,
+            "max_kld": 4.112191677093506,
+            "mean_kld": 0.042177579223107754,
+            "mean_ref_top1_prob": 0.591761624645417,
+            "median_context_kld": 0.04301762345270623,
+            "median_context_p99": 0.4865816831588745,
+            "p90_context_kld": 0.05590323249503212,
             "positions": 14329,
-            "top1_agreement": 0.9061344127294299,
+            "top1_agreement": 0.9046688533742759,
             "worst_context_id": 936,
-            "worst_context_kld": 0.056212005211155165
+            "worst_context_kld": 0.05590323249503212
           }
         },
         "key": "wikipedia_es",
         "label": "wikipedia_es",
-        "relative_to_run": 0.671544800122431
+        "relative_to_run": 0.6705144654852823
       },
       {
         "cells": {
           "deployed": {
             "contexts": 52,
-            "max_kld": 11.045836448669434,
-            "mean_kld": 0.04128254693956573,
-            "mean_ref_top1_prob": 0.8582575125835092,
-            "median_context_kld": 0.023671564396975233,
-            "median_context_p99": 0.29987481236457825,
-            "p90_context_kld": 0.08002233608295971,
+            "max_kld": 11.044264793395996,
+            "mean_kld": 0.041481205123543335,
+            "mean_ref_top1_prob": 0.8582941084526002,
+            "median_context_kld": 0.023669867633967098,
+            "median_context_p99": 0.30457040667533875,
+            "p90_context_kld": 0.0795575492788322,
             "positions": 106444,
-            "top1_agreement": 0.952209612566232,
+            "top1_agreement": 0.9524726616812597,
             "worst_context_id": 667,
-            "worst_context_kld": 0.30212707418198775
+            "worst_context_kld": 0.3038625150634299
           }
         },
         "key": "github_code",
         "label": "github_code",
-        "relative_to_run": 0.6554991011154124
+        "relative_to_run": 0.6594439176788919
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 8.12700366973877,
-            "mean_kld": 0.040566097360712344,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.032986131219142754,
-            "median_context_p99": 0.4861080050468445,
-            "p90_context_kld": 0.06758088454436428,
+            "max_kld": 8.626458168029785,
+            "mean_kld": 0.04060035080074585,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.0334252776297643,
+            "median_context_p99": 0.5161235928535461,
+            "p90_context_kld": 0.06731926693285788,
             "positions": 73692,
-            "top1_agreement": 0.8673261683764859,
+            "top1_agreement": 0.8677468381913912,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.17541399166167326
+            "worst_context_kld": 0.17577993098848518
           }
         },
         "key": "starcoder_structured",
         "label": "starcoder_structured",
-        "relative_to_run": 0.6441230574904779
+        "relative_to_run": 0.6454406112706055
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 3.1974751949310303,
-            "mean_kld": 0.03949404265133017,
-            "mean_ref_top1_prob": 0.6660615117323249,
-            "median_context_kld": 0.043608288440393606,
-            "median_context_p99": 0.3538494408130646,
-            "p90_context_kld": 0.052096443535001986,
+            "max_kld": 3.1497995853424072,
+            "mean_kld": 0.03953001194457589,
+            "mean_ref_top1_prob": 0.6660535411394803,
+            "median_context_kld": 0.04384944806761139,
+            "median_context_p99": 0.3535293936729431,
+            "p90_context_kld": 0.05227392440310183,
             "positions": 12282,
-            "top1_agreement": 0.9182543559680834,
+            "top1_agreement": 0.9176029962546817,
             "worst_context_id": 945,
-            "worst_context_kld": 0.052096443535001986
+            "worst_context_kld": 0.05227392440310183
           }
         },
         "key": "wikipedia_ru",
         "label": "wikipedia_ru",
-        "relative_to_run": 0.6271005879375382
+        "relative_to_run": 0.6284249906671419
       },
       {
         "cells": {
           "deployed": {
             "contexts": 26,
-            "max_kld": 5.230713844299316,
-            "mean_kld": 0.039179449811893155,
-            "mean_ref_top1_prob": 0.5182113024080622,
-            "median_context_kld": 0.034239836504682784,
-            "median_context_p99": 0.2907443940639496,
-            "p90_context_kld": 0.07070047592068328,
+            "max_kld": 5.3028950691223145,
+            "mean_kld": 0.0391037879601579,
+            "mean_ref_top1_prob": 0.5182496321630543,
+            "median_context_kld": 0.03424817697368274,
+            "median_context_p99": 0.29649659991264343,
+            "p90_context_kld": 0.07079835954222014,
             "positions": 53222,
-            "top1_agreement": 0.9109015068956446,
+            "top1_agreement": 0.9108263500056367,
             "worst_context_id": 312,
-            "worst_context_kld": 0.10701209945166093
+            "worst_context_kld": 0.10728550220238522
           }
         },
         "key": "public_domain_review",
         "label": "public_domain_review",
-        "relative_to_run": 0.6221053698912731
+        "relative_to_run": 0.6216491312566885
       },
       {
         "cells": {
           "deployed": {
             "contexts": 44,
-            "max_kld": 12.63759994506836,
-            "mean_kld": 0.036687029289166916,
-            "mean_ref_top1_prob": 0.720394649593742,
-            "median_context_kld": 0.025294000329121418,
-            "median_context_p99": 0.3360637426376343,
-            "p90_context_kld": 0.05499313809799162,
+            "max_kld": 12.795402526855469,
+            "mean_kld": 0.036702740503037984,
+            "mean_ref_top1_prob": 0.7203910827522294,
+            "median_context_kld": 0.025259543064091543,
+            "median_context_p99": 0.3422532379627228,
+            "p90_context_kld": 0.059046368797602125,
             "positions": 90068,
-            "top1_agreement": 0.9412777012923569,
+            "top1_agreement": 0.941344317626682,
             "worst_context_id": 582,
-            "worst_context_kld": 0.2905706972670119
+            "worst_context_kld": 0.28809825111830073
           }
         },
         "key": "stackv2",
         "label": "stackv2",
-        "relative_to_run": 0.5825298220298396
+        "relative_to_run": 0.5834786842568872
       },
       {
         "cells": {
           "deployed": {
             "contexts": 3,
-            "max_kld": 1.5788623094558716,
-            "mean_kld": 0.032687936449138914,
-            "mean_ref_top1_prob": 0.6023603953518782,
-            "median_context_kld": 0.03278231217327133,
-            "median_context_p99": 0.2191195785999298,
-            "p90_context_kld": 0.03937690705275581,
+            "max_kld": 1.9549686908721924,
+            "mean_kld": 0.03269768840975896,
+            "mean_ref_top1_prob": 0.6023746847155801,
+            "median_context_kld": 0.0328435920958054,
+            "median_context_p99": 0.22859367728233337,
+            "p90_context_kld": 0.03932901724185086,
             "positions": 6141,
-            "top1_agreement": 0.9145090376160234,
+            "top1_agreement": 0.9128806383325191,
             "worst_context_id": 967,
-            "worst_context_kld": 0.03937690705275581
+            "worst_context_kld": 0.03932901724185086
           }
         },
         "key": "wikipedia_fr",
         "label": "wikipedia_fr",
-        "relative_to_run": 0.5190307902052539
+        "relative_to_run": 0.5198087104691449
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 4.66934871673584,
-            "mean_kld": 0.02752632003290474,
-            "mean_ref_top1_prob": 0.5379292386352562,
-            "median_context_kld": 0.024710670167099987,
-            "median_context_p99": 0.231853187084198,
-            "p90_context_kld": 0.04196699435622174,
+            "max_kld": 4.567612171173096,
+            "mean_kld": 0.02747660751611676,
+            "mean_ref_top1_prob": 0.537879873360319,
+            "median_context_kld": 0.024121960420666787,
+            "median_context_p99": 0.22183915972709656,
+            "p90_context_kld": 0.04159375471674202,
             "positions": 47081,
-            "top1_agreement": 0.9219217943544105,
+            "top1_agreement": 0.9224952741020794,
             "worst_context_id": 343,
-            "worst_context_kld": 0.05140069161038806
+            "worst_context_kld": 0.05106101385348644
           }
         },
         "key": "open_news",
         "label": "open_news",
-        "relative_to_run": 0.43707279168115293
+        "relative_to_run": 0.43680702262599963
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 21.455047607421875,
-            "mean_kld": 0.026672020564325254,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.020993579961020265,
-            "median_context_p99": 0.16144904494285583,
-            "p90_context_kld": 0.040855219811793145,
+            "max_kld": 9.440866470336914,
+            "mean_kld": 0.02663927997817303,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.02102722273338084,
+            "median_context_p99": 0.1589260697364807,
+            "p90_context_kld": 0.04065772069359729,
             "positions": 196512,
-            "top1_agreement": 0.9386195245074093,
+            "top1_agreement": 0.9386958557238235,
             "worst_context_id": 825,
-            "worst_context_kld": 0.10989124282970031
+            "worst_context_kld": 0.11036421424426199
           }
         },
         "key": "libretexts",
         "label": "libretexts",
-        "relative_to_run": 0.4235079180177859
+        "relative_to_run": 0.4234956795645455
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.329400062561035,
-            "mean_kld": 0.026081798631463017,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.024741105384693707,
-            "median_context_p99": 0.19944342970848083,
-            "p90_context_kld": 0.037098149314752825,
+            "max_kld": 12.36015510559082,
+            "mean_kld": 0.026089785651106444,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.024915956636511808,
+            "median_context_p99": 0.2050503045320511,
+            "p90_context_kld": 0.036752015789522266,
             "positions": 196512,
-            "top1_agreement": 0.9310118466047875,
+            "top1_agreement": 0.9311543315420941,
             "worst_context_id": 246,
-            "worst_context_kld": 0.060966793708638514
+            "worst_context_kld": 0.06105584639032175
           }
         },
         "key": "scientific_papers",
         "label": "scientific_papers",
-        "relative_to_run": 0.41413616227276917
+        "relative_to_run": 0.4147601404040052
       }
     ],
     "stratum": [
@@ -390,217 +390,217 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 31.918502807617188,
-            "mean_kld": 0.14460519754814047,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.12319854523112893,
-            "median_context_p99": 2.259077787399292,
-            "p90_context_kld": 0.2993194405564322,
+            "max_kld": 31.67959976196289,
+            "mean_kld": 0.14393193833588852,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.11954766931313557,
+            "median_context_p99": 2.314512252807617,
+            "p90_context_kld": 0.319737868304473,
             "positions": 196512,
-            "top1_agreement": 0.9142037127503664,
+            "top1_agreement": 0.9143818189219997,
             "worst_context_id": 454,
-            "worst_context_kld": 1.0401913347215792
+            "worst_context_kld": 0.9887559796985864
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 2.2960932412475774
+        "relative_to_run": 2.2881457038832376
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.908188343048096,
-            "mean_kld": 0.1345311533987298,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.0800067759461001,
-            "median_context_p99": 0.6115172505378723,
-            "p90_context_kld": 0.2779199934738752,
+            "max_kld": 7.039397239685059,
+            "mean_kld": 0.13454855501209217,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.0797083946501763,
+            "median_context_p99": 0.6599892377853394,
+            "p90_context_kld": 0.2783823526198344,
             "positions": 147384,
-            "top1_agreement": 0.845030668186506,
+            "top1_agreement": 0.8443250284969874,
             "worst_context_id": 893,
-            "worst_context_kld": 0.5674233458273176
+            "worst_context_kld": 0.5681165746871794
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 2.1361339515699638
+        "relative_to_run": 2.1389741684445274
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 8.464103698730469,
-            "mean_kld": 0.06222418345240118,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.04821683398172381,
-            "median_context_p99": 0.541857123374939,
-            "p90_context_kld": 0.09363392489135787,
+            "max_kld": 8.426396369934082,
+            "mean_kld": 0.062162872968970516,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.0483626568274087,
+            "median_context_p99": 0.5371996760368347,
+            "p90_context_kld": 0.09342251370086758,
             "positions": 196512,
-            "top1_agreement": 0.9018736769255822,
+            "top1_agreement": 0.9020823155837812,
             "worst_context_id": 6,
-            "worst_context_kld": 0.351286367924311
+            "worst_context_kld": 0.3535688326255857
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.988017923903766
+        "relative_to_run": 0.9882289668958293
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 9.018062591552734,
-            "mean_kld": 0.05436721019998957,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.0455443970797962,
-            "median_context_p99": 0.4305042028427124,
-            "p90_context_kld": 0.10590312001808456,
+            "max_kld": 8.899713516235352,
+            "mean_kld": 0.054424187539078905,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.04558276349529022,
+            "median_context_p99": 0.42817598581314087,
+            "p90_context_kld": 0.10664729455358758,
             "positions": 73692,
-            "top1_agreement": 0.9045758019866471,
+            "top1_agreement": 0.9036937523747489,
             "worst_context_id": 947,
-            "worst_context_kld": 0.1165199785020918
+            "worst_context_kld": 0.1190153221496802
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.8632620818772753
+        "relative_to_run": 0.8652038758365567
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.404207706451416,
-            "mean_kld": 0.04707976949886068,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.03371621264780603,
-            "median_context_p99": 0.2907443940639496,
-            "p90_context_kld": 0.11144659651404887,
+            "max_kld": 7.351467609405518,
+            "mean_kld": 0.04704711827297485,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.033557425424993736,
+            "median_context_p99": 0.29649659991264343,
+            "p90_context_kld": 0.11091345794084626,
             "positions": 147384,
-            "top1_agreement": 0.9157710470607393,
+            "top1_agreement": 0.9159406719861043,
             "worst_context_id": 317,
-            "worst_context_kld": 0.15066450098557546
+            "worst_context_kld": 0.15121817439717355
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.747549482167406
+        "relative_to_run": 0.7479275468740905
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.460888862609863,
-            "mean_kld": 0.044350334816430914,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.03592069204934934,
-            "median_context_p99": 0.24048706889152527,
-            "p90_context_kld": 0.0533476602926932,
+            "max_kld": 8.17362117767334,
+            "mean_kld": 0.04437409335639619,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.03576088007500376,
+            "median_context_p99": 0.23678484559059143,
+            "p90_context_kld": 0.05390548673075974,
             "positions": 147384,
-            "top1_agreement": 0.899609184171959,
+            "top1_agreement": 0.8995549041958422,
             "worst_context_id": 443,
-            "worst_context_kld": 0.31145168263573736
+            "worst_context_kld": 0.3110997550139541
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.7042105383879657
+        "relative_to_run": 0.7054333614281264
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 8.12700366973877,
-            "mean_kld": 0.040566097360712344,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.032986131219142754,
-            "median_context_p99": 0.4861080050468445,
-            "p90_context_kld": 0.06758088454436428,
+            "max_kld": 8.626458168029785,
+            "mean_kld": 0.04060035080074585,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.0334252776297643,
+            "median_context_p99": 0.5161235928535461,
+            "p90_context_kld": 0.06731926693285788,
             "positions": 73692,
-            "top1_agreement": 0.8673261683764859,
+            "top1_agreement": 0.8677468381913912,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.17541399166167326
+            "worst_context_kld": 0.17577993098848518
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.6441230574904779
+        "relative_to_run": 0.6454406112706055
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.63759994506836,
-            "mean_kld": 0.03917626801646627,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.023835338446136917,
-            "median_context_p99": 0.3008233308792114,
-            "p90_context_kld": 0.07485065233115942,
+            "max_kld": 12.795402526855469,
+            "mean_kld": 0.03929107550581172,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.02376802223446685,
+            "median_context_p99": 0.30457040667533875,
+            "p90_context_kld": 0.07510747058223882,
             "positions": 196512,
-            "top1_agreement": 0.9471991532323726,
+            "top1_agreement": 0.9473721706562449,
             "worst_context_id": 667,
-            "worst_context_kld": 0.30212707418198775
+            "worst_context_kld": 0.3038625150634299
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.6220548482011916
+        "relative_to_run": 0.6246265190271398
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 21.455047607421875,
-            "mean_kld": 0.026672020564325254,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.020993579961020265,
-            "median_context_p99": 0.16144904494285583,
-            "p90_context_kld": 0.040855219811793145,
+            "max_kld": 9.440866470336914,
+            "mean_kld": 0.02663927997817303,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.02102722273338084,
+            "median_context_p99": 0.1589260697364807,
+            "p90_context_kld": 0.04065772069359729,
             "positions": 196512,
-            "top1_agreement": 0.9386195245074093,
+            "top1_agreement": 0.9386958557238235,
             "worst_context_id": 825,
-            "worst_context_kld": 0.10989124282970031
+            "worst_context_kld": 0.11036421424426199
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.4235079180177859
+        "relative_to_run": 0.4234956795645455
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.329400062561035,
-            "mean_kld": 0.026081798631463017,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.024741105384693707,
-            "median_context_p99": 0.19944342970848083,
-            "p90_context_kld": 0.037098149314752825,
+            "max_kld": 12.36015510559082,
+            "mean_kld": 0.026089785651106444,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.024915956636511808,
+            "median_context_p99": 0.2050503045320511,
+            "p90_context_kld": 0.036752015789522266,
             "positions": 196512,
-            "top1_agreement": 0.9310118466047875,
+            "top1_agreement": 0.9311543315420941,
             "worst_context_id": 246,
-            "worst_context_kld": 0.060966793708638514
+            "worst_context_kld": 0.06105584639032175
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.41413616227276917
+        "relative_to_run": 0.4147601404040052
       }
     ]
   },
   "overall": {
     "deployed": {
       "contexts": 768,
-      "max_kld": 31.918502807617188,
-      "mean_kld": 0.06297880022919694,
-      "mean_ref_top1_prob": 0.6415440866166368,
-      "median_context_kld": 0.03624156884367412,
-      "median_context_p99": 0.34093016386032104,
-      "p90_context_kld": 0.13468790532427413,
+      "max_kld": 31.67959976196289,
+      "mean_kld": 0.06290330991231023,
+      "mean_ref_top1_prob": 0.6415473983687051,
+      "median_context_kld": 0.03616175329293096,
+      "median_context_p99": 0.3476375341415405,
+      "p90_context_kld": 0.1364919523744591,
       "positions": 1572096,
-      "top1_agreement": 0.911584915933887,
+      "top1_agreement": 0.9116052709249308,
       "worst_context_id": 454,
-      "worst_context_kld": 1.0401913347215792
+      "worst_context_kld": 0.9887559796985864
     }
   },
   "primary": "deployed"

Qwen3.8-27B-INT4/strata.md

@@ -1,52 +1,52 @@
 # Fidelity by domain - Qwen3.8-27B / Qwen3.8-27B-INT4
 
-768 contexts, 1572096 scored positions, mean 0.06297880, reference top-1 64.2%, top-1 agreement 91.1585%.
+768 contexts, 1572096 scored positions, mean 0.06290331, reference top-1 64.2%, top-1 agreement 91.1605%.
 
-`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
+`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. The `deployed` cell is QxQ; `bxq` is the teacher-ID counterfactual on the same student weights. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
 
 ## Ranked by domain
 
 | Domain | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.14460520 | 2.30 | 0.12319855 | 0.29931944 | 1.040191 |
-| Chinese across several content types | 72 | 55.6% | 0.13453115 | 2.14 | 0.08000678 | 0.27791999 | 0.567423 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.06222418 | 0.99 | 0.04821683 | 0.09363392 | 0.351286 |
-| Other multilingual content | 36 | 64.5% | 0.05436721 | 0.86 | 0.04554440 | 0.10590312 | 0.116520 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.04707977 | 0.75 | 0.03371621 | 0.11144660 | 0.150665 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.04435033 | 0.70 | 0.03592069 | 0.05334766 | 0.311452 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.04056610 | 0.64 | 0.03298613 | 0.06758088 | 0.175414 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.03917627 | 0.62 | 0.02383534 | 0.07485065 | 0.302127 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.02667202 | 0.42 | 0.02099358 | 0.04085522 | 0.109891 |
-| Scientific and technical exposition | 96 | 60.3% | 0.02608180 | 0.41 | 0.02474111 | 0.03709815 | 0.060967 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.14393194 | 2.29 | 0.11954767 | 0.31973787 | 0.988756 |
+| Chinese across several content types | 72 | 55.6% | 0.13454856 | 2.14 | 0.07970839 | 0.27838235 | 0.568117 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.06216287 | 0.99 | 0.04836266 | 0.09342251 | 0.353569 |
+| Other multilingual content | 36 | 64.5% | 0.05442419 | 0.87 | 0.04558276 | 0.10664729 | 0.119015 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.04704712 | 0.75 | 0.03355743 | 0.11091346 | 0.151218 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.04437409 | 0.71 | 0.03576088 | 0.05390549 | 0.311100 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.04060035 | 0.65 | 0.03342528 | 0.06731927 | 0.175780 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.03929108 | 0.62 | 0.02376802 | 0.07510747 | 0.303863 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.02663928 | 0.42 | 0.02102722 | 0.04065772 | 0.110364 |
+| Scientific and technical exposition | 96 | 60.3% | 0.02608979 | 0.41 | 0.02491596 | 0.03675202 | 0.061056 |
 
 ## Ranked by source dataset
 
 | Source | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| wikisource_zh | 33 | 60.4% | 0.22439662 | 3.56 | 0.17356616 | 0.53231607 | 0.567423 |
-| wildchat | 96 | 73.5% | 0.14460520 | 2.30 | 0.12319855 | 0.29931944 | 1.040191 |
-| wikipedia_de | 7 | 73.6% | 0.09686725 | 1.54 | 0.10590312 | 0.11651998 | 0.116520 |
-| regulations | 23 | 72.3% | 0.07556402 | 1.20 | 0.09591328 | 0.13641989 | 0.150665 |
-| wikipedia_en | 96 | 59.5% | 0.06222418 | 0.99 | 0.04821683 | 0.09363392 | 0.351286 |
-| wikipedia_zh | 39 | 51.5% | 0.05849114 | 0.93 | 0.05639139 | 0.08040768 | 0.101248 |
-| wikipedia_cs | 6 | 68.9% | 0.05241203 | 0.83 | 0.04964427 | 0.06319860 | 0.063199 |
-| wikipedia_ja | 7 | 57.0% | 0.04765671 | 0.76 | 0.04595630 | 0.05639627 | 0.056396 |
-| public_domain_books | 72 | 50.6% | 0.04435033 | 0.70 | 0.03592069 | 0.05334766 | 0.311452 |
-| wikipedia_es | 7 | 59.2% | 0.04229309 | 0.67 | 0.04345389 | 0.05621201 | 0.056212 |
-| github_code | 52 | 85.8% | 0.04128255 | 0.66 | 0.02367156 | 0.08002234 | 0.302127 |
-| starcoder_structured | 36 | 70.7% | 0.04056610 | 0.64 | 0.03298613 | 0.06758088 | 0.175414 |
-| wikipedia_ru | 6 | 66.6% | 0.03949404 | 0.63 | 0.04360829 | 0.05209644 | 0.052096 |
-| public_domain_review | 26 | 51.8% | 0.03917945 | 0.62 | 0.03423984 | 0.07070048 | 0.107012 |
-| stackv2 | 44 | 72.0% | 0.03668703 | 0.58 | 0.02529400 | 0.05499314 | 0.290571 |
-| wikipedia_fr | 3 | 60.2% | 0.03268794 | 0.52 | 0.03278231 | 0.03937691 | 0.039377 |
-| open_news | 23 | 53.8% | 0.02752632 | 0.44 | 0.02471067 | 0.04196699 | 0.051401 |
-| libretexts | 96 | 65.8% | 0.02667202 | 0.42 | 0.02099358 | 0.04085522 | 0.109891 |
-| scientific_papers | 96 | 60.3% | 0.02608180 | 0.41 | 0.02474111 | 0.03709815 | 0.060967 |
+| wikisource_zh | 33 | 60.4% | 0.22442134 | 3.57 | 0.17347106 | 0.53327238 | 0.568117 |
+| wildchat | 96 | 73.5% | 0.14393194 | 2.29 | 0.11954767 | 0.31973787 | 0.988756 |
+| wikipedia_de | 7 | 73.6% | 0.09721811 | 1.55 | 0.10664729 | 0.11901532 | 0.119015 |
+| regulations | 23 | 72.3% | 0.07559705 | 1.20 | 0.09568556 | 0.13647847 | 0.151218 |
+| wikipedia_en | 96 | 59.5% | 0.06216287 | 0.99 | 0.04836266 | 0.09342251 | 0.353569 |
+| wikipedia_zh | 39 | 51.5% | 0.05850235 | 0.93 | 0.05616882 | 0.08070245 | 0.101207 |
+| wikipedia_cs | 6 | 68.9% | 0.05242830 | 0.83 | 0.05001897 | 0.06349262 | 0.063493 |
+| wikipedia_ja | 7 | 57.0% | 0.04766543 | 0.76 | 0.04594256 | 0.05671796 | 0.056718 |
+| public_domain_books | 72 | 50.6% | 0.04437409 | 0.71 | 0.03576088 | 0.05390549 | 0.311100 |
+| wikipedia_es | 7 | 59.2% | 0.04217758 | 0.67 | 0.04301762 | 0.05590323 | 0.055903 |
+| github_code | 52 | 85.8% | 0.04148121 | 0.66 | 0.02366987 | 0.07955755 | 0.303863 |
+| starcoder_structured | 36 | 70.7% | 0.04060035 | 0.65 | 0.03342528 | 0.06731927 | 0.175780 |
+| wikipedia_ru | 6 | 66.6% | 0.03953001 | 0.63 | 0.04384945 | 0.05227392 | 0.052274 |
+| public_domain_review | 26 | 51.8% | 0.03910379 | 0.62 | 0.03424818 | 0.07079836 | 0.107286 |
+| stackv2 | 44 | 72.0% | 0.03670274 | 0.58 | 0.02525954 | 0.05904637 | 0.288098 |
+| wikipedia_fr | 3 | 60.2% | 0.03269769 | 0.52 | 0.03284359 | 0.03932902 | 0.039329 |
+| open_news | 23 | 53.8% | 0.02747661 | 0.44 | 0.02412196 | 0.04159375 | 0.051061 |
+| libretexts | 96 | 65.8% | 0.02663928 | 0.42 | 0.02102722 | 0.04065772 | 0.110364 |
+| scientific_papers | 96 | 60.3% | 0.02608979 | 0.41 | 0.02491596 | 0.03675202 | 0.061056 |
 
 ## Reading
 
-- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.14460520, 2.30x the run mean of 0.06297880 over 96 context(s).
-- Strongest domain: **Scientific and technical exposition** at 0.02608180, 0.41x the run mean. The spread across domains is 5.5x.
+- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.14393194, 2.29x the run mean of 0.06290331 over 96 context(s).
+- Strongest domain: **Scientific and technical exposition** at 0.02608979, 0.41x the run mean. The spread across domains is 5.5x.
 - This is a strong finding: the reference was *more* certain on Natural dialogue, instruction following, and assistance (top-1 73.5% against 64.2% overall) and the candidate still diverged most there, so predictability does not explain it.
-- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 1.040191 against a median of 0.12319855 (context 454). Read the documents before treating the domain as weak.
+- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 0.988756 against a median of 0.11954767 (context 454). Read the documents before treating the domain as weak.
 

Qwen3.8-27B-NVFP4-BF16-LMHead/compliance.json

@@ -1,8 +1,9 @@
 {
   "attribution": null,
   "candidate": "/media/fmodels2/RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead",
-  "candidate_weights_sha256": null,
+  "candidate_weights_sha256": "2238b4ba183e9cc53757f6d7054ba998c02a87ed7c1842c27537fb802b1b09c0",
   "comparability_key": {
+    "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
     "context_length": 2048,
     "driver": "580.173.02",
     "gpu_names": [
@@ -12,19 +13,22 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 248044,
-    "laws_version": 9,
+    "kv_cache_dtype": "bfloat16",
+    "laws_version": 15,
     "model_runner_v2": false,
+    "numerics_digest": "251a9225b37415b91fc58d975b0670fc4b481e6f5a816b663573c42e161438f8",
     "reference_config_sha256": "191e0af232104ed8b65258cf3fb2b842e288008baca7633c11b82a1ac7203aab",
     "rows": 768,
     "score_from": 0,
     "stride": 2048,
+    "substituted_parameters": [],
     "suite_id": "qwen3.8-27b-fidelity-1024x2048-v1",
     "tensor_parallel_size": 1,
     "token_sha256": "9c935708bbffbf45d5baa2931fb4af7e7b22df1e041aa0b6f4ca44d3315e543b",
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T12:56:04.460087+00:00",
+  "evaluated_at": "2026-09-11T20:24:32.410217+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -52,13 +56,13 @@
       "title": "Real vocabulary"
     },
     {
-      "detail": "all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc",
+      "detail": "all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3",
       "law": 5,
       "status": "pass",
       "title": "Manifest binding"
     },
     {
-      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s)",
+      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s)",
       "law": 6,
       "status": "pass",
       "title": "Provenance"
@@ -70,13 +74,13 @@
       "title": "Storage integrity"
     },
     {
-      "detail": "trunk 0.05212951, deployed 0.05212951, delta 1.3275290033920584e-09",
+      "detail": "trunk 0.05156551, deployed 0.05156551, delta 1.1058352220039147e-09",
       "law": 8,
       "status": "pass",
       "title": "Head transparency"
     },
     {
-      "detail": "mean 0.05212951, median 0.01302997, max 27.48295784, 4 depth buckets",
+      "detail": "mean 0.05156551, median 0.01290934, max 36.40480804, 4 depth buckets",
       "law": 9,
       "status": "pass",
       "title": "Tail and depth disclosure"
@@ -103,38 +107,37 @@
       "detail": "reference declares no experts",
       "law": 14,
       "status": "not_applicable",
-      "title": "Component attribution"
+      "title": "Routed-model intervention"
     },
     {
-      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.14506069, strongest scientific_technical at 0.02322830, spread 6.2x",
+      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.14260814, strongest scientific_technical at 0.02302717, spread 6.2x",
       "law": 15,
       "status": "pass",
       "title": "Domain disclosure"
     },
     {
-      "approval": {
-        "approver": "Andy Kitzke",
-        "justification": "These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger.",
-        "timestamp": "2026-09-02T05:15:00Z"
-      },
-      "detail": "the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else",
+      "detail": "scored weights 2238b4ba183e9cc5 as inspected",
       "law": 16,
-      "status": "override",
+      "status": "pass",
       "title": "Candidate weight binding"
     },
     {
-      "detail": "1 override(s), each fully attributed",
+      "detail": "every scored layer used the checkpoint's own quantization parameters",
+      "law": 17,
+      "status": "pass",
+      "title": "Substitution disclosure"
+    },
+    {
+      "detail": "no overrides claimed",
       "law": 13,
       "status": "pass",
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 9,
-  "mean_kld": 0.05212951125207867,
+  "laws_version": 15,
+  "mean_kld": 0.05156550663959364,
   "nondeterminism_floor": 0.0,
-  "overridden_laws": [
-    16
-  ],
+  "overridden_laws": [],
   "partition": "analysis",
   "program": "Local Inference Lab \u2014 Distribution Fidelity",
   "ranking_floor": null,
@@ -147,16 +150,16 @@
     "overall": {
       "deployed": {
         "contexts": 768,
-        "max_kld": 27.48295783996582,
-        "mean_kld": 0.05212951125207867,
-        "mean_ref_top1_prob": 0.6415440866166368,
-        "median_context_kld": 0.03048191045903891,
-        "median_context_p99": 0.2586134076118469,
-        "p90_context_kld": 0.10388629113718155,
+        "max_kld": 36.404808044433594,
+        "mean_kld": 0.05156550663959363,
+        "mean_ref_top1_prob": 0.6415473983687051,
+        "median_context_kld": 0.030358945919737878,
+        "median_context_p99": 0.2575896382331848,
+        "p90_context_kld": 0.10094674048285278,
         "positions": 1572096,
-        "top1_agreement": 0.9151871132551702,
+        "top1_agreement": 0.9156470088340661,
         "worst_context_id": 454,
-        "worst_context_kld": 0.9351965383137365
+        "worst_context_kld": 0.8905369579241399
       }
     },
     "primary": "deployed",
@@ -165,201 +168,201 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 27.48295783996582,
-            "mean_kld": 0.1450606879671042,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.10672595977404836,
-            "median_context_p99": 2.366913318634033,
-            "p90_context_kld": 0.31745866158131525,
+            "max_kld": 36.404808044433594,
+            "mean_kld": 0.14260813963956548,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.10354972375019698,
+            "median_context_p99": 2.1813976764678955,
+            "p90_context_kld": 0.3451019763379924,
             "positions": 196512,
-            "top1_agreement": 0.9130485670086305,
+            "top1_agreement": 0.9151603973294252,
             "worst_context_id": 454,
-            "worst_context_kld": 0.9351965383137365
+            "worst_context_kld": 0.8905369579241399
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 2.7826980242658594
+        "relative_to_run": 2.765572355107366
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.188745975494385,
-            "mean_kld": 0.0800973933073713,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.0488142405969883,
-            "median_context_p99": 0.3909897208213806,
-            "p90_context_kld": 0.14562944384550067,
+            "max_kld": 7.264811038970947,
+            "mean_kld": 0.07907412001697732,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.04855310270042977,
+            "median_context_p99": 0.41729986667633057,
+            "p90_context_kld": 0.14487312296101867,
             "positions": 147384,
-            "top1_agreement": 0.8762687944417304,
+            "top1_agreement": 0.8768658741790154,
             "worst_context_id": 893,
-            "worst_context_kld": 0.33491145623420787
+            "worst_context_kld": 0.31717224464172616
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.5365076591654676
+        "relative_to_run": 1.5334692737462934
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 8.426796913146973,
-            "mean_kld": 0.05208049148082254,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.04317494681946243,
-            "median_context_p99": 0.43730735778808594,
-            "p90_context_kld": 0.0747909649283652,
+            "max_kld": 7.654315948486328,
+            "mean_kld": 0.05124551225377584,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.04196325648602333,
+            "median_context_p99": 0.4040193259716034,
+            "p90_context_kld": 0.07206766500842036,
             "positions": 196512,
-            "top1_agreement": 0.9043417195896434,
+            "top1_agreement": 0.9047081094284318,
             "worst_context_id": 6,
-            "worst_context_kld": 0.2575662736752676
+            "worst_context_kld": 0.253386224192646
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.9990596541176248
+        "relative_to_run": 0.9937944101266313
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 7.891758918762207,
-            "mean_kld": 0.03909128508337415,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.03482694495389481,
-            "median_context_p99": 0.2722674012184143,
-            "p90_context_kld": 0.07035002679922857,
+            "max_kld": 10.652168273925781,
+            "mean_kld": 0.03972506920105216,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.034178675689330404,
+            "median_context_p99": 0.2886667847633362,
+            "p90_context_kld": 0.06837731806777113,
             "positions": 73692,
-            "top1_agreement": 0.9148482874667535,
+            "top1_agreement": 0.9138033979265049,
             "worst_context_id": 953,
-            "worst_context_kld": 0.08191294001897026
+            "worst_context_kld": 0.08763029838216986
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.7498878110393828
+        "relative_to_run": 0.7703806631572974
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.322475910186768,
-            "mean_kld": 0.03826159344727907,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.031959890771641075,
-            "median_context_p99": 0.2690824270248413,
-            "p90_context_kld": 0.07946953402075285,
+            "max_kld": 6.221218109130859,
+            "mean_kld": 0.038112379624305213,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.03164991631323535,
+            "median_context_p99": 0.2747260630130768,
+            "p90_context_kld": 0.07574533150766247,
             "positions": 147384,
-            "top1_agreement": 0.9153232372577756,
+            "top1_agreement": 0.9153707322368778,
             "worst_context_id": 275,
-            "worst_context_kld": 0.10122025808876788
+            "worst_context_kld": 0.10094674048285278
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.7339718429789305
+        "relative_to_run": 0.7391060828837337
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 5.807681560516357,
-            "mean_kld": 0.03723338342274077,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.03198886603632158,
-            "median_context_p99": 0.19292420148849487,
-            "p90_context_kld": 0.04507155972030506,
+            "max_kld": 6.983285427093506,
+            "mean_kld": 0.03711085502771101,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.031371119738575946,
+            "median_context_p99": 0.20159181952476501,
+            "p90_context_kld": 0.04505553168422091,
             "positions": 147384,
-            "top1_agreement": 0.9027506377897194,
+            "top1_agreement": 0.9036666123866905,
             "worst_context_id": 443,
-            "worst_context_kld": 0.21792143969470823
+            "worst_context_kld": 0.21911474635649905
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.7142476982508844
+        "relative_to_run": 0.7196837080859033
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.813560485839844,
-            "mean_kld": 0.030346047259348898,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.02084674157560274,
-            "median_context_p99": 0.23820368945598602,
-            "p90_context_kld": 0.04975873418086208,
+            "max_kld": 14.599754333496094,
+            "mean_kld": 0.03018829355699219,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.02100278716601154,
+            "median_context_p99": 0.21924170851707458,
+            "p90_context_kld": 0.05378769088686552,
             "positions": 196512,
-            "top1_agreement": 0.9495908646800195,
+            "top1_agreement": 0.9502422243934213,
             "worst_context_id": 667,
-            "worst_context_kld": 0.2161889008647869
+            "worst_context_kld": 0.20473772003888704
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.5821279833721604
+        "relative_to_run": 0.5854357985462448
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 7.356241226196289,
-            "mean_kld": 0.027620313995326606,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.02249949949724309,
-            "median_context_p99": 0.40589118003845215,
-            "p90_context_kld": 0.05043523494828777,
+            "max_kld": 8.024088859558105,
+            "mean_kld": 0.02775163699079394,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.02158035745589829,
+            "median_context_p99": 0.34725338220596313,
+            "p90_context_kld": 0.04604335776204541,
             "positions": 73692,
-            "top1_agreement": 0.8654535092004559,
+            "top1_agreement": 0.8652906692721055,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.11659044195651949
+            "worst_context_kld": 0.11848209727913012
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.5298402638337655
+        "relative_to_run": 0.5381821841634996
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 17.052858352661133,
-            "mean_kld": 0.024609437161871247,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.020860553521432415,
-            "median_context_p99": 0.14598329365253448,
-            "p90_context_kld": 0.03323913724905251,
+            "max_kld": 13.109975814819336,
+            "mean_kld": 0.024428160646162753,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.020779478378882214,
+            "median_context_p99": 0.1441325843334198,
+            "p90_context_kld": 0.03284644349083438,
             "positions": 196512,
-            "top1_agreement": 0.9371743201432992,
+            "top1_agreement": 0.9363652092493079,
             "worst_context_id": 825,
-            "worst_context_kld": 0.08214911499231742
+            "worst_context_kld": 0.08643491137050359
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.4720826374693844
+        "relative_to_run": 0.4737306435656357
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.320075988769531,
-            "mean_kld": 0.023228298859926257,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.02226252774408313,
-            "median_context_p99": 0.16904449462890625,
-            "p90_context_kld": 0.031335149218943904,
+            "max_kld": 13.07381534576416,
+            "mean_kld": 0.023027166196565346,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.021995164691474722,
+            "median_context_p99": 0.16406352818012238,
+            "p90_context_kld": 0.030565308802914134,
             "positions": 196512,
-            "top1_agreement": 0.9289712587526462,
+            "top1_agreement": 0.929612440970526,
             "worst_context_id": 246,
-            "worst_context_kld": 0.04042488780931219
+            "worst_context_kld": 0.044429711784661804
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.44558827240107735
+        "relative_to_run": 0.44656142637187546
       }
     ]
   },

Qwen3.8-27B-NVFP4-BF16-LMHead/inspect.json

@@ -1,31 +1,461 @@
 {
+  "model": "/media/fmodels2/RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead",
+  "inspect_version": 5,
+  "weights_sha256": "2238b4ba183e9cc53757f6d7054ba998c02a87ed7c1842c27537fb802b1b09c0",
+  "weights_bytes": 23749332688,
+  "quant_method": "modelopt",
+  "declared": {},
+  "detected_scheme": "nvfp4",
+  "detected_block": 16,
+  "quant_algorithm": "round_to_nearest",
+  "scheme_mix": {
+    "format": null,
+    "groups": [
+      {
+        "group": "group_0",
+        "format": null,
+        "weight_bits": 8,
+        "activation_bits": 8,
+        "targets": [
+          "model.language_model.layers.0.linear_attn.in_proj_qkv",
+          "model.language_model.layers.0.linear_attn.in_proj_z",
+          "model.language_model.layers.0.linear_attn.out_proj",
+          "model.language_model.layers.1.linear_attn.in_proj_qkv",
+          "model.language_model.layers.1.linear_attn.in_proj_z",
+          "model.language_model.layers.1.linear_attn.out_proj",
+          "model.language_model.layers.10.linear_attn.in_proj_qkv",
+          "model.language_model.layers.10.linear_attn.in_proj_z",
+          "model.language_model.layers.10.linear_attn.out_proj",
+          "model.language_model.layers.11.self_attn.k_proj",
+          "model.language_model.layers.11.self_attn.o_proj",
+          "model.language_model.layers.11.self_attn.q_proj",
+          "model.language_model.layers.11.self_attn.v_proj",
+          "model.language_model.layers.12.linear_attn.in_proj_qkv",
+          "model.language_model.layers.12.linear_attn.in_proj_z",
+          "model.language_model.layers.12.linear_attn.out_proj",
+          "model.language_model.layers.13.linear_attn.in_proj_qkv",
+          "model.language_model.layers.13.linear_attn.in_proj_z",
+          "model.language_model.layers.13.linear_attn.out_proj",
+          "model.language_model.layers.14.linear_attn.in_proj_qkv",
+          "model.language_model.layers.14.linear_attn.in_proj_z",
+          "model.language_model.layers.14.linear_attn.out_proj",
+          "model.language_model.layers.15.self_attn.k_proj",
+          "model.language_model.layers.15.self_attn.o_proj",
+          "model.language_model.layers.15.self_attn.q_proj",
+          "model.language_model.layers.15.self_attn.v_proj",
+          "model.language_model.layers.16.linear_attn.in_proj_qkv",
+          "model.language_model.layers.16.linear_attn.in_proj_z",
+          "model.language_model.layers.16.linear_attn.out_proj",
+          "model.language_model.layers.17.linear_attn.in_proj_qkv",
+          "model.language_model.layers.17.linear_attn.in_proj_z",
+          "model.language_model.layers.17.linear_attn.out_proj",
+          "model.language_model.layers.18.linear_attn.in_proj_qkv",
+          "model.language_model.layers.18.linear_attn.in_proj_z",
+          "model.language_model.layers.18.linear_attn.out_proj",
+          "model.language_model.layers.19.self_attn.k_proj",
+          "model.language_model.layers.19.self_attn.o_proj",
+          "model.language_model.layers.19.self_attn.q_proj",
+          "model.language_model.layers.19.self_attn.v_proj",
+          "model.language_model.layers.2.linear_attn.in_proj_qkv",
+          "model.language_model.layers.2.linear_attn.in_proj_z",
+          "model.language_model.layers.2.linear_attn.out_proj",
+          "model.language_model.layers.20.linear_attn.in_proj_qkv",
+          "model.language_model.layers.20.linear_attn.in_proj_z",
+          "model.language_model.layers.20.linear_attn.out_proj",
+          "model.language_model.layers.21.linear_attn.in_proj_qkv",
+          "model.language_model.layers.21.linear_attn.in_proj_z",
+          "model.language_model.layers.21.linear_attn.out_proj",
+          "model.language_model.layers.22.linear_attn.in_proj_qkv",
+          "model.language_model.layers.22.linear_attn.in_proj_z",
+          "model.language_model.layers.22.linear_attn.out_proj",
+          "model.language_model.layers.23.self_attn.k_proj",
+          "model.language_model.layers.23.self_attn.o_proj",
+          "model.language_model.layers.23.self_attn.q_proj",
+          "model.language_model.layers.23.self_attn.v_proj",
+          "model.language_model.layers.24.linear_attn.in_proj_qkv",
+          "model.language_model.layers.24.linear_attn.in_proj_z",
+          "model.language_model.layers.24.linear_attn.out_proj",
+          "model.language_model.layers.25.linear_attn.in_proj_qkv",
+          "model.language_model.layers.25.linear_attn.in_proj_z",
+          "model.language_model.layers.25.linear_attn.out_proj",
+          "model.language_model.layers.26.linear_attn.in_proj_qkv",
+          "model.language_model.layers.26.linear_attn.in_proj_z",
+          "model.language_model.layers.26.linear_attn.out_proj",
+          "model.language_model.layers.27.self_attn.k_proj",
+          "model.language_model.layers.27.self_attn.o_proj",
+          "model.language_model.layers.27.self_attn.q_proj",
+          "model.language_model.layers.27.self_attn.v_proj",
+          "model.language_model.layers.28.linear_attn.in_proj_qkv",
+          "model.language_model.layers.28.linear_attn.in_proj_z",
+          "model.language_model.layers.28.linear_attn.out_proj",
+          "model.language_model.layers.29.linear_attn.in_proj_qkv",
+          "model.language_model.layers.29.linear_attn.in_proj_z",
+          "model.language_model.layers.29.linear_attn.out_proj",
+          "model.language_model.layers.3.self_attn.k_proj",
+          "model.language_model.layers.3.self_attn.o_proj",
+          "model.language_model.layers.3.self_attn.q_proj",
+          "model.language_model.layers.3.self_attn.v_proj",
+          "model.language_model.layers.30.linear_attn.in_proj_qkv",
+          "model.language_model.layers.30.linear_attn.in_proj_z",
+          "model.language_model.layers.30.linear_attn.out_proj",
+          "model.language_model.layers.31.self_attn.k_proj",
+          "model.language_model.layers.31.self_attn.o_proj",
+          "model.language_model.layers.31.self_attn.q_proj",
+          "model.language_model.layers.31.self_attn.v_proj",
+          "model.language_model.layers.32.linear_attn.in_proj_qkv",
+          "model.language_model.layers.32.linear_attn.in_proj_z",
+          "model.language_model.layers.32.linear_attn.out_proj",
+          "model.language_model.layers.33.linear_attn.in_proj_qkv",
+          "model.language_model.layers.33.linear_attn.in_proj_z",
+          "model.language_model.layers.33.linear_attn.out_proj",
+          "model.language_model.layers.34.linear_attn.in_proj_qkv",
+          "model.language_model.layers.34.linear_attn.in_proj_z",
+          "model.language_model.layers.34.linear_attn.out_proj",
+          "model.language_model.layers.35.self_attn.k_proj",
+          "model.language_model.layers.35.self_attn.o_proj",
+          "model.language_model.layers.35.self_attn.q_proj",
+          "model.language_model.layers.35.self_attn.v_proj",
+          "model.language_model.layers.36.linear_attn.in_proj_qkv",
+          "model.language_model.layers.36.linear_attn.in_proj_z",
+          "model.language_model.layers.36.linear_attn.out_proj",
+          "model.language_model.layers.37.linear_attn.in_proj_qkv",
+          "model.language_model.layers.37.linear_attn.in_proj_z",
+          "model.language_model.layers.37.linear_attn.out_proj",
+          "model.language_model.layers.38.linear_attn.in_proj_qkv",
+          "model.language_model.layers.38.linear_attn.in_proj_z",
+          "model.language_model.layers.38.linear_attn.out_proj",
+          "model.language_model.layers.39.self_attn.k_proj",
+          "model.language_model.layers.39.self_attn.o_proj",
+          "model.language_model.layers.39.self_attn.q_proj",
+          "model.language_model.layers.39.self_attn.v_proj",
+          "model.language_model.layers.4.linear_attn.in_proj_qkv",
+          "model.language_model.layers.4.linear_attn.in_proj_z",
+          "model.language_model.layers.4.linear_attn.out_proj",
+          "model.language_model.layers.40.linear_attn.in_proj_qkv",
+          "model.language_model.layers.40.linear_attn.in_proj_z",
+          "model.language_model.layers.40.linear_attn.out_proj",
+          "model.language_model.layers.41.linear_attn.in_proj_qkv",
+          "model.language_model.layers.41.linear_attn.in_proj_z",
+          "model.language_model.layers.41.linear_attn.out_proj",
+          "model.language_model.layers.42.linear_attn.in_proj_qkv",
+          "model.language_model.layers.42.linear_attn.in_proj_z",
+          "model.language_model.layers.42.linear_attn.out_proj",
+          "model.language_model.layers.43.self_attn.k_proj",
+          "model.language_model.layers.43.self_attn.o_proj",
+          "model.language_model.layers.43.self_attn.q_proj",
+          "model.language_model.layers.43.self_attn.v_proj",
+          "model.language_model.layers.44.linear_attn.in_proj_qkv",
+          "model.language_model.layers.44.linear_attn.in_proj_z",
+          "model.language_model.layers.44.linear_attn.out_proj",
+          "model.language_model.layers.45.linear_attn.in_proj_qkv",
+          "model.language_model.layers.45.linear_attn.in_proj_z",
+          "model.language_model.layers.45.linear_attn.out_proj",
+          "model.language_model.layers.46.linear_attn.in_proj_qkv",
+          "model.language_model.layers.46.linear_attn.in_proj_z",
+          "model.language_model.layers.46.linear_attn.out_proj",
+          "model.language_model.layers.47.self_attn.k_proj",
+          "model.language_model.layers.47.self_attn.o_proj",
+          "model.language_model.layers.47.self_attn.q_proj",
+          "model.language_model.layers.47.self_attn.v_proj",
+          "model.language_model.layers.48.linear_attn.in_proj_qkv",
+          "model.language_model.layers.48.linear_attn.in_proj_z",
+          "model.language_model.layers.48.linear_attn.out_proj",
+          "model.language_model.layers.49.linear_attn.in_proj_qkv",
+          "model.language_model.layers.49.linear_attn.in_proj_z",
+          "model.language_model.layers.49.linear_attn.out_proj",
+          "model.language_model.layers.5.linear_attn.in_proj_qkv",
+          "model.language_model.layers.5.linear_attn.in_proj_z",
+          "model.language_model.layers.5.linear_attn.out_proj",
+          "model.language_model.layers.50.linear_attn.in_proj_qkv",
+          "model.language_model.layers.50.linear_attn.in_proj_z",
+          "model.language_model.layers.50.linear_attn.out_proj",
+          "model.language_model.layers.51.self_attn.k_proj",
+          "model.language_model.layers.51.self_attn.o_proj",
+          "model.language_model.layers.51.self_attn.q_proj",
+          "model.language_model.layers.51.self_attn.v_proj",
+          "model.language_model.layers.52.linear_attn.in_proj_qkv",
+          "model.language_model.layers.52.linear_attn.in_proj_z",
+          "model.language_model.layers.52.linear_attn.out_proj",
+          "model.language_model.layers.53.linear_attn.in_proj_qkv",
+          "model.language_model.layers.53.linear_attn.in_proj_z",
+          "model.language_model.layers.53.linear_attn.out_proj",
+          "model.language_model.layers.54.linear_attn.in_proj_qkv",
+          "model.language_model.layers.54.linear_attn.in_proj_z",
+          "model.language_model.layers.54.linear_attn.out_proj",
+          "model.language_model.layers.55.self_attn.k_proj",
+          "model.language_model.layers.55.self_attn.o_proj",
+          "model.language_model.layers.55.self_attn.q_proj",
+          "model.language_model.layers.55.self_attn.v_proj",
+          "model.language_model.layers.56.linear_attn.in_proj_qkv",
+          "model.language_model.layers.56.linear_attn.in_proj_z",
+          "model.language_model.layers.56.linear_attn.out_proj",
+          "model.language_model.layers.57.linear_attn.in_proj_qkv",
+          "model.language_model.layers.57.linear_attn.in_proj_z",
+          "model.language_model.layers.57.linear_attn.out_proj",
+          "model.language_model.layers.58.linear_attn.in_proj_qkv",
+          "model.language_model.layers.58.linear_attn.in_proj_z",
+          "model.language_model.layers.58.linear_attn.out_proj",
+          "model.language_model.layers.59.self_attn.k_proj",
+          "model.language_model.layers.59.self_attn.o_proj",
+          "model.language_model.layers.59.self_attn.q_proj",
+          "model.language_model.layers.59.self_attn.v_proj",
+          "model.language_model.layers.6.linear_attn.in_proj_qkv",
+          "model.language_model.layers.6.linear_attn.in_proj_z",
+          "model.language_model.layers.6.linear_attn.out_proj",
+          "model.language_model.layers.60.linear_attn.in_proj_qkv",
+          "model.language_model.layers.60.linear_attn.in_proj_z",
+          "model.language_model.layers.60.linear_attn.out_proj",
+          "model.language_model.layers.61.linear_attn.in_proj_qkv",
+          "model.language_model.layers.61.linear_attn.in_proj_z",
+          "model.language_model.layers.61.linear_attn.out_proj",
+          "model.language_model.layers.62.linear_attn.in_proj_qkv",
+          "model.language_model.layers.62.linear_attn.in_proj_z",
+          "model.language_model.layers.62.linear_attn.out_proj",
+          "model.language_model.layers.63.self_attn.k_proj",
+          "model.language_model.layers.63.self_attn.o_proj",
+          "model.language_model.layers.63.self_attn.q_proj",
+          "model.language_model.layers.63.self_attn.v_proj",
+          "model.language_model.layers.7.self_attn.k_proj",
+          "model.language_model.layers.7.self_attn.o_proj",
+          "model.language_model.layers.7.self_attn.q_proj",
+          "model.language_model.layers.7.self_attn.v_proj",
+          "model.language_model.layers.8.linear_attn.in_proj_qkv",
+          "model.language_model.layers.8.linear_attn.in_proj_z",
+          "model.language_model.layers.8.linear_attn.out_proj",
+          "model.language_model.layers.9.linear_attn.in_proj_qkv",
+          "model.language_model.layers.9.linear_attn.in_proj_z",
+          "model.language_model.layers.9.linear_attn.out_proj"
+        ]
+      },
+      {
+        "group": "group_1",
+        "format": null,
+        "weight_bits": 4,
+        "activation_bits": 4,
+        "targets": [
+          "model.language_model.layers.0.mlp.down_proj",
+          "model.language_model.layers.0.mlp.gate_proj",
+          "model.language_model.layers.0.mlp.up_proj",
+          "model.language_model.layers.1.mlp.down_proj",
+          "model.language_model.layers.1.mlp.gate_proj",
+          "model.language_model.layers.1.mlp.up_proj",
+          "model.language_model.layers.10.mlp.down_proj",
+          "model.language_model.layers.10.mlp.gate_proj",
+          "model.language_model.layers.10.mlp.up_proj",
+          "model.language_model.layers.11.mlp.down_proj",
+          "model.language_model.layers.11.mlp.gate_proj",
+          "model.language_model.layers.11.mlp.up_proj",
+          "model.language_model.layers.12.mlp.down_proj",
+          "model.language_model.layers.12.mlp.gate_proj",
+          "model.language_model.layers.12.mlp.up_proj",
+          "model.language_model.layers.13.mlp.down_proj",
+          "model.language_model.layers.13.mlp.gate_proj",
+          "model.language_model.layers.13.mlp.up_proj",
+          "model.language_model.layers.14.mlp.down_proj",
+          "model.language_model.layers.14.mlp.gate_proj",
+          "model.language_model.layers.14.mlp.up_proj",
+          "model.language_model.layers.15.mlp.down_proj",
+          "model.language_model.layers.15.mlp.gate_proj",
+          "model.language_model.layers.15.mlp.up_proj",
+          "model.language_model.layers.16.mlp.down_proj",
+          "model.language_model.layers.16.mlp.gate_proj",
+          "model.language_model.layers.16.mlp.up_proj",
+          "model.language_model.layers.17.mlp.down_proj",
+          "model.language_model.layers.17.mlp.gate_proj",
+          "model.language_model.layers.17.mlp.up_proj",
+          "model.language_model.layers.18.mlp.down_proj",
+          "model.language_model.layers.18.mlp.gate_proj",
+          "model.language_model.layers.18.mlp.up_proj",
+          "model.language_model.layers.19.mlp.down_proj",
+          "model.language_model.layers.19.mlp.gate_proj",
+          "model.language_model.layers.19.mlp.up_proj",
+          "model.language_model.layers.2.mlp.down_proj",
+          "model.language_model.layers.2.mlp.gate_proj",
+          "model.language_model.layers.2.mlp.up_proj",
+          "model.language_model.layers.20.mlp.down_proj",
+          "model.language_model.layers.20.mlp.gate_proj",
+          "model.language_model.layers.20.mlp.up_proj",
+          "model.language_model.layers.21.mlp.down_proj",
+          "model.language_model.layers.21.mlp.gate_proj",
+          "model.language_model.layers.21.mlp.up_proj",
+          "model.language_model.layers.22.mlp.down_proj",
+          "model.language_model.layers.22.mlp.gate_proj",
+          "model.language_model.layers.22.mlp.up_proj",
+          "model.language_model.layers.23.mlp.down_proj",
+          "model.language_model.layers.23.mlp.gate_proj",
+          "model.language_model.layers.23.mlp.up_proj",
+          "model.language_model.layers.24.mlp.down_proj",
+          "model.language_model.layers.24.mlp.gate_proj",
+          "model.language_model.layers.24.mlp.up_proj",
+          "model.language_model.layers.25.mlp.down_proj",
+          "model.language_model.layers.25.mlp.gate_proj",
+          "model.language_model.layers.25.mlp.up_proj",
+          "model.language_model.layers.26.mlp.down_proj",
+          "model.language_model.layers.26.mlp.gate_proj",
+          "model.language_model.layers.26.mlp.up_proj",
+          "model.language_model.layers.27.mlp.down_proj",
+          "model.language_model.layers.27.mlp.gate_proj",
+          "model.language_model.layers.27.mlp.up_proj",
+          "model.language_model.layers.28.mlp.down_proj",
+          "model.language_model.layers.28.mlp.gate_proj",
+          "model.language_model.layers.28.mlp.up_proj",
+          "model.language_model.layers.29.mlp.down_proj",
+          "model.language_model.layers.29.mlp.gate_proj",
+          "model.language_model.layers.29.mlp.up_proj",
+          "model.language_model.layers.3.mlp.down_proj",
+          "model.language_model.layers.3.mlp.gate_proj",
+          "model.language_model.layers.3.mlp.up_proj",
+          "model.language_model.layers.30.mlp.down_proj",
+          "model.language_model.layers.30.mlp.gate_proj",
+          "model.language_model.layers.30.mlp.up_proj",
+          "model.language_model.layers.31.mlp.down_proj",
+          "model.language_model.layers.31.mlp.gate_proj",
+          "model.language_model.layers.31.mlp.up_proj",
+          "model.language_model.layers.32.mlp.down_proj",
+          "model.language_model.layers.32.mlp.gate_proj",
+          "model.language_model.layers.32.mlp.up_proj",
+          "model.language_model.layers.33.mlp.down_proj",
+          "model.language_model.layers.33.mlp.gate_proj",
+          "model.language_model.layers.33.mlp.up_proj",
+          "model.language_model.layers.34.mlp.down_proj",
+          "model.language_model.layers.34.mlp.gate_proj",
+          "model.language_model.layers.34.mlp.up_proj",
+          "model.language_model.layers.35.mlp.down_proj",
+          "model.language_model.layers.35.mlp.gate_proj",
+          "model.language_model.layers.35.mlp.up_proj",
+          "model.language_model.layers.36.mlp.down_proj",
+          "model.language_model.layers.36.mlp.gate_proj",
+          "model.language_model.layers.36.mlp.up_proj",
+          "model.language_model.layers.37.mlp.down_proj",
+          "model.language_model.layers.37.mlp.gate_proj",
+          "model.language_model.layers.37.mlp.up_proj",
+          "model.language_model.layers.38.mlp.down_proj",
+          "model.language_model.layers.38.mlp.gate_proj",
+          "model.language_model.layers.38.mlp.up_proj",
+          "model.language_model.layers.39.mlp.down_proj",
+          "model.language_model.layers.39.mlp.gate_proj",
+          "model.language_model.layers.39.mlp.up_proj",
+          "model.language_model.layers.4.mlp.down_proj",
+          "model.language_model.layers.4.mlp.gate_proj",
+          "model.language_model.layers.4.mlp.up_proj",
+          "model.language_model.layers.40.mlp.down_proj",
+          "model.language_model.layers.40.mlp.gate_proj",
+          "model.language_model.layers.40.mlp.up_proj",
+          "model.language_model.layers.41.mlp.down_proj",
+          "model.language_model.layers.41.mlp.gate_proj",
+          "model.language_model.layers.41.mlp.up_proj",
+          "model.language_model.layers.42.mlp.down_proj",
+          "model.language_model.layers.42.mlp.gate_proj",
+          "model.language_model.layers.42.mlp.up_proj",
+          "model.language_model.layers.43.mlp.down_proj",
+          "model.language_model.layers.43.mlp.gate_proj",
+          "model.language_model.layers.43.mlp.up_proj",
+          "model.language_model.layers.44.mlp.down_proj",
+          "model.language_model.layers.44.mlp.gate_proj",
+          "model.language_model.layers.44.mlp.up_proj",
+          "model.language_model.layers.45.mlp.down_proj",
+          "model.language_model.layers.45.mlp.gate_proj",
+          "model.language_model.layers.45.mlp.up_proj",
+          "model.language_model.layers.46.mlp.down_proj",
+          "model.language_model.layers.46.mlp.gate_proj",
+          "model.language_model.layers.46.mlp.up_proj",
+          "model.language_model.layers.47.mlp.down_proj",
+          "model.language_model.layers.47.mlp.gate_proj",
+          "model.language_model.layers.47.mlp.up_proj",
+          "model.language_model.layers.48.mlp.down_proj",
+          "model.language_model.layers.48.mlp.gate_proj",
+          "model.language_model.layers.48.mlp.up_proj",
+          "model.language_model.layers.49.mlp.down_proj",
+          "model.language_model.layers.49.mlp.gate_proj",
+          "model.language_model.layers.49.mlp.up_proj",
+          "model.language_model.layers.5.mlp.down_proj",
+          "model.language_model.layers.5.mlp.gate_proj",
+          "model.language_model.layers.5.mlp.up_proj",
+          "model.language_model.layers.50.mlp.down_proj",
+          "model.language_model.layers.50.mlp.gate_proj",
+          "model.language_model.layers.50.mlp.up_proj",
+          "model.language_model.layers.51.mlp.down_proj",
+          "model.language_model.layers.51.mlp.gate_proj",
+          "model.language_model.layers.51.mlp.up_proj",
+          "model.language_model.layers.52.mlp.down_proj",
+          "model.language_model.layers.52.mlp.gate_proj",
+          "model.language_model.layers.52.mlp.up_proj",
+          "model.language_model.layers.53.mlp.down_proj",
+          "model.language_model.layers.53.mlp.gate_proj",
+          "model.language_model.layers.53.mlp.up_proj",
+          "model.language_model.layers.54.mlp.down_proj",
+          "model.language_model.layers.54.mlp.gate_proj",
+          "model.language_model.layers.54.mlp.up_proj",
+          "model.language_model.layers.55.mlp.down_proj",
+          "model.language_model.layers.55.mlp.gate_proj",
+          "model.language_model.layers.55.mlp.up_proj",
+          "model.language_model.layers.56.mlp.down_proj",
+          "model.language_model.layers.56.mlp.gate_proj",
+          "model.language_model.layers.56.mlp.up_proj",
+          "model.language_model.layers.57.mlp.down_proj",
+          "model.language_model.layers.57.mlp.gate_proj",
+          "model.language_model.layers.57.mlp.up_proj",
+          "model.language_model.layers.58.mlp.down_proj",
+          "model.language_model.layers.58.mlp.gate_proj",
+          "model.language_model.layers.58.mlp.up_proj",
+          "model.language_model.layers.59.mlp.down_proj",
+          "model.language_model.layers.59.mlp.gate_proj",
+          "model.language_model.layers.59.mlp.up_proj",
+          "model.language_model.layers.6.mlp.down_proj",
+          "model.language_model.layers.6.mlp.gate_proj",
+          "model.language_model.layers.6.mlp.up_proj",
+          "model.language_model.layers.60.mlp.down_proj",
+          "model.language_model.layers.60.mlp.gate_proj",
+          "model.language_model.layers.60.mlp.up_proj",
+          "model.language_model.layers.61.mlp.down_proj",
+          "model.language_model.layers.61.mlp.gate_proj",
+          "model.language_model.layers.61.mlp.up_proj",
+          "model.language_model.layers.62.mlp.down_proj",
+          "model.language_model.layers.62.mlp.gate_proj",
+          "model.language_model.layers.62.mlp.up_proj",
+          "model.language_model.layers.63.mlp.down_proj",
+          "model.language_model.layers.63.mlp.gate_proj",
+          "model.language_model.layers.63.mlp.up_proj",
+          "model.language_model.layers.7.mlp.down_proj",
+          "model.language_model.layers.7.mlp.gate_proj",
+          "model.language_model.layers.7.mlp.up_proj",
+          "model.language_model.layers.8.mlp.down_proj",
+          "model.language_model.layers.8.mlp.gate_proj",
+          "model.language_model.layers.8.mlp.up_proj",
+          "model.language_model.layers.9.mlp.down_proj",
+          "model.language_model.layers.9.mlp.gate_proj",
+          "model.language_model.layers.9.mlp.up_proj"
+        ]
+      }
+    ]
+  },
+  "kv_cache_scheme": {
+    "num_bits": 8,
+    "type": "float",
+    "strategy": null
+  },
   "coverage": {
-    "attention": {
-      "quantized": 64,
-      "weights": 68
-    },
-    "dense_mlp": {
-      "quantized": 192,
-      "weights": 195
+    "router": {
+      "weights": 0,
+      "quantized": 0
     },
     "experts": {
-      "quantized": 0,
-      "weights": 0
-    },
-    "router": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
     },
     "shared_expert": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
+    },
+    "attention": {
+      "weights": 68,
+      "quantized": 64
+    },
+    "dense_mlp": {
+      "weights": 195,
+      "quantized": 192
     }
   },
-  "declared": {},
-  "detected_block": null,
-  "detected_scheme": "fp8_per_tensor",
-  "model": "/media/fmodels2/RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead",
-  "quant_method": "modelopt",
-  "weights_bytes": 23749332688,
-  "weights_bytes_source": "hub"
+  "unloadable_reason": null,
+  "quantized_names_sha256": "eccbe240ef5caa18a896c8a311d4a9b4257c68423f0d8f9ae1c982d53797e657"
 }

Qwen3.8-27B-NVFP4-BF16-LMHead/report.md

@@ -1,12 +1,12 @@
 # Qwen3.8-27B / Qwen3.8-27B-NVFP4-BF16-LMHead: distribution fidelity
 
-**Mean KLD(reference || candidate) = 0.05212951** over 1572096 scored positions.
+**Mean KLD(reference || candidate) = 0.05156551** over 1572096 scored positions.
 
 | Mean | Median | p90 | p99 | Max | Top-1 agreement |
 |---|---|---|---|---|---|
-| 0.05212951 | 0.01302997 | 0.07506841 | 0.67170381 | 27.48295784 | 91.5187% |
+| 0.05156551 | 0.01290934 | 0.07437129 | 0.65636102 | 36.40480804 | 91.5647% |
 
-Reverse direction, KLD(candidate || reference): 0.05476134.
+Reverse direction, KLD(candidate || reference): 0.05437684.
 
 ## Identity
 
@@ -15,10 +15,10 @@ Reverse direction, KLD(candidate || reference): 0.05476134.
 | Reference checkpoint | /media/fmodels2/Qwen/Qwen3.8-27B |
 | Reference config SHA-256 | 191e0af23210 |
 | Candidate checkpoint | /media/fmodels2/RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead |
-| Candidate weights SHA-256 | unbound (Law 16 gap) |
+| Candidate weights SHA-256 | 2238b4ba183e9cc5 |
 | Suite | qwen3.8-27b-fidelity-1024x2048-v1 |
 | Suite token SHA-256 | 9c935708bbffbf45 |
-| Capture manifest SHA-256 | 304794b4e35b326d |
+| Capture manifest SHA-256 | dbfacc9cbb6d3e06 |
 | Tokenizer | None |
 | Scored vocabulary | 248044 |
 | Declared vocabulary | 248320 |
@@ -27,32 +27,39 @@ Reverse direction, KLD(candidate || reference): 0.05476134.
 | Model runner | V1 |
 | Tensor parallel | 1 |
 | Eager enforced | True |
+| KV cache | bfloat16 |
 | Prefix caching | False |
 | max_num_seqs | 1 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45 |
+| vLLM commit | 60071d1ab732 |
+| vLLM dirty digest | e3b0c44298fc1c14 |
+| Numerics digest | 251a9225b37415b9 |
+| Compiled extensions | f2fbc7537b0f01f6 |
+| FlashInfer | 0.6.17 |
 | torch | 2.13.0+cu132 |
 | Driver | 580.173.02 |
 | GPUs | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
-| Laws version | 9 |
+| Laws version | 15 |
 | Partition | analysis |
 
+The commit says when this number was taken. The numerics digest says what took it: a hash of the runtime and the scorer, which moves only when code that can change a logit changes. Comparability is bounded by the digest, so a result stays current across commits that cannot reach a number, and two results carrying the same digest were computed by the same code whatever their commits say.
+
 ## Fidelity by domain
 
-The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
+The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `deployed` is QxQ (natural student routing). `bxq` is the teacher-ID counterfactual on the same weights. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
 
 | Domain | Contexts | Reference top-1 | Mean KLD | x run | Worst context |
 |---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.14506069 | 2.78 | 0.93519654 |
-| Chinese across several content types | 72 | 55.6% | 0.08009739 | 1.54 | 0.33491146 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.05208049 | 1.00 | 0.25756627 |
-| Other multilingual content | 36 | 64.5% | 0.03909129 | 0.75 | 0.08191294 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.03826159 | 0.73 | 0.10122026 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.03723338 | 0.71 | 0.21792144 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.03034605 | 0.58 | 0.21618890 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.02762031 | 0.53 | 0.11659044 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.02460944 | 0.47 | 0.08214911 |
-| Scientific and technical exposition | 96 | 60.3% | 0.02322830 | 0.45 | 0.04042489 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.14260814 | 2.77 | 0.89053696 |
+| Chinese across several content types | 72 | 55.6% | 0.07907412 | 1.53 | 0.31717224 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.05124551 | 0.99 | 0.25338622 |
+| Other multilingual content | 36 | 64.5% | 0.03972507 | 0.77 | 0.08763030 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.03811238 | 0.74 | 0.10094674 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.03711086 | 0.72 | 0.21911475 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.03018829 | 0.59 | 0.20473772 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.02775164 | 0.54 | 0.11848210 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.02442816 | 0.47 | 0.08643491 |
+| Scientific and technical exposition | 96 | 60.3% | 0.02302717 | 0.45 | 0.04442971 |
 
 **Natural dialogue, instruction following, and assistance** is this candidate's weakest domain and **Scientific and technical exposition** its strongest, a spread of 6.2x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
 
@@ -60,8 +67,8 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Component | Value |
 |---|---|
-| Trunk (candidate hidden states, reference head) | 0.05212951 |
-| Deployed (candidate's own head) | 0.05212951 |
+| Trunk (candidate hidden states, reference head) | 0.05156551 |
+| Deployed (candidate's own head) | 0.05156551 |
 | Head-associated delta (not additive) | 0.00000000 |
 | Reference head | {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]} |
 | Candidate head | {'runtime': {'state': 'quantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedLinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedLinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'quantized'}]}, 'state': 'quantized', 'static': {'ignored': False, 'lm_head_dtypes': {'lm_head.weight': 'BF16', 'model.language_model.embed_tokens.weight': 'BF16'}, 'lm_head_keys': ['lm_head.weight'], 'output_weight_keys': ['lm_head.weight'], 'packed_keys': [], 'quant_method': 'modelopt', 'state': 'unquantized', 'tie_word_embeddings': False}} |
@@ -70,26 +77,26 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Position range | Positions | Mean KLD |
 |---|---|---|
-| 0–511 | 393216 | 0.04442098 |
-| 512–1023 | 393216 | 0.04805359 |
-| 1024–1535 | 393216 | 0.05569485 |
-| 1536–2046 | 392448 | 0.06036471 |
+| 0–511 | 393216 | 0.04401547 |
+| 512–1023 | 393216 | 0.04793031 |
+| 1024–1535 | 393216 | 0.05523113 |
+| 1536–2046 | 392448 | 0.05909983 |
 
 ## Error by reference confidence
 
 | Reference top-1 probability | Positions | Share | Mean KLD |
 |---|---|---|---|
-| [0.00, 0.25) | 240569 | 15.3% | 0.05929329 |
-| [0.25, 0.50) | 346390 | 22.0% | 0.07293711 |
-| [0.50, 0.75) | 273384 | 17.4% | 0.07885677 |
-| [0.75, 0.95) | 249939 | 15.9% | 0.05907473 |
-| [0.95, 1.00) | 461814 | 29.4% | 0.01320992 |
+| [0.00, 0.25) | 240522 | 15.3% | 0.05898762 |
+| [0.25, 0.50) | 346508 | 22.0% | 0.07237716 |
+| [0.50, 0.75) | 273204 | 17.4% | 0.07748764 |
+| [0.75, 0.95) | 250041 | 15.9% | 0.05845658 |
+| [0.95, 1.00) | 461821 | 29.4% | 0.01301883 |
 
 ## Top-K set agreement
 
 | K=1 | K=2 | K=3 | K=4 | K=5 |
 |---|---|---|---|---|
-| 91.3861% | 73.6066% | 53.9226% | 36.5832% | 23.7458% |
+| 91.4269% | 73.6996% | 54.1363% | 36.7871% | 23.9870% |
 
 ## Law compliance
 
@@ -101,22 +108,19 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | 2 | Determinism | PASS | eager enforced; prefix caching off, max_num_seqs=1 |
 | 3 | Frozen input | PASS | token hash matches suite qwen3.8-27b-fidelity-1024x2048-v1 [analysis]: 9c935708bbffbf45 |
 | 4 | Real vocabulary | PASS | scored 248044 real tokens of 248320 declared (276 padding rows) |
-| 5 | Manifest binding | PASS | all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc |
-| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s) |
+| 5 | Manifest binding | PASS | all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3 |
+| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s) |
 | 7 | Storage integrity | PASS | hidden storage with bitwise-exact replay |
-| 8 | Head transparency | PASS | trunk 0.05212951, deployed 0.05212951, delta 1.3275290033920584e-09 |
-| 9 | Tail and depth disclosure | PASS | mean 0.05212951, median 0.01302997, max 27.48295784, 4 depth buckets |
+| 8 | Head transparency | PASS | trunk 0.05156551, deployed 0.05156551, delta 1.1058352220039147e-09 |
+| 9 | Tail and depth disclosure | PASS | mean 0.05156551, median 0.01290934, max 36.40480804, 4 depth buckets |
 | 10 | Comparability | PASS | comparability key fully resolved |
 | 11 | Freeze before qualification | NOT_APPLICABLE | partition is 'analysis' |
 | 12 | Reusable reference | PASS | suite, reference, head, and checksums present; published reference matches the scored capture on every bound field |
-| 14 | Component attribution | NOT_APPLICABLE | reference declares no experts |
-| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.14506069, strongest scientific_technical at 0.02322830, spread 6.2x |
-| 16 | Candidate weight binding | OVERRIDE | the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else |
-| 13 | Recorded deviation | PASS | 1 override(s), each fully attributed |
-
-### Recorded deviations
-
-- **Law 16 (Candidate weight binding)** overridden by Andy Kitzke at 2026-09-02T05:15:00Z. Justification: These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger. Underlying finding: the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else.
+| 14 | Routed-model intervention | NOT_APPLICABLE | reference declares no experts |
+| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.14260814, strongest scientific_technical at 0.02302717, spread 6.2x |
+| 16 | Candidate weight binding | PASS | scored weights 2238b4ba183e9cc5 as inspected |
+| 17 | Substitution disclosure | PASS | every scored layer used the checkpoint's own quantization parameters |
+| 13 | Recorded deviation | PASS | no overrides claimed |
 
 ## Environment
 
@@ -126,7 +130,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | Platform | Linux-6.8.0-137-generic-x86_64-with-glibc2.39 |
 | Python | 3.12.3 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45fbe0b0205e70893bcd28fa928a88 |
+| vLLM commit | 60071d1ab73229321712254b1350dcaaac1c125a |
 | torch | 2.13.0+cu132 |
 | torch CUDA runtime | 13.2 |
 | cuDNN | 9.20.0 (92000) |
@@ -134,7 +138,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | CUDA arch list | sm_75, sm_80, sm_86, sm_90, sm_100, sm_120 |
 | nvcc | Cuda compilation tools, release 13.0, V13.0.88 |
 | gcc | gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0 |
-| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.8) 2.39 |
+| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.9) 2.39 |
 | NVIDIA driver | 580.173.02 |
 | float32 matmul precision | highest |
 | TF32 (matmul / cuDNN) | False / True |
@@ -157,8 +161,18 @@ Credential values are never published. Set but redacted: `HF_TOKEN`.
 
 | Variable | Value |
 |---|---|
+| `CUBLAS_WORKSPACE_CONFIG` | `:4096:8` |
 | `CUDA_HOME` | `/usr/local/cuda-13.0` |
 | `HF_TOKEN` | `<redacted>` |
+| `NCCL_DETERMINISTIC` | `1` |
+| `PYTORCH_NVML_BASED_CUDA_CHECK` | `1` |
+| `TORCHINDUCTOR_CACHE_DIR` | `/tmp/torchinductor_phaedawg` |
+| `TORCHINDUCTOR_COMPILE_THREADS` | `1` |
+| `TRITON_CACHE_AUTOTUNING` | `1` |
+| `TRITON_PTXAS_BLACKWELL_PATH` | `/usr/local/cuda-13.0/bin/ptxas` |
+| `VLLM_BATCH_INVARIANT` | `1` |
+| `VLLM_MARLIN_USE_ATOMIC_ADD` | `0` |
+| `VLLM_MOE_USE_DEEP_GEMM` | `0` |
 
 ## Files in this artifact
 
@@ -166,26 +180,26 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 
 | Path | Size | What it is |
 |---|---|---|
-| `Qwen3.8-27B-NVFP4-BF16-LMHead/report.md` | 12.37 KiB | This document. |
-| `Qwen3.8-27B-NVFP4-BF16-LMHead/report.json` | 254.28 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
-| `Qwen3.8-27B-NVFP4-BF16-LMHead/manifest.json` | 85.43 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
-| `Qwen3.8-27B-NVFP4-BF16-LMHead/compliance.json` | 13.03 KiB | The law-by-law receipt, including the comparability key. |
-| `baselines/self-kld.json` | 4.74 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
+| `Qwen3.8-27B-NVFP4-BF16-LMHead/report.md` | 13.21 KiB | This document. |
+| `Qwen3.8-27B-NVFP4-BF16-LMHead/report.json` | 256.72 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
+| `Qwen3.8-27B-NVFP4-BF16-LMHead/manifest.json` | 86.99 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
+| `Qwen3.8-27B-NVFP4-BF16-LMHead/compliance.json` | 12.30 KiB | The law-by-law receipt, including the comparability key. |
+| `baselines/self-kld.json` | 7.06 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
 | `suite/suite-manifest.json` | 382.68 KiB | The frozen evaluation input's identity: token hashes per context and per partition, sources, strata, and the analysis/qualification split. |
 | `suite/tokens` | 15.49 MiB in 1024 files | The token IDs themselves. These are the evaluation input, not a description of it; retokenizing source text does not reproduce them. |
 | `suite/sources.json` | 737.42 KiB | Per-context provenance: dataset, revision, licence, source unit, and the deterministic token offset chosen within the document. |
 | `suite/validation/capability-overlap.json` | 1.48 KiB | The benchmark-contamination scan and every document it blocked. |
-| `reference/manifest.json` | 85.43 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
+| `reference/manifest.json` | 86.99 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
 | `reference` | 19.25 GiB in 771 files | The reusable reference distributions. Pass this directory as `--reference-logits` to score a new candidate against the same reference without loading the reference checkpoint. |
 | `reference/lm_head.safetensors` | 2.37 GiB | The reference language-model head, which turns the stored hidden states back into reference logits. |
-| `environment/runtime.json` | 2.26 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
+| `environment/runtime.json` | 5.42 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
 | `environment/summary.md` | 1.18 KiB | The same provenance as prose, plus an index of every captured file. |
 | `environment/toolchain-nvcc.txt` | 243 B | `nvcc --version` verbatim; `toolchain-gcc.txt` and `toolchain-ldd.txt` sit beside it. |
 | `environment/gpu-smi-query.txt` | 54.73 KiB | `nvidia-smi -q` verbatim: ECC state, persistence mode, clocks, and throttle reasons, any of which can move a bitwise result. |
-| `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
-| `environment/models` | 7.22 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
-| `checksums.txt` | 182.99 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
+| `environment/pip-freeze.txt` | 4.47 KiB | Every installed package version in the scoring environment. |
+| `environment/models` | 19.46 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
+| `checksums.txt` | 183.22 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
+| `LAWS.md` | 40.20 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

Qwen3.8-27B-NVFP4-BF16-LMHead/strata.json

@@ -8,381 +8,381 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 27.48295783996582,
-            "mean_kld": 0.1450606879671042,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.10672595977404836,
-            "median_context_p99": 2.366913318634033,
-            "p90_context_kld": 0.31745866158131525,
+            "max_kld": 36.404808044433594,
+            "mean_kld": 0.14260813963956548,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.10354972375019698,
+            "median_context_p99": 2.1813976764678955,
+            "p90_context_kld": 0.3451019763379924,
             "positions": 196512,
-            "top1_agreement": 0.9130485670086305,
+            "top1_agreement": 0.9151603973294252,
             "worst_context_id": 454,
-            "worst_context_kld": 0.9351965383137365
+            "worst_context_kld": 0.8905369579241399
           }
         },
         "key": "wildchat",
         "label": "wildchat",
-        "relative_to_run": 2.7826980242658594
+        "relative_to_run": 2.765572355107366
       },
       {
         "cells": {
           "deployed": {
             "contexts": 33,
-            "max_kld": 7.188745975494385,
-            "mean_kld": 0.1293620515682006,
-            "mean_ref_top1_prob": 0.6040606719329803,
-            "median_context_kld": 0.09606564325586857,
-            "median_context_p99": 0.7667441964149475,
-            "p90_context_kld": 0.304599996166998,
+            "max_kld": 6.398060321807861,
+            "mean_kld": 0.1270748634639472,
+            "mean_ref_top1_prob": 0.6040861921682756,
+            "median_context_kld": 0.09748326356933214,
+            "median_context_p99": 0.7841387987136841,
+            "p90_context_kld": 0.2912568851459713,
             "positions": 67551,
-            "top1_agreement": 0.8494914953146512,
+            "top1_agreement": 0.8525854539533093,
             "worst_context_id": 893,
-            "worst_context_kld": 0.33491145623420787
+            "worst_context_kld": 0.31717224464172616
           }
         },
         "key": "wikisource_zh",
         "label": "wikisource_zh",
-        "relative_to_run": 2.4815512070054617
+        "relative_to_run": 2.4643385034905307
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 7.891758918762207,
-            "mean_kld": 0.06534968393771708,
-            "mean_ref_top1_prob": 0.7357366218042676,
-            "median_context_kld": 0.07035002679922857,
-            "median_context_p99": 1.1989408731460571,
-            "p90_context_kld": 0.08191294001897026,
+            "max_kld": 10.652168273925781,
+            "mean_kld": 0.06826552331233253,
+            "mean_ref_top1_prob": 0.7358200961573914,
+            "median_context_kld": 0.06837731806777113,
+            "median_context_p99": 1.382102608680725,
+            "p90_context_kld": 0.08763029838216986,
             "positions": 14329,
             "top1_agreement": 0.9057156814851002,
             "worst_context_id": 953,
-            "worst_context_kld": 0.08191294001897026
+            "worst_context_kld": 0.08763029838216986
           }
         },
         "key": "wikipedia_de",
         "label": "wikipedia_de",
-        "relative_to_run": 1.2536024675487678
+        "relative_to_run": 1.323860226749253
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 96,
-            "max_kld": 8.426796913146973,
-            "mean_kld": 0.05208049148082254,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.04317494681946243,
-            "median_context_p99": 0.43730735778808594,
-            "p90_context_kld": 0.0747909649283652,
-            "positions": 196512,
-            "top1_agreement": 0.9043417195896434,
-            "worst_context_id": 6,
-            "worst_context_kld": 0.2575662736752676
+            "contexts": 23,
+            "max_kld": 6.221218109130859,
+            "mean_kld": 0.05175295408977827,
+            "mean_ref_top1_prob": 0.7233191001139795,
+            "median_context_kld": 0.06760430901394703,
+            "median_context_p99": 0.8653435707092285,
+            "p90_context_kld": 0.08414925132163899,
+            "positions": 47081,
+            "top1_agreement": 0.9238121535226524,
+            "worst_context_id": 275,
+            "worst_context_kld": 0.10094674048285278
           }
         },
-        "key": "wikipedia_en",
-        "label": "wikipedia_en",
-        "relative_to_run": 0.9990596541176248
+        "key": "regulations",
+        "label": "regulations",
+        "relative_to_run": 1.0036351325217217
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 23,
-            "max_kld": 6.322475910186768,
-            "mean_kld": 0.05162461144194454,
-            "mean_ref_top1_prob": 0.7232628548554616,
-            "median_context_kld": 0.06074220332847508,
-            "median_context_p99": 0.7631682753562927,
-            "p90_context_kld": 0.08217725732574639,
-            "positions": 47081,
-            "top1_agreement": 0.9227076740086234,
-            "worst_context_id": 275,
-            "worst_context_kld": 0.10122025808876788
+            "contexts": 96,
+            "max_kld": 7.654315948486328,
+            "mean_kld": 0.05124551225377584,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.04196325648602333,
+            "median_context_p99": 0.4040193259716034,
+            "p90_context_kld": 0.07206766500842036,
+            "positions": 196512,
+            "top1_agreement": 0.9047081094284318,
+            "worst_context_id": 6,
+            "worst_context_kld": 0.253386224192646
           }
         },
-        "key": "regulations",
-        "label": "regulations",
-        "relative_to_run": 0.990314510955366
+        "key": "wikipedia_en",
+        "label": "wikipedia_en",
+        "relative_to_run": 0.9937944101266313
       },
       {
         "cells": {
           "deployed": {
             "contexts": 39,
-            "max_kld": 5.228215217590332,
-            "mean_kld": 0.038411913240515724,
-            "mean_ref_top1_prob": 0.515358929042611,
-            "median_context_kld": 0.037433911971596914,
-            "median_context_p99": 0.29926493763923645,
-            "p90_context_kld": 0.052808316067491694,
+            "max_kld": 7.264811038970947,
+            "mean_kld": 0.03845810633107973,
+            "mean_ref_top1_prob": 0.5153760832971095,
+            "median_context_kld": 0.03694213221466098,
+            "median_context_p99": 0.29629912972450256,
+            "p90_context_kld": 0.05189267066783427,
             "positions": 79833,
-            "top1_agreement": 0.8989265090877206,
-            "worst_context_id": 886,
-            "worst_context_kld": 0.05916575064199425
+            "top1_agreement": 0.8974108451392281,
+            "worst_context_id": 890,
+            "worst_context_kld": 0.05947408993629915
           }
         },
         "key": "wikipedia_zh",
         "label": "wikipedia_zh",
-        "relative_to_run": 0.7368554263777803
+        "relative_to_run": 0.7458106947319388
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 5.807681560516357,
-            "mean_kld": 0.03723338342274077,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.03198886603632158,
-            "median_context_p99": 0.19292420148849487,
-            "p90_context_kld": 0.04507155972030506,
+            "max_kld": 6.983285427093506,
+            "mean_kld": 0.03711085502771101,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.031371119738575946,
+            "median_context_p99": 0.20159181952476501,
+            "p90_context_kld": 0.04505553168422091,
             "positions": 147384,
-            "top1_agreement": 0.9027506377897194,
+            "top1_agreement": 0.9036666123866905,
             "worst_context_id": 443,
-            "worst_context_kld": 0.21792143969470823
+            "worst_context_kld": 0.21911474635649905
           }
         },
         "key": "public_domain_books",
         "label": "public_domain_books",
-        "relative_to_run": 0.7142476982508844
+        "relative_to_run": 0.7196837080859033
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 2.8669304847717285,
-            "mean_kld": 0.037127001025355474,
-            "mean_ref_top1_prob": 0.5702914711488575,
-            "median_context_kld": 0.03723822674918608,
-            "median_context_p99": 0.2575864791870117,
-            "p90_context_kld": 0.04216571407522604,
+            "max_kld": 1.9876776933670044,
+            "mean_kld": 0.036838090958891075,
+            "mean_ref_top1_prob": 0.5701150598950927,
+            "median_context_kld": 0.03802288941157957,
+            "median_context_p99": 0.2619818150997162,
+            "p90_context_kld": 0.041282101737260615,
             "positions": 14329,
-            "top1_agreement": 0.906273989810873,
-            "worst_context_id": 968,
-            "worst_context_kld": 0.04216571407522604
+            "top1_agreement": 0.9052969502407705,
+            "worst_context_id": 950,
+            "worst_context_kld": 0.041282101737260615
           }
         },
         "key": "wikipedia_ja",
         "label": "wikipedia_ja",
-        "relative_to_run": 0.7122069655674171
+        "relative_to_run": 0.71439404670962
       },
       {
         "cells": {
           "deployed": {
             "contexts": 26,
-            "max_kld": 4.580113887786865,
-            "mean_kld": 0.03648647187529798,
-            "mean_ref_top1_prob": 0.5182113024080622,
-            "median_context_kld": 0.03274300100939314,
-            "median_context_p99": 0.28775858879089355,
-            "p90_context_kld": 0.05933113081028002,
+            "max_kld": 5.564556121826172,
+            "mean_kld": 0.036061618806001935,
+            "mean_ref_top1_prob": 0.5182496321630543,
+            "median_context_kld": 0.03299574269935542,
+            "median_context_p99": 0.26128193736076355,
+            "p90_context_kld": 0.06031771898826233,
             "positions": 53222,
-            "top1_agreement": 0.9078200744053211,
+            "top1_agreement": 0.907200030062756,
             "worst_context_id": 312,
-            "worst_context_kld": 0.08784422891990816
+            "worst_context_kld": 0.08485144357051137
           }
         },
         "key": "public_domain_review",
         "label": "public_domain_review",
-        "relative_to_run": 0.6999196999730758
+        "relative_to_run": 0.6993360708748022
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 2.774463653564453,
-            "mean_kld": 0.03460012152143339,
-            "mean_ref_top1_prob": 0.591765751203999,
-            "median_context_kld": 0.033330434864274824,
-            "median_context_p99": 0.3238604664802551,
-            "p90_context_kld": 0.04504018613805693,
+            "max_kld": 2.439326047897339,
+            "mean_kld": 0.03428308194879127,
+            "mean_ref_top1_prob": 0.591761624645417,
+            "median_context_kld": 0.03246858632634692,
+            "median_context_p99": 0.308083176612854,
+            "p90_context_kld": 0.04459574729516285,
             "positions": 14329,
-            "top1_agreement": 0.9142996719938586,
+            "top1_agreement": 0.9115779189057157,
             "worst_context_id": 936,
-            "worst_context_kld": 0.04504018613805693
+            "worst_context_kld": 0.04459574729516285
           }
         },
         "key": "wikipedia_es",
         "label": "wikipedia_es",
-        "relative_to_run": 0.6637338561284459
+        "relative_to_run": 0.6648452460363811
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 4.131490707397461,
-            "mean_kld": 0.03329140193577505,
-            "mean_ref_top1_prob": 0.6892386900239509,
-            "median_context_kld": 0.035325973994472015,
-            "median_context_p99": 0.2722674012184143,
-            "p90_context_kld": 0.03764456319544782,
+            "max_kld": 3.7520291805267334,
+            "mean_kld": 0.03395948911061795,
+            "mean_ref_top1_prob": 0.689239984531695,
+            "median_context_kld": 0.034178675689330404,
+            "median_context_p99": 0.3189646601676941,
+            "p90_context_kld": 0.0410441593638348,
             "positions": 12282,
-            "top1_agreement": 0.9235466536394724,
+            "top1_agreement": 0.9212668946425664,
             "worst_context_id": 958,
-            "worst_context_kld": 0.03764456319544782
+            "worst_context_kld": 0.0410441593638348
           }
         },
         "key": "wikipedia_cs",
         "label": "wikipedia_cs",
-        "relative_to_run": 0.6386286987190496
+        "relative_to_run": 0.6585698720652688
       },
       {
         "cells": {
           "deployed": {
             "contexts": 52,
-            "max_kld": 12.813560485839844,
-            "mean_kld": 0.030658742448252164,
-            "mean_ref_top1_prob": 0.8582575125835092,
-            "median_context_kld": 0.019792192856613437,
-            "median_context_p99": 0.22031214833259583,
-            "p90_context_kld": 0.05239252369494203,
+            "max_kld": 14.599754333496094,
+            "mean_kld": 0.030492619988162543,
+            "mean_ref_top1_prob": 0.8582941084526002,
+            "median_context_kld": 0.019386495896305277,
+            "median_context_p99": 0.2122478187084198,
+            "p90_context_kld": 0.05378769088686552,
             "positions": 106444,
-            "top1_agreement": 0.9569538912479801,
+            "top1_agreement": 0.9571511780842509,
             "worst_context_id": 667,
-            "worst_context_kld": 0.2161889008647869
+            "worst_context_kld": 0.20473772003888704
           }
         },
         "key": "github_code",
         "label": "github_code",
-        "relative_to_run": 0.5881264126954507
+        "relative_to_run": 0.5913375427741718
       },
       {
         "cells": {
           "deployed": {
             "contexts": 44,
-            "max_kld": 9.661121368408203,
-            "mean_kld": 0.029976498399735953,
-            "mean_ref_top1_prob": 0.720394649593742,
-            "median_context_kld": 0.022446578193640192,
-            "median_context_p99": 0.24618467688560486,
-            "p90_context_kld": 0.04893623020139342,
+            "max_kld": 12.78420352935791,
+            "mean_kld": 0.02982863504742722,
+            "mean_ref_top1_prob": 0.7203910827522294,
+            "median_context_kld": 0.02299396894002985,
+            "median_context_p99": 0.2256595939397812,
+            "p90_context_kld": 0.05378093218949216,
             "positions": 90068,
-            "top1_agreement": 0.9408891060087934,
+            "top1_agreement": 0.942077097304259,
             "worst_context_id": 582,
-            "worst_context_kld": 0.2090905349448391
+            "worst_context_kld": 0.1888053465058198
           }
         },
         "key": "stackv2",
         "label": "stackv2",
-        "relative_to_run": 0.5750389305355446
+        "relative_to_run": 0.5784610099132401
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 1.9787611961364746,
-            "mean_kld": 0.02863895871709025,
-            "mean_ref_top1_prob": 0.6660615117323249,
-            "median_context_kld": 0.02861696461493997,
-            "median_context_p99": 0.26254552602767944,
-            "p90_context_kld": 0.03984703824302733,
+            "max_kld": 3.0671863555908203,
+            "mean_kld": 0.028917358048958376,
+            "mean_ref_top1_prob": 0.6660535411394803,
+            "median_context_kld": 0.030182035840909158,
+            "median_context_p99": 0.2886667847633362,
+            "p90_context_kld": 0.03903039077527934,
             "positions": 12282,
-            "top1_agreement": 0.923383813711122,
+            "top1_agreement": 0.9246051131737502,
             "worst_context_id": 945,
-            "worst_context_kld": 0.03984703824302733
+            "worst_context_kld": 0.03903039077527934
           }
         },
         "key": "wikipedia_ru",
         "label": "wikipedia_ru",
-        "relative_to_run": 0.5493809174347145
+        "relative_to_run": 0.5607887895113729
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 7.356241226196289,
-            "mean_kld": 0.027620313995326606,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.02249949949724309,
-            "median_context_p99": 0.40589118003845215,
-            "p90_context_kld": 0.05043523494828777,
+            "max_kld": 8.024088859558105,
+            "mean_kld": 0.02775163699079394,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.02158035745589829,
+            "median_context_p99": 0.34725338220596313,
+            "p90_context_kld": 0.04604335776204541,
             "positions": 73692,
-            "top1_agreement": 0.8654535092004559,
+            "top1_agreement": 0.8652906692721055,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.11659044195651949
+            "worst_context_kld": 0.11848209727913012
           }
         },
         "key": "starcoder_structured",
         "label": "starcoder_structured",
-        "relative_to_run": 0.5298402638337655
+        "relative_to_run": 0.5381821841634996
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 5.2930073738098145,
-            "mean_kld": 0.026905234620940046,
-            "mean_ref_top1_prob": 0.5379292386352562,
-            "median_context_kld": 0.02633995404131651,
-            "median_context_p99": 0.18736493587493896,
-            "p90_context_kld": 0.03581834205987438,
+            "max_kld": 4.237236976623535,
+            "mean_kld": 0.02679005651865327,
+            "mean_ref_top1_prob": 0.537879873360319,
+            "median_context_kld": 0.02628154840440571,
+            "median_context_p99": 0.1890130341053009,
+            "p90_context_kld": 0.03704089324845825,
             "positions": 47081,
-            "top1_agreement": 0.9164206367749198,
+            "top1_agreement": 0.9161657568870669,
             "worst_context_id": 343,
-            "worst_context_kld": 0.04333593684776235
+            "worst_context_kld": 0.0388510133394851
           }
         },
         "key": "open_news",
         "label": "open_news",
-        "relative_to_run": 0.5161229018786782
+        "relative_to_run": 0.5195344381254079
       },
       {
         "cells": {
           "deployed": {
             "contexts": 3,
-            "max_kld": 1.2579565048217773,
-            "mean_kld": 0.025388817897578672,
-            "mean_ref_top1_prob": 0.6023603953518782,
-            "median_context_kld": 0.025382461787984437,
-            "median_context_p99": 0.1844366490840912,
-            "p90_context_kld": 0.029254880163471348,
+            "max_kld": 1.023111343383789,
+            "mean_kld": 0.025711511580105272,
+            "mean_ref_top1_prob": 0.6023746847155801,
+            "median_context_kld": 0.026561403553094273,
+            "median_context_p99": 0.23045402765274048,
+            "p90_context_kld": 0.02939077877766676,
             "positions": 6141,
-            "top1_agreement": 0.9229767138902459,
+            "top1_agreement": 0.9211854746783912,
             "worst_context_id": 967,
-            "worst_context_kld": 0.029254880163471348
+            "worst_context_kld": 0.02939077877766676
           }
         },
         "key": "wikipedia_fr",
         "label": "wikipedia_fr",
-        "relative_to_run": 0.48703349192759393
+        "relative_to_run": 0.4986184225786925
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 17.052858352661133,
-            "mean_kld": 0.024609437161871247,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.020860553521432415,
-            "median_context_p99": 0.14598329365253448,
-            "p90_context_kld": 0.03323913724905251,
+            "max_kld": 13.109975814819336,
+            "mean_kld": 0.024428160646162753,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.020779478378882214,
+            "median_context_p99": 0.1441325843334198,
+            "p90_context_kld": 0.03284644349083438,
             "positions": 196512,
-            "top1_agreement": 0.9371743201432992,
+            "top1_agreement": 0.9363652092493079,
             "worst_context_id": 825,
-            "worst_context_kld": 0.08214911499231742
+            "worst_context_kld": 0.08643491137050359
           }
         },
         "key": "libretexts",
         "label": "libretexts",
-        "relative_to_run": 0.4720826374693844
+        "relative_to_run": 0.4737306435656357
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.320075988769531,
-            "mean_kld": 0.023228298859926257,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.02226252774408313,
-            "median_context_p99": 0.16904449462890625,
-            "p90_context_kld": 0.031335149218943904,
+            "max_kld": 13.07381534576416,
+            "mean_kld": 0.023027166196565346,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.021995164691474722,
+            "median_context_p99": 0.16406352818012238,
+            "p90_context_kld": 0.030565308802914134,
             "positions": 196512,
-            "top1_agreement": 0.9289712587526462,
+            "top1_agreement": 0.929612440970526,
             "worst_context_id": 246,
-            "worst_context_kld": 0.04042488780931219
+            "worst_context_kld": 0.044429711784661804
           }
         },
         "key": "scientific_papers",
         "label": "scientific_papers",
-        "relative_to_run": 0.44558827240107735
+        "relative_to_run": 0.44656142637187546
       }
     ],
     "stratum": [
@@ -390,217 +390,217 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 27.48295783996582,
-            "mean_kld": 0.1450606879671042,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.10672595977404836,
-            "median_context_p99": 2.366913318634033,
-            "p90_context_kld": 0.31745866158131525,
+            "max_kld": 36.404808044433594,
+            "mean_kld": 0.14260813963956548,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.10354972375019698,
+            "median_context_p99": 2.1813976764678955,
+            "p90_context_kld": 0.3451019763379924,
             "positions": 196512,
-            "top1_agreement": 0.9130485670086305,
+            "top1_agreement": 0.9151603973294252,
             "worst_context_id": 454,
-            "worst_context_kld": 0.9351965383137365
+            "worst_context_kld": 0.8905369579241399
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 2.7826980242658594
+        "relative_to_run": 2.765572355107366
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.188745975494385,
-            "mean_kld": 0.0800973933073713,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.0488142405969883,
-            "median_context_p99": 0.3909897208213806,
-            "p90_context_kld": 0.14562944384550067,
+            "max_kld": 7.264811038970947,
+            "mean_kld": 0.07907412001697732,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.04855310270042977,
+            "median_context_p99": 0.41729986667633057,
+            "p90_context_kld": 0.14487312296101867,
             "positions": 147384,
-            "top1_agreement": 0.8762687944417304,
+            "top1_agreement": 0.8768658741790154,
             "worst_context_id": 893,
-            "worst_context_kld": 0.33491145623420787
+            "worst_context_kld": 0.31717224464172616
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.5365076591654676
+        "relative_to_run": 1.5334692737462934
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 8.426796913146973,
-            "mean_kld": 0.05208049148082254,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.04317494681946243,
-            "median_context_p99": 0.43730735778808594,
-            "p90_context_kld": 0.0747909649283652,
+            "max_kld": 7.654315948486328,
+            "mean_kld": 0.05124551225377584,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.04196325648602333,
+            "median_context_p99": 0.4040193259716034,
+            "p90_context_kld": 0.07206766500842036,
             "positions": 196512,
-            "top1_agreement": 0.9043417195896434,
+            "top1_agreement": 0.9047081094284318,
             "worst_context_id": 6,
-            "worst_context_kld": 0.2575662736752676
+            "worst_context_kld": 0.253386224192646
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.9990596541176248
+        "relative_to_run": 0.9937944101266313
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 7.891758918762207,
-            "mean_kld": 0.03909128508337415,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.03482694495389481,
-            "median_context_p99": 0.2722674012184143,
-            "p90_context_kld": 0.07035002679922857,
+            "max_kld": 10.652168273925781,
+            "mean_kld": 0.03972506920105216,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.034178675689330404,
+            "median_context_p99": 0.2886667847633362,
+            "p90_context_kld": 0.06837731806777113,
             "positions": 73692,
-            "top1_agreement": 0.9148482874667535,
+            "top1_agreement": 0.9138033979265049,
             "worst_context_id": 953,
-            "worst_context_kld": 0.08191294001897026
+            "worst_context_kld": 0.08763029838216986
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.7498878110393828
+        "relative_to_run": 0.7703806631572974
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.322475910186768,
-            "mean_kld": 0.03826159344727907,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.031959890771641075,
-            "median_context_p99": 0.2690824270248413,
-            "p90_context_kld": 0.07946953402075285,
+            "max_kld": 6.221218109130859,
+            "mean_kld": 0.038112379624305213,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.03164991631323535,
+            "median_context_p99": 0.2747260630130768,
+            "p90_context_kld": 0.07574533150766247,
             "positions": 147384,
-            "top1_agreement": 0.9153232372577756,
+            "top1_agreement": 0.9153707322368778,
             "worst_context_id": 275,
-            "worst_context_kld": 0.10122025808876788
+            "worst_context_kld": 0.10094674048285278
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.7339718429789305
+        "relative_to_run": 0.7391060828837337
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 5.807681560516357,
-            "mean_kld": 0.03723338342274077,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.03198886603632158,
-            "median_context_p99": 0.19292420148849487,
-            "p90_context_kld": 0.04507155972030506,
+            "max_kld": 6.983285427093506,
+            "mean_kld": 0.03711085502771101,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.031371119738575946,
+            "median_context_p99": 0.20159181952476501,
+            "p90_context_kld": 0.04505553168422091,
             "positions": 147384,
-            "top1_agreement": 0.9027506377897194,
+            "top1_agreement": 0.9036666123866905,
             "worst_context_id": 443,
-            "worst_context_kld": 0.21792143969470823
+            "worst_context_kld": 0.21911474635649905
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.7142476982508844
+        "relative_to_run": 0.7196837080859033
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.813560485839844,
-            "mean_kld": 0.030346047259348898,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.02084674157560274,
-            "median_context_p99": 0.23820368945598602,
-            "p90_context_kld": 0.04975873418086208,
+            "max_kld": 14.599754333496094,
+            "mean_kld": 0.03018829355699219,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.02100278716601154,
+            "median_context_p99": 0.21924170851707458,
+            "p90_context_kld": 0.05378769088686552,
             "positions": 196512,
-            "top1_agreement": 0.9495908646800195,
+            "top1_agreement": 0.9502422243934213,
             "worst_context_id": 667,
-            "worst_context_kld": 0.2161889008647869
+            "worst_context_kld": 0.20473772003888704
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.5821279833721604
+        "relative_to_run": 0.5854357985462448
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 7.356241226196289,
-            "mean_kld": 0.027620313995326606,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.02249949949724309,
-            "median_context_p99": 0.40589118003845215,
-            "p90_context_kld": 0.05043523494828777,
+            "max_kld": 8.024088859558105,
+            "mean_kld": 0.02775163699079394,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.02158035745589829,
+            "median_context_p99": 0.34725338220596313,
+            "p90_context_kld": 0.04604335776204541,
             "positions": 73692,
-            "top1_agreement": 0.8654535092004559,
+            "top1_agreement": 0.8652906692721055,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.11659044195651949
+            "worst_context_kld": 0.11848209727913012
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.5298402638337655
+        "relative_to_run": 0.5381821841634996
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 17.052858352661133,
-            "mean_kld": 0.024609437161871247,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.020860553521432415,
-            "median_context_p99": 0.14598329365253448,
-            "p90_context_kld": 0.03323913724905251,
+            "max_kld": 13.109975814819336,
+            "mean_kld": 0.024428160646162753,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.020779478378882214,
+            "median_context_p99": 0.1441325843334198,
+            "p90_context_kld": 0.03284644349083438,
             "positions": 196512,
-            "top1_agreement": 0.9371743201432992,
+            "top1_agreement": 0.9363652092493079,
             "worst_context_id": 825,
-            "worst_context_kld": 0.08214911499231742
+            "worst_context_kld": 0.08643491137050359
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.4720826374693844
+        "relative_to_run": 0.4737306435656357
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.320075988769531,
-            "mean_kld": 0.023228298859926257,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.02226252774408313,
-            "median_context_p99": 0.16904449462890625,
-            "p90_context_kld": 0.031335149218943904,
+            "max_kld": 13.07381534576416,
+            "mean_kld": 0.023027166196565346,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.021995164691474722,
+            "median_context_p99": 0.16406352818012238,
+            "p90_context_kld": 0.030565308802914134,
             "positions": 196512,
-            "top1_agreement": 0.9289712587526462,
+            "top1_agreement": 0.929612440970526,
             "worst_context_id": 246,
-            "worst_context_kld": 0.04042488780931219
+            "worst_context_kld": 0.044429711784661804
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.44558827240107735
+        "relative_to_run": 0.44656142637187546
       }
     ]
   },
   "overall": {
     "deployed": {
       "contexts": 768,
-      "max_kld": 27.48295783996582,
-      "mean_kld": 0.05212951125207867,
-      "mean_ref_top1_prob": 0.6415440866166368,
-      "median_context_kld": 0.03048191045903891,
-      "median_context_p99": 0.2586134076118469,
-      "p90_context_kld": 0.10388629113718155,
+      "max_kld": 36.404808044433594,
+      "mean_kld": 0.05156550663959363,
+      "mean_ref_top1_prob": 0.6415473983687051,
+      "median_context_kld": 0.030358945919737878,
+      "median_context_p99": 0.2575896382331848,
+      "p90_context_kld": 0.10094674048285278,
       "positions": 1572096,
-      "top1_agreement": 0.9151871132551702,
+      "top1_agreement": 0.9156470088340661,
       "worst_context_id": 454,
-      "worst_context_kld": 0.9351965383137365
+      "worst_context_kld": 0.8905369579241399
     }
   },
   "primary": "deployed"

Qwen3.8-27B-NVFP4-BF16-LMHead/strata.md

@@ -1,52 +1,52 @@
 # Fidelity by domain - Qwen3.8-27B / Qwen3.8-27B-NVFP4-BF16-LMHead
 
-768 contexts, 1572096 scored positions, mean 0.05212951, reference top-1 64.2%, top-1 agreement 91.5187%.
+768 contexts, 1572096 scored positions, mean 0.05156551, reference top-1 64.2%, top-1 agreement 91.5647%.
 
-`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
+`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. The `deployed` cell is QxQ; `bxq` is the teacher-ID counterfactual on the same student weights. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
 
 ## Ranked by domain
 
 | Domain | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.14506069 | 2.78 | 0.10672596 | 0.31745866 | 0.935197 |
-| Chinese across several content types | 72 | 55.6% | 0.08009739 | 1.54 | 0.04881424 | 0.14562944 | 0.334911 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.05208049 | 1.00 | 0.04317495 | 0.07479096 | 0.257566 |
-| Other multilingual content | 36 | 64.5% | 0.03909129 | 0.75 | 0.03482694 | 0.07035003 | 0.081913 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.03826159 | 0.73 | 0.03195989 | 0.07946953 | 0.101220 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.03723338 | 0.71 | 0.03198887 | 0.04507156 | 0.217921 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.03034605 | 0.58 | 0.02084674 | 0.04975873 | 0.216189 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.02762031 | 0.53 | 0.02249950 | 0.05043523 | 0.116590 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.02460944 | 0.47 | 0.02086055 | 0.03323914 | 0.082149 |
-| Scientific and technical exposition | 96 | 60.3% | 0.02322830 | 0.45 | 0.02226253 | 0.03133515 | 0.040425 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.14260814 | 2.77 | 0.10354972 | 0.34510198 | 0.890537 |
+| Chinese across several content types | 72 | 55.6% | 0.07907412 | 1.53 | 0.04855310 | 0.14487312 | 0.317172 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.05124551 | 0.99 | 0.04196326 | 0.07206767 | 0.253386 |
+| Other multilingual content | 36 | 64.5% | 0.03972507 | 0.77 | 0.03417868 | 0.06837732 | 0.087630 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.03811238 | 0.74 | 0.03164992 | 0.07574533 | 0.100947 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.03711086 | 0.72 | 0.03137112 | 0.04505553 | 0.219115 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.03018829 | 0.59 | 0.02100279 | 0.05378769 | 0.204738 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.02775164 | 0.54 | 0.02158036 | 0.04604336 | 0.118482 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.02442816 | 0.47 | 0.02077948 | 0.03284644 | 0.086435 |
+| Scientific and technical exposition | 96 | 60.3% | 0.02302717 | 0.45 | 0.02199516 | 0.03056531 | 0.044430 |
 
 ## Ranked by source dataset
 
 | Source | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| wildchat | 96 | 73.5% | 0.14506069 | 2.78 | 0.10672596 | 0.31745866 | 0.935197 |
-| wikisource_zh | 33 | 60.4% | 0.12936205 | 2.48 | 0.09606564 | 0.30460000 | 0.334911 |
-| wikipedia_de | 7 | 73.6% | 0.06534968 | 1.25 | 0.07035003 | 0.08191294 | 0.081913 |
-| wikipedia_en | 96 | 59.5% | 0.05208049 | 1.00 | 0.04317495 | 0.07479096 | 0.257566 |
-| regulations | 23 | 72.3% | 0.05162461 | 0.99 | 0.06074220 | 0.08217726 | 0.101220 |
-| wikipedia_zh | 39 | 51.5% | 0.03841191 | 0.74 | 0.03743391 | 0.05280832 | 0.059166 |
-| public_domain_books | 72 | 50.6% | 0.03723338 | 0.71 | 0.03198887 | 0.04507156 | 0.217921 |
-| wikipedia_ja | 7 | 57.0% | 0.03712700 | 0.71 | 0.03723823 | 0.04216571 | 0.042166 |
-| public_domain_review | 26 | 51.8% | 0.03648647 | 0.70 | 0.03274300 | 0.05933113 | 0.087844 |
-| wikipedia_es | 7 | 59.2% | 0.03460012 | 0.66 | 0.03333043 | 0.04504019 | 0.045040 |
-| wikipedia_cs | 6 | 68.9% | 0.03329140 | 0.64 | 0.03532597 | 0.03764456 | 0.037645 |
-| github_code | 52 | 85.8% | 0.03065874 | 0.59 | 0.01979219 | 0.05239252 | 0.216189 |
-| stackv2 | 44 | 72.0% | 0.02997650 | 0.58 | 0.02244658 | 0.04893623 | 0.209091 |
-| wikipedia_ru | 6 | 66.6% | 0.02863896 | 0.55 | 0.02861696 | 0.03984704 | 0.039847 |
-| starcoder_structured | 36 | 70.7% | 0.02762031 | 0.53 | 0.02249950 | 0.05043523 | 0.116590 |
-| open_news | 23 | 53.8% | 0.02690523 | 0.52 | 0.02633995 | 0.03581834 | 0.043336 |
-| wikipedia_fr | 3 | 60.2% | 0.02538882 | 0.49 | 0.02538246 | 0.02925488 | 0.029255 |
-| libretexts | 96 | 65.8% | 0.02460944 | 0.47 | 0.02086055 | 0.03323914 | 0.082149 |
-| scientific_papers | 96 | 60.3% | 0.02322830 | 0.45 | 0.02226253 | 0.03133515 | 0.040425 |
+| wildchat | 96 | 73.5% | 0.14260814 | 2.77 | 0.10354972 | 0.34510198 | 0.890537 |
+| wikisource_zh | 33 | 60.4% | 0.12707486 | 2.46 | 0.09748326 | 0.29125689 | 0.317172 |
+| wikipedia_de | 7 | 73.6% | 0.06826552 | 1.32 | 0.06837732 | 0.08763030 | 0.087630 |
+| regulations | 23 | 72.3% | 0.05175295 | 1.00 | 0.06760431 | 0.08414925 | 0.100947 |
+| wikipedia_en | 96 | 59.5% | 0.05124551 | 0.99 | 0.04196326 | 0.07206767 | 0.253386 |
+| wikipedia_zh | 39 | 51.5% | 0.03845811 | 0.75 | 0.03694213 | 0.05189267 | 0.059474 |
+| public_domain_books | 72 | 50.6% | 0.03711086 | 0.72 | 0.03137112 | 0.04505553 | 0.219115 |
+| wikipedia_ja | 7 | 57.0% | 0.03683809 | 0.71 | 0.03802289 | 0.04128210 | 0.041282 |
+| public_domain_review | 26 | 51.8% | 0.03606162 | 0.70 | 0.03299574 | 0.06031772 | 0.084851 |
+| wikipedia_es | 7 | 59.2% | 0.03428308 | 0.66 | 0.03246859 | 0.04459575 | 0.044596 |
+| wikipedia_cs | 6 | 68.9% | 0.03395949 | 0.66 | 0.03417868 | 0.04104416 | 0.041044 |
+| github_code | 52 | 85.8% | 0.03049262 | 0.59 | 0.01938650 | 0.05378769 | 0.204738 |
+| stackv2 | 44 | 72.0% | 0.02982864 | 0.58 | 0.02299397 | 0.05378093 | 0.188805 |
+| wikipedia_ru | 6 | 66.6% | 0.02891736 | 0.56 | 0.03018204 | 0.03903039 | 0.039030 |
+| starcoder_structured | 36 | 70.7% | 0.02775164 | 0.54 | 0.02158036 | 0.04604336 | 0.118482 |
+| open_news | 23 | 53.8% | 0.02679006 | 0.52 | 0.02628155 | 0.03704089 | 0.038851 |
+| wikipedia_fr | 3 | 60.2% | 0.02571151 | 0.50 | 0.02656140 | 0.02939078 | 0.029391 |
+| libretexts | 96 | 65.8% | 0.02442816 | 0.47 | 0.02077948 | 0.03284644 | 0.086435 |
+| scientific_papers | 96 | 60.3% | 0.02302717 | 0.45 | 0.02199516 | 0.03056531 | 0.044430 |
 
 ## Reading
 
-- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.14506069, 2.78x the run mean of 0.05212951 over 96 context(s).
-- Strongest domain: **Scientific and technical exposition** at 0.02322830, 0.45x the run mean. The spread across domains is 6.2x.
+- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.14260814, 2.77x the run mean of 0.05156551 over 96 context(s).
+- Strongest domain: **Scientific and technical exposition** at 0.02302717, 0.45x the run mean. The spread across domains is 6.2x.
 - This is a strong finding: the reference was *more* certain on Natural dialogue, instruction following, and assistance (top-1 73.5% against 64.2% overall) and the candidate still diverged most there, so predictability does not explain it.
-- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 0.935197 against a median of 0.10672596 (context 454). Read the documents before treating the domain as weak.
+- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 0.890537 against a median of 0.10354972 (context 454). Read the documents before treating the domain as weak.
 

Qwen3.8-27B-NVFP4-RTX5090/compliance.json

@@ -1,8 +1,9 @@
 {
   "attribution": null,
   "candidate": "/media/fmodels2/gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090",
-  "candidate_weights_sha256": null,
+  "candidate_weights_sha256": "380fc3ae60e1e6524cbfa31862cc686215120f7d7b93308f3356e5bf72c0e801",
   "comparability_key": {
+    "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
     "context_length": 2048,
     "driver": "580.173.02",
     "gpu_names": [
@@ -12,19 +13,22 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 248044,
-    "laws_version": 9,
+    "kv_cache_dtype": "bfloat16",
+    "laws_version": 15,
     "model_runner_v2": false,
+    "numerics_digest": "251a9225b37415b91fc58d975b0670fc4b481e6f5a816b663573c42e161438f8",
     "reference_config_sha256": "191e0af232104ed8b65258cf3fb2b842e288008baca7633c11b82a1ac7203aab",
     "rows": 768,
     "score_from": 0,
     "stride": 2048,
+    "substituted_parameters": [],
     "suite_id": "qwen3.8-27b-fidelity-1024x2048-v1",
     "tensor_parallel_size": 1,
     "token_sha256": "9c935708bbffbf45d5baa2931fb4af7e7b22df1e041aa0b6f4ca44d3315e543b",
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T12:56:04.368291+00:00",
+  "evaluated_at": "2026-09-11T20:24:29.719493+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -52,13 +56,13 @@
       "title": "Real vocabulary"
     },
     {
-      "detail": "all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc",
+      "detail": "all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3",
       "law": 5,
       "status": "pass",
       "title": "Manifest binding"
     },
     {
-      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s)",
+      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s)",
       "law": 6,
       "status": "pass",
       "title": "Provenance"
@@ -70,13 +74,13 @@
       "title": "Storage integrity"
     },
     {
-      "detail": "trunk 0.07603511, deployed 0.08884902, delta 0.012813907739867667",
+      "detail": "trunk 0.07615896, deployed 0.08890054, delta 0.012741578000413284",
       "law": 8,
       "status": "pass",
       "title": "Head transparency"
     },
     {
-      "detail": "mean 0.08884902, median 0.03153922, max 29.62161064, 4 depth buckets",
+      "detail": "mean 0.08890054, median 0.03146119, max 29.41000557, 4 depth buckets",
       "law": 9,
       "status": "pass",
       "title": "Tail and depth disclosure"
@@ -103,38 +107,37 @@
       "detail": "reference declares no experts",
       "law": 14,
       "status": "not_applicable",
-      "title": "Component attribution"
+      "title": "Routed-model intervention"
     },
     {
-      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.21268638, strongest worked_math_reasoning at 0.04891121, spread 4.3x",
+      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.21324364, strongest worked_math_reasoning at 0.04852616, spread 4.4x",
       "law": 15,
       "status": "pass",
       "title": "Domain disclosure"
     },
     {
-      "approval": {
-        "approver": "Andy Kitzke",
-        "justification": "These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger.",
-        "timestamp": "2026-09-02T05:15:00Z"
-      },
-      "detail": "the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else",
+      "detail": "scored weights 380fc3ae60e1e652 as inspected",
       "law": 16,
-      "status": "override",
+      "status": "pass",
       "title": "Candidate weight binding"
     },
     {
-      "detail": "1 override(s), each fully attributed",
+      "detail": "every scored layer used the checkpoint's own quantization parameters",
+      "law": 17,
+      "status": "pass",
+      "title": "Substitution disclosure"
+    },
+    {
+      "detail": "no overrides claimed",
       "law": 13,
       "status": "pass",
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 9,
-  "mean_kld": 0.08884901732903162,
+  "laws_version": 15,
+  "mean_kld": 0.08890053681422792,
   "nondeterminism_floor": 0.0,
-  "overridden_laws": [
-    16
-  ],
+  "overridden_laws": [],
   "partition": "analysis",
   "program": "Local Inference Lab \u2014 Distribution Fidelity",
   "ranking_floor": null,
@@ -147,16 +150,16 @@
     "overall": {
       "deployed": {
         "contexts": 768,
-        "max_kld": 29.621610641479492,
-        "mean_kld": 0.08884901732903162,
-        "mean_ref_top1_prob": 0.6415440866166368,
-        "median_context_kld": 0.06137484148042411,
-        "median_context_p99": 0.41598302125930786,
-        "p90_context_kld": 0.16160134971982643,
+        "max_kld": 29.410005569458008,
+        "mean_kld": 0.08890053681422791,
+        "mean_ref_top1_prob": 0.6415473983687051,
+        "median_context_kld": 0.06094675094259141,
+        "median_context_p99": 0.4184585213661194,
+        "p90_context_kld": 0.169365564213705,
         "positions": 1572096,
-        "top1_agreement": 0.879409399934864,
+        "top1_agreement": 0.8795868700130272,
         "worst_context_id": 454,
-        "worst_context_kld": 1.254965819110839
+        "worst_context_kld": 1.2597357640464442
       }
     },
     "primary": "deployed",
@@ -165,201 +168,201 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 29.621610641479492,
-            "mean_kld": 0.21268638446455335,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.16980764884602795,
-            "median_context_p99": 3.194436550140381,
-            "p90_context_kld": 0.4663941209945985,
+            "max_kld": 29.410005569458008,
+            "mean_kld": 0.21324363685885994,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.17013793919462264,
+            "median_context_p99": 3.1346852779388428,
+            "p90_context_kld": 0.4793873904592046,
             "positions": 196512,
-            "top1_agreement": 0.8831114639309559,
+            "top1_agreement": 0.8833353688324377,
             "worst_context_id": 454,
-            "worst_context_kld": 1.254965819110839
+            "worst_context_kld": 1.2597357640464442
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 2.3937955743159085
+        "relative_to_run": 2.398676594096019
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 9.221136093139648,
-            "mean_kld": 0.13804625823415112,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.0941144168045219,
-            "median_context_p99": 0.6546816229820251,
-            "p90_context_kld": 0.23829632700468703,
+            "max_kld": 8.557491302490234,
+            "mean_kld": 0.13763110758662334,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.09366887791148955,
+            "median_context_p99": 0.6111054420471191,
+            "p90_context_kld": 0.24518771943300743,
             "positions": 147384,
-            "top1_agreement": 0.826616186288878,
-            "worst_context_id": 905,
-            "worst_context_kld": 0.48569205838970464
+            "top1_agreement": 0.8276067958530098,
+            "worst_context_id": 893,
+            "worst_context_kld": 0.4890878310734861
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.5537173328876446
+        "relative_to_run": 1.5481470924548617
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 9.237215042114258,
-            "mean_kld": 0.08939194390567347,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.07517420846258792,
-            "median_context_p99": 0.6442986130714417,
-            "p90_context_kld": 0.12549267043511111,
+            "max_kld": 8.673469543457031,
+            "mean_kld": 0.08926161203012813,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.07524436844303807,
+            "median_context_p99": 0.6247144937515259,
+            "p90_context_kld": 0.12359177821272875,
             "positions": 196512,
-            "top1_agreement": 0.8639675948542582,
+            "top1_agreement": 0.8644510258915487,
             "worst_context_id": 6,
-            "worst_context_kld": 0.37588815477459775
+            "worst_context_kld": 0.38041751277782426
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 1.006110664956836
+        "relative_to_run": 1.0040615639549482
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.085050582885742,
-            "mean_kld": 0.07346640922786896,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.0650303690739286,
-            "median_context_p99": 0.35447409749031067,
-            "p90_context_kld": 0.08656176264974488,
+            "max_kld": 7.7061967849731445,
+            "mean_kld": 0.0738154133464014,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.06468236685300296,
+            "median_context_p99": 0.3413587808609009,
+            "p90_context_kld": 0.08678552501587226,
             "positions": 147384,
-            "top1_agreement": 0.8569519079411605,
+            "top1_agreement": 0.8566194430874451,
             "worst_context_id": 443,
-            "worst_context_kld": 0.3243600094405916
+            "worst_context_kld": 0.3221895584307333
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.8268679996291151
+        "relative_to_run": 0.8303145964197121
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.182098388671875,
-            "mean_kld": 0.07235635445357978,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.06563072477178752,
-            "median_context_p99": 0.4368785619735718,
-            "p90_context_kld": 0.1307554214582166,
+            "max_kld": 7.552076816558838,
+            "mean_kld": 0.07310515835266779,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.06447523888587277,
+            "median_context_p99": 0.4327349364757538,
+            "p90_context_kld": 0.13155629979250485,
             "positions": 147384,
-            "top1_agreement": 0.8739211854746783,
+            "top1_agreement": 0.8737515605493134,
             "worst_context_id": 275,
-            "worst_context_kld": 0.15489253509157094
+            "worst_context_kld": 0.1603293014694022
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.8143742792970337
+        "relative_to_run": 0.8223252746542226
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 7.345405578613281,
-            "mean_kld": 0.07083341233029045,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.06525897825527403,
-            "median_context_p99": 0.44970738887786865,
-            "p90_context_kld": 0.11245465481627331,
+            "max_kld": 11.588532447814941,
+            "mean_kld": 0.07121532314429206,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.06519233793250864,
+            "median_context_p99": 0.46661657094955444,
+            "p90_context_kld": 0.1105094893175383,
             "positions": 73692,
-            "top1_agreement": 0.8756852846984747,
+            "top1_agreement": 0.8746403951582261,
             "worst_context_id": 953,
-            "worst_context_kld": 0.11817462742460928
+            "worst_context_kld": 0.12065877383606818
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.7972334918232739
+        "relative_to_run": 0.801067414172178
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 11.309407234191895,
-            "mean_kld": 0.05240112493687464,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.0419624148682241,
-            "median_context_p99": 0.3964604139328003,
-            "p90_context_kld": 0.07997063031452036,
+            "max_kld": 14.120194435119629,
+            "mean_kld": 0.05264429081489764,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.040841334767735375,
+            "median_context_p99": 0.4023333787918091,
+            "p90_context_kld": 0.08157172921411385,
             "positions": 196512,
-            "top1_agreement": 0.9284572952287901,
+            "top1_agreement": 0.9283351652825272,
             "worst_context_id": 667,
-            "worst_context_kld": 0.27689472241669605
+            "worst_context_kld": 0.29321267733423706
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.5897772030817099
+        "relative_to_run": 0.5921706741198473
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 16.185749053955078,
-            "mean_kld": 0.04949621323585394,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.048019773744291346,
-            "median_context_p99": 0.30111250281333923,
-            "p90_context_kld": 0.06206171818080255,
+            "max_kld": 11.157588958740234,
+            "mean_kld": 0.049079914368717295,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.048805704917541055,
+            "median_context_p99": 0.312124103307724,
+            "p90_context_kld": 0.05848004086686995,
             "positions": 196512,
-            "top1_agreement": 0.8900728708679368,
-            "worst_context_id": 246,
-            "worst_context_kld": 0.07372579095666683
+            "top1_agreement": 0.8908056505455137,
+            "worst_context_id": 174,
+            "worst_context_kld": 0.07139858353297991
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.5570822809728582
+        "relative_to_run": 0.5520766929819303
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 8.847333908081055,
-            "mean_kld": 0.0491758979962088,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.04248309454994046,
-            "median_context_p99": 0.494755357503891,
-            "p90_context_kld": 0.08263404999373626,
+            "max_kld": 10.949919700622559,
+            "mean_kld": 0.04887779489384767,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.04098168850353924,
+            "median_context_p99": 0.5531740188598633,
+            "p90_context_kld": 0.07973034153401946,
             "positions": 73692,
-            "top1_agreement": 0.8469304673505944,
+            "top1_agreement": 0.8472968571893829,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.1645537792651185
+            "worst_context_kld": 0.17824096230327108
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.553477117412535
+        "relative_to_run": 0.5498031468132274
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 17.194229125976562,
-            "mean_kld": 0.04891121428016043,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.04473202064081297,
-            "median_context_p99": 0.26329120993614197,
-            "p90_context_kld": 0.05904413538005142,
+            "max_kld": 17.83475685119629,
+            "mean_kld": 0.04852616171264854,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.04458567627830201,
+            "median_context_p99": 0.2575961947441101,
+            "p90_context_kld": 0.05704854838084238,
             "positions": 196512,
-            "top1_agreement": 0.9055681078000326,
+            "top1_agreement": 0.9055579303045107,
             "worst_context_id": 825,
-            "worst_context_kld": 0.12918918401151672
+            "worst_context_kld": 0.1312607970175872
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.5504980893489136
+        "relative_to_run": 0.5458477918311317
       }
     ]
   },

Qwen3.8-27B-NVFP4-RTX5090/inspect.json

@@ -1,31 +1,41 @@
 {
+  "model": "/media/fmodels2/gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090",
+  "inspect_version": 5,
+  "weights_sha256": "380fc3ae60e1e6524cbfa31862cc686215120f7d7b93308f3356e5bf72c0e801",
+  "weights_bytes": 18765513016,
+  "quant_method": "modelopt",
+  "declared": {},
+  "detected_scheme": "nvfp4",
+  "detected_block": 16,
+  "quant_algorithm": "modelopt",
+  "scheme_mix": null,
+  "kv_cache_scheme": {
+    "num_bits": 8,
+    "type": "float",
+    "strategy": null
+  },
   "coverage": {
-    "attention": {
-      "quantized": 64,
-      "weights": 68
-    },
-    "dense_mlp": {
-      "quantized": 192,
-      "weights": 195
+    "router": {
+      "weights": 0,
+      "quantized": 0
     },
     "experts": {
-      "quantized": 0,
-      "weights": 0
-    },
-    "router": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
     },
     "shared_expert": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
+    },
+    "attention": {
+      "weights": 68,
+      "quantized": 64
+    },
+    "dense_mlp": {
+      "weights": 195,
+      "quantized": 192
     }
   },
-  "declared": {},
-  "detected_block": null,
-  "detected_scheme": "fp8_per_tensor",
-  "model": "/media/fmodels2/gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090",
-  "quant_method": "modelopt",
-  "weights_bytes": 18765513016,
-  "weights_bytes_source": "hub"
+  "unloadable_reason": null,
+  "quantized_names_sha256": "eccbe240ef5caa18a896c8a311d4a9b4257c68423f0d8f9ae1c982d53797e657"
 }

Qwen3.8-27B-NVFP4-RTX5090/report.md

@@ -1,12 +1,12 @@
 # Qwen3.8-27B / Qwen3.8-27B-NVFP4-RTX5090: distribution fidelity
 
-**Mean KLD(reference || candidate) = 0.08884902** over 1572096 scored positions.
+**Mean KLD(reference || candidate) = 0.08890054** over 1572096 scored positions.
 
 | Mean | Median | p90 | p99 | Max | Top-1 agreement |
 |---|---|---|---|---|---|
-| 0.08884902 | 0.03153922 | 0.14679415 | 1.04055545 | 29.62161064 | 87.9409% |
+| 0.08890054 | 0.03146119 | 0.14654309 | 1.03599295 | 29.41000557 | 87.9587% |
 
-Reverse direction, KLD(candidate || reference): 0.09722847.
+Reverse direction, KLD(candidate || reference): 0.09720515.
 
 ## Identity
 
@@ -15,10 +15,10 @@ Reverse direction, KLD(candidate || reference): 0.09722847.
 | Reference checkpoint | /media/fmodels2/Qwen/Qwen3.8-27B |
 | Reference config SHA-256 | 191e0af23210 |
 | Candidate checkpoint | /media/fmodels2/gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 |
-| Candidate weights SHA-256 | unbound (Law 16 gap) |
+| Candidate weights SHA-256 | 380fc3ae60e1e652 |
 | Suite | qwen3.8-27b-fidelity-1024x2048-v1 |
 | Suite token SHA-256 | 9c935708bbffbf45 |
-| Capture manifest SHA-256 | 304794b4e35b326d |
+| Capture manifest SHA-256 | dbfacc9cbb6d3e06 |
 | Tokenizer | None |
 | Scored vocabulary | 248044 |
 | Declared vocabulary | 248320 |
@@ -27,42 +27,58 @@ Reverse direction, KLD(candidate || reference): 0.09722847.
 | Model runner | V1 |
 | Tensor parallel | 1 |
 | Eager enforced | True |
+| KV cache | bfloat16 |
 | Prefix caching | False |
 | max_num_seqs | 1 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45 |
+| vLLM commit | 60071d1ab732 |
+| vLLM dirty digest | e3b0c44298fc1c14 |
+| Numerics digest | 251a9225b37415b9 |
+| Compiled extensions | f2fbc7537b0f01f6 |
+| FlashInfer | 0.6.17 |
 | torch | 2.13.0+cu132 |
 | Driver | 580.173.02 |
 | GPUs | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
-| Laws version | 9 |
+| Laws version | 15 |
 | Partition | analysis |
 
+The commit says when this number was taken. The numerics digest says what took it: a hash of the runtime and the scorer, which moves only when code that can change a logit changes. Comparability is bounded by the digest, so a result stays current across commits that cannot reach a number, and two results carrying the same digest were computed by the same code whatever their commits say.
+
+## Expert kernels: declared against built
+
+| Property | Value |
+|---|---|
+| Declared for its experts | `4-bit float` |
+| Expert implementation built | n/a |
+| Expert kernel built | n/a |
+| Expert layers carrying an activation scale | n/a |
+
 ## Fidelity by domain
 
-The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
+The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `deployed` is QxQ (natural student routing). `bxq` is the teacher-ID counterfactual on the same weights. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
 
 | Domain | Contexts | Reference top-1 | Mean KLD | x run | Worst context |
 |---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.21268638 | 2.39 | 1.25496582 |
-| Chinese across several content types | 72 | 55.6% | 0.13804626 | 1.55 | 0.48569206 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.08939194 | 1.01 | 0.37588815 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.07346641 | 0.83 | 0.32436001 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.07235635 | 0.81 | 0.15489254 |
-| Other multilingual content | 36 | 64.5% | 0.07083341 | 0.80 | 0.11817463 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.05240112 | 0.59 | 0.27689472 |
-| Scientific and technical exposition | 96 | 60.3% | 0.04949621 | 0.56 | 0.07372579 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.04917590 | 0.55 | 0.16455378 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.04891121 | 0.55 | 0.12918918 |
-
-**Natural dialogue, instruction following, and assistance** is this candidate's weakest domain and **Worked mathematics, science, and formal reasoning** its strongest, a spread of 4.3x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.21324364 | 2.40 | 1.25973576 |
+| Chinese across several content types | 72 | 55.6% | 0.13763111 | 1.55 | 0.48908783 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.08926161 | 1.00 | 0.38041751 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.07381541 | 0.83 | 0.32218956 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.07310516 | 0.82 | 0.16032930 |
+| Other multilingual content | 36 | 64.5% | 0.07121532 | 0.80 | 0.12065877 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.05264429 | 0.59 | 0.29321268 |
+| Scientific and technical exposition | 96 | 60.3% | 0.04907991 | 0.55 | 0.07139858 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.04887779 | 0.55 | 0.17824096 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.04852616 | 0.55 | 0.13126080 |
+
+**Natural dialogue, instruction following, and assistance** is this candidate's weakest domain and **Worked mathematics, science, and formal reasoning** its strongest, a spread of 4.4x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
 
 ## Trunk versus head
 
 | Component | Value |
 |---|---|
-| Trunk (candidate hidden states, reference head) | 0.07603511 |
-| Deployed (candidate's own head) | 0.08884902 |
-| Head-associated delta (not additive) | 0.01281391 |
+| Trunk (candidate hidden states, reference head) | 0.07615896 |
+| Deployed (candidate's own head) | 0.08890054 |
+| Head-associated delta (not additive) | 0.01274158 |
 | Reference head | {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]} |
 | Candidate head | {'runtime': {'state': 'quantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'ModelOptNvFp4LinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.uint8'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'ModelOptNvFp4LinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.uint8'}, 'state': 'quantized'}]}, 'state': 'quantized', 'static': {'ignored': False, 'lm_head_dtypes': {'lm_head.input_scale': 'F32', 'lm_head.weight': 'U8', 'lm_head.weight_scale': 'F8_E4M3', 'lm_head.weight_scale_2': 'F32', 'model.language_model.embed_tokens.weight': 'BF16'}, 'lm_head_keys': ['lm_head.input_scale', 'lm_head.weight', 'lm_head.weight_scale', 'lm_head.weight_scale_2'], 'output_weight_keys': ['lm_head.input_scale', 'lm_head.weight', 'lm_head.weight_scale', 'lm_head.weight_scale_2'], 'packed_keys': ['lm_head.weight_scale', 'lm_head.weight_scale_2'], 'quant_method': 'modelopt', 'state': 'quantized', 'tie_word_embeddings': False}} |
 
@@ -70,26 +86,26 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Position range | Positions | Mean KLD |
 |---|---|---|
-| 0–511 | 393216 | 0.07970075 |
-| 512–1023 | 393216 | 0.08368177 |
-| 1024–1535 | 393216 | 0.09329681 |
-| 1536–2046 | 392448 | 0.09873604 |
+| 0–511 | 393216 | 0.07967133 |
+| 512–1023 | 393216 | 0.08388078 |
+| 1024–1535 | 393216 | 0.09256314 |
+| 1536–2046 | 392448 | 0.09950761 |
 
 ## Error by reference confidence
 
 | Reference top-1 probability | Positions | Share | Mean KLD |
 |---|---|---|---|
-| [0.00, 0.25) | 240569 | 15.3% | 0.11477444 |
-| [0.25, 0.50) | 346390 | 22.0% | 0.12619287 |
-| [0.50, 0.75) | 273384 | 17.4% | 0.12732539 |
-| [0.75, 0.95) | 249939 | 15.9% | 0.09405716 |
-| [0.95, 1.00) | 461814 | 29.4% | 0.02173773 |
+| [0.00, 0.25) | 240522 | 15.3% | 0.11444036 |
+| [0.25, 0.50) | 346508 | 22.0% | 0.12688756 |
+| [0.50, 0.75) | 273204 | 17.4% | 0.12535953 |
+| [0.75, 0.95) | 250041 | 15.9% | 0.09468442 |
+| [0.95, 1.00) | 461821 | 29.4% | 0.02239718 |
 
 ## Top-K set agreement
 
 | K=1 | K=2 | K=3 | K=4 | K=5 |
 |---|---|---|---|---|
-| 87.7338% | 64.6300% | 41.9884% | 24.6266% | 13.5939% |
+| 87.7500% | 64.7214% | 42.0533% | 24.6980% | 13.6359% |
 
 ## Law compliance
 
@@ -101,22 +117,19 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | 2 | Determinism | PASS | eager enforced; prefix caching off, max_num_seqs=1 |
 | 3 | Frozen input | PASS | token hash matches suite qwen3.8-27b-fidelity-1024x2048-v1 [analysis]: 9c935708bbffbf45 |
 | 4 | Real vocabulary | PASS | scored 248044 real tokens of 248320 declared (276 padding rows) |
-| 5 | Manifest binding | PASS | all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc |
-| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s) |
+| 5 | Manifest binding | PASS | all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3 |
+| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s) |
 | 7 | Storage integrity | PASS | hidden storage with bitwise-exact replay |
-| 8 | Head transparency | PASS | trunk 0.07603511, deployed 0.08884902, delta 0.012813907739867667 |
-| 9 | Tail and depth disclosure | PASS | mean 0.08884902, median 0.03153922, max 29.62161064, 4 depth buckets |
+| 8 | Head transparency | PASS | trunk 0.07615896, deployed 0.08890054, delta 0.012741578000413284 |
+| 9 | Tail and depth disclosure | PASS | mean 0.08890054, median 0.03146119, max 29.41000557, 4 depth buckets |
 | 10 | Comparability | PASS | comparability key fully resolved |
 | 11 | Freeze before qualification | NOT_APPLICABLE | partition is 'analysis' |
 | 12 | Reusable reference | PASS | suite, reference, head, and checksums present; published reference matches the scored capture on every bound field |
-| 14 | Component attribution | NOT_APPLICABLE | reference declares no experts |
-| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.21268638, strongest worked_math_reasoning at 0.04891121, spread 4.3x |
-| 16 | Candidate weight binding | OVERRIDE | the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else |
-| 13 | Recorded deviation | PASS | 1 override(s), each fully attributed |
-
-### Recorded deviations
-
-- **Law 16 (Candidate weight binding)** overridden by Andy Kitzke at 2026-09-02T05:15:00Z. Justification: These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger. Underlying finding: the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else.
+| 14 | Routed-model intervention | NOT_APPLICABLE | reference declares no experts |
+| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.21324364, strongest worked_math_reasoning at 0.04852616, spread 4.4x |
+| 16 | Candidate weight binding | PASS | scored weights 380fc3ae60e1e652 as inspected |
+| 17 | Substitution disclosure | PASS | every scored layer used the checkpoint's own quantization parameters |
+| 13 | Recorded deviation | PASS | no overrides claimed |
 
 ## Environment
 
@@ -126,7 +139,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | Platform | Linux-6.8.0-137-generic-x86_64-with-glibc2.39 |
 | Python | 3.12.3 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45fbe0b0205e70893bcd28fa928a88 |
+| vLLM commit | 60071d1ab73229321712254b1350dcaaac1c125a |
 | torch | 2.13.0+cu132 |
 | torch CUDA runtime | 13.2 |
 | cuDNN | 9.20.0 (92000) |
@@ -134,7 +147,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | CUDA arch list | sm_75, sm_80, sm_86, sm_90, sm_100, sm_120 |
 | nvcc | Cuda compilation tools, release 13.0, V13.0.88 |
 | gcc | gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0 |
-| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.8) 2.39 |
+| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.9) 2.39 |
 | NVIDIA driver | 580.173.02 |
 | float32 matmul precision | highest |
 | TF32 (matmul / cuDNN) | False / True |
@@ -157,8 +170,18 @@ Credential values are never published. Set but redacted: `HF_TOKEN`.
 
 | Variable | Value |
 |---|---|
+| `CUBLAS_WORKSPACE_CONFIG` | `:4096:8` |
 | `CUDA_HOME` | `/usr/local/cuda-13.0` |
 | `HF_TOKEN` | `<redacted>` |
+| `NCCL_DETERMINISTIC` | `1` |
+| `PYTORCH_NVML_BASED_CUDA_CHECK` | `1` |
+| `TORCHINDUCTOR_CACHE_DIR` | `/tmp/torchinductor_phaedawg` |
+| `TORCHINDUCTOR_COMPILE_THREADS` | `1` |
+| `TRITON_CACHE_AUTOTUNING` | `1` |
+| `TRITON_PTXAS_BLACKWELL_PATH` | `/usr/local/cuda-13.0/bin/ptxas` |
+| `VLLM_BATCH_INVARIANT` | `1` |
+| `VLLM_MARLIN_USE_ATOMIC_ADD` | `0` |
+| `VLLM_MOE_USE_DEEP_GEMM` | `0` |
 
 ## Files in this artifact
 
@@ -166,26 +189,26 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 
 | Path | Size | What it is |
 |---|---|---|
-| `Qwen3.8-27B-NVFP4-RTX5090/report.md` | 12.65 KiB | This document. |
-| `Qwen3.8-27B-NVFP4-RTX5090/report.json` | 253.93 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
-| `Qwen3.8-27B-NVFP4-RTX5090/manifest.json` | 85.43 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
-| `Qwen3.8-27B-NVFP4-RTX5090/compliance.json` | 13.02 KiB | The law-by-law receipt, including the comparability key. |
-| `baselines/self-kld.json` | 4.74 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
+| `Qwen3.8-27B-NVFP4-RTX5090/report.md` | 13.49 KiB | This document. |
+| `Qwen3.8-27B-NVFP4-RTX5090/report.json` | 256.45 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
+| `Qwen3.8-27B-NVFP4-RTX5090/manifest.json` | 86.99 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
+| `Qwen3.8-27B-NVFP4-RTX5090/compliance.json` | 12.29 KiB | The law-by-law receipt, including the comparability key. |
+| `baselines/self-kld.json` | 7.06 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
 | `suite/suite-manifest.json` | 382.68 KiB | The frozen evaluation input's identity: token hashes per context and per partition, sources, strata, and the analysis/qualification split. |
 | `suite/tokens` | 15.49 MiB in 1024 files | The token IDs themselves. These are the evaluation input, not a description of it; retokenizing source text does not reproduce them. |
 | `suite/sources.json` | 737.42 KiB | Per-context provenance: dataset, revision, licence, source unit, and the deterministic token offset chosen within the document. |
 | `suite/validation/capability-overlap.json` | 1.48 KiB | The benchmark-contamination scan and every document it blocked. |
-| `reference/manifest.json` | 85.43 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
+| `reference/manifest.json` | 86.99 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
 | `reference` | 19.25 GiB in 771 files | The reusable reference distributions. Pass this directory as `--reference-logits` to score a new candidate against the same reference without loading the reference checkpoint. |
 | `reference/lm_head.safetensors` | 2.37 GiB | The reference language-model head, which turns the stored hidden states back into reference logits. |
-| `environment/runtime.json` | 2.26 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
+| `environment/runtime.json` | 5.42 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
 | `environment/summary.md` | 1.18 KiB | The same provenance as prose, plus an index of every captured file. |
 | `environment/toolchain-nvcc.txt` | 243 B | `nvcc --version` verbatim; `toolchain-gcc.txt` and `toolchain-ldd.txt` sit beside it. |
 | `environment/gpu-smi-query.txt` | 54.73 KiB | `nvidia-smi -q` verbatim: ECC state, persistence mode, clocks, and throttle reasons, any of which can move a bitwise result. |
-| `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
-| `environment/models` | 7.22 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
-| `checksums.txt` | 182.99 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
+| `environment/pip-freeze.txt` | 4.47 KiB | Every installed package version in the scoring environment. |
+| `environment/models` | 19.46 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
+| `checksums.txt` | 183.22 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
+| `LAWS.md` | 40.20 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

Qwen3.8-27B-NVFP4-RTX5090/strata.json

@@ -8,381 +8,381 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 29.621610641479492,
-            "mean_kld": 0.21268638446455335,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.16980764884602795,
-            "median_context_p99": 3.194436550140381,
-            "p90_context_kld": 0.4663941209945985,
+            "max_kld": 29.410005569458008,
+            "mean_kld": 0.21324363685885994,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.17013793919462264,
+            "median_context_p99": 3.1346852779388428,
+            "p90_context_kld": 0.4793873904592046,
             "positions": 196512,
-            "top1_agreement": 0.8831114639309559,
+            "top1_agreement": 0.8833353688324377,
             "worst_context_id": 454,
-            "worst_context_kld": 1.254965819110839
+            "worst_context_kld": 1.2597357640464442
           }
         },
         "key": "wildchat",
         "label": "wildchat",
-        "relative_to_run": 2.3937955743159085
+        "relative_to_run": 2.398676594096019
       },
       {
         "cells": {
           "deployed": {
             "contexts": 33,
-            "max_kld": 9.221136093139648,
-            "mean_kld": 0.2106804432508119,
-            "mean_ref_top1_prob": 0.6040606719329803,
-            "median_context_kld": 0.16749470140762515,
-            "median_context_p99": 1.1817240715026855,
-            "p90_context_kld": 0.4601693628589735,
+            "max_kld": 8.557491302490234,
+            "mean_kld": 0.20942829459093157,
+            "mean_ref_top1_prob": 0.6040861921682756,
+            "median_context_kld": 0.169365564213705,
+            "median_context_p99": 1.2919573783874512,
+            "p90_context_kld": 0.4482288233449379,
             "positions": 67551,
-            "top1_agreement": 0.7997068881289692,
-            "worst_context_id": 905,
-            "worst_context_kld": 0.48569205838970464
+            "top1_agreement": 0.8017497890482747,
+            "worst_context_id": 893,
+            "worst_context_kld": 0.4890878310734861
           }
         },
         "key": "wikisource_zh",
         "label": "wikisource_zh",
-        "relative_to_run": 2.37121860864939
+        "relative_to_run": 2.355759617386405
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 7.206441402435303,
-            "mean_kld": 0.10192228244303433,
-            "mean_ref_top1_prob": 0.7357366218042676,
-            "median_context_kld": 0.11245465481627331,
-            "median_context_p99": 1.8527274131774902,
-            "p90_context_kld": 0.11817462742460928,
+            "max_kld": 11.588532447814941,
+            "mean_kld": 0.1031817371159509,
+            "mean_ref_top1_prob": 0.7358200961573914,
+            "median_context_kld": 0.1105094893175383,
+            "median_context_p99": 1.737134575843811,
+            "p90_context_kld": 0.12065877383606818,
             "positions": 14329,
-            "top1_agreement": 0.8715890850722311,
+            "top1_agreement": 0.868727754902645,
             "worst_context_id": 953,
-            "worst_context_kld": 0.11817462742460928
+            "worst_context_kld": 0.12065877383606818
           }
         },
         "key": "wikipedia_de",
         "label": "wikipedia_de",
-        "relative_to_run": 1.1471402330268765
+        "relative_to_run": 1.1606424529422794
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 96,
-            "max_kld": 9.237215042114258,
-            "mean_kld": 0.08939194390567347,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.07517420846258792,
-            "median_context_p99": 0.6442986130714417,
-            "p90_context_kld": 0.12549267043511111,
-            "positions": 196512,
-            "top1_agreement": 0.8639675948542582,
-            "worst_context_id": 6,
-            "worst_context_kld": 0.37588815477459775
+            "contexts": 23,
+            "max_kld": 7.552076816558838,
+            "mean_kld": 0.09111800091462494,
+            "mean_ref_top1_prob": 0.7233191001139795,
+            "median_context_kld": 0.10723593518216033,
+            "median_context_p99": 1.4357880353927612,
+            "p90_context_kld": 0.14780949122751222,
+            "positions": 47081,
+            "top1_agreement": 0.8949682462139716,
+            "worst_context_id": 275,
+            "worst_context_kld": 0.1603293014694022
           }
         },
-        "key": "wikipedia_en",
-        "label": "wikipedia_en",
-        "relative_to_run": 1.006110664956836
+        "key": "regulations",
+        "label": "regulations",
+        "relative_to_run": 1.0249432025931495
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 23,
-            "max_kld": 7.488497734069824,
-            "mean_kld": 0.08888439895968081,
-            "mean_ref_top1_prob": 0.7232628548554616,
-            "median_context_kld": 0.10944161648564113,
-            "median_context_p99": 1.2938449382781982,
-            "p90_context_kld": 0.138309287310408,
-            "positions": 47081,
-            "top1_agreement": 0.8938212867186338,
-            "worst_context_id": 275,
-            "worst_context_kld": 0.15489253509157094
+            "contexts": 96,
+            "max_kld": 8.673469543457031,
+            "mean_kld": 0.08926161203012813,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.07524436844303807,
+            "median_context_p99": 0.6247144937515259,
+            "p90_context_kld": 0.12359177821272875,
+            "positions": 196512,
+            "top1_agreement": 0.8644510258915487,
+            "worst_context_id": 6,
+            "worst_context_kld": 0.38041751277782426
           }
         },
-        "key": "regulations",
-        "label": "regulations",
-        "relative_to_run": 1.000398221969278
+        "key": "wikipedia_en",
+        "label": "wikipedia_en",
+        "relative_to_run": 1.0040615639549482
       },
       {
         "cells": {
           "deployed": {
             "contexts": 39,
-            "max_kld": 6.928176403045654,
-            "mean_kld": 0.07658656322005357,
-            "mean_ref_top1_prob": 0.515358929042611,
-            "median_context_kld": 0.07334535190566364,
-            "median_context_p99": 0.4606652557849884,
-            "p90_context_kld": 0.10122490028374392,
+            "max_kld": 6.9802069664001465,
+            "mean_kld": 0.076879641659901,
+            "mean_ref_top1_prob": 0.5153760832971095,
+            "median_context_kld": 0.07373860768803792,
+            "median_context_p99": 0.46467477083206177,
+            "p90_context_kld": 0.10373904431787762,
             "positions": 79833,
-            "top1_agreement": 0.8493855924241855,
+            "top1_agreement": 0.8494858016108626,
             "worst_context_id": 890,
-            "worst_context_kld": 0.10990622245183433
+            "worst_context_kld": 0.10933838389260316
           }
         },
         "key": "wikipedia_zh",
         "label": "wikipedia_zh",
-        "relative_to_run": 0.8619854841661679
+        "relative_to_run": 0.8647826482820173
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.085050582885742,
-            "mean_kld": 0.07346640922786896,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.0650303690739286,
-            "median_context_p99": 0.35447409749031067,
-            "p90_context_kld": 0.08656176264974488,
+            "max_kld": 7.7061967849731445,
+            "mean_kld": 0.0738154133464014,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.06468236685300296,
+            "median_context_p99": 0.3413587808609009,
+            "p90_context_kld": 0.08678552501587226,
             "positions": 147384,
-            "top1_agreement": 0.8569519079411605,
+            "top1_agreement": 0.8566194430874451,
             "worst_context_id": 443,
-            "worst_context_kld": 0.3243600094405916
+            "worst_context_kld": 0.3221895584307333
           }
         },
         "key": "public_domain_books",
         "label": "public_domain_books",
-        "relative_to_run": 0.8268679996291151
+        "relative_to_run": 0.8303145964197121
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 3.718787431716919,
-            "mean_kld": 0.07299386253552488,
-            "mean_ref_top1_prob": 0.5702914711488575,
-            "median_context_kld": 0.07676595262747475,
-            "median_context_p99": 0.4214121401309967,
-            "p90_context_kld": 0.08001700776501085,
+            "max_kld": 3.6783010959625244,
+            "mean_kld": 0.07229253937155643,
+            "mean_ref_top1_prob": 0.5701150598950927,
+            "median_context_kld": 0.07461241188006175,
+            "median_context_p99": 0.43328213691711426,
+            "p90_context_kld": 0.08007905997068493,
             "positions": 14329,
-            "top1_agreement": 0.8627957289413078,
+            "top1_agreement": 0.8609114383418243,
             "worst_context_id": 950,
-            "worst_context_kld": 0.08001700776501085
+            "worst_context_kld": 0.08007905997068493
           }
         },
         "key": "wikipedia_ja",
         "label": "wikipedia_ja",
-        "relative_to_run": 0.821549463684096
+        "relative_to_run": 0.8131845089149845
       },
       {
         "cells": {
           "deployed": {
             "contexts": 26,
-            "max_kld": 8.182098388671875,
-            "mean_kld": 0.07178554333186508,
-            "mean_ref_top1_prob": 0.5182113024080622,
-            "median_context_kld": 0.06815590338023643,
-            "median_context_p99": 0.4050147831439972,
-            "p90_context_kld": 0.09782585941008454,
+            "max_kld": 4.8773674964904785,
+            "mean_kld": 0.07159772192369603,
+            "mean_ref_top1_prob": 0.5182496321630543,
+            "median_context_kld": 0.06766348511317534,
+            "median_context_p99": 0.3930010199546814,
+            "p90_context_kld": 0.104331674890945,
             "positions": 53222,
-            "top1_agreement": 0.8588177821201759,
+            "top1_agreement": 0.8578783209950772,
             "worst_context_id": 312,
-            "worst_context_kld": 0.1451335772405673
+            "worst_context_kld": 0.14655173578548103
           }
         },
         "key": "public_domain_review",
         "label": "public_domain_review",
-        "relative_to_run": 0.8079497724327559
+        "relative_to_run": 0.8053688367856662
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 4.2509942054748535,
-            "mean_kld": 0.06601945572480945,
-            "mean_ref_top1_prob": 0.591765751203999,
-            "median_context_kld": 0.06525897825527403,
-            "median_context_p99": 0.43164998292922974,
-            "p90_context_kld": 0.0802859432000463,
+            "max_kld": 6.276042461395264,
+            "mean_kld": 0.06658863195079154,
+            "mean_ref_top1_prob": 0.591761624645417,
+            "median_context_kld": 0.0636619725031808,
+            "median_context_p99": 0.4681042432785034,
+            "p90_context_kld": 0.07747506305285463,
             "positions": 14329,
-            "top1_agreement": 0.8619582664526485,
+            "top1_agreement": 0.862586363319143,
             "worst_context_id": 936,
-            "worst_context_kld": 0.0802859432000463
+            "worst_context_kld": 0.07747506305285463
           }
         },
         "key": "wikipedia_es",
         "label": "wikipedia_es",
-        "relative_to_run": 0.7430521767091895
+        "relative_to_run": 0.7490239579760838
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 7.345405578613281,
-            "mean_kld": 0.06272890207822825,
-            "mean_ref_top1_prob": 0.6892386900239509,
-            "median_context_kld": 0.06199259017115178,
-            "median_context_p99": 0.4519658088684082,
-            "p90_context_kld": 0.07546921805391646,
+            "max_kld": 7.353713035583496,
+            "mean_kld": 0.06374343751347408,
+            "mean_ref_top1_prob": 0.689239984531695,
+            "median_context_kld": 0.0644250611690942,
+            "median_context_p99": 0.5090126991271973,
+            "p90_context_kld": 0.07887764799120904,
             "positions": 12282,
-            "top1_agreement": 0.8907344080768604,
+            "top1_agreement": 0.888291809151604,
             "worst_context_id": 958,
-            "worst_context_kld": 0.07546921805391646
+            "worst_context_kld": 0.07887764799120904
           }
         },
         "key": "wikipedia_cs",
         "label": "wikipedia_cs",
-        "relative_to_run": 0.7060168357960154
+        "relative_to_run": 0.7170197143654636
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 4.502773761749268,
-            "mean_kld": 0.05647357469376497,
-            "mean_ref_top1_prob": 0.5379292386352562,
-            "median_context_kld": 0.05530742046289816,
-            "median_context_p99": 0.3166840970516205,
-            "p90_context_kld": 0.06890052055988576,
+            "max_kld": 4.442514419555664,
+            "mean_kld": 0.05679637436259177,
+            "mean_ref_top1_prob": 0.537879873360319,
+            "median_context_kld": 0.0549388505062868,
+            "median_context_p99": 0.3124707341194153,
+            "p90_context_kld": 0.07257355053062066,
             "positions": 47081,
-            "top1_agreement": 0.8710944967184214,
+            "top1_agreement": 0.8704785369894438,
             "worst_context_id": 343,
-            "worst_context_kld": 0.08539307900748275
+            "worst_context_kld": 0.08359138831170872
           }
         },
         "key": "open_news",
         "label": "open_news",
-        "relative_to_run": 0.6356128226452775
+        "relative_to_run": 0.6388754938710551
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 3,
-            "max_kld": 3.8121566772460938,
-            "mean_kld": 0.0552400798556026,
-            "mean_ref_top1_prob": 0.6023603953518782,
-            "median_context_kld": 0.056000113507537676,
-            "median_context_p99": 0.3778476119041443,
-            "p90_context_kld": 0.05916260943105591,
-            "positions": 6141,
-            "top1_agreement": 0.8791727731639798,
-            "worst_context_id": 967,
-            "worst_context_kld": 0.05916260943105591
+            "contexts": 44,
+            "max_kld": 14.120194435119629,
+            "mean_kld": 0.054789466668457965,
+            "mean_ref_top1_prob": 0.7203910827522294,
+            "median_context_kld": 0.04732764899256564,
+            "median_context_p99": 0.4036661684513092,
+            "p90_context_kld": 0.08157172921411385,
+            "positions": 90068,
+            "top1_agreement": 0.913787360660834,
+            "worst_context_id": 582,
+            "worst_context_kld": 0.27315084096214337
           }
         },
-        "key": "wikipedia_fr",
-        "label": "wikipedia_fr",
-        "relative_to_run": 0.6217297784063704
+        "key": "stackv2",
+        "label": "stackv2",
+        "relative_to_run": 0.6163007404887716
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 44,
-            "max_kld": 10.251813888549805,
-            "mean_kld": 0.05440211262874329,
-            "mean_ref_top1_prob": 0.720394649593742,
-            "median_context_kld": 0.04682076122968642,
-            "median_context_p99": 0.3964604139328003,
-            "p90_context_kld": 0.07820085520112714,
-            "positions": 90068,
-            "top1_agreement": 0.9138206688279966,
-            "worst_context_id": 582,
-            "worst_context_kld": 0.2682417939115386
+            "contexts": 3,
+            "max_kld": 2.9037251472473145,
+            "mean_kld": 0.05439633696763774,
+            "mean_ref_top1_prob": 0.6023746847155801,
+            "median_context_kld": 0.053121687734998244,
+            "median_context_p99": 0.3306625187397003,
+            "p90_context_kld": 0.05942542538622225,
+            "positions": 6141,
+            "top1_agreement": 0.8824295717309885,
+            "worst_context_id": 967,
+            "worst_context_kld": 0.05942542538622225
           }
         },
-        "key": "stackv2",
-        "label": "stackv2",
-        "relative_to_run": 0.6122984166192603
+        "key": "wikipedia_fr",
+        "label": "wikipedia_fr",
+        "relative_to_run": 0.6118786108267007
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 6.290522575378418,
-            "mean_kld": 0.053559997821783036,
-            "mean_ref_top1_prob": 0.6660615117323249,
-            "median_context_kld": 0.053893820669890874,
-            "median_context_p99": 0.4115641713142395,
-            "p90_context_kld": 0.06853919860015303,
+            "max_kld": 4.571195602416992,
+            "mean_kld": 0.0539436063571107,
+            "mean_ref_top1_prob": 0.6660535411394803,
+            "median_context_kld": 0.055948375007732136,
+            "median_context_p99": 0.45503899455070496,
+            "p90_context_kld": 0.0703836061073812,
             "positions": 12282,
-            "top1_agreement": 0.894723986321446,
+            "top1_agreement": 0.8940726266080443,
             "worst_context_id": 945,
-            "worst_context_kld": 0.06853919860015303
+            "worst_context_kld": 0.0703836061073812
           }
         },
         "key": "wikipedia_ru",
         "label": "wikipedia_ru",
-        "relative_to_run": 0.6028203736169199
+        "relative_to_run": 0.6067860587820141
       },
       {
         "cells": {
           "deployed": {
             "contexts": 52,
-            "max_kld": 11.309407234191895,
-            "mean_kld": 0.05070798150529347,
-            "mean_ref_top1_prob": 0.8582575125835092,
-            "median_context_kld": 0.03662647812839685,
-            "median_context_p99": 0.40668925642967224,
-            "p90_context_kld": 0.07997063031452036,
+            "max_kld": 11.164253234863281,
+            "mean_kld": 0.050829142015731216,
+            "mean_ref_top1_prob": 0.8582941084526002,
+            "median_context_kld": 0.036396497022982435,
+            "median_context_p99": 0.4023333787918091,
+            "p90_context_kld": 0.07490327471272329,
             "positions": 106444,
-            "top1_agreement": 0.9408421329525384,
+            "top1_agreement": 0.9406448461162678,
             "worst_context_id": 667,
-            "worst_context_kld": 0.27689472241669605
+            "worst_context_kld": 0.29321267733423706
           }
         },
         "key": "github_code",
         "label": "github_code",
-        "relative_to_run": 0.5707207916268594
+        "relative_to_run": 0.5717529256538344
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 16.185749053955078,
-            "mean_kld": 0.04949621323585394,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.048019773744291346,
-            "median_context_p99": 0.30111250281333923,
-            "p90_context_kld": 0.06206171818080255,
+            "max_kld": 11.157588958740234,
+            "mean_kld": 0.049079914368717295,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.048805704917541055,
+            "median_context_p99": 0.312124103307724,
+            "p90_context_kld": 0.05848004086686995,
             "positions": 196512,
-            "top1_agreement": 0.8900728708679368,
-            "worst_context_id": 246,
-            "worst_context_kld": 0.07372579095666683
+            "top1_agreement": 0.8908056505455137,
+            "worst_context_id": 174,
+            "worst_context_kld": 0.07139858353297991
           }
         },
         "key": "scientific_papers",
         "label": "scientific_papers",
-        "relative_to_run": 0.5570822809728582
+        "relative_to_run": 0.5520766929819303
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 8.847333908081055,
-            "mean_kld": 0.0491758979962088,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.04248309454994046,
-            "median_context_p99": 0.494755357503891,
-            "p90_context_kld": 0.08263404999373626,
+            "max_kld": 10.949919700622559,
+            "mean_kld": 0.04887779489384767,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.04098168850353924,
+            "median_context_p99": 0.5531740188598633,
+            "p90_context_kld": 0.07973034153401946,
             "positions": 73692,
-            "top1_agreement": 0.8469304673505944,
+            "top1_agreement": 0.8472968571893829,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.1645537792651185
+            "worst_context_kld": 0.17824096230327108
           }
         },
         "key": "starcoder_structured",
         "label": "starcoder_structured",
-        "relative_to_run": 0.553477117412535
+        "relative_to_run": 0.5498031468132274
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 17.194229125976562,
-            "mean_kld": 0.04891121428016043,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.04473202064081297,
-            "median_context_p99": 0.26329120993614197,
-            "p90_context_kld": 0.05904413538005142,
+            "max_kld": 17.83475685119629,
+            "mean_kld": 0.04852616171264854,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.04458567627830201,
+            "median_context_p99": 0.2575961947441101,
+            "p90_context_kld": 0.05704854838084238,
             "positions": 196512,
-            "top1_agreement": 0.9055681078000326,
+            "top1_agreement": 0.9055579303045107,
             "worst_context_id": 825,
-            "worst_context_kld": 0.12918918401151672
+            "worst_context_kld": 0.1312607970175872
           }
         },
         "key": "libretexts",
         "label": "libretexts",
-        "relative_to_run": 0.5504980893489136
+        "relative_to_run": 0.5458477918311317
       }
     ],
     "stratum": [
@@ -390,217 +390,217 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 29.621610641479492,
-            "mean_kld": 0.21268638446455335,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.16980764884602795,
-            "median_context_p99": 3.194436550140381,
-            "p90_context_kld": 0.4663941209945985,
+            "max_kld": 29.410005569458008,
+            "mean_kld": 0.21324363685885994,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.17013793919462264,
+            "median_context_p99": 3.1346852779388428,
+            "p90_context_kld": 0.4793873904592046,
             "positions": 196512,
-            "top1_agreement": 0.8831114639309559,
+            "top1_agreement": 0.8833353688324377,
             "worst_context_id": 454,
-            "worst_context_kld": 1.254965819110839
+            "worst_context_kld": 1.2597357640464442
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 2.3937955743159085
+        "relative_to_run": 2.398676594096019
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 9.221136093139648,
-            "mean_kld": 0.13804625823415112,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.0941144168045219,
-            "median_context_p99": 0.6546816229820251,
-            "p90_context_kld": 0.23829632700468703,
+            "max_kld": 8.557491302490234,
+            "mean_kld": 0.13763110758662334,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.09366887791148955,
+            "median_context_p99": 0.6111054420471191,
+            "p90_context_kld": 0.24518771943300743,
             "positions": 147384,
-            "top1_agreement": 0.826616186288878,
-            "worst_context_id": 905,
-            "worst_context_kld": 0.48569205838970464
+            "top1_agreement": 0.8276067958530098,
+            "worst_context_id": 893,
+            "worst_context_kld": 0.4890878310734861
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.5537173328876446
+        "relative_to_run": 1.5481470924548617
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 9.237215042114258,
-            "mean_kld": 0.08939194390567347,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.07517420846258792,
-            "median_context_p99": 0.6442986130714417,
-            "p90_context_kld": 0.12549267043511111,
+            "max_kld": 8.673469543457031,
+            "mean_kld": 0.08926161203012813,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.07524436844303807,
+            "median_context_p99": 0.6247144937515259,
+            "p90_context_kld": 0.12359177821272875,
             "positions": 196512,
-            "top1_agreement": 0.8639675948542582,
+            "top1_agreement": 0.8644510258915487,
             "worst_context_id": 6,
-            "worst_context_kld": 0.37588815477459775
+            "worst_context_kld": 0.38041751277782426
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 1.006110664956836
+        "relative_to_run": 1.0040615639549482
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.085050582885742,
-            "mean_kld": 0.07346640922786896,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.0650303690739286,
-            "median_context_p99": 0.35447409749031067,
-            "p90_context_kld": 0.08656176264974488,
+            "max_kld": 7.7061967849731445,
+            "mean_kld": 0.0738154133464014,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.06468236685300296,
+            "median_context_p99": 0.3413587808609009,
+            "p90_context_kld": 0.08678552501587226,
             "positions": 147384,
-            "top1_agreement": 0.8569519079411605,
+            "top1_agreement": 0.8566194430874451,
             "worst_context_id": 443,
-            "worst_context_kld": 0.3243600094405916
+            "worst_context_kld": 0.3221895584307333
           }
         },
         "key": "literary_narrative",
         "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.8268679996291151
+        "relative_to_run": 0.8303145964197121
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 8.182098388671875,
-            "mean_kld": 0.07235635445357978,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.06563072477178752,
-            "median_context_p99": 0.4368785619735718,
-            "p90_context_kld": 0.1307554214582166,
+            "max_kld": 7.552076816558838,
+            "mean_kld": 0.07310515835266779,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.06447523888587277,
+            "median_context_p99": 0.4327349364757538,
+            "p90_context_kld": 0.13155629979250485,
             "positions": 147384,
-            "top1_agreement": 0.8739211854746783,
+            "top1_agreement": 0.8737515605493134,
             "worst_context_id": 275,
-            "worst_context_kld": 0.15489253509157094
+            "worst_context_kld": 0.1603293014694022
           }
         },
         "key": "news_history_legal_essays",
         "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.8143742792970337
+        "relative_to_run": 0.8223252746542226
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 7.345405578613281,
-            "mean_kld": 0.07083341233029045,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.06525897825527403,
-            "median_context_p99": 0.44970738887786865,
-            "p90_context_kld": 0.11245465481627331,
+            "max_kld": 11.588532447814941,
+            "mean_kld": 0.07121532314429206,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.06519233793250864,
+            "median_context_p99": 0.46661657094955444,
+            "p90_context_kld": 0.1105094893175383,
             "positions": 73692,
-            "top1_agreement": 0.8756852846984747,
+            "top1_agreement": 0.8746403951582261,
             "worst_context_id": 953,
-            "worst_context_kld": 0.11817462742460928
+            "worst_context_kld": 0.12065877383606818
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.7972334918232739
+        "relative_to_run": 0.801067414172178
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 11.309407234191895,
-            "mean_kld": 0.05240112493687464,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.0419624148682241,
-            "median_context_p99": 0.3964604139328003,
-            "p90_context_kld": 0.07997063031452036,
+            "max_kld": 14.120194435119629,
+            "mean_kld": 0.05264429081489764,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.040841334767735375,
+            "median_context_p99": 0.4023333787918091,
+            "p90_context_kld": 0.08157172921411385,
             "positions": 196512,
-            "top1_agreement": 0.9284572952287901,
+            "top1_agreement": 0.9283351652825272,
             "worst_context_id": 667,
-            "worst_context_kld": 0.27689472241669605
+            "worst_context_kld": 0.29321267733423706
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.5897772030817099
+        "relative_to_run": 0.5921706741198473
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 16.185749053955078,
-            "mean_kld": 0.04949621323585394,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.048019773744291346,
-            "median_context_p99": 0.30111250281333923,
-            "p90_context_kld": 0.06206171818080255,
+            "max_kld": 11.157588958740234,
+            "mean_kld": 0.049079914368717295,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.048805704917541055,
+            "median_context_p99": 0.312124103307724,
+            "p90_context_kld": 0.05848004086686995,
             "positions": 196512,
-            "top1_agreement": 0.8900728708679368,
-            "worst_context_id": 246,
-            "worst_context_kld": 0.07372579095666683
+            "top1_agreement": 0.8908056505455137,
+            "worst_context_id": 174,
+            "worst_context_kld": 0.07139858353297991
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.5570822809728582
+        "relative_to_run": 0.5520766929819303
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 8.847333908081055,
-            "mean_kld": 0.0491758979962088,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.04248309454994046,
-            "median_context_p99": 0.494755357503891,
-            "p90_context_kld": 0.08263404999373626,
+            "max_kld": 10.949919700622559,
+            "mean_kld": 0.04887779489384767,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.04098168850353924,
+            "median_context_p99": 0.5531740188598633,
+            "p90_context_kld": 0.07973034153401946,
             "positions": 73692,
-            "top1_agreement": 0.8469304673505944,
+            "top1_agreement": 0.8472968571893829,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.1645537792651185
+            "worst_context_kld": 0.17824096230327108
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.553477117412535
+        "relative_to_run": 0.5498031468132274
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 17.194229125976562,
-            "mean_kld": 0.04891121428016043,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.04473202064081297,
-            "median_context_p99": 0.26329120993614197,
-            "p90_context_kld": 0.05904413538005142,
+            "max_kld": 17.83475685119629,
+            "mean_kld": 0.04852616171264854,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.04458567627830201,
+            "median_context_p99": 0.2575961947441101,
+            "p90_context_kld": 0.05704854838084238,
             "positions": 196512,
-            "top1_agreement": 0.9055681078000326,
+            "top1_agreement": 0.9055579303045107,
             "worst_context_id": 825,
-            "worst_context_kld": 0.12918918401151672
+            "worst_context_kld": 0.1312607970175872
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.5504980893489136
+        "relative_to_run": 0.5458477918311317
       }
     ]
   },
   "overall": {
     "deployed": {
       "contexts": 768,
-      "max_kld": 29.621610641479492,
-      "mean_kld": 0.08884901732903162,
-      "mean_ref_top1_prob": 0.6415440866166368,
-      "median_context_kld": 0.06137484148042411,
-      "median_context_p99": 0.41598302125930786,
-      "p90_context_kld": 0.16160134971982643,
+      "max_kld": 29.410005569458008,
+      "mean_kld": 0.08890053681422791,
+      "mean_ref_top1_prob": 0.6415473983687051,
+      "median_context_kld": 0.06094675094259141,
+      "median_context_p99": 0.4184585213661194,
+      "p90_context_kld": 0.169365564213705,
       "positions": 1572096,
-      "top1_agreement": 0.879409399934864,
+      "top1_agreement": 0.8795868700130272,
       "worst_context_id": 454,
-      "worst_context_kld": 1.254965819110839
+      "worst_context_kld": 1.2597357640464442
     }
   },
   "primary": "deployed"

Qwen3.8-27B-NVFP4-RTX5090/strata.md

@@ -1,52 +1,52 @@
 # Fidelity by domain - Qwen3.8-27B / Qwen3.8-27B-NVFP4-RTX5090
 
-768 contexts, 1572096 scored positions, mean 0.08884902, reference top-1 64.2%, top-1 agreement 87.9409%.
+768 contexts, 1572096 scored positions, mean 0.08890054, reference top-1 64.2%, top-1 agreement 87.9587%.
 
-`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
+`x run` is the domain's mean divided by the run's mean, so 1.00 is a domain that degrades exactly as much as the model as a whole. The `deployed` cell is QxQ; `bxq` is the teacher-ID counterfactual on the same student weights. `Median ctx` and `p90 ctx` are the spread of per-context means within the domain.
 
 ## Ranked by domain
 
 | Domain | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.21268638 | 2.39 | 0.16980765 | 0.46639412 | 1.254966 |
-| Chinese across several content types | 72 | 55.6% | 0.13804626 | 1.55 | 0.09411442 | 0.23829633 | 0.485692 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.08939194 | 1.01 | 0.07517421 | 0.12549267 | 0.375888 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.07346641 | 0.83 | 0.06503037 | 0.08656176 | 0.324360 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.07235635 | 0.81 | 0.06563072 | 0.13075542 | 0.154893 |
-| Other multilingual content | 36 | 64.5% | 0.07083341 | 0.80 | 0.06525898 | 0.11245465 | 0.118175 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.05240112 | 0.59 | 0.04196241 | 0.07997063 | 0.276895 |
-| Scientific and technical exposition | 96 | 60.3% | 0.04949621 | 0.56 | 0.04801977 | 0.06206172 | 0.073726 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.04917590 | 0.55 | 0.04248309 | 0.08263405 | 0.164554 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.04891121 | 0.55 | 0.04473202 | 0.05904414 | 0.129189 |
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.21324364 | 2.40 | 0.17013794 | 0.47938739 | 1.259736 |
+| Chinese across several content types | 72 | 55.6% | 0.13763111 | 1.55 | 0.09366888 | 0.24518772 | 0.489088 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.08926161 | 1.00 | 0.07524437 | 0.12359178 | 0.380418 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.07381541 | 0.83 | 0.06468237 | 0.08678553 | 0.322190 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.07310516 | 0.82 | 0.06447524 | 0.13155630 | 0.160329 |
+| Other multilingual content | 36 | 64.5% | 0.07121532 | 0.80 | 0.06519234 | 0.11050949 | 0.120659 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.05264429 | 0.59 | 0.04084133 | 0.08157173 | 0.293213 |
+| Scientific and technical exposition | 96 | 60.3% | 0.04907991 | 0.55 | 0.04880570 | 0.05848004 | 0.071399 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.04887779 | 0.55 | 0.04098169 | 0.07973034 | 0.178241 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.04852616 | 0.55 | 0.04458568 | 0.05704855 | 0.131261 |
 
 ## Ranked by source dataset
 
 | Source | Contexts | Ref top-1 | Mean KLD | x run | Median ctx | p90 ctx | Worst ctx |
 |---|---|---|---|---|---|---|---|
-| wildchat | 96 | 73.5% | 0.21268638 | 2.39 | 0.16980765 | 0.46639412 | 1.254966 |
-| wikisource_zh | 33 | 60.4% | 0.21068044 | 2.37 | 0.16749470 | 0.46016936 | 0.485692 |
-| wikipedia_de | 7 | 73.6% | 0.10192228 | 1.15 | 0.11245465 | 0.11817463 | 0.118175 |
-| wikipedia_en | 96 | 59.5% | 0.08939194 | 1.01 | 0.07517421 | 0.12549267 | 0.375888 |
-| regulations | 23 | 72.3% | 0.08888440 | 1.00 | 0.10944162 | 0.13830929 | 0.154893 |
-| wikipedia_zh | 39 | 51.5% | 0.07658656 | 0.86 | 0.07334535 | 0.10122490 | 0.109906 |
-| public_domain_books | 72 | 50.6% | 0.07346641 | 0.83 | 0.06503037 | 0.08656176 | 0.324360 |
-| wikipedia_ja | 7 | 57.0% | 0.07299386 | 0.82 | 0.07676595 | 0.08001701 | 0.080017 |
-| public_domain_review | 26 | 51.8% | 0.07178554 | 0.81 | 0.06815590 | 0.09782586 | 0.145134 |
-| wikipedia_es | 7 | 59.2% | 0.06601946 | 0.74 | 0.06525898 | 0.08028594 | 0.080286 |
-| wikipedia_cs | 6 | 68.9% | 0.06272890 | 0.71 | 0.06199259 | 0.07546922 | 0.075469 |
-| open_news | 23 | 53.8% | 0.05647357 | 0.64 | 0.05530742 | 0.06890052 | 0.085393 |
-| wikipedia_fr | 3 | 60.2% | 0.05524008 | 0.62 | 0.05600011 | 0.05916261 | 0.059163 |
-| stackv2 | 44 | 72.0% | 0.05440211 | 0.61 | 0.04682076 | 0.07820086 | 0.268242 |
-| wikipedia_ru | 6 | 66.6% | 0.05356000 | 0.60 | 0.05389382 | 0.06853920 | 0.068539 |
-| github_code | 52 | 85.8% | 0.05070798 | 0.57 | 0.03662648 | 0.07997063 | 0.276895 |
-| scientific_papers | 96 | 60.3% | 0.04949621 | 0.56 | 0.04801977 | 0.06206172 | 0.073726 |
-| starcoder_structured | 36 | 70.7% | 0.04917590 | 0.55 | 0.04248309 | 0.08263405 | 0.164554 |
-| libretexts | 96 | 65.8% | 0.04891121 | 0.55 | 0.04473202 | 0.05904414 | 0.129189 |
+| wildchat | 96 | 73.5% | 0.21324364 | 2.40 | 0.17013794 | 0.47938739 | 1.259736 |
+| wikisource_zh | 33 | 60.4% | 0.20942829 | 2.36 | 0.16936556 | 0.44822882 | 0.489088 |
+| wikipedia_de | 7 | 73.6% | 0.10318174 | 1.16 | 0.11050949 | 0.12065877 | 0.120659 |
+| regulations | 23 | 72.3% | 0.09111800 | 1.02 | 0.10723594 | 0.14780949 | 0.160329 |
+| wikipedia_en | 96 | 59.5% | 0.08926161 | 1.00 | 0.07524437 | 0.12359178 | 0.380418 |
+| wikipedia_zh | 39 | 51.5% | 0.07687964 | 0.86 | 0.07373861 | 0.10373904 | 0.109338 |
+| public_domain_books | 72 | 50.6% | 0.07381541 | 0.83 | 0.06468237 | 0.08678553 | 0.322190 |
+| wikipedia_ja | 7 | 57.0% | 0.07229254 | 0.81 | 0.07461241 | 0.08007906 | 0.080079 |
+| public_domain_review | 26 | 51.8% | 0.07159772 | 0.81 | 0.06766349 | 0.10433167 | 0.146552 |
+| wikipedia_es | 7 | 59.2% | 0.06658863 | 0.75 | 0.06366197 | 0.07747506 | 0.077475 |
+| wikipedia_cs | 6 | 68.9% | 0.06374344 | 0.72 | 0.06442506 | 0.07887765 | 0.078878 |
+| open_news | 23 | 53.8% | 0.05679637 | 0.64 | 0.05493885 | 0.07257355 | 0.083591 |
+| stackv2 | 44 | 72.0% | 0.05478947 | 0.62 | 0.04732765 | 0.08157173 | 0.273151 |
+| wikipedia_fr | 3 | 60.2% | 0.05439634 | 0.61 | 0.05312169 | 0.05942543 | 0.059425 |
+| wikipedia_ru | 6 | 66.6% | 0.05394361 | 0.61 | 0.05594838 | 0.07038361 | 0.070384 |
+| github_code | 52 | 85.8% | 0.05082914 | 0.57 | 0.03639650 | 0.07490327 | 0.293213 |
+| scientific_papers | 96 | 60.3% | 0.04907991 | 0.55 | 0.04880570 | 0.05848004 | 0.071399 |
+| starcoder_structured | 36 | 70.7% | 0.04887779 | 0.55 | 0.04098169 | 0.07973034 | 0.178241 |
+| libretexts | 96 | 65.8% | 0.04852616 | 0.55 | 0.04458568 | 0.05704855 | 0.131261 |
 
 ## Reading
 
-- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.21268638, 2.39x the run mean of 0.08884902 over 96 context(s).
-- Strongest domain: **Worked mathematics, science, and formal reasoning** at 0.04891121, 0.55x the run mean. The spread across domains is 4.3x.
+- Weakest domain: **Natural dialogue, instruction following, and assistance** at 0.21324364, 2.40x the run mean of 0.08890054 over 96 context(s).
+- Strongest domain: **Worked mathematics, science, and formal reasoning** at 0.04852616, 0.55x the run mean. The spread across domains is 4.4x.
 - This is a strong finding: the reference was *more* certain on Natural dialogue, instruction following, and assistance (top-1 73.5% against 64.2% overall) and the candidate still diverged most there, so predictability does not explain it.
-- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 1.254966 against a median of 0.16980765 (context 454). Read the documents before treating the domain as weak.
+- Natural dialogue, instruction following, and assistance is driven by outlier contexts: its worst context is 1.259736 against a median of 0.17013794 (context 454). Read the documents before treating the domain as weak.
 

Qwen3.8-27B-NVFP4/compliance.json

@@ -1,8 +1,9 @@
 {
   "attribution": null,
   "candidate": "/media/fmodels2/unsloth/Qwen3.8-27B-NVFP4",
-  "candidate_weights_sha256": null,
+  "candidate_weights_sha256": "529539fc371093824cdabaab7afc083f24f46e2e92bb91ab29c32d6542bbcfb3",
   "comparability_key": {
+    "compiled_extensions_sha256": "f2fbc7537b0f01f65341030cfa90741c2ed7ad46c88b0faa3dd0b46d311ccdc8",
     "context_length": 2048,
     "driver": "580.173.02",
     "gpu_names": [
@@ -12,19 +13,22 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 248044,
-    "laws_version": 9,
+    "kv_cache_dtype": "bfloat16",
+    "laws_version": 15,
     "model_runner_v2": false,
+    "numerics_digest": "251a9225b37415b91fc58d975b0670fc4b481e6f5a816b663573c42e161438f8",
     "reference_config_sha256": "191e0af232104ed8b65258cf3fb2b842e288008baca7633c11b82a1ac7203aab",
     "rows": 768,
     "score_from": 0,
     "stride": 2048,
+    "substituted_parameters": [],
     "suite_id": "qwen3.8-27b-fidelity-1024x2048-v1",
     "tensor_parallel_size": 1,
     "token_sha256": "9c935708bbffbf45d5baa2931fb4af7e7b22df1e041aa0b6f4ca44d3315e543b",
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T12:56:04.087010+00:00",
+  "evaluated_at": "2026-09-11T20:24:21.646814+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -52,13 +56,13 @@
       "title": "Real vocabulary"
     },
     {
-      "detail": "all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc",
+      "detail": "all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3",
       "law": 5,
       "status": "pass",
       "title": "Manifest binding"
     },
     {
-      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s)",
+      "detail": "torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s)",
       "law": 6,
       "status": "pass",
       "title": "Provenance"
@@ -70,13 +74,13 @@
       "title": "Storage integrity"
     },
     {
-      "detail": "trunk 0.03966656, deployed 0.04096872, delta 0.0013021660852515493",
+      "detail": "trunk 0.03923024, deployed 0.04053080, delta 0.0013005567674299265",
       "law": 8,
       "status": "pass",
       "title": "Head transparency"
     },
     {
-      "detail": "mean 0.04096872, median 0.00944244, max 36.54566574, 4 depth buckets",
+      "detail": "mean 0.04053080, median 0.00929746, max 28.66598129, 4 depth buckets",
       "law": 9,
       "status": "pass",
       "title": "Tail and depth disclosure"
@@ -103,38 +107,37 @@
       "detail": "reference declares no experts",
       "law": 14,
       "status": "not_applicable",
-      "title": "Component attribution"
+      "title": "Routed-model intervention"
     },
     {
-      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.12457770, strongest scientific_technical at 0.01691031, spread 7.4x",
+      "detail": "10 domains disclosed; weakest dialogue_instruction at 0.12158332, strongest scientific_technical at 0.01656302, spread 7.3x",
       "law": 15,
       "status": "pass",
       "title": "Domain disclosure"
     },
     {
-      "approval": {
-        "approver": "Andy Kitzke",
-        "justification": "These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger.",
-        "timestamp": "2026-09-02T05:15:00Z"
-      },
-      "detail": "the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else",
+      "detail": "scored weights 529539fc37109382 as inspected",
       "law": 16,
-      "status": "override",
+      "status": "pass",
       "title": "Candidate weight binding"
     },
     {
-      "detail": "1 override(s), each fully attributed",
+      "detail": "every scored layer used the checkpoint's own quantization parameters",
+      "law": 17,
+      "status": "pass",
+      "title": "Substitution disclosure"
+    },
+    {
+      "detail": "no overrides claimed",
       "law": 13,
       "status": "pass",
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 9,
-  "mean_kld": 0.040968724321885384,
+  "laws_version": 15,
+  "mean_kld": 0.04053079722531637,
   "nondeterminism_floor": 0.0,
-  "overridden_laws": [
-    16
-  ],
+  "overridden_laws": [],
   "partition": "analysis",
   "program": "Local Inference Lab \u2014 Distribution Fidelity",
   "ranking_floor": null,
@@ -147,16 +150,16 @@
     "overall": {
       "deployed": {
         "contexts": 768,
-        "max_kld": 36.5456657409668,
-        "mean_kld": 0.040968724321885384,
-        "mean_ref_top1_prob": 0.6415440866166368,
-        "median_context_kld": 0.021978546635872478,
-        "median_context_p99": 0.17513030767440796,
-        "p90_context_kld": 0.08278915535830515,
+        "max_kld": 28.66598129272461,
+        "mean_kld": 0.040530797225316356,
+        "mean_ref_top1_prob": 0.6415473983687051,
+        "median_context_kld": 0.021778975352759924,
+        "median_context_p99": 0.17481790482997894,
+        "p90_context_kld": 0.08018104969267713,
         "positions": 1572096,
-        "top1_agreement": 0.926059858939912,
+        "top1_agreement": 0.9268944135727081,
         "worst_context_id": 454,
-        "worst_context_kld": 0.8362269388706521
+        "worst_context_kld": 0.8104930786551602
       }
     },
     "primary": "deployed",
@@ -165,201 +168,201 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 36.5456657409668,
-            "mean_kld": 0.12457769541074566,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.09409491196048908,
-            "median_context_p99": 2.0393991470336914,
-            "p90_context_kld": 0.2799708803626206,
+            "max_kld": 28.66598129272461,
+            "mean_kld": 0.12158332374682436,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.08575015208991663,
+            "median_context_p99": 1.6861377954483032,
+            "p90_context_kld": 0.29358947054916235,
             "positions": 196512,
-            "top1_agreement": 0.923175175052923,
+            "top1_agreement": 0.9250681892199968,
             "worst_context_id": 454,
-            "worst_context_kld": 0.8362269388706521
+            "worst_context_kld": 0.8104930786551602
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 3.0407999632099014
+        "relative_to_run": 2.999776270644905
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.911944389343262,
-            "mean_kld": 0.05745101151734411,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.034198077382997016,
-            "median_context_p99": 0.27472054958343506,
-            "p90_context_kld": 0.12171763135253476,
+            "max_kld": 6.782408714294434,
+            "mean_kld": 0.058009695747318554,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.03501158816479015,
+            "median_context_p99": 0.27062809467315674,
+            "p90_context_kld": 0.1256960806689686,
             "positions": 147384,
-            "top1_agreement": 0.8939437116647668,
-            "worst_context_id": 893,
-            "worst_context_kld": 0.23709843126137567
+            "top1_agreement": 0.8944390164468328,
+            "worst_context_id": 917,
+            "worst_context_kld": 0.2366614644814825
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.4023138984255346
+        "relative_to_run": 1.4312498080122769
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 5.581716537475586,
-            "mean_kld": 0.03972593213279645,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.03066756860698419,
-            "median_context_p99": 0.291412353515625,
-            "p90_context_kld": 0.05708804087610574,
+            "max_kld": 6.517783164978027,
+            "mean_kld": 0.03968051556345783,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.030948956898133834,
+            "median_context_p99": 0.3017843961715698,
+            "p90_context_kld": 0.05457949171108933,
             "positions": 196512,
-            "top1_agreement": 0.9163511643054877,
+            "top1_agreement": 0.9173027601367856,
             "worst_context_id": 65,
-            "worst_context_kld": 0.2408369537170143
+            "worst_context_kld": 0.24843901303775767
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.9696648550898366
+        "relative_to_run": 0.9790213437665266
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.997079849243164,
-            "mean_kld": 0.029428653535056526,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.02369674456730006,
-            "median_context_p99": 0.15645328164100647,
-            "p90_context_kld": 0.034802640181948945,
+            "max_kld": 4.948120594024658,
+            "mean_kld": 0.0291750717041567,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.023173652596112884,
+            "median_context_p99": 0.17660361528396606,
+            "p90_context_kld": 0.0621877665506871,
             "positions": 147384,
-            "top1_agreement": 0.9151671823264398,
-            "worst_context_id": 443,
-            "worst_context_kld": 0.21936384601423634
+            "top1_agreement": 0.9259281875915975,
+            "worst_context_id": 312,
+            "worst_context_kld": 0.0811365691669795
           }
         },
-        "key": "literary_narrative",
-        "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.7183199873112919
+        "key": "news_history_legal_essays",
+        "label": "News, history, economics, legal analysis, and essays",
+        "relative_to_run": 0.7198247678664795
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.266860008239746,
-            "mean_kld": 0.029052629707756242,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.021822250241878293,
-            "median_context_p99": 0.17231622338294983,
-            "p90_context_kld": 0.06020746195574988,
+            "max_kld": 6.705466270446777,
+            "mean_kld": 0.029120922214834796,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.023827455457346176,
+            "median_context_p99": 0.14993515610694885,
+            "p90_context_kld": 0.033321719956789224,
             "positions": 147384,
-            "top1_agreement": 0.9257585626662325,
-            "worst_context_id": 275,
-            "worst_context_kld": 0.0819730113510625
+            "top1_agreement": 0.9155810671443304,
+            "worst_context_id": 443,
+            "worst_context_kld": 0.21794198408233922
           }
         },
-        "key": "news_history_legal_essays",
-        "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.709141673035604
+        "key": "literary_narrative",
+        "label": "Literary, narrative, and creative writing",
+        "relative_to_run": 0.7184887593734594
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 7.721675872802734,
-            "mean_kld": 0.025570533792631602,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.0227388942645489,
-            "median_context_p99": 0.1798495501279831,
-            "p90_context_kld": 0.04305093457533655,
+            "max_kld": 6.416473865509033,
+            "mean_kld": 0.025013050454758354,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.022881467166419225,
+            "median_context_p99": 0.17793317139148712,
+            "p90_context_kld": 0.04138496706841051,
             "positions": 73692,
-            "top1_agreement": 0.9274819519079411,
+            "top1_agreement": 0.9287303913586278,
             "worst_context_id": 953,
-            "worst_context_kld": 0.051774189189553635
+            "worst_context_kld": 0.0452351423921429
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.6241476691274932
+        "relative_to_run": 0.617136897547495
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 8.132791519165039,
-            "mean_kld": 0.02250031790751401,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.015097981818138071,
-            "median_context_p99": 0.16505374014377594,
-            "p90_context_kld": 0.03532434663282433,
+            "max_kld": 12.169252395629883,
+            "mean_kld": 0.02270613721635169,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.015219370374946644,
+            "median_context_p99": 0.176174134016037,
+            "p90_context_kld": 0.03640113065999465,
             "positions": 196512,
-            "top1_agreement": 0.9556871844976388,
+            "top1_agreement": 0.9565929815990881,
             "worst_context_id": 667,
-            "worst_context_kld": 0.1707121626802731
+            "worst_context_kld": 0.17765063071981094
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.5492071886527938
+        "relative_to_run": 0.5602193583838262
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 4.752742767333984,
-            "mean_kld": 0.020421348473034326,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.01775127215107961,
-            "median_context_p99": 0.22656621038913727,
-            "p90_context_kld": 0.03592128721473434,
+            "max_kld": 5.577369689941406,
+            "mean_kld": 0.020248958642327196,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.015101154141245148,
+            "median_context_p99": 0.2136673778295517,
+            "p90_context_kld": 0.03390113755388437,
             "positions": 73692,
-            "top1_agreement": 0.882931661510069,
+            "top1_agreement": 0.8844786408293981,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.09928303706595111
+            "worst_context_kld": 0.10268640378926563
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.4984619074928163
+        "relative_to_run": 0.4995943832478895
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.103275299072266,
-            "mean_kld": 0.01983936171590355,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.016081149073808487,
-            "median_context_p99": 0.11479724943637848,
-            "p90_context_kld": 0.03028211425308765,
+            "max_kld": 11.945074081420898,
+            "mean_kld": 0.01951085562769419,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.015771377783579774,
+            "median_context_p99": 0.11416608095169067,
+            "p90_context_kld": 0.029937825294863343,
             "positions": 196512,
-            "top1_agreement": 0.9445988031265267,
+            "top1_agreement": 0.9449143054877056,
             "worst_context_id": 825,
-            "worst_context_kld": 0.07883124659302003
+            "worst_context_kld": 0.07172076344828876
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.48425627217553896
+        "relative_to_run": 0.48138346549737526
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 4.14390754699707,
-            "mean_kld": 0.01691031048838098,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.016672932853405074,
-            "median_context_p99": 0.1243298128247261,
-            "p90_context_kld": 0.021205758610799045,
+            "max_kld": 3.9671146869659424,
+            "mean_kld": 0.01656302498706322,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.016550733119089625,
+            "median_context_p99": 0.12118519842624664,
+            "p90_context_kld": 0.021278483650238936,
             "positions": 196512,
-            "top1_agreement": 0.9386093470118874,
-            "worst_context_id": 246,
-            "worst_context_kld": 0.031466900875017814
+            "top1_agreement": 0.9393624816805081,
+            "worst_context_id": 174,
+            "worst_context_kld": 0.027381622246614497
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.4127614605599896
+        "relative_to_run": 0.40865282996993746
       }
     ]
   },

Qwen3.8-27B-NVFP4/inspect.json

@@ -1,31 +1,66 @@
 {
+  "model": "/media/fmodels2/unsloth/Qwen3.8-27B-NVFP4",
+  "inspect_version": 5,
+  "weights_sha256": "529539fc371093824cdabaab7afc083f24f46e2e92bb91ab29c32d6542bbcfb3",
+  "weights_bytes": 23417592488,
+  "quant_method": "compressed-tensors",
+  "declared": {},
+  "detected_scheme": "nvfp4",
+  "detected_block": 16,
+  "quant_algorithm": "round_to_nearest",
+  "scheme_mix": {
+    "format": "mixed-precision",
+    "groups": [
+      {
+        "group": "group_0",
+        "format": "float-quantized",
+        "weight_bits": 8,
+        "activation_bits": 8,
+        "targets": [
+          "re:.*self_attn\\.(q|k|v|o)_proj$",
+          "re:.*linear_attn\\.(in_proj_qkv|in_proj_z|out_proj)$",
+          "re:.*lm_head",
+          "re:.*layers\\.(56|57|58|59|60|61|62|63)\\.mlp\\.(gate|up|down)_proj$"
+        ]
+      },
+      {
+        "group": "group_1",
+        "format": "nvfp4-pack-quantized",
+        "weight_bits": 4,
+        "activation_bits": 4,
+        "targets": [
+          "re:.*mlp\\.(gate|up|down)_proj$"
+        ]
+      }
+    ]
+  },
+  "kv_cache_scheme": {
+    "num_bits": 8,
+    "type": "float",
+    "strategy": "tensor"
+  },
   "coverage": {
-    "attention": {
-      "quantized": 64,
-      "weights": 68
-    },
-    "dense_mlp": {
-      "quantized": 192,
-      "weights": 27
+    "router": {
+      "weights": 0,
+      "quantized": 0
     },
     "experts": {
-      "quantized": 0,
-      "weights": 0
-    },
-    "router": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
     },
     "shared_expert": {
-      "quantized": 0,
-      "weights": 0
+      "weights": 0,
+      "quantized": 0
+    },
+    "attention": {
+      "weights": 68,
+      "quantized": 64
+    },
+    "dense_mlp": {
+      "weights": 195,
+      "quantized": 192
     }
   },
-  "declared": {},
-  "detected_block": null,
-  "detected_scheme": "fp8_per_channel",
-  "model": "/media/fmodels2/unsloth/Qwen3.8-27B-NVFP4",
-  "quant_method": "compressed-tensors",
-  "weights_bytes": 23417592488,
-  "weights_bytes_source": "hub"
+  "unloadable_reason": null,
+  "quantized_names_sha256": "eccbe240ef5caa18a896c8a311d4a9b4257c68423f0d8f9ae1c982d53797e657"
 }

Qwen3.8-27B-NVFP4/report.md

@@ -1,12 +1,12 @@
 # Qwen3.8-27B / Qwen3.8-27B-NVFP4: distribution fidelity
 
-**Mean KLD(reference || candidate) = 0.04096872** over 1572096 scored positions.
+**Mean KLD(reference || candidate) = 0.04053080** over 1572096 scored positions.
 
 | Mean | Median | p90 | p99 | Max | Top-1 agreement |
 |---|---|---|---|---|---|
-| 0.04096872 | 0.00944244 | 0.05542825 | 0.50567393 | 36.54566574 | 92.6060% |
+| 0.04053080 | 0.00929746 | 0.05469899 | 0.50382712 | 28.66598129 | 92.6894% |
 
-Reverse direction, KLD(candidate || reference): 0.04331330.
+Reverse direction, KLD(candidate || reference): 0.04259346.
 
 ## Identity
 
@@ -15,10 +15,10 @@ Reverse direction, KLD(candidate || reference): 0.04331330.
 | Reference checkpoint | /media/fmodels2/Qwen/Qwen3.8-27B |
 | Reference config SHA-256 | 191e0af23210 |
 | Candidate checkpoint | /media/fmodels2/unsloth/Qwen3.8-27B-NVFP4 |
-| Candidate weights SHA-256 | unbound (Law 16 gap) |
+| Candidate weights SHA-256 | 529539fc37109382 |
 | Suite | qwen3.8-27b-fidelity-1024x2048-v1 |
 | Suite token SHA-256 | 9c935708bbffbf45 |
-| Capture manifest SHA-256 | 304794b4e35b326d |
+| Capture manifest SHA-256 | dbfacc9cbb6d3e06 |
 | Tokenizer | None |
 | Scored vocabulary | 248044 |
 | Declared vocabulary | 248320 |
@@ -27,42 +27,49 @@ Reverse direction, KLD(candidate || reference): 0.04331330.
 | Model runner | V1 |
 | Tensor parallel | 1 |
 | Eager enforced | True |
+| KV cache | bfloat16 |
 | Prefix caching | False |
 | max_num_seqs | 1 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45 |
+| vLLM commit | 60071d1ab732 |
+| vLLM dirty digest | e3b0c44298fc1c14 |
+| Numerics digest | 251a9225b37415b9 |
+| Compiled extensions | f2fbc7537b0f01f6 |
+| FlashInfer | 0.6.17 |
 | torch | 2.13.0+cu132 |
 | Driver | 580.173.02 |
 | GPUs | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
-| Laws version | 9 |
+| Laws version | 15 |
 | Partition | analysis |
 
+The commit says when this number was taken. The numerics digest says what took it: a hash of the runtime and the scorer, which moves only when code that can change a logit changes. Comparability is bounded by the digest, so a result stays current across commits that cannot reach a number, and two results carrying the same digest were computed by the same code whatever their commits say.
+
 ## Fidelity by domain
 
-The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
+The suite is stratified, so the mean above is an average over kinds of text that do not degrade equally. `deployed` is QxQ (natural student routing). `bxq` is the teacher-ID counterfactual on the same weights. `x run` is a domain's mean divided by the run's, so 1.00 degrades exactly as much as the model overall. The reference's own top-1 probability is shown because a domain the reference finds harder will diverge more for that reason alone.
 
 | Domain | Contexts | Reference top-1 | Mean KLD | x run | Worst context |
 |---|---|---|---|---|---|
-| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.12457770 | 3.04 | 0.83622694 |
-| Chinese across several content types | 72 | 55.6% | 0.05745101 | 1.40 | 0.23709843 |
-| Encyclopedic and factual reference | 96 | 59.5% | 0.03972593 | 0.97 | 0.24083695 |
-| Literary, narrative, and creative writing | 72 | 50.6% | 0.02942865 | 0.72 | 0.21936385 |
-| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.02905263 | 0.71 | 0.08197301 |
-| Other multilingual content | 36 | 64.5% | 0.02557053 | 0.62 | 0.05177419 |
-| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.02250032 | 0.55 | 0.17071216 |
-| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.02042135 | 0.50 | 0.09928304 |
-| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.01983936 | 0.48 | 0.07883125 |
-| Scientific and technical exposition | 96 | 60.3% | 0.01691031 | 0.41 | 0.03146690 |
-
-**Natural dialogue, instruction following, and assistance** is this candidate's weakest domain and **Scientific and technical exposition** its strongest, a spread of 7.4x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
+| Natural dialogue, instruction following, and assistance | 96 | 73.5% | 0.12158332 | 3.00 | 0.81049308 |
+| Chinese across several content types | 72 | 55.6% | 0.05800970 | 1.43 | 0.23666146 |
+| Encyclopedic and factual reference | 96 | 59.5% | 0.03968052 | 0.98 | 0.24843901 |
+| News, history, economics, legal analysis, and essays | 72 | 59.0% | 0.02917507 | 0.72 | 0.08113657 |
+| Literary, narrative, and creative writing | 72 | 50.6% | 0.02912092 | 0.72 | 0.21794198 |
+| Other multilingual content | 36 | 64.5% | 0.02501305 | 0.62 | 0.04523514 |
+| Source code, tests, technical documentation, and issue discussions | 96 | 79.5% | 0.02270614 | 0.56 | 0.17765063 |
+| Structured data, tool calls, APIs, JSON, and tables | 36 | 70.7% | 0.02024896 | 0.50 | 0.10268640 |
+| Worked mathematics, science, and formal reasoning | 96 | 65.8% | 0.01951086 | 0.48 | 0.07172076 |
+| Scientific and technical exposition | 96 | 60.3% | 0.01656302 | 0.41 | 0.02738162 |
+
+**Natural dialogue, instruction following, and assistance** is this candidate's weakest domain and **Scientific and technical exposition** its strongest, a spread of 7.3x. A deployment weighted toward the weakest domain sees more divergence than the headline mean implies; the per-source breakdown and the reading are in [strata.md](strata.md) and [strata.json](strata.json).
 
 ## Trunk versus head
 
 | Component | Value |
 |---|---|
-| Trunk (candidate hidden states, reference head) | 0.03966656 |
-| Deployed (candidate's own head) | 0.04096872 |
-| Head-associated delta (not additive) | 0.00130217 |
+| Trunk (candidate hidden states, reference head) | 0.03923024 |
+| Deployed (candidate's own head) | 0.04053080 |
+| Head-associated delta (not additive) | 0.00130056 |
 | Reference head | {'state': 'unquantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'UnquantizedEmbeddingMethod', 'state': 'unquantized', 'weight_dtype': 'torch.bfloat16'}, 'state': 'unquantized'}]} |
 | Candidate head | {'runtime': {'state': 'quantized', 'workers': [{'heads': [{'name': 'language_model.lm_head', 'org_vocab_size': 248320, 'quant_method': 'CompressedTensorsLinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.float8_e4m3fn'}], 'logits_processor': {'head_dtype': 'torch.bfloat16', 'org_vocab_size': 248320, 'scale': 1.0, 'soft_cap': None, 'type': 'LogitsProcessor', 'vocab_size': 248320}, 'primary': {'org_vocab_size': 248320, 'quant_method': 'CompressedTensorsLinearMethod', 'state': 'quantized', 'weight_dtype': 'torch.float8_e4m3fn'}, 'state': 'quantized'}]}, 'state': 'quantized', 'static': {'ignored': False, 'lm_head_dtypes': {'lm_head.weight': 'F8_E4M3', 'lm_head.weight_scale': 'BF16', 'model.language_model.embed_tokens.weight': 'BF16'}, 'lm_head_keys': ['lm_head.weight', 'lm_head.weight_scale'], 'output_weight_keys': ['lm_head.weight', 'lm_head.weight_scale'], 'packed_keys': ['lm_head.weight_scale'], 'quant_method': 'compressed-tensors', 'state': 'quantized', 'tie_word_embeddings': False}} |
 
@@ -70,26 +77,26 @@ The suite is stratified, so the mean above is an average over kinds of text that
 
 | Position range | Positions | Mean KLD |
 |---|---|---|
-| 0–511 | 393216 | 0.03419946 |
-| 512–1023 | 393216 | 0.03730702 |
-| 1024–1535 | 393216 | 0.04390492 |
-| 1536–2046 | 392448 | 0.04847816 |
+| 0–511 | 393216 | 0.03439430 |
+| 512–1023 | 393216 | 0.03809782 |
+| 1024–1535 | 393216 | 0.04331275 |
+| 1536–2046 | 392448 | 0.04632964 |
 
 ## Error by reference confidence
 
 | Reference top-1 probability | Positions | Share | Mean KLD |
 |---|---|---|---|
-| [0.00, 0.25) | 240569 | 15.3% | 0.04418532 |
-| [0.25, 0.50) | 346390 | 22.0% | 0.05743935 |
-| [0.50, 0.75) | 273384 | 17.4% | 0.06343079 |
-| [0.75, 0.95) | 249939 | 15.9% | 0.04697047 |
-| [0.95, 1.00) | 461814 | 29.4% | 0.01039384 |
+| [0.00, 0.25) | 240522 | 15.3% | 0.04417294 |
+| [0.25, 0.50) | 346508 | 22.0% | 0.05791929 |
+| [0.50, 0.75) | 273204 | 17.4% | 0.06115178 |
+| [0.75, 0.95) | 250041 | 15.9% | 0.04707342 |
+| [0.95, 1.00) | 461821 | 29.4% | 0.00984591 |
 
 ## Top-K set agreement
 
 | K=1 | K=2 | K=3 | K=4 | K=5 |
 |---|---|---|---|---|
-| 92.4652% | 76.5753% | 58.2991% | 41.4397% | 28.2154% |
+| 92.5520% | 76.7247% | 58.5141% | 41.6387% | 28.4032% |
 
 ## Law compliance
 
@@ -101,22 +108,19 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | 2 | Determinism | PASS | eager enforced; prefix caching off, max_num_seqs=1 |
 | 3 | Frozen input | PASS | token hash matches suite qwen3.8-27b-fidelity-1024x2048-v1 [analysis]: 9c935708bbffbf45 |
 | 4 | Real vocabulary | PASS | scored 248044 real tokens of 248320 declared (276 padding rows) |
-| 5 | Manifest binding | PASS | all bound fields present; manifest 304794b4e35b326dee2486603652ea94215f869ba009c43ecaeab18938a22adc |
-| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 398f63d84a45, driver 580.173.02, 4 GPU(s) |
+| 5 | Manifest binding | PASS | all bound fields present; manifest dbfacc9cbb6d3e06edb143112389e9792da61d28bc8ceccd9b32da216ba3b7e3 |
+| 6 | Provenance | PASS | torch 2.13.0+cu132, vLLM 0.1.dev20446+gb2bc9171d @ 60071d1ab732, driver 580.173.02, 4 GPU(s) |
 | 7 | Storage integrity | PASS | hidden storage with bitwise-exact replay |
-| 8 | Head transparency | PASS | trunk 0.03966656, deployed 0.04096872, delta 0.0013021660852515493 |
-| 9 | Tail and depth disclosure | PASS | mean 0.04096872, median 0.00944244, max 36.54566574, 4 depth buckets |
+| 8 | Head transparency | PASS | trunk 0.03923024, deployed 0.04053080, delta 0.0013005567674299265 |
+| 9 | Tail and depth disclosure | PASS | mean 0.04053080, median 0.00929746, max 28.66598129, 4 depth buckets |
 | 10 | Comparability | PASS | comparability key fully resolved |
 | 11 | Freeze before qualification | NOT_APPLICABLE | partition is 'analysis' |
 | 12 | Reusable reference | PASS | suite, reference, head, and checksums present; published reference matches the scored capture on every bound field |
-| 14 | Component attribution | NOT_APPLICABLE | reference declares no experts |
-| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.12457770, strongest scientific_technical at 0.01691031, spread 7.4x |
-| 16 | Candidate weight binding | OVERRIDE | the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else |
-| 13 | Recorded deviation | PASS | 1 override(s), each fully attributed |
-
-### Recorded deviations
-
-- **Law 16 (Candidate weight binding)** overridden by Andy Kitzke at 2026-09-02T05:15:00Z. Justification: These candidates were scored under laws version 7, before Law 16 required a digest of the weights each report read. The campaign runs with fetch=lease, so every candidate's weights were released immediately after scoring and no digest can be recovered from them now. The measurement itself is unaffected: the tokens, the capture, and the reference remain bound under Laws 3, 5, and 12, and each candidate still pins its hf_repo and revision. What these results cannot prove is that the directory scored held the repo it names. They publish with their weights marked unbound, and every measurement taken after laws version 8 is bound at score time. A second case is covered by the same reasoning: a report bound to a digest at score time whose inspection cannot be re-taken, because the leased weights were released before assembly and the inspection record was cleared. Such a report states which weights it read and is simply uncorroborated, which is the weaker of the two positions here, not the stronger. Underlying finding: the report names a checkpoint path but records no digest of its weights, so nothing rules out a directory that held something else.
+| 14 | Routed-model intervention | NOT_APPLICABLE | reference declares no experts |
+| 15 | Domain disclosure | PASS | 10 domains disclosed; weakest dialogue_instruction at 0.12158332, strongest scientific_technical at 0.01656302, spread 7.3x |
+| 16 | Candidate weight binding | PASS | scored weights 529539fc37109382 as inspected |
+| 17 | Substitution disclosure | PASS | every scored layer used the checkpoint's own quantization parameters |
+| 13 | Recorded deviation | PASS | no overrides claimed |
 
 ## Environment
 
@@ -126,7 +130,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | Platform | Linux-6.8.0-137-generic-x86_64-with-glibc2.39 |
 | Python | 3.12.3 |
 | vLLM | 0.1.dev20446+gb2bc9171d |
-| vLLM commit | 398f63d84a45fbe0b0205e70893bcd28fa928a88 |
+| vLLM commit | 60071d1ab73229321712254b1350dcaaac1c125a |
 | torch | 2.13.0+cu132 |
 | torch CUDA runtime | 13.2 |
 | cuDNN | 9.20.0 (92000) |
@@ -134,7 +138,7 @@ Zero baseline: reference against itself scored **0.00000000** over 2047 position
 | CUDA arch list | sm_75, sm_80, sm_86, sm_90, sm_100, sm_120 |
 | nvcc | Cuda compilation tools, release 13.0, V13.0.88 |
 | gcc | gcc (Ubuntu 13.3.0-6ubuntu2~24.04.1) 13.3.0 |
-| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.8) 2.39 |
+| glibc / ldd | ldd (Ubuntu GLIBC 2.39-0ubuntu8.9) 2.39 |
 | NVIDIA driver | 580.173.02 |
 | float32 matmul precision | highest |
 | TF32 (matmul / cuDNN) | False / True |
@@ -157,8 +161,18 @@ Credential values are never published. Set but redacted: `HF_TOKEN`.
 
 | Variable | Value |
 |---|---|
+| `CUBLAS_WORKSPACE_CONFIG` | `:4096:8` |
 | `CUDA_HOME` | `/usr/local/cuda-13.0` |
 | `HF_TOKEN` | `<redacted>` |
+| `NCCL_DETERMINISTIC` | `1` |
+| `PYTORCH_NVML_BASED_CUDA_CHECK` | `1` |
+| `TORCHINDUCTOR_CACHE_DIR` | `/tmp/torchinductor_phaedawg` |
+| `TORCHINDUCTOR_COMPILE_THREADS` | `1` |
+| `TRITON_CACHE_AUTOTUNING` | `1` |
+| `TRITON_PTXAS_BLACKWELL_PATH` | `/usr/local/cuda-13.0/bin/ptxas` |
+| `VLLM_BATCH_INVARIANT` | `1` |
+| `VLLM_MARLIN_USE_ATOMIC_ADD` | `0` |
+| `VLLM_MOE_USE_DEEP_GEMM` | `0` |
 
 ## Files in this artifact
 
@@ -166,26 +180,26 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 
 | Path | Size | What it is |
 |---|---|---|
-| `Qwen3.8-27B-NVFP4/report.md` | 12.43 KiB | This document. |
-| `Qwen3.8-27B-NVFP4/report.json` | 254.73 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
-| `Qwen3.8-27B-NVFP4/manifest.json` | 85.43 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
-| `Qwen3.8-27B-NVFP4/compliance.json` | 13.02 KiB | The law-by-law receipt, including the comparability key. |
-| `baselines/self-kld.json` | 4.74 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
+| `Qwen3.8-27B-NVFP4/report.md` | 13.27 KiB | This document. |
+| `Qwen3.8-27B-NVFP4/report.json` | 257.26 KiB | Every statistic behind it, machine-readable: per-bucket means, percentiles, agreement rates, and the phase timings. |
+| `Qwen3.8-27B-NVFP4/manifest.json` | 86.99 KiB | The capture manifest this result is bound to (Law 5). Scoring refuses to run if the live configuration differs from it. |
+| `Qwen3.8-27B-NVFP4/compliance.json` | 12.29 KiB | The law-by-law receipt, including the comparability key. |
+| `baselines/self-kld.json` | 7.06 KiB | The zero-baseline proof required by Law 1: the reference scored against a capture of itself. |
 | `suite/suite-manifest.json` | 382.68 KiB | The frozen evaluation input's identity: token hashes per context and per partition, sources, strata, and the analysis/qualification split. |
 | `suite/tokens` | 15.49 MiB in 1024 files | The token IDs themselves. These are the evaluation input, not a description of it; retokenizing source text does not reproduce them. |
 | `suite/sources.json` | 737.42 KiB | Per-context provenance: dataset, revision, licence, source unit, and the deterministic token offset chosen within the document. |
 | `suite/validation/capability-overlap.json` | 1.48 KiB | The benchmark-contamination scan and every document it blocked. |
-| `reference/manifest.json` | 85.43 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
+| `reference/manifest.json` | 86.99 KiB | The reference capture's own manifest: geometry, vocabulary, storage mode, and a hash for every tensor file. |
 | `reference` | 19.25 GiB in 771 files | The reusable reference distributions. Pass this directory as `--reference-logits` to score a new candidate against the same reference without loading the reference checkpoint. |
 | `reference/lm_head.safetensors` | 2.37 GiB | The reference language-model head, which turns the stored hidden states back into reference logits. |
-| `environment/runtime.json` | 2.26 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
+| `environment/runtime.json` | 5.42 KiB | Machine-readable provenance: torch, CUDA, cuDNN, NCCL, driver, devices, and the captured environment variables. |
 | `environment/summary.md` | 1.18 KiB | The same provenance as prose, plus an index of every captured file. |
 | `environment/toolchain-nvcc.txt` | 243 B | `nvcc --version` verbatim; `toolchain-gcc.txt` and `toolchain-ldd.txt` sit beside it. |
 | `environment/gpu-smi-query.txt` | 54.73 KiB | `nvidia-smi -q` verbatim: ECC state, persistence mode, clocks, and throttle reasons, any of which can move a bitwise result. |
-| `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
-| `environment/models` | 7.22 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
-| `checksums.txt` | 182.99 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
+| `environment/pip-freeze.txt` | 4.47 KiB | Every installed package version in the scoring environment. |
+| `environment/models` | 19.46 KiB in 10 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
+| `checksums.txt` | 183.22 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
+| `LAWS.md` | 40.20 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

Qwen3.8-27B-NVFP4/strata.json

@@ -8,381 +8,381 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 36.5456657409668,
-            "mean_kld": 0.12457769541074566,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.09409491196048908,
-            "median_context_p99": 2.0393991470336914,
-            "p90_context_kld": 0.2799708803626206,
+            "max_kld": 28.66598129272461,
+            "mean_kld": 0.12158332374682436,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.08575015208991663,
+            "median_context_p99": 1.6861377954483032,
+            "p90_context_kld": 0.29358947054916235,
             "positions": 196512,
-            "top1_agreement": 0.923175175052923,
+            "top1_agreement": 0.9250681892199968,
             "worst_context_id": 454,
-            "worst_context_kld": 0.8362269388706521
+            "worst_context_kld": 0.8104930786551602
           }
         },
         "key": "wildchat",
         "label": "wildchat",
-        "relative_to_run": 3.0407999632099014
+        "relative_to_run": 2.999776270644905
       },
       {
         "cells": {
           "deployed": {
             "contexts": 33,
-            "max_kld": 6.911944389343262,
-            "mean_kld": 0.09430849689233974,
-            "mean_ref_top1_prob": 0.6040606719329803,
-            "median_context_kld": 0.06800715166529955,
-            "median_context_p99": 0.5470696687698364,
-            "p90_context_kld": 0.2165044876798134,
+            "max_kld": 6.782408714294434,
+            "mean_kld": 0.09551845590935071,
+            "mean_ref_top1_prob": 0.6040861921682756,
+            "median_context_kld": 0.07310532158913022,
+            "median_context_p99": 0.5301501750946045,
+            "p90_context_kld": 0.21919990539863463,
             "positions": 67551,
-            "top1_agreement": 0.8718597800180604,
-            "worst_context_id": 893,
-            "worst_context_kld": 0.23709843126137567
+            "top1_agreement": 0.8712084203046587,
+            "worst_context_id": 917,
+            "worst_context_kld": 0.2366614644814825
           }
         },
         "key": "wikisource_zh",
         "label": "wikisource_zh",
-        "relative_to_run": 2.3019632281291313
+        "relative_to_run": 2.3566883073715594
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 6.266860008239746,
-            "mean_kld": 0.0422661248729009,
-            "mean_ref_top1_prob": 0.7232628548554616,
-            "median_context_kld": 0.05356200102035192,
-            "median_context_p99": 0.696960985660553,
-            "p90_context_kld": 0.07344522910392678,
+            "max_kld": 4.948120594024658,
+            "mean_kld": 0.042288741216326774,
+            "mean_ref_top1_prob": 0.7233191001139795,
+            "median_context_kld": 0.0524429853328287,
+            "median_context_p99": 0.6635613441467285,
+            "p90_context_kld": 0.07412235440737637,
             "positions": 47081,
-            "top1_agreement": 0.9291009111955991,
+            "top1_agreement": 0.9314160701769291,
             "worst_context_id": 275,
-            "worst_context_kld": 0.0819730113510625
+            "worst_context_kld": 0.07787789349897044
           }
         },
         "key": "regulations",
         "label": "regulations",
-        "relative_to_run": 1.0316680729627319
+        "relative_to_run": 1.0433730425098169
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 7,
-            "max_kld": 7.721675872802734,
-            "mean_kld": 0.04007121415198143,
-            "mean_ref_top1_prob": 0.7357366218042676,
-            "median_context_kld": 0.04305093457533655,
-            "median_context_p99": 0.6485264897346497,
-            "p90_context_kld": 0.051774189189553635,
-            "positions": 14329,
-            "top1_agreement": 0.9239304906134412,
-            "worst_context_id": 953,
-            "worst_context_kld": 0.051774189189553635
+            "contexts": 96,
+            "max_kld": 6.517783164978027,
+            "mean_kld": 0.03968051556345783,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.030948956898133834,
+            "median_context_p99": 0.3017843961715698,
+            "p90_context_kld": 0.05457949171108933,
+            "positions": 196512,
+            "top1_agreement": 0.9173027601367856,
+            "worst_context_id": 65,
+            "worst_context_kld": 0.24843901303775767
           }
         },
-        "key": "wikipedia_de",
-        "label": "wikipedia_de",
-        "relative_to_run": 0.9780927967673011
+        "key": "wikipedia_en",
+        "label": "wikipedia_en",
+        "relative_to_run": 0.9790213437665266
       },
       {
         "cells": {
           "deployed": {
-            "contexts": 96,
-            "max_kld": 5.581716537475586,
-            "mean_kld": 0.03972593213279645,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.03066756860698419,
-            "median_context_p99": 0.291412353515625,
-            "p90_context_kld": 0.05708804087610574,
-            "positions": 196512,
-            "top1_agreement": 0.9163511643054877,
-            "worst_context_id": 65,
-            "worst_context_kld": 0.2408369537170143
+            "contexts": 7,
+            "max_kld": 6.416473865509033,
+            "mean_kld": 0.037692824229436095,
+            "mean_ref_top1_prob": 0.7358200961573914,
+            "median_context_kld": 0.04138496706841051,
+            "median_context_p99": 0.6894661784172058,
+            "p90_context_kld": 0.0452351423921429,
+            "positions": 14329,
+            "top1_agreement": 0.9207900062809686,
+            "worst_context_id": 953,
+            "worst_context_kld": 0.0452351423921429
           }
         },
-        "key": "wikipedia_en",
-        "label": "wikipedia_en",
-        "relative_to_run": 0.9696648550898366
+        "key": "wikipedia_de",
+        "label": "wikipedia_de",
+        "relative_to_run": 0.9299798377983148
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.997079849243164,
-            "mean_kld": 0.029428653535056526,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.02369674456730006,
-            "median_context_p99": 0.15645328164100647,
-            "p90_context_kld": 0.034802640181948945,
+            "max_kld": 6.705466270446777,
+            "mean_kld": 0.029120922214834796,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.023827455457346176,
+            "median_context_p99": 0.14993515610694885,
+            "p90_context_kld": 0.033321719956789224,
             "positions": 147384,
-            "top1_agreement": 0.9151671823264398,
+            "top1_agreement": 0.9155810671443304,
             "worst_context_id": 443,
-            "worst_context_kld": 0.21936384601423634
+            "worst_context_kld": 0.21794198408233922
           }
         },
         "key": "public_domain_books",
         "label": "public_domain_books",
-        "relative_to_run": 0.7183199873112919
+        "relative_to_run": 0.7184887593734594
       },
       {
         "cells": {
           "deployed": {
             "contexts": 26,
-            "max_kld": 3.1836166381835938,
-            "mean_kld": 0.026675995630818197,
-            "mean_ref_top1_prob": 0.5182113024080622,
-            "median_context_kld": 0.023769710039784613,
-            "median_context_p99": 0.174808531999588,
-            "p90_context_kld": 0.04163838728856443,
+            "max_kld": 4.11423921585083,
+            "mean_kld": 0.027032613734328758,
+            "mean_ref_top1_prob": 0.5182496321630543,
+            "median_context_kld": 0.023448965647329817,
+            "median_context_p99": 0.17007294297218323,
+            "p90_context_kld": 0.044098667591478326,
             "positions": 53222,
-            "top1_agreement": 0.9193190785765285,
+            "top1_agreement": 0.9194318139115404,
             "worst_context_id": 312,
-            "worst_context_kld": 0.07495626607201038
+            "worst_context_kld": 0.0811365691669795
           }
         },
         "key": "public_domain_review",
         "label": "public_domain_review",
-        "relative_to_run": 0.6511307362472097
+        "relative_to_run": 0.6669647671633668
       },
       {
         "cells": {
           "deployed": {
             "contexts": 39,
-            "max_kld": 6.2669878005981445,
-            "mean_kld": 0.026263908507732442,
-            "mean_ref_top1_prob": 0.515358929042611,
-            "median_context_kld": 0.024238450137424036,
-            "median_context_p99": 0.18913322687149048,
-            "p90_context_kld": 0.03705976399824896,
+            "max_kld": 6.276115894317627,
+            "mean_kld": 0.02627151407175288,
+            "mean_ref_top1_prob": 0.5153760832971095,
+            "median_context_kld": 0.023902755651442852,
+            "median_context_p99": 0.1772077977657318,
+            "p90_context_kld": 0.038946215648654386,
             "positions": 79833,
-            "top1_agreement": 0.9126301153658262,
-            "worst_context_id": 886,
-            "worst_context_kld": 0.04330919491649525
+            "top1_agreement": 0.9140956747209801,
+            "worst_context_id": 880,
+            "worst_context_kld": 0.04410864426378436
           }
         },
         "key": "wikipedia_zh",
         "label": "wikipedia_zh",
-        "relative_to_run": 0.641072157907107
+        "relative_to_run": 0.6481864624005758
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 2.5077576637268066,
-            "mean_kld": 0.024812117374526384,
-            "mean_ref_top1_prob": 0.5702914711488575,
-            "median_context_kld": 0.025088542724129696,
-            "median_context_p99": 0.17169347405433655,
-            "p90_context_kld": 0.029367452228230048,
+            "max_kld": 1.6534563302993774,
+            "mean_kld": 0.024447840130642398,
+            "mean_ref_top1_prob": 0.5701150598950927,
+            "median_context_kld": 0.024018356789416235,
+            "median_context_p99": 0.17541679739952087,
+            "p90_context_kld": 0.02978178680875595,
             "positions": 14329,
-            "top1_agreement": 0.9192546583850931,
-            "worst_context_id": 950,
-            "worst_context_kld": 0.029367452228230048
+            "top1_agreement": 0.9228138739618954,
+            "worst_context_id": 968,
+            "worst_context_kld": 0.02978178680875595
           }
         },
         "key": "wikipedia_ja",
         "label": "wikipedia_ja",
-        "relative_to_run": 0.6056355862970284
+        "relative_to_run": 0.6031916913633216
       },
       {
         "cells": {
           "deployed": {
             "contexts": 52,
-            "max_kld": 8.132791519165039,
-            "mean_kld": 0.02405582642049702,
-            "mean_ref_top1_prob": 0.8582575125835092,
-            "median_context_kld": 0.014228856650553021,
-            "median_context_p99": 0.1700565665960312,
-            "p90_context_kld": 0.04196825792757604,
+            "max_kld": 10.22834300994873,
+            "mean_kld": 0.024163990682836575,
+            "mean_ref_top1_prob": 0.8582941084526002,
+            "median_context_kld": 0.014683369537724708,
+            "median_context_p99": 0.18004028499126434,
+            "p90_context_kld": 0.04043856255151532,
             "positions": 106444,
-            "top1_agreement": 0.9608150764721356,
+            "top1_agreement": 0.9615572507609635,
             "worst_context_id": 667,
-            "worst_context_kld": 0.1707121626802731
+            "worst_context_kld": 0.17765063071981094
           }
         },
         "key": "github_code",
         "label": "github_code",
-        "relative_to_run": 0.5871753836290788
+        "relative_to_run": 0.5961883885112247
       },
       {
         "cells": {
           "deployed": {
             "contexts": 7,
-            "max_kld": 3.2778303623199463,
-            "mean_kld": 0.02355379620898713,
-            "mean_ref_top1_prob": 0.591765751203999,
-            "median_context_kld": 0.021873228884030954,
-            "median_context_p99": 0.1554984450340271,
-            "p90_context_kld": 0.031223083698208594,
+            "max_kld": 2.3562636375427246,
+            "mean_kld": 0.023291095160471002,
+            "mean_ref_top1_prob": 0.591761624645417,
+            "median_context_kld": 0.02178277997013687,
+            "median_context_p99": 0.17793317139148712,
+            "p90_context_kld": 0.02853840221116572,
             "positions": 14329,
-            "top1_agreement": 0.9241398562356061,
+            "top1_agreement": 0.9272105520273571,
             "worst_context_id": 966,
-            "worst_context_kld": 0.031223083698208594
+            "worst_context_kld": 0.02853840221116572
           }
         },
         "key": "wikipedia_es",
         "label": "wikipedia_es",
-        "relative_to_run": 0.5749213967202966
+        "relative_to_run": 0.5746517896253694
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 3.428466796875,
-            "mean_kld": 0.02215582366887136,
-            "mean_ref_top1_prob": 0.6892386900239509,
-            "median_context_kld": 0.02272664081976165,
-            "median_context_p99": 0.18951201438903809,
-            "p90_context_kld": 0.02691429370744579,
+            "max_kld": 2.180680274963379,
+            "mean_kld": 0.02231558741892687,
+            "mean_ref_top1_prob": 0.689239984531695,
+            "median_context_kld": 0.023257077772700495,
+            "median_context_p99": 0.2136688232421875,
+            "p90_context_kld": 0.02581685553262255,
             "positions": 12282,
-            "top1_agreement": 0.9307930304510666,
+            "top1_agreement": 0.9338869890897248,
             "worst_context_id": 958,
-            "worst_context_kld": 0.02691429370744579
+            "worst_context_kld": 0.02581685553262255
           }
         },
         "key": "wikipedia_cs",
         "label": "wikipedia_cs",
-        "relative_to_run": 0.5407984758030597
+        "relative_to_run": 0.5505834808743437
       },
       {
         "cells": {
           "deployed": {
             "contexts": 44,
-            "max_kld": 7.012263298034668,
-            "mean_kld": 0.020661989664897726,
-            "mean_ref_top1_prob": 0.720394649593742,
-            "median_context_kld": 0.016385709625882167,
-            "median_context_p99": 0.16171938180923462,
-            "p90_context_kld": 0.029012030879484348,
+            "max_kld": 12.169252395629883,
+            "mean_kld": 0.020983219483233187,
+            "mean_ref_top1_prob": 0.7203910827522294,
+            "median_context_kld": 0.01634585862576508,
+            "median_context_p99": 0.16236217319965363,
+            "p90_context_kld": 0.033880059242524525,
             "positions": 90068,
-            "top1_agreement": 0.949626948527779,
+            "top1_agreement": 0.9507261180441444,
             "worst_context_id": 582,
-            "worst_context_kld": 0.12324286582467237
+            "worst_context_kld": 0.1300324749810704
           }
         },
         "key": "stackv2",
         "label": "stackv2",
-        "relative_to_run": 0.5043356854990027
+        "relative_to_run": 0.5177105045969005
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 4.752742767333984,
-            "mean_kld": 0.020421348473034326,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.01775127215107961,
-            "median_context_p99": 0.22656621038913727,
-            "p90_context_kld": 0.03592128721473434,
+            "max_kld": 5.577369689941406,
+            "mean_kld": 0.020248958642327196,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.015101154141245148,
+            "median_context_p99": 0.2136673778295517,
+            "p90_context_kld": 0.03390113755388437,
             "positions": 73692,
-            "top1_agreement": 0.882931661510069,
+            "top1_agreement": 0.8844786408293981,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.09928303706595111
+            "worst_context_kld": 0.10268640378926563
           }
         },
         "key": "starcoder_structured",
         "label": "starcoder_structured",
-        "relative_to_run": 0.4984619074928163
+        "relative_to_run": 0.4995943832478895
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.103275299072266,
-            "mean_kld": 0.01983936171590355,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.016081149073808487,
-            "median_context_p99": 0.11479724943637848,
-            "p90_context_kld": 0.03028211425308765,
+            "max_kld": 11.945074081420898,
+            "mean_kld": 0.01951085562769419,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.015771377783579774,
+            "median_context_p99": 0.11416608095169067,
+            "p90_context_kld": 0.029937825294863343,
             "positions": 196512,
-            "top1_agreement": 0.9445988031265267,
+            "top1_agreement": 0.9449143054877056,
             "worst_context_id": 825,
-            "worst_context_kld": 0.07883124659302003
+            "worst_context_kld": 0.07172076344828876
           }
         },
         "key": "libretexts",
         "label": "libretexts",
-        "relative_to_run": 0.48425627217553896
+        "relative_to_run": 0.48138346549737526
       },
       {
         "cells": {
           "deployed": {
             "contexts": 6,
-            "max_kld": 1.49409019947052,
-            "mean_kld": 0.019274730503398098,
-            "mean_ref_top1_prob": 0.6660615117323249,
-            "median_context_kld": 0.018592186564097727,
-            "median_context_p99": 0.15590962767601013,
-            "p90_context_kld": 0.026276644368793366,
+            "max_kld": 3.389317512512207,
+            "mean_kld": 0.019194425117889727,
+            "mean_ref_top1_prob": 0.6660535411394803,
+            "median_context_kld": 0.018331660641119785,
+            "median_context_p99": 0.14996421337127686,
+            "p90_context_kld": 0.02629231236218796,
             "positions": 12282,
-            "top1_agreement": 0.936248168050806,
+            "top1_agreement": 0.9363295880149812,
             "worst_context_id": 945,
-            "worst_context_kld": 0.026276644368793366
+            "worst_context_kld": 0.02629231236218796
           }
         },
         "key": "wikipedia_ru",
         "label": "wikipedia_ru",
-        "relative_to_run": 0.4704742659780985
+        "relative_to_run": 0.47357630325367744
       },
       {
         "cells": {
           "deployed": {
             "contexts": 23,
-            "max_kld": 3.4781346321105957,
-            "mean_kld": 0.018525764368715465,
-            "mean_ref_top1_prob": 0.5379292386352562,
-            "median_context_kld": 0.018225347523296576,
-            "median_context_p99": 0.127279132604599,
-            "p90_context_kld": 0.02545460794392238,
+            "max_kld": 3.6872079372406006,
+            "mean_kld": 0.01848331120135734,
+            "mean_ref_top1_prob": 0.537879873360319,
+            "median_context_kld": 0.017542628717267538,
+            "median_context_p99": 0.11397124081850052,
+            "p90_context_kld": 0.024723821286509063,
             "positions": 47081,
-            "top1_agreement": 0.9296956309339224,
+            "top1_agreement": 0.927784031775026,
             "worst_context_id": 343,
-            "worst_context_kld": 0.02797805131289826
+            "worst_context_kld": 0.029905186844627056
           }
         },
         "key": "open_news",
         "label": "open_news",
-        "relative_to_run": 0.45219285382579144
+        "relative_to_run": 0.4560312766266608
       },
       {
         "cells": {
           "deployed": {
             "contexts": 3,
-            "max_kld": 1.2447932958602905,
-            "mean_kld": 0.017631999117552123,
-            "mean_ref_top1_prob": 0.6023603953518782,
-            "median_context_kld": 0.018434458582606405,
-            "median_context_p99": 0.1548728197813034,
-            "p90_context_kld": 0.01930741002503341,
+            "max_kld": 1.0351228713989258,
+            "mean_kld": 0.017795808168851582,
+            "mean_ref_top1_prob": 0.6023746847155801,
+            "median_context_kld": 0.018796216898064363,
+            "median_context_p99": 0.15429577231407166,
+            "p90_context_kld": 0.019356053005890286,
             "positions": 6141,
-            "top1_agreement": 0.9386093470118874,
+            "top1_agreement": 0.9390978667969386,
             "worst_context_id": 967,
-            "worst_context_kld": 0.01930741002503341
+            "worst_context_kld": 0.019356053005890286
           }
         },
         "key": "wikipedia_fr",
         "label": "wikipedia_fr",
-        "relative_to_run": 0.4303770598034744
+        "relative_to_run": 0.439068791810884
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 4.14390754699707,
-            "mean_kld": 0.01691031048838098,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.016672932853405074,
-            "median_context_p99": 0.1243298128247261,
-            "p90_context_kld": 0.021205758610799045,
+            "max_kld": 3.9671146869659424,
+            "mean_kld": 0.01656302498706322,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.016550733119089625,
+            "median_context_p99": 0.12118519842624664,
+            "p90_context_kld": 0.021278483650238936,
             "positions": 196512,
-            "top1_agreement": 0.9386093470118874,
-            "worst_context_id": 246,
-            "worst_context_kld": 0.031466900875017814
+            "top1_agreement": 0.9393624816805081,
+            "worst_context_id": 174,
+            "worst_context_kld": 0.027381622246614497
           }
         },
         "key": "scientific_papers",
         "label": "scientific_papers",
-        "relative_to_run": 0.4127614605599896
+        "relative_to_run": 0.40865282996993746
       }
     ],
     "stratum": [
@@ -390,217 +390,217 @@
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 36.5456657409668,
-            "mean_kld": 0.12457769541074566,
-            "mean_ref_top1_prob": 0.7352292693891235,
-            "median_context_kld": 0.09409491196048908,
-            "median_context_p99": 2.0393991470336914,
-            "p90_context_kld": 0.2799708803626206,
+            "max_kld": 28.66598129272461,
+            "mean_kld": 0.12158332374682436,
+            "mean_ref_top1_prob": 0.735242277628431,
+            "median_context_kld": 0.08575015208991663,
+            "median_context_p99": 1.6861377954483032,
+            "p90_context_kld": 0.29358947054916235,
             "positions": 196512,
-            "top1_agreement": 0.923175175052923,
+            "top1_agreement": 0.9250681892199968,
             "worst_context_id": 454,
-            "worst_context_kld": 0.8362269388706521
+            "worst_context_kld": 0.8104930786551602
           }
         },
         "key": "dialogue_instruction",
         "label": "Natural dialogue, instruction following, and assistance",
-        "relative_to_run": 3.0407999632099014
+        "relative_to_run": 2.999776270644905
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.911944389343262,
-            "mean_kld": 0.05745101151734411,
-            "mean_ref_top1_prob": 0.5560138945340303,
-            "median_context_kld": 0.034198077382997016,
-            "median_context_p99": 0.27472054958343506,
-            "p90_context_kld": 0.12171763135253476,
+            "max_kld": 6.782408714294434,
+            "mean_kld": 0.058009695747318554,
+            "mean_ref_top1_prob": 0.5560348831963939,
+            "median_context_kld": 0.03501158816479015,
+            "median_context_p99": 0.27062809467315674,
+            "p90_context_kld": 0.1256960806689686,
             "positions": 147384,
-            "top1_agreement": 0.8939437116647668,
-            "worst_context_id": 893,
-            "worst_context_kld": 0.23709843126137567
+            "top1_agreement": 0.8944390164468328,
+            "worst_context_id": 917,
+            "worst_context_kld": 0.2366614644814825
           }
         },
         "key": "chinese",
         "label": "Chinese across several content types",
-        "relative_to_run": 1.4023138984255346
+        "relative_to_run": 1.4312498080122769
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 5.581716537475586,
-            "mean_kld": 0.03972593213279645,
-            "mean_ref_top1_prob": 0.5946008380921044,
-            "median_context_kld": 0.03066756860698419,
-            "median_context_p99": 0.291412353515625,
-            "p90_context_kld": 0.05708804087610574,
+            "max_kld": 6.517783164978027,
+            "mean_kld": 0.03968051556345783,
+            "mean_ref_top1_prob": 0.5945716883648833,
+            "median_context_kld": 0.030948956898133834,
+            "median_context_p99": 0.3017843961715698,
+            "p90_context_kld": 0.05457949171108933,
             "positions": 196512,
-            "top1_agreement": 0.9163511643054877,
+            "top1_agreement": 0.9173027601367856,
             "worst_context_id": 65,
-            "worst_context_kld": 0.2408369537170143
+            "worst_context_kld": 0.24843901303775767
           }
         },
         "key": "encyclopedic_reference",
         "label": "Encyclopedic and factual reference",
-        "relative_to_run": 0.9696648550898366
+        "relative_to_run": 0.9790213437665266
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 7.997079849243164,
-            "mean_kld": 0.029428653535056526,
-            "mean_ref_top1_prob": 0.5056158503386161,
-            "median_context_kld": 0.02369674456730006,
-            "median_context_p99": 0.15645328164100647,
-            "p90_context_kld": 0.034802640181948945,
+            "max_kld": 4.948120594024658,
+            "mean_kld": 0.0291750717041567,
+            "mean_ref_top1_prob": 0.590028705918726,
+            "median_context_kld": 0.023173652596112884,
+            "median_context_p99": 0.17660361528396606,
+            "p90_context_kld": 0.0621877665506871,
             "positions": 147384,
-            "top1_agreement": 0.9151671823264398,
-            "worst_context_id": 443,
-            "worst_context_kld": 0.21936384601423634
+            "top1_agreement": 0.9259281875915975,
+            "worst_context_id": 312,
+            "worst_context_kld": 0.0811365691669795
           }
         },
-        "key": "literary_narrative",
-        "label": "Literary, narrative, and creative writing",
-        "relative_to_run": 0.7183199873112919
+        "key": "news_history_legal_essays",
+        "label": "News, history, economics, legal analysis, and essays",
+        "relative_to_run": 0.7198247678664795
       },
       {
         "cells": {
           "deployed": {
             "contexts": 72,
-            "max_kld": 6.266860008239746,
-            "mean_kld": 0.029052629707756242,
-            "mean_ref_top1_prob": 0.5900126668457796,
-            "median_context_kld": 0.021822250241878293,
-            "median_context_p99": 0.17231622338294983,
-            "p90_context_kld": 0.06020746195574988,
+            "max_kld": 6.705466270446777,
+            "mean_kld": 0.029120922214834796,
+            "mean_ref_top1_prob": 0.5055996513127425,
+            "median_context_kld": 0.023827455457346176,
+            "median_context_p99": 0.14993515610694885,
+            "p90_context_kld": 0.033321719956789224,
             "positions": 147384,
-            "top1_agreement": 0.9257585626662325,
-            "worst_context_id": 275,
-            "worst_context_kld": 0.0819730113510625
+            "top1_agreement": 0.9155810671443304,
+            "worst_context_id": 443,
+            "worst_context_kld": 0.21794198408233922
           }
         },
-        "key": "news_history_legal_essays",
-        "label": "News, history, economics, legal analysis, and essays",
-        "relative_to_run": 0.709141673035604
+        "key": "literary_narrative",
+        "label": "Literary, narrative, and creative writing",
+        "relative_to_run": 0.7184887593734594
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 7.721675872802734,
-            "mean_kld": 0.025570533792631602,
-            "mean_ref_top1_prob": 0.6450955362692544,
-            "median_context_kld": 0.0227388942645489,
-            "median_context_p99": 0.1798495501279831,
-            "p90_context_kld": 0.04305093457533655,
+            "max_kld": 6.416473865509033,
+            "mean_kld": 0.025013050454758354,
+            "mean_ref_top1_prob": 0.6450767409183084,
+            "median_context_kld": 0.022881467166419225,
+            "median_context_p99": 0.17793317139148712,
+            "p90_context_kld": 0.04138496706841051,
             "positions": 73692,
-            "top1_agreement": 0.9274819519079411,
+            "top1_agreement": 0.9287303913586278,
             "worst_context_id": 953,
-            "worst_context_kld": 0.051774189189553635
+            "worst_context_kld": 0.0452351423921429
           }
         },
         "key": "other_multilingual",
         "label": "Other multilingual content",
-        "relative_to_run": 0.6241476691274932
+        "relative_to_run": 0.617136897547495
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 8.132791519165039,
-            "mean_kld": 0.02250031790751401,
-            "mean_ref_top1_prob": 0.7950703670465326,
-            "median_context_kld": 0.015097981818138071,
-            "median_context_p99": 0.16505374014377594,
-            "p90_context_kld": 0.03532434663282433,
+            "max_kld": 12.169252395629883,
+            "mean_kld": 0.02270613721635169,
+            "mean_ref_top1_prob": 0.7950885550065969,
+            "median_context_kld": 0.015219370374946644,
+            "median_context_p99": 0.176174134016037,
+            "p90_context_kld": 0.03640113065999465,
             "positions": 196512,
-            "top1_agreement": 0.9556871844976388,
+            "top1_agreement": 0.9565929815990881,
             "worst_context_id": 667,
-            "worst_context_kld": 0.1707121626802731
+            "worst_context_kld": 0.17765063071981094
           }
         },
         "key": "code_docs_issues",
         "label": "Source code, tests, technical documentation, and issue discussions",
-        "relative_to_run": 0.5492071886527938
+        "relative_to_run": 0.5602193583838262
       },
       {
         "cells": {
           "deployed": {
             "contexts": 36,
-            "max_kld": 4.752742767333984,
-            "mean_kld": 0.020421348473034326,
-            "mean_ref_top1_prob": 0.7071241246993897,
-            "median_context_kld": 0.01775127215107961,
-            "median_context_p99": 0.22656621038913727,
-            "p90_context_kld": 0.03592128721473434,
+            "max_kld": 5.577369689941406,
+            "mean_kld": 0.020248958642327196,
+            "mean_ref_top1_prob": 0.7071550040111672,
+            "median_context_kld": 0.015101154141245148,
+            "median_context_p99": 0.2136673778295517,
+            "p90_context_kld": 0.03390113755388437,
             "positions": 73692,
-            "top1_agreement": 0.882931661510069,
+            "top1_agreement": 0.8844786408293981,
             "worst_context_id": 1004,
-            "worst_context_kld": 0.09928303706595111
+            "worst_context_kld": 0.10268640378926563
           }
         },
         "key": "structured_data_tools",
         "label": "Structured data, tool calls, APIs, JSON, and tables",
-        "relative_to_run": 0.4984619074928163
+        "relative_to_run": 0.4995943832478895
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 12.103275299072266,
-            "mean_kld": 0.01983936171590355,
-            "mean_ref_top1_prob": 0.6582818745125177,
-            "median_context_kld": 0.016081149073808487,
-            "median_context_p99": 0.11479724943637848,
-            "p90_context_kld": 0.03028211425308765,
+            "max_kld": 11.945074081420898,
+            "mean_kld": 0.01951085562769419,
+            "mean_ref_top1_prob": 0.6582834184806026,
+            "median_context_kld": 0.015771377783579774,
+            "median_context_p99": 0.11416608095169067,
+            "p90_context_kld": 0.029937825294863343,
             "positions": 196512,
-            "top1_agreement": 0.9445988031265267,
+            "top1_agreement": 0.9449143054877056,
             "worst_context_id": 825,
-            "worst_context_kld": 0.07883124659302003
+            "worst_context_kld": 0.07172076344828876
           }
         },
         "key": "worked_math_reasoning",
         "label": "Worked mathematics, science, and formal reasoning",
-        "relative_to_run": 0.48425627217553896
+        "relative_to_run": 0.48138346549737526
       },
       {
         "cells": {
           "deployed": {
             "contexts": 96,
-            "max_kld": 4.14390754699707,
-            "mean_kld": 0.01691031048838098,
-            "mean_ref_top1_prob": 0.6033561622407548,
-            "median_context_kld": 0.016672932853405074,
-            "median_context_p99": 0.1243298128247261,
-            "p90_context_kld": 0.021205758610799045,
+            "max_kld": 3.9671146869659424,
+            "mean_kld": 0.01656302498706322,
+            "mean_ref_top1_prob": 0.6033589127996771,
+            "median_context_kld": 0.016550733119089625,
+            "median_context_p99": 0.12118519842624664,
+            "p90_context_kld": 0.021278483650238936,
             "positions": 196512,
-            "top1_agreement": 0.9386093470118874,
-            "worst_context_id": 246,
-            "worst_context_kld": 0.031466900875017814
+            "top1_agreement": 0.9393624816805081,
+            "worst_context_id": 174,
+            "worst_context_kld": 0.027381622246614497
           }
         },
         "key": "scientific_technical",
         "label": "Scientific and technical exposition",
-        "relative_to_run": 0.4127614605599896
+        "relative_to_run": 0.40865282996993746
       }
     ]
   },
   "overall": {
     "deployed": {
       "contexts": 768,
-      "max_kld": 36.5456657409668,
-      "mean_kld": 0.040968724321885384,
-      "mean_ref_top1_prob": 0.6415440866166368,
-      "median_context_kld": 0.021978546635872478,
-      "median_context_p99": 0.17513030767440796,
-      "p90_context_kld": 0.08278915535830515,
+      "max_kld": 28.66598129272461,
+      "mean_kld": 0.040530797225316356,
+      "mean_ref_top1_prob": 0.6415473983687051,
+      "median_context_kld": 0.021778975352759924,
+      "median_context_p99": 0.17481790482997894,
+      "p90_context_kld": 0.08018104969267713,
       "positions": 1572096,
-      "top1_agreement": 0.926059858939912,
+      "top1_agreement": 0.9268944135727081,
       "worst_context_id": 454,
-      "worst_context_kld": 0.8362269388706521
+      "worst_context_kld": 0.8104930786551602
     }
   },
   "primary": "deployed"