Publish gemma-4-26B-A4B-it distribution-fidelity artifact

Gemma-4-26B-A4B-it-NVFP4/compliance.json

@@ -110,7 +110,7 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 262144,
-    "laws_version": 8,
+    "laws_version": 9,
     "model_runner_v2": false,
     "reference_config_sha256": "ed0c1eb3633de771906e9ba004a44cc5635bcc06ee2062077c3d2e88a50707d3",
     "rows": 768,
@@ -122,7 +122,7 @@
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T02:29:06.763180+00:00",
+  "evaluated_at": "2026-09-03T12:42:12.205642+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -222,7 +222,7 @@
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 8,
+  "laws_version": 9,
   "mean_kld": 1.8214520923049085,
   "nondeterminism_floor": 0.0,
   "overridden_laws": [],

Gemma-4-26B-A4B-it-NVFP4/report.md

@@ -34,7 +34,7 @@ Reverse direction, KLD(candidate || reference): 2.11838833.
 | 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 | 8 |
+| Laws version | 9 |
 | Partition | analysis |
 
 ## Component attribution
@@ -287,7 +287,7 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 | `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
 | `environment/models` | 11.35 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
 | `checksums.txt` | 191.21 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 29.81 KiB | The laws this artifact was produced under, including the override procedure. |
+| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

LAWS.md

@@ -62,6 +62,19 @@ identical mean. A version-7 report whose weights were released cannot be bound
 after the fact; it publishes under a recorded Law 13 deviation saying so, or it
 is rescored.
 
+Version 9 states what agreement between tensor-parallel workers means, under
+Law 8. Versions 1 through 8 required the workers to return identical floats for
+the trunk KLD, which no correct multi-worker run can do: each rank multiplies its
+own slice of the vocabulary on its own device, and a sum over the vocabulary
+amplifies the last-bit difference in the gathered logits. Every tensor-parallel
+candidate was therefore refused for a fault it did not have, and the refusal
+reported no numbers, so the cause stayed hidden. Version 9 requires agreement to
+a stated bound, publishes the observed divergence beside the figure so no reader
+takes more precision from the digits than was measured, and refuses a real
+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.
+
 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,
@@ -257,6 +270,25 @@ and the detected head state for reference and candidate.
 bit-identical to the reference's, in which case the delta is necessarily zero and
 the second scoring pass buys no information.
 
+### Agreement, not identity, across tensor-parallel workers
+
+The trunk figure is computed on every tensor-parallel worker from the same three
+files, so the workers cross-check each other: if a rank installed the reference
+head wrongly or holds the wrong shard, its answer diverges. But the workers are
+not required to produce identical floats, and demanding that they do is wrong.
+Each rank multiplies its own slice of the vocabulary on its own device, and the
+gathered logits differ in their last bits; a sum over a quarter-million-token
+vocabulary amplifies that difference into the KLD.
+
+So the contract is agreement to a stated bound — currently `1e-5` absolute per
+position and `1e-7` relative on the mean. Within it, the run proceeds, the trunk
+figure is rank 0's, and the one-pager states how far the ranks actually diverged,
+so nobody reads more precision into the digits than was measured. Beyond it, the
+ranks disagree about the model rather than about rounding, and scoring refuses
+naming the field, the position, and both values. A top-1 flip between ranks is
+disclosed and counted rather than refused: an argmax over two near-tied logits
+can flip on last-bit noise without either rank being wrong.
+
 ## Law 9 — Tail and depth disclosure
 
 **Required.** A published mean is always accompanied by median, p90, p99, and

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/compliance.json

@@ -109,7 +109,7 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 262144,
-    "laws_version": 8,
+    "laws_version": 9,
     "model_runner_v2": false,
     "reference_config_sha256": "ed0c1eb3633de771906e9ba004a44cc5635bcc06ee2062077c3d2e88a50707d3",
     "rows": 768,
@@ -121,7 +121,7 @@
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T02:29:07.106596+00:00",
+  "evaluated_at": "2026-09-03T12:42:12.559969+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -221,7 +221,7 @@
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 8,
+  "laws_version": 9,
   "mean_kld": 1.8129804764368658,
   "nondeterminism_floor": 0.0,
   "overridden_laws": [],

Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/report.md

@@ -34,7 +34,7 @@ Reverse direction, KLD(candidate || reference): 2.10446417.
 | 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 | 8 |
+| Laws version | 9 |
 | Partition | analysis |
 
 ## Component attribution
@@ -287,7 +287,7 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 | `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
 | `environment/models` | 11.35 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
 | `checksums.txt` | 191.21 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 29.81 KiB | The laws this artifact was produced under, including the override procedure. |
+| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

README.md

@@ -0,0 +1,36 @@
+---
+pretty_name: gemma-4-26B-A4B-it quantization analysis
+tags:
+- quantization
+- kl-divergence
+- distribution-fidelity
+- evaluation
+---
+
+# gemma-4-26B-A4B-it quantization analysis
+
+Mean KL divergence against on-disk size
+
+Scored under the distribution-fidelity laws, version 9. Read [`LAWS.md`](LAWS.md) first: these numbers are comparable only within this artifact's token suite, geometry, and runtime identity, and not against any number produced elsewhere.
+
+Each candidate directory holds its one-pager (`report.md`), its raw report, its compliance receipt, and its Law 14 attribution where one was produced. `reference/` carries the reusable teacher tensors and head, so a third party can score a new candidate without loading the reference checkpoint. `checksums.txt` covers every file here.
+
+## Candidates by mean KLD
+
+| Candidate | Scheme | On disk | Mean KLD |
+|---|---|---|---|
+| RedHatAI/gemma-4-26B-A4B-it-FP8-dynamic | fp8_per_channel | 26.68 GiB | 0.68045608 |
+| cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit | unquantized | 16.01 GiB | 1.01984639 |
+| dhruvil237/gemma-4-26B-A4B-it-W4A16 | unquantized | 14.58 GiB | 1.15699481 |
+| unsloth/gemma-4-26B-A4B-it-NVFP4 | fp8_per_channel | 15.75 GiB | 1.16528327 |
+| lcu0312/gemma-4-26B-A4B-it-AWQ-4bit | unquantized | 16.01 GiB | 1.23145061 |
+| cklaus/gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 16.78 GiB | 1.57775650 |
+| RedHatAI/gemma-4-26B-A4B-it-NVFP4 | unquantized | 15.30 GiB | 1.75937795 |
+| Neural-ICE/Gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 15.29 GiB | 1.81298048 |
+| bg-digitalservices/Gemma-4-26B-A4B-it-NVFP4 | nvfp4 | 15.29 GiB | 1.82145209 |
+
+## Fidelity against size
+
+![mean KLD against on-disk size](kld-vs-size.png)
+
+The numbers behind the chart are in [`kld-vs-size.json`](kld-vs-size.json), because a picture is not evidence.

checksums.txt

@@ -4,29 +4,30 @@
 1a8b0ec72384883c491c05f1ffac90f2f8e0e4ddef15cf250ea82d0a9c7b8130  Gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json
 ba3fba0921da11d3d75fe9efa5acd137e9cfeee830bbe6da236f1e4fa5094f7c  Gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4-qdq-manifest.json
 163128bf034eae96b43f5d338f508508edfd882ed08f6ac2047bb54118d9d38f  Gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json
-60bb8a810891bdb3183e4d1115669dd6ff4e127ad71fbc1a87325c692abd7eee  Gemma-4-26B-A4B-it-NVFP4/compliance.json
+403ce17bba6bfcae13d089cf50f85d6172c19f40f4293632550a3f95882718c9  Gemma-4-26B-A4B-it-NVFP4/compliance.json
 30e6d9f47587ad1a2b83748f170aa16c5b2ef82b4907e0d7667f790ed715a970  Gemma-4-26B-A4B-it-NVFP4/inspect.json
 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097  Gemma-4-26B-A4B-it-NVFP4/manifest.json
 c175fd2c40b96586dfb956c6fd44315dbe3538252646845597a986bd0d9a2965  Gemma-4-26B-A4B-it-NVFP4/provenance.json
 1f9c734f10fbe07ee040dd5421782150f1066ffd4eec79f0bc46aab604be6725  Gemma-4-26B-A4B-it-NVFP4/report.json
-28d9743f7de9b181eaf1e3e75b1e0a5bfa069ec0dcf757e7a1147f2ae9d97e7f  Gemma-4-26B-A4B-it-NVFP4/report.md
+c600caf8c43f4f7c4360454996211ec29fdfaf2689fb5a8a5e6f7ef4afdce4eb  Gemma-4-26B-A4B-it-NVFP4/report.md
 da111dfb9cbde291b68303f3134e622e26887567981acddf5ef929c79cd764d3  Gemma-4-26B-A4B-it-NVFP4/strata.json
 9f4800509649cdd3f07172b269979e925377d8675637953c8ec05dc314aae908  Gemma-4-26B-A4B-it-NVFP4/strata.md
-7fd1e1917c9d839ca37c46221540476e592a6355163ae367c9e17bb49286c3db  LAWS.md
+b27401dca11a7c7f1c00df4be4b71f5a4afc902f569ebf5dcb26d1b2fec175ec  LAWS.md
 610cd330ee854ee54fc0acb117cc6179e96ea5ab7f7cc2ec7d3ca482a79e4161  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution.json
 568b126a313f817bbf7b5b8c8ead57aeeb6e96002fd26ce462442c17dd4c1fd3  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-attention-dense_mlp-nvfp4-qdq-manifest.json
 3ad6c58045e3e7af296473f996a73c9ff0be77be50192e96a5dba427db277377  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-attention-dense_mlp-nvfp4.json
 1a8b0ec72384883c491c05f1ffac90f2f8e0e4ddef15cf250ea82d0a9c7b8130  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json
 ba3fba0921da11d3d75fe9efa5acd137e9cfeee830bbe6da236f1e4fa5094f7c  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4-qdq-manifest.json
 163128bf034eae96b43f5d338f508508edfd882ed08f6ac2047bb54118d9d38f  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json
-c0b695ee48ccb19f436a5fe7840849fef8652a3031ef9472f4fbe7da449a1d4d  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/compliance.json
+a39486ff38e2789591ad598cb93b42cf7489ac408f28e097d8b5f1b26e381318  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/compliance.json
 f0501e9acd4aa9bd8d4e9fcba9a1f9614193d86555b6108ab837010b3791244b  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/inspect.json
 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/manifest.json
 dbfd855639d775b58d0de2c805c0af7c8909843cc2c86b0a94449125cc07c071  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/provenance.json
 2a91fb88aedf76b79c67f7f742f78de8335a5a1b8eca2da378b3ba5bcacf32e9  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/report.json
-c6c4062235328cb84b853d0736dc8d315a9400814cc4ad053d5ade33652d0cdd  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/report.md
+b0e681ff040c92c92f2dd02434c3700670bf9f13f63e05be6704f74b498c21e7  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/report.md
 4771a0793ba1b21b908c75ff71e78c828d98d6aea510c7a023af666156309745  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/strata.json
 74b466b0bb05e4e7b9ed595b37cedd1c06db76f8dbdc95444f0bbbc6e598b1a7  Neural-ICE-Gemma-4-26B-A4B-it-NVFP4/strata.md
+67f4008ff8f10b0e13b4118b07fde769a4534e5cfde2ae457b6c49f5819241b4  README.md
 1cf434675aa0d523430405fd4dc7b1541db74dca917d3e9f5bd6889153c0e26c  baselines/self-kld.json
 35fd27a1281d34345ff16eceaa77de26d7eedb09c6e8ed7ae3a25382c0d2457c  cklaus-gemma-4-26B-A4B-it-NVFP4/attribution.json
 0a289c8a6da1af4fc45c672404c92b7a7fc179309fa8865020e7ce4dccdd3e58  cklaus-gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-dense_mlp-nvfp4-qdq-manifest.json
@@ -34,12 +35,12 @@ c6c4062235328cb84b853d0736dc8d315a9400814cc4ad053d5ade33652d0cdd  Neural-ICE-Gem
 1a8b0ec72384883c491c05f1ffac90f2f8e0e4ddef15cf250ea82d0a9c7b8130  cklaus-gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json
 ba3fba0921da11d3d75fe9efa5acd137e9cfeee830bbe6da236f1e4fa5094f7c  cklaus-gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4-qdq-manifest.json
 163128bf034eae96b43f5d338f508508edfd882ed08f6ac2047bb54118d9d38f  cklaus-gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json
-be13058508ecb7e086cfdebee40f22440cca37c34c8760d5aa85f4cb67be7535  cklaus-gemma-4-26B-A4B-it-NVFP4/compliance.json
+3922985b812b8e2a95684c0826e8c15c64e2993042583274639e3425d016a04f  cklaus-gemma-4-26B-A4B-it-NVFP4/compliance.json
 8a80ea8531551ca2bfd0da58f5ff09d23ddeb724bf0f31b7c9ae597359444a04  cklaus-gemma-4-26B-A4B-it-NVFP4/inspect.json
 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097  cklaus-gemma-4-26B-A4B-it-NVFP4/manifest.json
 471797082acd6ab2868518df277455b69ab51f2cb0a65f5e15ec52652543f3e5  cklaus-gemma-4-26B-A4B-it-NVFP4/provenance.json
 0c0924616f9b7e6afefe42177a275e39215789c35ccda34b9841f5dbb395f954  cklaus-gemma-4-26B-A4B-it-NVFP4/report.json
-43450dacabcdf3eb114c4c0f356b54241e44b436332c86ce7c598bc8b591b11e  cklaus-gemma-4-26B-A4B-it-NVFP4/report.md
+a53ceecf46107e53dfe513abf46b80dc578957852a32e5d98ce6ae7e27d441cd  cklaus-gemma-4-26B-A4B-it-NVFP4/report.md
 d99e3caa051736f5a920693c2ff17656f38b2a3397c054e66d04144ffc594ff3  cklaus-gemma-4-26B-A4B-it-NVFP4/strata.json
 01ce90ffe22f0aa02e9f6ad5b0e84b57ae947dd9007f5da944c969a548bda353  cklaus-gemma-4-26B-A4B-it-NVFP4/strata.md
 22828d684d66b21814b9887afc4d057b1c448b1064230a385b33d5d576616af3  environment/gpu-clocks.txt
@@ -75,12 +76,12 @@ e85f485c3013bfe8beba7d8268ebe7fd03b307b91a90e25074943d27af599610  gemma-4-26B-A4
 736eb608a26cde2882bc9b3f9b797cec1c7b6222c729be87cd9ed483c4ab22f4  gemma-4-26B-A4B-it-AWQ-4bit/attribution/gemma-4-26B-A4B-it-qdq-experts-int4_g32_sym.json
 1a8b0ec72384883c491c05f1ffac90f2f8e0e4ddef15cf250ea82d0a9c7b8130  gemma-4-26B-A4B-it-AWQ-4bit/attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json
 163128bf034eae96b43f5d338f508508edfd882ed08f6ac2047bb54118d9d38f  gemma-4-26B-A4B-it-AWQ-4bit/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json
-30712a16b0d3642175555c9275a1b43a4926fb335389f72df77059f6f0840855  gemma-4-26B-A4B-it-AWQ-4bit/compliance.json
+a8bb2aa2e3a9fa38dc0188d6dabc78a6319fbcedfdd77eec72d160212aaf7e9a  gemma-4-26B-A4B-it-AWQ-4bit/compliance.json
 ce9711f798577f815464e4d47d965d874d9908be9257623c93e1cc3fdc566f17  gemma-4-26B-A4B-it-AWQ-4bit/inspect.json
 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097  gemma-4-26B-A4B-it-AWQ-4bit/manifest.json
 8ef16987cdccd9a222111fabd0bdc837c41b97bbe1c0aa01a7641ff537e674a5  gemma-4-26B-A4B-it-AWQ-4bit/provenance.json
 4c31b02fc443b83ef5a1e189b3c94dbd432fadb7e9f32e142be87590e44f3900  gemma-4-26B-A4B-it-AWQ-4bit/report.json
-9557cc8db5fde6d00b7c8246d487c2836548370c76e680da96110b5797a2f4cb  gemma-4-26B-A4B-it-AWQ-4bit/report.md
+59c1dee79df9192e2ceefc85f82968862ae4c16530e658f04e95447d4e303d73  gemma-4-26B-A4B-it-AWQ-4bit/report.md
 9cd3c00cf017925fd4ac64b7bef251e9749219ba081417a4e74f99e498735d44  gemma-4-26B-A4B-it-AWQ-4bit/strata.json
 e95dbd3588e5fcda4b681db524c124d8b81cf66b5858ea6fb57864d5dc1822b9  gemma-4-26B-A4B-it-AWQ-4bit/strata.md
 59967321f1768520a3ee2cedbec241ac2a141e77b134a29e8abdd6fce4b17e80  gemma-4-26B-A4B-it-FP8-dynamic/attribution.json
@@ -90,12 +91,12 @@ fd55d9f6c5f97980d3bc3eb74e37f4217a002aebfb054bfd8ee5c31ab0418bda  gemma-4-26B-A4
 fb7e31911e569b2cb7cb1d187d977145ded2f6d11a3a21a58ba65002bb612a15  gemma-4-26B-A4B-it-FP8-dynamic/attribution/gemma-4-26B-A4B-it-qdq-experts-fp8_per_channel.json
 1a8b0ec72384883c491c05f1ffac90f2f8e0e4ddef15cf250ea82d0a9c7b8130  gemma-4-26B-A4B-it-FP8-dynamic/attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json
 163128bf034eae96b43f5d338f508508edfd882ed08f6ac2047bb54118d9d38f  gemma-4-26B-A4B-it-FP8-dynamic/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json
-c2fdbbcc4f3aeb4b6b68100ff0d757b39334ef780276c88eeb1e39f2216e898b  gemma-4-26B-A4B-it-FP8-dynamic/compliance.json
+32680e2ad8355f9c698742d0f52bdabb5b05bf00b8c8184ba21c7e741b8e612f  gemma-4-26B-A4B-it-FP8-dynamic/compliance.json
 c249f5af56cffb0bed1fc574b730ae3fa2038bba76e0ee7230c70e0e1ea54c8f  gemma-4-26B-A4B-it-FP8-dynamic/inspect.json
 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097  gemma-4-26B-A4B-it-FP8-dynamic/manifest.json
 012c69a12da1e3df6777194a5f39440a3db002c13b963e1348f2741ff2bfcced  gemma-4-26B-A4B-it-FP8-dynamic/provenance.json
 49c9f8ae786409474213ae270f09fc8d16354208c94af1356e7cada9f82f418b  gemma-4-26B-A4B-it-FP8-dynamic/report.json
-59f5fbfa23facf1430f6acc46d7b4cd7331941bcd7e8ed158c47188b379af5dd  gemma-4-26B-A4B-it-FP8-dynamic/report.md
+4723a2df6b833e56eda63ae88e9c49bec1032dffb959dc5e970b01eefeef2720  gemma-4-26B-A4B-it-FP8-dynamic/report.md
 e38a0af1f3c7e088df7f3e6fba302bd14dd3a76cf4c49f86a7b5d077f73a78c9  gemma-4-26B-A4B-it-FP8-dynamic/strata.json
 ea6767455518c824e265781242599e5184d1069a30582f77e79f9376fafacc26  gemma-4-26B-A4B-it-FP8-dynamic/strata.md
 ea7d6b4a1db5f065b7daa4d555254a668f39693a12dc0613616fa0cdf36a912c  gemma-4-26B-A4B-it-NVFP4/attribution.json
@@ -104,12 +105,12 @@ ea7d6b4a1db5f065b7daa4d555254a668f39693a12dc0613616fa0cdf36a912c  gemma-4-26B-A4
 1a8b0ec72384883c491c05f1ffac90f2f8e0e4ddef15cf250ea82d0a9c7b8130  gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json
 ba3fba0921da11d3d75fe9efa5acd137e9cfeee830bbe6da236f1e4fa5094f7c  gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4-qdq-manifest.json
 163128bf034eae96b43f5d338f508508edfd882ed08f6ac2047bb54118d9d38f  gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json
-9dfad71fc2022a2a33d75954f98a993a25929cac999803f1a6d110a35dc465f3  gemma-4-26B-A4B-it-NVFP4/compliance.json
+1663debb942beca8d31faebeff83e0f18865854aa9104888a17f7f9c9a603383  gemma-4-26B-A4B-it-NVFP4/compliance.json
 1370e8d65d75aa151145524b06a681bb28a24f0c0aaf84a690bfad964f421a60  gemma-4-26B-A4B-it-NVFP4/inspect.json
 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097  gemma-4-26B-A4B-it-NVFP4/manifest.json
 b912c1094667e6d6d42b62b2f9b4ada71ce7e157824a4edbb97411063ea25240  gemma-4-26B-A4B-it-NVFP4/provenance.json
 1b1c6cec01dcf7b7ce35544a63d630c49a58209af9b7bdcb138a846e312ed810  gemma-4-26B-A4B-it-NVFP4/report.json
-b1e2efe053579214a5efbaec3216cb773d44a49ee9479e3023b0a25384bafeb7  gemma-4-26B-A4B-it-NVFP4/report.md
+5233c21eeb92953042db01374a41cbb05a073f65808147b24b8e831ed10def99  gemma-4-26B-A4B-it-NVFP4/report.md
 ebcd3e0b2446271e5324553a6e858cb68b86c078ec069f5c49c1319c8b559c0a  gemma-4-26B-A4B-it-NVFP4/strata.json
 3ea3b5c42a2f2bfe594cb318583ac3a262b5ecb2504330019ae360d30faf5ed8  gemma-4-26B-A4B-it-NVFP4/strata.md
 de85e3dbba6bbff3e77add3702e7b2092143b8bfdaefe284b4f72eaf36756fe2  gemma-4-26B-A4B-it-W4A16/attribution.json
@@ -119,12 +120,12 @@ f3a3ec38cecd73f2c32eb138042456de48f04388d05d994ce2d59ffa544224a9  gemma-4-26B-A4
 032f0f146545640c9018ca6d02d0a43e82f6d56d127ca0685c2fe8bef3dac5ab  gemma-4-26B-A4B-it-W4A16/attribution/gemma-4-26B-A4B-it-qdq-experts-int4_g64_sym.json
 1a8b0ec72384883c491c05f1ffac90f2f8e0e4ddef15cf250ea82d0a9c7b8130  gemma-4-26B-A4B-it-W4A16/attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json
 163128bf034eae96b43f5d338f508508edfd882ed08f6ac2047bb54118d9d38f  gemma-4-26B-A4B-it-W4A16/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json
-9ef551d66936fc8225a768ce2e57044e63eeab81c38da410ae1161b35799cf30  gemma-4-26B-A4B-it-W4A16/compliance.json
+ea9db3ba3543747224f9c1ff4b6404a8c3c3fca4475ad9a85ae86a44dadade9a  gemma-4-26B-A4B-it-W4A16/compliance.json
 9724fbe1877e03ab207cbda96a2a79166aba0b0f700647292970e71871d77154  gemma-4-26B-A4B-it-W4A16/inspect.json
 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097  gemma-4-26B-A4B-it-W4A16/manifest.json
 22d3392fd18f290e3ca668a07b7ce4b421bd0eb318f4d6a9960a3ece5127c3d2  gemma-4-26B-A4B-it-W4A16/provenance.json
 e92c54a717d7f25b4adfa65ace260a42726095d66360668fd9229772dd965222  gemma-4-26B-A4B-it-W4A16/report.json
-1e39a039ce0804a0400e4fcb873c0881cbc845a6aaf547ebf77a86030bba390f  gemma-4-26B-A4B-it-W4A16/report.md
+7dee3c5e6375471c46ad2cf26ed582e1426c14e795116c8c353e7eb421fc1f93  gemma-4-26B-A4B-it-W4A16/report.md
 cd267299cccb1640ec7897234f7de6e1709cabb0cc09c3270c8ab671368fdea8  gemma-4-26B-A4B-it-W4A16/strata.json
 1c9b9735fd9c8bf0afbfea1f7acec7410d460a4584d4a2fb4e37972a1d09a2a8  gemma-4-26B-A4B-it-W4A16/strata.md
 2839db725cf7f936c8fa5c4d2b1f7a349c16c103dbdcd66a10d292e4f5d0aad6  kld-vs-size.json
@@ -136,12 +137,12 @@ e85f485c3013bfe8beba7d8268ebe7fd03b307b91a90e25074943d27af599610  lcu0312-gemma-
 736eb608a26cde2882bc9b3f9b797cec1c7b6222c729be87cd9ed483c4ab22f4  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/attribution/gemma-4-26B-A4B-it-qdq-experts-int4_g32_sym.json
 1a8b0ec72384883c491c05f1ffac90f2f8e0e4ddef15cf250ea82d0a9c7b8130  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json
 163128bf034eae96b43f5d338f508508edfd882ed08f6ac2047bb54118d9d38f  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json
-1d0df6809f2f97f89e0f05ee85edfdd8c11e17329916722c1d62ea5b8d0b23d3  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/compliance.json
+84c48c7f36fe14b736206872b8132bc196cb9f3aaf19d73ac3d0432ae5dab524  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/compliance.json
 198feddebe39ec856cde543ee931812dafd322bb8f2fbb4454f8b446b4a90370  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/inspect.json
 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/manifest.json
 c67582b0e1581873e3bfddb1cebb6d7dfd66018fd007e61e1aa1f6bfa57b6c19  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/provenance.json
 f384f8b65e009962d740402d21ef1394cd1ad07ee528ae8a950838a2626d6b36  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/report.json
-7ef6524c4b947815e24e1139d89e826b8d11edc001258ac626d3894bb119c690  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/report.md
+dfe0c1c595d290e47eee79991ed7352d39a15a90853a144173b76088ef4ad465  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/report.md
 97a2ec5f9d9e88c764144ccbb192a265f29371f576f68cf665b46f8d554f2385  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/strata.json
 0a9d2ee9fe80e87eb667a6aeb1194d258a1db79cc475579ba586fdefef1e60e5  lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/strata.md
 766c5ac8eb7de58bd325abe9779a4c5c03b6eb1ffce2599a157f01ed39955c03  reference/hidden_0.safetensors
@@ -1950,11 +1951,11 @@ e570315115fbb04b5e4328a899a786c69127bfcb842342dda3bef92e71dfd201  suite/tokens/c
 1a8b0ec72384883c491c05f1ffac90f2f8e0e4ddef15cf250ea82d0a9c7b8130  unsloth-gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-mxfp8.json
 ba3fba0921da11d3d75fe9efa5acd137e9cfeee830bbe6da236f1e4fa5094f7c  unsloth-gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4-qdq-manifest.json
 163128bf034eae96b43f5d338f508508edfd882ed08f6ac2047bb54118d9d38f  unsloth-gemma-4-26B-A4B-it-NVFP4/attribution/gemma-4-26B-A4B-it-qdq-experts-nvfp4.json
-03f8170691f2934466b6c9b9668a3efdb00372613d5420e984231c11131a3c31  unsloth-gemma-4-26B-A4B-it-NVFP4/compliance.json
+8e30a0638b3e3ffc5fbe6e2089974a57bb13dd14797eb1415f7304efc06460a3  unsloth-gemma-4-26B-A4B-it-NVFP4/compliance.json
 4698dcdbfa4c412d535e7c452df48856a571c1155a2984a0409f49482a5e431b  unsloth-gemma-4-26B-A4B-it-NVFP4/inspect.json
 395c569193357bb7f32e750fa94d4156797bef41dc8f46f13370aca8764f4097  unsloth-gemma-4-26B-A4B-it-NVFP4/manifest.json
 657a736ad94e7849bc398130fd6df6fe666f877ec1561dcb49ccb13a01703c38  unsloth-gemma-4-26B-A4B-it-NVFP4/provenance.json
 585cc0b5e165ec0a3300f32f9376278378aa48548e3c38e15a086124e772bd55  unsloth-gemma-4-26B-A4B-it-NVFP4/report.json
-4fc2586e7e17c76e4c51519131c30a7cd784d630c009942fed81f7b428cd09f3  unsloth-gemma-4-26B-A4B-it-NVFP4/report.md
+7a4121a691f7caa75f2a2989e1d3ee1415d32afbe9d24519f63306e71629a288  unsloth-gemma-4-26B-A4B-it-NVFP4/report.md
 df81c2cd44b35d7408def7ce1e6bdde82522a64a9963dade1e0b2878e1732fb2  unsloth-gemma-4-26B-A4B-it-NVFP4/strata.json
 22bd4bc93a0e2afa985211cc12178782a9665b82e696ed8abd90f299816efc34  unsloth-gemma-4-26B-A4B-it-NVFP4/strata.md

cklaus-gemma-4-26B-A4B-it-NVFP4/compliance.json

@@ -108,7 +108,7 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 262144,
-    "laws_version": 8,
+    "laws_version": 9,
     "model_runner_v2": false,
     "reference_config_sha256": "ed0c1eb3633de771906e9ba004a44cc5635bcc06ee2062077c3d2e88a50707d3",
     "rows": 768,
@@ -120,7 +120,7 @@
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T02:29:06.998559+00:00",
+  "evaluated_at": "2026-09-03T12:42:12.448217+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -220,7 +220,7 @@
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 8,
+  "laws_version": 9,
   "mean_kld": 1.5777564987346007,
   "nondeterminism_floor": 0.0,
   "overridden_laws": [],

cklaus-gemma-4-26B-A4B-it-NVFP4/report.md

@@ -34,7 +34,7 @@ Reverse direction, KLD(candidate || reference): 1.79113895.
 | 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 | 8 |
+| Laws version | 9 |
 | Partition | analysis |
 
 ## Component attribution
@@ -287,7 +287,7 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 | `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
 | `environment/models` | 11.35 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
 | `checksums.txt` | 191.21 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 29.81 KiB | The laws this artifact was produced under, including the override procedure. |
+| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

gemma-4-26B-A4B-it-AWQ-4bit/compliance.json

@@ -114,7 +114,7 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 262144,
-    "laws_version": 8,
+    "laws_version": 9,
     "model_runner_v2": false,
     "reference_config_sha256": "ed0c1eb3633de771906e9ba004a44cc5635bcc06ee2062077c3d2e88a50707d3",
     "rows": 768,
@@ -126,7 +126,7 @@
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T02:29:07.225928+00:00",
+  "evaluated_at": "2026-09-03T12:42:12.678121+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -231,7 +231,7 @@
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 8,
+  "laws_version": 9,
   "mean_kld": 1.0198463905464685,
   "nondeterminism_floor": 0.0,
   "overridden_laws": [

gemma-4-26B-A4B-it-AWQ-4bit/report.md

@@ -34,7 +34,7 @@ Reverse direction, KLD(candidate || reference): 1.03887223.
 | 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 | 8 |
+| Laws version | 9 |
 | Partition | analysis |
 
 ## Component attribution
@@ -295,7 +295,7 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 | `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
 | `environment/models` | 11.35 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
 | `checksums.txt` | 191.21 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 29.81 KiB | The laws this artifact was produced under, including the override procedure. |
+| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

gemma-4-26B-A4B-it-FP8-dynamic/compliance.json

@@ -114,7 +114,7 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 262144,
-    "laws_version": 8,
+    "laws_version": 9,
     "model_runner_v2": false,
     "reference_config_sha256": "ed0c1eb3633de771906e9ba004a44cc5635bcc06ee2062077c3d2e88a50707d3",
     "rows": 768,
@@ -126,7 +126,7 @@
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T02:29:06.543868+00:00",
+  "evaluated_at": "2026-09-03T12:42:11.981791+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -231,7 +231,7 @@
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 8,
+  "laws_version": 9,
   "mean_kld": 0.6804560779754232,
   "nondeterminism_floor": 0.0,
   "overridden_laws": [

gemma-4-26B-A4B-it-FP8-dynamic/report.md

@@ -34,7 +34,7 @@ Reverse direction, KLD(candidate || reference): 0.68247164.
 | 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 | 8 |
+| Laws version | 9 |
 | Partition | analysis |
 
 ## Component attribution
@@ -295,7 +295,7 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 | `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
 | `environment/models` | 11.35 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
 | `checksums.txt` | 191.21 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 29.81 KiB | The laws this artifact was produced under, including the override procedure. |
+| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

gemma-4-26B-A4B-it-NVFP4/compliance.json

@@ -109,7 +109,7 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 262144,
-    "laws_version": 8,
+    "laws_version": 9,
     "model_runner_v2": false,
     "reference_config_sha256": "ed0c1eb3633de771906e9ba004a44cc5635bcc06ee2062077c3d2e88a50707d3",
     "rows": 768,
@@ -121,7 +121,7 @@
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T02:29:06.654195+00:00",
+  "evaluated_at": "2026-09-03T12:42:12.092823+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -226,7 +226,7 @@
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 8,
+  "laws_version": 9,
   "mean_kld": 1.7593779469283644,
   "nondeterminism_floor": 0.0,
   "overridden_laws": [

gemma-4-26B-A4B-it-NVFP4/report.md

@@ -34,7 +34,7 @@ Reverse direction, KLD(candidate || reference): 2.03008919.
 | 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 | 8 |
+| Laws version | 9 |
 | Partition | analysis |
 
 ## Component attribution
@@ -292,7 +292,7 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 | `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
 | `environment/models` | 11.35 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
 | `checksums.txt` | 191.21 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 29.81 KiB | The laws this artifact was produced under, including the override procedure. |
+| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

gemma-4-26B-A4B-it-W4A16/compliance.json

@@ -115,7 +115,7 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 262144,
-    "laws_version": 8,
+    "laws_version": 9,
     "model_runner_v2": false,
     "reference_config_sha256": "ed0c1eb3633de771906e9ba004a44cc5635bcc06ee2062077c3d2e88a50707d3",
     "rows": 768,
@@ -127,7 +127,7 @@
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T02:29:07.344639+00:00",
+  "evaluated_at": "2026-09-03T12:42:12.796245+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -232,7 +232,7 @@
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 8,
+  "laws_version": 9,
   "mean_kld": 1.1569948142271558,
   "nondeterminism_floor": 0.0,
   "overridden_laws": [

gemma-4-26B-A4B-it-W4A16/report.md

@@ -34,7 +34,7 @@ Reverse direction, KLD(candidate || reference): 1.26012456.
 | 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 | 8 |
+| Laws version | 9 |
 | Partition | analysis |
 
 ## Component attribution
@@ -295,7 +295,7 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 | `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
 | `environment/models` | 11.35 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
 | `checksums.txt` | 191.21 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 29.81 KiB | The laws this artifact was produced under, including the override procedure. |
+| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope
 

lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/compliance.json

@@ -114,7 +114,7 @@
       "NVIDIA RTX PRO 6000 Blackwell Workstation Edition"
     ],
     "kld_vocab_size": 262144,
-    "laws_version": 8,
+    "laws_version": 9,
     "model_runner_v2": false,
     "reference_config_sha256": "ed0c1eb3633de771906e9ba004a44cc5635bcc06ee2062077c3d2e88a50707d3",
     "rows": 768,
@@ -126,7 +126,7 @@
     "torch": "2.13.0+cu132"
   },
   "compliant": true,
-  "evaluated_at": "2026-09-03T02:29:07.460597+00:00",
+  "evaluated_at": "2026-09-03T12:42:12.916024+00:00",
   "failed_laws": [],
   "findings": [
     {
@@ -231,7 +231,7 @@
       "title": "Recorded deviation"
     }
   ],
-  "laws_version": 8,
+  "laws_version": 9,
   "mean_kld": 1.231450608676746,
   "nondeterminism_floor": 0.0,
   "overridden_laws": [

lcu0312-gemma-4-26B-A4B-it-AWQ-4bit/report.md

@@ -34,7 +34,7 @@ Reverse direction, KLD(candidate || reference): 1.36951483.
 | 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 | 8 |
+| Laws version | 9 |
 | Partition | analysis |
 
 ## Component attribution
@@ -295,7 +295,7 @@ Paths are relative to the artifact root, the same paths `checksums.txt` uses. Ve
 | `environment/pip-freeze.txt` | 4.31 KiB | Every installed package version in the scoring environment. |
 | `environment/models` | 11.35 KiB in 7 files | Checkpoint fingerprints: file listing, sizes, config and tokenizer hashes, and `config.json` verbatim for each model scored. |
 | `checksums.txt` | 191.21 KiB | `sha256sum --check` compatible over every other file here. This is authoritative for integrity (Law 12). |
-| `LAWS.md` | 29.81 KiB | The laws this artifact was produced under, including the override procedure. |
+| `LAWS.md` | 31.90 KiB | The laws this artifact was produced under, including the override procedure. |
 
 ## Scope