Release measured v1.1 quality improvement

ATTRIBUTIONS.md

@@ -12,3 +12,14 @@ Intent utterances retain their source labels; training converts them into varied
 AG News and DBpedia-14 are used only in evaluation and excluded from custom training. No raw evaluation text is included in a model bundle. Exclusion from custom training does not establish absence from base-model pretraining.
 
 Sol inherits 200 primary backbone updates, 100 decision-head warmup updates, 800 bucket-mixed Stage2 updates and 200 Stage3 updates. Nox inherits 748 primary backbone updates, 100 decision-head warmup updates, 200 Stage2 updates and 800 Stage3 updates. Head-only warmup does not update the backbone. These are different training histories, not a controlled size-only experiment. Discarded training branches are not part of either released checkpoint. Stage3 contains 47,000 newly generated training rows and 22,000 replay rows; a separate 1,000 generated rows form internal validation.
+
+## Subsequent decision adaptation
+
+The subsequent registered curriculum adds MultiNLI human-labeled evidence judgments to programmatically verified decision tasks and replay data. The scheduled MultiNLI training subset contains 12,000 rows from the government, slate, telephone and travel genres; fiction is excluded. Original entailment/neutral/contradiction annotations are preserved. Selection and probability calibration use separate premise groups. Scheduled rows are not a claim that every selected intermediate checkpoint has seen the complete schedule.
+
+MultiNLI is by Adina Williams, Nikita Nangia and Samuel R. Bowman, *A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference* (NAACL 2018). The pinned dataset card describes the majority of its material under the Open American National Corpus's permissive terms, with separate terms for fiction. This update uses the stated non-fiction subset and retains this attribution; no source corpus text is redistributed in the model package.
+
+- [Pinned MultiNLI dataset and license description](https://huggingface.co/datasets/nyu-mll/multi_nli/blob/da70db2af9d09693783c3320c4249840212ee221/README.md).
+- [MultiNLI paper](https://aclanthology.org/N18-1101/).
+
+MASSIVE and SLURP remain excluded from custom training, checkpoint selection and calibration. MASSIVE is used only for evaluation. Official Jev outputs are never used as training labels. The initial-release training histories above remain historical; the subsequent selected model and evaluation provenance identify the actual update.

EVALUATION.md

@@ -1,147 +1,107 @@
-# Decision 1.0 decoder evaluation
+# Sol: measured quality update
 
-Measured on 21 September 2026. These results describe the first released Sol and Nox checkpoints, compared on the same frozen requests with official Jev, released open decision models and their untuned Qwen parents. They are a bounded evaluation, not proof of universal superiority or recovery of Jev's internal architecture.
+The primary score averages 20 task-family accuracies equally: one half the original 880-question panel and one half the independently constructed 880-question confirmation panel. The original panel was already observed during development. The added panel was excluded from custom training, selection and calibration; candidate weights, prompt and temperature were fixed before its inference. Once unsealed for the first candidate, it remains an observed regression panel. A later release does not make it fresh again.
 
-## What the main score means
+| Model | Overall accuracy ↑ | Choice ↑ | Noul ↑ | Score ↑ |
+| --- | ---: | ---: | ---: | ---: |
+| Jev · 1.13.0 | **72.74** | **71.77** | **80.36** | **68.06** |
+| Nox · 4B · v1.1 | 67.05 | 70.55 | 60.27 | 55.56 |
+| Sol · 2B · v1.1 | 60.14 | 66.67 | 48.66 | 40.28 |
+| Qwen3.5 · 4B · untuned | 56.61 | 60.06 | 53.57 | 52.08 |
+| Decider · 2B | 55.30 | 57.26 | 58.93 | 50.00 |
+| Qwen3.5 · 2B · untuned | 48.06 | 50.36 | 45.09 | 47.22 |
+| Laya · EN/ML | 47.38 | 53.02 | 52.23 | 19.44 |
 
-The main score is **native decision accuracy, averaged equally across ten task families**. The core has **880 questions, 432 semantic groups, English and Chinese, and 2–14 candidates**: 752 Choice, 64 Noul and 64 Score questions. It combines 640 constructed decision tasks, 128 official AG News test examples and 112 official DBpedia-14 test examples. The natural-intent slice has only 16 underlying groups and 64 views. No benchmark-specific adaptation was performed on these evaluation rows.
+All values are percentages. Choice, Noul and Score columns pool correct/total questions of that type across both panels; their mean is not the overall family-macro score. Score accuracy tests the highest-probability rubric level; expected-value error is reported separately. Missing or invalid predictions count as incorrect. Confidence intervals use 10,000 paired cluster-bootstrap resamples. The two language views of each MASSIVE source utterance share draws within menu strata; synthetic counterfactual blocks remain intact. The original and confirmation panels are resampled independently, then combined with equal weight.
 
-The same core examples and family weights apply to every model. Choice and Score use the selected/max-probability category; Noul uses P(true) ≥ 0.5. Score accuracy is ordinal-bin accuracy, not accuracy of a rounded expected scalar. The Choice/Noul/Score columns pool examples of each type and therefore do not average to the ten-family overall score. Missing or invalid decisions count as incorrect. Different accepted subsets are not silently substituted.
+## Change from the released weights
 
-Confidence intervals use 10,000 percentile bootstrap resamples of semantic groups **within each family**, followed by the equal-weight family mean. Related translations and option variants remain grouped. Paired differences reuse the same sampled groups for both models. These are not simultaneous confidence intervals for every family.
-
-| Model | Overall accuracy ↑ | 95% CI | Choice ↑ | Noul ↑ | Score ↑ |
-| --- | ---: | ---: | ---: | ---: | ---: |
-| Nox · 4B | 79.32 | 75.74–82.98 | 76.06 | 90.62 | 87.50 |
-| Jev 1.13.0 | 79.10 | 76.17–82.05 | 72.61 | 100.00 | 100.00 |
-| Qwen3.5 · 4B, untuned | 69.89 | 66.23–73.40 | 68.48 | 62.50 | 90.62 |
-| Sol · 2B | 66.25 | 61.94–70.64 | 71.41 | 42.19 | 43.75 |
-| Decider · 2B | 64.01 | 59.88–68.11 | 60.77 | 71.88 | 84.38 |
-| Qwen3.5 · 2B, untuned | 57.12 | 53.45–60.65 | 56.38 | 37.50 | 81.25 |
-| Laya · EN/ML routed | 57.01 | 53.16–60.88 | 64.10 | 43.75 | 18.75 |
-| Laya · English | 56.54 | 52.60–60.49 | 63.70 | 43.75 | 18.75 |
-| Laya · Multilingual | 47.25 | 43.44–51.12 | 53.72 | 39.06 | 12.50 |
-
-All entries are accuracy percentages. “Untuned” means the exact **post-trained Qwen3.5 parent before our decision adaptation**, not a Qwen `-Base` checkpoint. Its frozen adapter scores A–Z with the pretrained LM head, with thinking disabled; no random classifier is used. The tested adapter supports at most 26 candidates. These are adapter-specific zero-additional-training results, not an upper bound on everything the parent language model can do. Comparing the release with this baseline measures the whole adaptation package (new head plus supervised backbone updates), not a pure head-only or backbone-only ablation.
-
-Laya routed uses its published input-based English/multilingual routing policy, fixed before scoring, with automatic task detection disabled. The two individual Laya checkpoints are shown separately. The official Jev API reported version 1.13.0; hosted internals and server-side truncation are not observable. We do not infer an encoder/decoder attention mask from response behavior.
-
-## Every family
-
-| Family | Questions | Nox · 4B | Jev 1.13.0 | Qwen3.5 · 4B, untuned | Sol · 2B | Decider · 2B | Qwen3.5 · 2B, untuned | Laya · EN/ML routed |
-| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
-| Natural intents | 64 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | 96.88 | 85.94 |
-| AG News | 128 | 82.03 | 85.16 | 84.38 | 82.81 | 86.72 | 80.47 | 91.41 |
-| DBpedia-14 | 112 | 95.54 | 96.43 | 97.32 | 93.75 | 98.21 | 92.86 | 83.93 |
-| Boolean constraints | 64 | 90.62 | 100.00 | 62.50 | 42.19 | 71.88 | 37.50 | 43.75 |
-| Ordered rubrics | 64 | 87.50 | 100.00 | 90.62 | 43.75 | 84.38 | 81.25 | 18.75 |
-| Relational composition | 96 | 41.67 | 56.25 | 51.04 | 45.83 | 54.17 | 48.96 | 25.00 |
-| Scoped evidence | 96 | 72.92 | 89.58 | 51.04 | 54.17 | 47.92 | 36.46 | 37.50 |
-| State tracking | 96 | 35.42 | 33.33 | 28.12 | 18.75 | 29.17 | 25.00 | 23.96 |
-| Unknown rejection | 64 | 87.50 | 100.00 | 60.94 | 81.25 | 59.38 | 62.50 | 64.06 |
-| Option carriers | 96 | 100.00 | 30.21 | 72.92 | 100.00 | 8.33 | 9.38 | 95.83 |
-
-The option-carrier family deliberately stresses binding evidence and arbitrary candidate representations. Its weight is one tenth of the headline score; this is **not** an estimate of its frequency in user workloads. AG News and DBpedia-14 favor several reference models. Relational composition and state tracking remain difficult for both releases. A high aggregate must not be read as an all-family win. The unknown-rejection slice tests whether a known value is inside or outside the supplied menu; it does not establish recognition of unknown real-world facts.
-
-## Paired differences and release scope
+| Panel | Accuracy change (points) | Paired 95% interval |
+| --- | ---: | ---: |
+| Old | +7.07 | [+3.89, +10.33] |
+| Fresh | +2.04 | [-0.62, +4.58] |
+| Joint | +4.56 | [+2.47, +6.62] |
 
-| Candidate | Reference | Difference (pp) | Paired 95% CI (pp) |
-| --- | ---: | ---: | ---: |
-| Nox · 4B | Jev 1.13.0 | +0.22 | -3.58 to +4.05 |
-| Nox · 4B | Laya · EN/ML routed | +22.31 | +17.50 to +27.08 |
-| Nox · 4B | Decider · 2B | +15.31 | +11.28 to +19.51 |
-| Nox · 4B | Qwen3.5 · 4B, untuned | +9.43 | +4.82 to +14.15 |
-| Sol · 2B | Jev 1.13.0 | -12.85 | -17.71 to -8.00 |
-| Sol · 2B | Laya · EN/ML routed | +9.24 | +4.20 to +14.22 |
-| Sol · 2B | Decider · 2B | +2.24 | -1.83 to +6.39 |
-| Sol · 2B | Qwen3.5 · 2B, untuned | +9.13 | +4.45 to +13.90 |
+Release eligibility follows the prospectively fixed positive joint point improvement plus complete, finite, valid candidate execution. It does not require every family to improve or the confidence interval to exclude zero. An eligible update is not evidence of universal superiority.
 
-Nox has a higher average than Laya and Decider on this panel. Its mean is close to Jev's, but the interval does not establish equivalence and its calibration and several task families remain behind. Sol's point estimate exceeds Laya and Decider, but its paired interval against Decider includes zero; it is materially below Jev overall.
+## Native API coverage
 
-The original preregistered v2 policy required stricter per-family/calibration conditions; **neither candidate passed that policy**. After seeing these results, the project owner authorized first releases based on average superiority over Laya and Decider, independently for each model. This publication decision did not change the examples, weights, candidate selection, temperatures or reported measurements. The original failed gates remain in [the aggregate record](metrics/quality-aggregate.json). These now-observed panels become regression tests; later generalization claims require new independent evaluations.
+Each separate native panel has 68 requests and 220 requested answers, including repeated packing and isolation probes. These answers are descriptive diagnostics, not 220 independent questions and not part of the primary score. Unsupported requests are reported with the full denominator rather than relabeled as wrong semantic predictions.
 
-## Probabilities and typed decisions
+| Model | Original: valid / 220 | Original: correct / 220 | New: valid / 220 | New: correct / 220 |
+| --- | ---: | ---: | ---: | ---: |
+| Nox · 4B · v1.1 | 220 | 200 | 220 | 138 |
+| Sol · 2B · v1.1 | 220 | 138 | 220 | 127 |
+| Jev · 1.13.0 | 220 | 203 | 220 | 168 |
+| Laya · EN/ML | 208 | 88 | 208 | 98 |
+| Decider · 2B | 220 | 111 | 220 | 96 |
+| Qwen3.5 · 2B · untuned | 196 | 86 | 196 | 71 |
+| Qwen3.5 · 4B · untuned | 196 | 112 | 196 | 104 |
 
-Brier below is the equal-family mean of the multiclass squared-probability error (sum across classes, not divided by K). NLL uses natural logarithms without epsilon clipping; a returned zero probability for the correct class yields infinity under the reported-distribution metric. For hosted APIs with rounded outputs, this does not prove that the unobserved internal probability is exactly zero. RPS averages squared cumulative errors over K−1 ordinal boundaries. Native Score MAE divides the absolute expected-index error by K−1. All core targets are hard labels, so this release does not establish quality on soft business targets.
+## Scope and interpretation
 
-| Model | Macro Brier ↓ | Choice NLL ↓ | False recall ↑ | True recall ↑ | Noul balanced ↑ | Score RPS ↓ | Score MAE ↓ |
-| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
-| Nox · 4B | 0.3577 | 2.0021 | 85.00 | 100.00 | 92.50 | 0.0291 | 0.0302 |
-| Jev 1.13.0 | 0.2481 | infinity | 100.00 | 100.00 | 100.00 | 0.0000 | 0.0000 |
-| Qwen3.5 · 4B, untuned | 0.4271 | 0.8927 | 40.00 | 100.00 | 70.00 | 0.0214 | 0.0484 |
-| Sol · 2B | 0.5187 | 1.4094 | 10.00 | 95.83 | 52.92 | 0.1946 | 0.2638 |
-| Decider · 2B | 0.4425 | 0.8561 | 55.00 | 100.00 | 77.50 | 0.0730 | 0.1744 |
-| Qwen3.5 · 2B, untuned | 0.5724 | 1.1525 | 0.00 | 100.00 | 50.00 | 0.0560 | 0.1033 |
-| Laya · EN/ML routed | 0.5710 | infinity | 20.00 | 83.33 | 51.67 | 0.2003 | 0.3122 |
-| Laya · English | 0.5708 | infinity | 20.00 | 83.33 | 51.67 | 0.2003 | 0.3122 |
-| Laya · Multilingual | 0.6314 | 1.0724 | 7.50 | 91.67 | 49.58 | 0.2817 | 0.4104 |
+All comparators receive the same frozen requests through documented native or fixed adaptation interfaces. The untuned baselines are post-trained Qwen3.5 parents before decision adaptation, evaluated with a fixed LM-head decision adapter. Laya uses fixed English/multilingual routing. Jev is the official API; server-side truncation and internal model architecture cannot be independently inspected.
 
-Noul has 40 false and 24 true cases. In particular, Sol's false recall is only 10% here: a syntactically valid yes/no answer is not a correctness guarantee. Do not use its raw confidence as an automatic escalation threshold without application-specific validation.
+| Model | Original reported truncations | New reported truncations |
+| --- | ---: | ---: |
+| Nox · 4B · v1.1 | 0 | 0 |
+| Sol · 2B · v1.1 | 0 | 0 |
+| Jev · 1.13.0 | 0 | 0 |
+| Laya · EN/ML | 112 | 40 |
+| Decider · 2B | 0 | 0 |
+| Qwen3.5 · 2B · untuned | 0 | 0 |
+| Qwen3.5 · 4B · untuned | 0 | 0 |
 
-Both checkpoints keep the **single temperature fitted on development data before final evaluation**: Sol T=0.7033302993804421; Nox T=0.5332910931023557. On this final panel, the shipped temperatures worsen family-macro Brier relative to raw T=1 (Sol 0.5187 vs 0.4857; Nox 0.3577 vs 0.3253). They were not switched after seeing final results. Jev's corresponding Brier is 0.2481. The aggregate preserves raw and shipped proper scores separately. “Confidence” in the public Choice/Score schema is (K·max(p)−1)/(K−1), a concentration statistic, not a proven calibrated probability of correctness or a recovered Jev formula.
+Zero reported truncations for a remote service is not proof of complete server-side input processing. The Decision candidate additionally passes explicit complete-token and overflow checks.
 
-An analytic uniform-probability control and its fixed-tie versus expected-random accuracy are included in the aggregate. A global training-class prior is not meaningful for request-specific arbitrary candidate IDs and is omitted with that reason.
+The confirmation panel includes unseen MASSIVE English/Chinese intents and eight independently generated decision families: temporal exclusion, constraint assignment, transaction recovery, conflicting-rule reasoning, multiset reconciliation, record identity, capacity constraints and service-loss scoring. Training, selection and calibration exclude MASSIVE and SLURP. Exact overlap was zero; three moderate lexical near-pairs with training were disclosed and retained. Gold labels were not supplied by Jev.
 
-## Input support and separate native probe
+MASSIVE uses the [official 1.1 archive](https://amazon-massive-nlu-dataset.s3.amazonaws.com/amazon-massive-dataset-1.1.tar.gz), CC BY 4.0. Exposure during base-model pretraining is unknown. Synthetic English/Chinese views share symbolic JSON state; they do not establish natural multilingual document comprehension.
 
-Core probabilities are valid on all 880 questions for every primary model. Sol, Nox, Decider and the untuned Qwen adapters reported no core truncations. Each Laya variant reported truncation on 112 core questions; this is a comparison of the released interfaces on identical requests, not a claim that every implementation consumed identical token content. Jev does not expose enough information to verify server-side truncation.
+The [complete aggregate measurements](quality-metrics.json) retain every family, raw and calibrated probability diagnostics, native coverage, failure counts, Brier/NLL and ordinal metrics. Proper probability scores are conditional on valid probability coverage; zero probability on the true answer produces explicit infinite NLL rather than clipping. Per-family regressions remain visible in the capability matrix.
 
-The independent native probe has **68 requests containing 220 questions**. It includes structured inputs, multiple questions, lengths and candidate counts up to 255. It is reported separately and does not contribute to the main ranking. Validity and semantic correctness are different measurements; unsupported requests remain in the all-requested denominator.
+The question-scaling figure retains measurements from v1.0 weights and is labeled accordingly. The architecture and answer semantics are unchanged. Sol v1.1 additionally binds an explicit FLA normalization profile after portable reload exposed a real autotuning difference. Its selected weights, prompt and temperature are unchanged by this runtime binding. The older curve is not a timing claim for these weights or this profile.
 
-| Model | Accepted requests / 68 | Valid questions / 220 | Correct questions / 220 |
-| --- | ---: | ---: | ---: |
-| Nox · 4B | 68 | 220 | 162 |
-| Jev 1.13.0 | 68 | 220 | 203 |
-| Qwen3.5 · 4B, untuned | 44 | 196 | 112 |
-| Sol · 2B | 68 | 220 | 116 |
-| Decider · 2B | 68 | 220 | 111 |
-| Qwen3.5 · 2B, untuned | 44 | 196 | 86 |
-| Laya · EN/ML routed | 56 | 208 | 88 |
-| Laya · English | 56 | 208 | 88 |
-| Laya · Multilingual | 62 | 214 | 58 |
+Evaluation policy SHA-256: `359de2b6b81be91a66c6c538b3655bed06f4728357328cd318b1a765e64565cd`. Scoring receipt SHA-256: `bca099653d5b92cf282a5a82c19998c65dae0ac89d6fd039123b60b3680e85b3`.
 
-Sol and Nox both accepted all questions without truncation. Each scored 23/42 on the large-candidate probe and 9/9 on the length probe. Nox scored 162/220 and Sol 116/220 overall, versus Jev's 203/220. This is a substantial remaining native-capability gap despite Nox's similar core macro accuracy. The untuned Qwen adapter rejected candidate sets beyond A–Z explicitly. Laya's accepted subsets differ and are not interchangeable paired comparisons.
+## Peer revision and aggregate provenance
 
-The Decision engine supports Choice 2–255, Noul false/true, and Score 2–10 ordered descriptions. Score returns an expected **index**; arbitrary supplied numeric values are not implemented in this release. Each complete question—including state, instructions, all candidates and special tokens—must fit 16,384 tokens. Overflow rejects the complete call before inference; it is never silently truncated. This is a verified input budget, not a claim of uniformly strong reasoning across the full context. Questions execute independently in batches of eight with no cross-question state-prefix cache.
+The displayed Nox peer is the already published [Nox v1.1](https://huggingface.co/llm-semantic-router/Decision-1.0-Nox/tree/5bee3061a7abf675813963e4e05267cb1a28f92f). Its 67.05% overall score, family counts, native coverage and absolute confidence intervals come from its separately completed evaluation on the exact same frozen inputs, family weights and bootstrap implementation. This replaces the historical Nox v1.0 peer only in this public comparison projection. No new Sol-versus-Nox paired significance claim is made.
 
-## Additional open references
+Sol's qualification remains its original comparison against Sol v1.0: 55.58% to 60.14%, +4.56 points, paired 95% interval [+2.47,+6.62]. Its original-panel change is +7.07 points; its later-panel change is +2.04 points with interval [−0.63,+4.58]. The global first exposure occurred for Nox before Sol's selected checkpoint was frozen, so both panels are observed regression evidence for this Sol release.
 
-These models used the same core and the matched FLA runtime where relevant. They are additional measured references, not additional release gates or evidence of significant wins against every model.
+[Projection provenance](peer-projection-provenance.json) binds both source reports and Nox's verified publication. Sol evaluation receipt SHA-256: `bca099653d5b92cf282a5a82c19998c65dae0ac89d6fd039123b60b3680e85b3`. Nox evaluation receipt SHA-256: `927c92121c2df534f2ba12523e466ecebd8bd8936f949c7792de3e747c92a47c`. The original qualification, receipt and statistics remain unchanged.
 
-| Open reference | Macro accuracy (%) | 95% CI |
-| --- | ---: | ---: |
-| kev-0.8b | 60.70 | 56.88–64.40 |
-| kev-4b | 71.28 | 67.41–75.10 |
-| kev-9b | 76.36 | 72.94–79.81 |
-| nimble-9b | 78.85 | 75.65–82.02 |
-| llm2jev-2b | 62.60 | 59.08–66.08 |
-| llm2jev-4b | 69.91 | 66.36–73.41 |
 
-## Training, selection and provenance
+## Reading supplied evidence: observed V4 supplement
 
-The text-only backbones are Qwen/Qwen3.5-2B at `15852e8c16360a2fea060d615a32b45270f8a8fc` and Qwen/Qwen3.5-4B at `851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a`. Both are post-trained parents. The shared candidate head is trained with the text backbone. Selected checkpoints were frozen before final scoring: Sol at Stage3 step 200 and Nox at Stage3 step 800, selected among the scheduled checkpoints using production batch-size-eight development accuracy. They have unequal inherited training histories; the comparison does not isolate model size as the sole cause of differences.
+Sol v1.1 scores **73.44%**, up **4.06 percentage points** from Sol v1.0; paired 95% CI **[+1.10, +7.03]pp**. This is observed regression evidence, supplementary to the unchanged V3 publication qualification. Reading remains a gap against Jev, Decider and the untuned 4B LM-head baseline.
 
-Stage3 uses 69,000 training rows: 47,000 new generated rows plus 22,000 replay rows; a separate 1,000 generated rows are internal validation. Training mixes programmatically verified decision tasks with licensed BANKING77 and CLINC150 training data, using varied instructions, structured views and candidate permutations. Reserved intent labels/domains are excluded from adaptation. AG News and DBpedia-14 are evaluation-only. The Stage3 overlap audit found no exact or token-Jaccard ≥0.5 matches against the final core; short overlaps with earlier development intent examples are a separate development limitation. Upstream pretraining exposure remains unknown. No official Jev predictions were used as training labels. Source licenses and exact training lineage are in [ATTRIBUTIONS.md](ATTRIBUTIONS.md).
+The fixed score is **50% BoolQ + 25% Belebele English + 25% Belebele Chinese**. Each family has 160 questions; BoolQ contains 80 false and 80 true labels. All 480 requested examples remain in the denominator. Bold marks column leaders. Weighted accuracy differs from raw correct/480.
 
-The actual architecture and numerical inference source are shipped with each model. Standalone export/reload with network access disabled reproduced all 256 development logits bit-exactly; emitted probabilities differed by at most 2.384×10⁻⁷ and selected decisions did not change. Real packaged Choice/Noul/Score inference also passed in the publicly rebuildable ROCm runtime; see [RUNTIME.md](RUNTIME.md). That small runtime equivalence probe is distinct from a complete benchmark rerun.
+| Model | BoolQ % (correct) | Belebele EN % (correct) | Belebele ZH % (correct) | Weighted % | Weighted 95% CI | Raw correct |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Jev · 1.13.0 | **92.50 (148/160)** | **96.88 (155/160)** | **96.25 (154/160)** | **94.53** | [92.01, 96.76] | 457/480 |
+| Decider · 2B | 91.88 (147/160) | 93.12 (149/160) | 91.25 (146/160) | 92.03 | [89.17, 94.65] | 442/480 |
+| Nox · 4B · v1.1 | 87.50 (140/160) | 71.25 (114/160) | 67.50 (108/160) | 78.44 | [73.99, 82.65] | 362/480 |
+| Sol · 2B · v1.1 | 79.38 (127/160) | 68.12 (109/160) | 66.88 (107/160) | 73.44 | [68.69, 77.96] | 343/480 |
+| Qwen3.5 · 4B · untuned | 83.75 (134/160) | 93.12 (149/160) | 91.25 (146/160) | 87.97 | [84.53, 91.25] | 429/480 |
+| Qwen3.5 · 2B · untuned | 65.00 (104/160) | 83.75 (134/160) | 81.25 (130/160) | 73.75 | [69.18, 78.23] | 368/480 |
+| Laya · EN/ML | 69.38 (111/160) | 38.12 (61/160) | 28.12 (45/160) | 51.25 | [46.60, 55.84] | 217/480 |
 
-The Laya checkpoint is `convaiinnovations/laya` at `1c5edc17a7acd8701df6fc341c0d179f1c62c982` (English root and multilingual subfolder). Its native router follows source commit `42626c348753fbb17572a813127df2278a1ec527`. Decider is `Mapika/decider-2b` at `b37f7e1ba3fbc9238004cf531fabbee2619973fd`; its released numerical interface is retained. Full aggregate metrics and immutable source/prediction fingerprints are in [metrics/quality-aggregate.json](metrics/quality-aggregate.json).
+## Sol version change
 
-## Efficiency
+| Version | BoolQ | Belebele EN | Belebele ZH | Natural Choice mean | Weighted |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Sol v1.0 | 119/160 · 74.38% | 106/160 · 66.25% | 100/160 · 62.50% | 64.38% | 69.38% |
+| Sol v1.1 | 127/160 · 79.38% | 109/160 · 68.12% | 107/160 · 66.88% | 67.50% | 73.44% |
 
-The completed quiet benchmark uses one AMD gfx942 accelerator with 261824 MiB visible memory. It includes all six local primary models, independently randomized model order in three blocks, 30 repetitions per supported cell, and separate Choice and Noul/Score panels. All 36 quality/runtime identity checks match. The 114 complete-input cells provide 3,420 measured request times; truncated and unsupported inputs are explicit exclusions rather than short-input substitutions.
+The exact weighted delta is **13/320**. Calibration is the frozen production temperature; no parameters or prompts were changed using this panel. This Sol run used the default public `from_pretrained` entry point with its automatically bound, validated normalization profile. Nox is the already published **v1.1** peer, from the identical frozen panel and statistical implementation; its separate source receipt is linked in the aggregate evidence. Untuned Qwen models use their original vocabulary heads and fixed LM-head adapters, not random decision heads.
 
-Latency includes rendering, tokenization, device transfers, model forward and answer assembly, but excludes loading, diagnostics, warmup, network and post-response shape checks. No other training/inference GPU process ran during measurement. CPU affinity and clocks were not locked, and telemetry counters on otherwise idle devices occasionally reported 1–3%; this is an empirical quiet request benchmark, not a noise-free hardware microbenchmark. After a transient process was caught by preflight before job 26, only the remaining jobs resumed; no timing job was repeated.
+All seven models returned **480/480 finite, valid answers**. The six local models retained all inputs, with zero truncations; Jev's internal retention is unknown. These concurrent-workload measurements establish quality only and provide no latency comparison.
 
-| Model | Actual tokens | p50 / p95 (ms) | Peak allocated GiB |
-| --- | ---: | ---: | ---: |
-| Nox · 4B | 309 | 27.37 / 28.95 | 8.06 |
-| Qwen3.5 · 4B, untuned | 296 | 27.47 / 28.76 | 8.04 |
-| Sol · 2B | 309 | 21.04 / 21.41 | 3.61 |
-| Decider · 2B | 250 | 23.35 / 23.91 | 3.62 |
-| Qwen3.5 · 2B, untuned | 296 | 20.51 / 20.71 | 3.60 |
-| Laya · EN/ML routed | 228 | 11.54 / 12.21 | 3.56 |
+Intervals use 10,000 paired passage-cluster bootstrap draws. BoolQ has 160 passage parents; Belebele has 144 shared passage parents. All questions and both language translations belonging to each Belebele passage are resampled together. Identical draws are shared across models. The machine-readable [aggregate](metrics/natural-confirmation.json) includes family intervals, exact weighted fractions, Noul/Choice probability metrics, coverage and source hashes. Zero gold probability produces infinite NLL; no epsilon clipping is used.
 
-This table uses the same short English state, one question and four options. Native token counts vary with the prompt/template. Laya uses its English branch. Sol is faster than Decider here, but Decider is faster on multiquestion Choice and K=255; neither Decision release establishes universal speed superiority. Nox trades higher quality on the measured panel for more memory and latency. Hosted Jev wall time includes service and network cost and is not included in this local timing comparison.
+[BoolQ validation](https://huggingface.co/datasets/google/boolq/tree/35b264d03638db9f4ce671b711558bf7ff0f80d5) (CC BY-SA 3.0) and [Belebele test](https://huggingface.co/datasets/facebook/belebele/tree/7899cdfa4e1e0d733fd77c848e2c273cb1d32be2) (CC BY-SA 4.0) were excluded from custom TRAIN/SELECT/CAL. Original human labels and parallel alignment were independently checked; they were not newly adjudicated. The lexical audit found no exact or character-5-gram Jaccard ≥0.5 overlaps across 51 earlier data/panel files. This does not prove semantic independence or exclude unknown upstream-pretraining exposure. BoolQ lacks article-title metadata, so its clustering is by passage, not article.
 
-See [TIMING.md](TIMING.md) for every length/question-count/candidate-count/type cell, questions/s and peak-memory definitions, and [metrics/timing-aggregate.json](metrics/timing-aggregate.json) for machine-readable measurements and runtime pairing. The full request benchmark and the small publicly rebuilt runtime parity test are separate evidence.
+The permanent first-exposure ledger predates this Sol candidate freeze: this is **observed regression**, not a newly unseen confirmation result for Sol. This panel covers evidence yes/no and four-choice comprehension; it does not measure ordinal Score or establish universal decision capability. It adds no release gate and leaves the existing qualification receipt unchanged.

FIGURE-NOTICES.md

@@ -1,10 +0,0 @@
-# Figure provenance
-
-The architecture figures describe the shipped Decision implementation. Backbone operator details were checked against the actual Qwen3.5 implementation in Transformers 5.17.0 and the pinned parent configurations. They do not claim to reveal Jev's unpublished architecture.
-
-- [Qwen3.5-2B configuration](https://huggingface.co/Qwen/Qwen3.5-2B/blob/15852e8c16360a2fea060d615a32b45270f8a8fc/config.json)
-- [Qwen3.5-4B configuration](https://huggingface.co/Qwen/Qwen3.5-4B/blob/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a/config.json)
-- [Transformers source and Apache 2.0 license](https://github.com/huggingface/transformers)
-- [Decision family visual guide](https://gist.github.com/Xunzhuo/4020f574e3d38e5e5eae00063bdf9dce)
-
-The rank and capability figures are generated from the same aggregate evidence linked in EVALUATION.md. The supplied Decision mark is included unchanged; chart backgrounds are white. SVG and PDF preserve the chart/architecture geometry and typography as vectors; the branding mark is a raster asset.

NORMALIZATION_RUNTIME.md

@@ -0,0 +1,9 @@
+# Validated normalization runtime
+
+This Sol bundle automatically loads an explicit FLA normalization launch profile before importing FLA. The validated environment is the recorded ROCm runtime on gfx942 GPUs. No private compilation cache or manual environment setting is required. Use a fresh Python process; repeated loading of this same profile is allowed, while mixing profiled and unprofiled models in one process is rejected.
+
+The profile fixes only the FLA l2norm forward launch settings for BF16 input/output, FP32 reciprocal norms, head dimension 128, and normalization buckets 1–32. This covers the production batch size of at most 8 and complete-question limit of 16,384 tokens with 16 key heads. Unmatched shapes/dtypes and alternate normalization kernels fail closed. Other FLA kernels keep the recorded runtime behavior; this is not a claim that every FLA kernel is deterministic.
+
+Buckets previously observed during production preserve their chosen configurations. Previously unobserved buckets use the prospectively specified configuration in the profile. Boundary tests establish finite execution and dispatch coverage, not numerical equivalence to an unobserved previous automatic choice. Weights, tokenizer, prompt, readout and calibration are unchanged from the selected candidate.
+
+This runtime has its own numerical validation. Earlier release latency measurements must not be treated as measurements of this profile.

QUESTION-SCALING.md

@@ -1,4 +1,6 @@
-# Fixed-question latency scaling
+# Fixed-question latency scaling — v1.0
+
+These measurements use the initial Sol/Nox v1.0 weights. They are retained as versioned evidence and do not measure the updated weights.
 
 A fixed English state (819 characters), identical question text and a four-choice menu are repeated Q times in one API request. Native rendered rows, including state and question, contain 309 tokens for Sol/Nox, 228 for Laya and 250 for Decider. Tokenizers and templates differ. Every model uses the same physical AMD gfx942 GPU and matches its quality/runtime fingerprint. Native question packing is preserved, with no cross-request cache or concurrent-client load.
 
@@ -8,7 +10,7 @@ Q1/Q8/Q32 come from the original GPU-quiet round; Q2/Q4/Q16 are an additive roun
 
 Base-model appendix retains the original measured Q1/Q8/Q32 points only.
 
-The machine exposes 261824 MiB of device memory. Parameters, native precision, shipped temperature and model/runtime fingerprints are unchanged from the quality benchmark. No throughput saturation claim is made. p95 is an empirical percentile, not a confidence interval.
+The machine exposes 261824 MiB of device memory. Parameters, native precision, shipped temperature and model/runtime fingerprints are those of the original v1.0 quality benchmark. No throughput saturation claim is made. p95 is an empirical percentile, not a confidence interval.
 
 Full aggregate evidence: [question-scaling.json](metrics/question-scaling.json).
 

READING-SUPPLEMENT.md

@@ -0,0 +1,32 @@
+# Reading supplied evidence: observed V4 supplement
+
+Sol v1.1 scores **73.44%**, up **4.06 percentage points** from Sol v1.0; paired 95% CI **[+1.10, +7.03]pp**. This is observed regression evidence, supplementary to the unchanged V3 publication qualification. Reading remains a gap against Jev, Decider and the untuned 4B LM-head baseline.
+
+The fixed score is **50% BoolQ + 25% Belebele English + 25% Belebele Chinese**. Each family has 160 questions; BoolQ contains 80 false and 80 true labels. All 480 requested examples remain in the denominator. Bold marks column leaders. Weighted accuracy differs from raw correct/480.
+
+| Model | BoolQ % (correct) | Belebele EN % (correct) | Belebele ZH % (correct) | Weighted % | Weighted 95% CI | Raw correct |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Jev · 1.13.0 | **92.50 (148/160)** | **96.88 (155/160)** | **96.25 (154/160)** | **94.53** | [92.01, 96.76] | 457/480 |
+| Decider · 2B | 91.88 (147/160) | 93.12 (149/160) | 91.25 (146/160) | 92.03 | [89.17, 94.65] | 442/480 |
+| Nox · 4B · v1.1 | 87.50 (140/160) | 71.25 (114/160) | 67.50 (108/160) | 78.44 | [73.99, 82.65] | 362/480 |
+| Sol · 2B · v1.1 | 79.38 (127/160) | 68.12 (109/160) | 66.88 (107/160) | 73.44 | [68.69, 77.96] | 343/480 |
+| Qwen3.5 · 4B · untuned | 83.75 (134/160) | 93.12 (149/160) | 91.25 (146/160) | 87.97 | [84.53, 91.25] | 429/480 |
+| Qwen3.5 · 2B · untuned | 65.00 (104/160) | 83.75 (134/160) | 81.25 (130/160) | 73.75 | [69.18, 78.23] | 368/480 |
+| Laya · EN/ML | 69.38 (111/160) | 38.12 (61/160) | 28.12 (45/160) | 51.25 | [46.60, 55.84] | 217/480 |
+
+## Sol version change
+
+| Version | BoolQ | Belebele EN | Belebele ZH | Natural Choice mean | Weighted |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Sol v1.0 | 119/160 · 74.38% | 106/160 · 66.25% | 100/160 · 62.50% | 64.38% | 69.38% |
+| Sol v1.1 | 127/160 · 79.38% | 109/160 · 68.12% | 107/160 · 66.88% | 67.50% | 73.44% |
+
+The exact weighted delta is **13/320**. Calibration is the frozen production temperature; no parameters or prompts were changed using this panel. This Sol run used the default public `from_pretrained` entry point with its automatically bound, validated normalization profile. Nox is the already published **v1.1** peer, from the identical frozen panel and statistical implementation; its separate source receipt is linked in the aggregate evidence. Untuned Qwen models use their original vocabulary heads and fixed LM-head adapters, not random decision heads.
+
+All seven models returned **480/480 finite, valid answers**. The six local models retained all inputs, with zero truncations; Jev's internal retention is unknown. These concurrent-workload measurements establish quality only and provide no latency comparison.
+
+Intervals use 10,000 paired passage-cluster bootstrap draws. BoolQ has 160 passage parents; Belebele has 144 shared passage parents. All questions and both language translations belonging to each Belebele passage are resampled together. Identical draws are shared across models. The machine-readable [aggregate](metrics/natural-confirmation.json) includes family intervals, exact weighted fractions, Noul/Choice probability metrics, coverage and source hashes. Zero gold probability produces infinite NLL; no epsilon clipping is used.
+
+[BoolQ validation](https://huggingface.co/datasets/google/boolq/tree/35b264d03638db9f4ce671b711558bf7ff0f80d5) (CC BY-SA 3.0) and [Belebele test](https://huggingface.co/datasets/facebook/belebele/tree/7899cdfa4e1e0d733fd77c848e2c273cb1d32be2) (CC BY-SA 4.0) were excluded from custom TRAIN/SELECT/CAL. Original human labels and parallel alignment were independently checked; they were not newly adjudicated. The lexical audit found no exact or character-5-gram Jaccard ≥0.5 overlaps across 51 earlier data/panel files. This does not prove semantic independence or exclude unknown upstream-pretraining exposure. BoolQ lacks article-title metadata, so its clustering is by passage, not article.
+
+The permanent first-exposure ledger predates this Sol candidate freeze: this is **observed regression**, not a newly unseen confirmation result for Sol. This panel covers evidence yes/no and four-choice comprehension; it does not measure ordinal Score or establish universal decision capability. It adds no release gate and leaves the existing qualification receipt unchanged.

README.md

@@ -18,6 +18,7 @@ tags:
 datasets:
 - PolyAI/banking77
 - clinc/clinc_oos
+- nyu-mll/multi_nli
 ---
 
 ![Decision 1.0 — Your move.](assets/decision-family-header.png)
@@ -46,35 +47,49 @@ Your question names and candidate IDs are preserved. Labels are defined at runti
 
 ## Measured capability
 
-**66.25% overall**, +9.13 points over its untuned parent. Its mean exceeds Laya and Decider; the paired interval against Decider includes zero.
+**60.14% overall**, up 4.56 points from Sol v1.0 across 20 equally weighted capabilities.
 
 | Model | Overall accuracy ↑ | Choice ↑ | Noul ↑ | Score ↑ |
 | --- | ---: | ---: | ---: | ---: |
-| Nox · 4B | **79.32** | **76.06** | 90.62 | 87.50 |
-| Jev 1.13.0 | 79.10 | 72.61 | **100.00** | **100.00** |
-| Qwen3.5 · 4B, untuned | 69.89 | 68.48 | 62.50 | 90.62 |
-| Sol · 2B | 66.25 | 71.41 | 42.19 | 43.75 |
-| Decider · 2B | 64.01 | 60.77 | 71.88 | 84.38 |
-| Qwen3.5 · 2B, untuned | 57.12 | 56.38 | 37.50 | 81.25 |
-| Laya · EN/ML | 57.01 | 64.10 | 43.75 | 18.75 |
+| Jev · 1.13.0 | **72.74** | **71.77** | **80.36** | **68.06** |
+| Nox · 4B · v1.1 | 67.05 | 70.55 | 60.27 | 55.56 |
+| Sol · 2B · v1.1 | 60.14 | 66.67 | 48.66 | 40.28 |
+| Qwen3.5 · 4B · untuned | 56.61 | 60.06 | 53.57 | 52.08 |
+| Decider · 2B | 55.30 | 57.26 | 58.93 | 50.00 |
+| Qwen3.5 · 2B · untuned | 48.06 | 50.36 | 45.09 | 47.22 |
+| Laya · EN/ML | 47.38 | 53.02 | 52.23 | 19.44 |
 
-Accuracy (%). Overall averages **10 families / 880 questions** equally; type columns pool their questions. Same requests for every model. Untuned means the post-trained parent before decision adaptation. Laya's native interface truncated 112 inputs. [Methods and uncertainty](EVALUATION.md).
+Accuracy (%). Overall averages **20 task families / 1,760 questions** equally; type columns pool their questions. Both Decision peers are v1.1. Laya's native interface reported 152 truncated inputs. [Methods, coverage and uncertainty](EVALUATION.md).
 
 ![Decision quality ranking](assets/decision-quality.png)
 
-![Capability comparison across ten families](assets/decision-capabilities.png)
+![Capability comparison across twenty families](assets/decision-capabilities.png)
 
-False judgments, rubric scoring and state tracking remain weak; Jev is substantially stronger overall.
+Sol improves overall, while evidence judgments, rubric scoring and state reasoning remain uneven. See the evaluation for observed-panel limitations.
+
+## Reading supplied evidence
+
+| Model | BoolQ ↑ | Natural Choice ↑ | Weighted ↑ |
+| --- | ---: | ---: | ---: |
+| Jev · 1.13.0 | **92.50** | **96.56** | **94.53** |
+| Decider · 2B | 91.88 | 92.19 | 92.03 |
+| Nox · 4B · v1.1 | 87.50 | 69.38 | 78.44 |
+| Sol · 2B · v1.1 | 79.38 | 67.50 | 73.44 |
+| Qwen3.5 · 4B · untuned | 83.75 | 92.19 | 87.97 |
+| Qwen3.5 · 2B · untuned | 65.00 | 82.50 | 73.75 |
+| Laya · EN/ML | 69.38 | 33.12 | 51.25 |
+
+Accuracy (%), **480 human-annotated questions**. Natural Choice averages English and Chinese Belebele; weighted = 50% BoolQ + 50% Natural Choice. All local inputs were retained; Jev's internal retention is unknown. Public pretraining exposure is unknown. [Counts, uncertainty and limitations](READING-SUPPLEMENT.md).
 
 ## More questions, measured
 
 ![Latency as questions per request increase](assets/decision-question-scaling.png)
 
-Same state, same question length, four choices; only the number of questions changes. Each model keeps its native interface (Sol: 309 tokens/question). Median of 30 requests per point on one AMD gfx942 GPU; local Python request time includes tokenization and inference, excluding loading and network. [p95, drift checks and full methods](QUESTION-SCALING.md).
+**Measured with v1.0 weights and runtime.** Same state, same question length, four choices; only the number of questions changes. Each model keeps its native interface (Sol: 309 tokens/question). Median of 30 requests per point on one AMD gfx942 GPU; local Python request time includes tokenization and inference, excluding loading and network. [p95, drift checks and full methods](QUESTION-SCALING.md).
 
 ## Try it
 
-Download the fixed release with `hf download llm-semantic-router/Decision-1.0-Sol --revision v1.0 --local-dir decision-model`, then follow the [ROCm setup](RUNTIME.md). Inside that container, with the model mounted at `/model`:
+Download the fixed release with `hf download llm-semantic-router/Decision-1.0-Sol --revision v1.1 --local-dir decision-model`, then follow the [ROCm setup](RUNTIME.md). Inside that container, with the model mounted at `/model`:
 
 ```python
 from decision import DecisionModel
@@ -86,7 +101,7 @@ print(model.decide(**REQUEST)["answers"])
 
 [Actual request and measured output](model-card-example.json) · [Install and API guide](USAGE.md)
 
-The complete state, question and candidates must fit 16,384 tokens. Overflow is rejected. Score returns an expected ordinal index. Runtime: AMD gfx942 validated; CPU/MPS unsupported; NVIDIA unqualified.
+The complete state, question and candidates must fit 16,384 tokens. Overflow is rejected. Score returns an expected ordinal index. Runtime: AMD gfx942 validated; CPU/MPS unsupported; NVIDIA unqualified. This release automatically binds its validated normalization profile; use a fresh Python process. [Runtime profile](NORMALIZATION_RUNTIME.md).
 
 ## Architecture
 

RUNTIME.md

@@ -2,7 +2,9 @@
 
 The package has a public, digest-pinned installation path. `Dockerfile.runtime` starts from `vllm/vllm-openai-rocm@sha256:1fd21abe66455b4df5a2e83629e97cdcc9d58913b16052d8118b92b239792339`, adds two hash-checked FLA wheels, and installs this repository's loading wrapper. It keeps the base image's ROCm PyTorch and Triton builds. The vLLM server is not used by Decision inference.
 
-**Validation boundary:** the public registry manifest, base-image ancestry, package metadata and critical PyTorch binary hashes have been checked. The recipe built successfully and passed CPU imports and real AMD ROCm gfx942 GPU GPU inference for both released bundles. On the packaged three-question Choice/Noul/Score example, its complete responses matched the qualified research runtime exactly; the wrapper matched the direct engine and rejected an oversized complete input. Evidence is in `runtime-build-provenance.json`. This example establishes a working public installation path; it is not a full rerun of the quality or timing benchmark. Published benchmark results use the qualified runtime in each bundle's `runtime.json`.
+**Validation boundary:** the public registry manifest, base-image ancestry, package metadata and critical PyTorch binary hashes have been checked. The recipe built successfully and passed CPU imports and real AMD ROCm gfx942 GPU inference for both initial v1.0 bundles. On the packaged three-question Choice/Noul/Score example, its complete responses matched the qualified research runtime exactly; the wrapper matched the direct engine and rejected an oversized complete input. The initial v1.0 build evidence is in `runtime-build-provenance.json`. Updated weights retain the same installation recipe and are separately checked during release; their actual request and output are in `model-card-example.json`. This example establishes a working public installation path; it is not a full rerun of the quality or timing benchmark. Published benchmark results use the qualified runtime in each bundle's `runtime.json`.
+
+The Sol v1.1 default entrypoint additionally binds its bundled normalization profile before importing FLA. It has passed a fresh-cache, network-disabled reload on all 1,000 calibration questions, plus the three-type example and overflow rejection. Use a fresh Python process; see [normalization runtime](NORMALIZATION_RUNTIME.md). The historical image-build evidence below identifies its original v1.0 wrappers and bundles; it does not substitute for v1.1 validation.
 
 ## Build and run
 
@@ -50,3 +52,5 @@ The official [FLA installation guide](https://github.com/fla-org/flash-linear-at
 The public [PyTorch ROCm 7.2 wheel index](https://download.pytorch.org/whl/rocm7.2/torch/) contains ordinary release wheels such as `2.12.0+rocm7.2`. They are a different artifact from the qualified `2.12.0+git6bbd260` build and are not interchangeable evidence. The latter's [source commit is public](https://github.com/pytorch/pytorch/commit/6bbd26020da1c6dc198625dfcdd968b1e4e6b1c5), but a source commit alone does not reproduce compiler flags, linked libraries and binary behavior.
 
 An alternative runtime must recheck the installed versions, actual FLA Gated DeltaNet dispatch, BF16 backbone plus FP32 head, fixed prompt rendering, batch size eight, and model output agreement. Runtime changes can shift probabilities near a decision boundary even with identical weights. The supplied engine uses FLA Gated DeltaNet, reference PyTorch causal convolution and SDPA; selecting a different kernel is a new runtime configuration.
+
+For this profiled Sol release, `allow_unvalidated_runtime=True` does not bypass the normalization profile version, GPU architecture or supported-shape checks.

RUNTIME_BINDING.json

@@ -0,0 +1,54 @@
+{
+  "source_bundle_manifest_sha256": "9242b7fd4ce226f638108810ec5156dca6f270cce5a16efcf5ef091180ad75be",
+  "profile_sha256": "6b03450d42dbb68f0ffe14945ffcf3e6ea043e1033a819fb7211a8176a51722f",
+  "unchanged_files": [
+    {
+      "file": "backbone/config.json",
+      "bytes": 1789,
+      "sha256": "e7bed6f1a1a4d5f029a28040e8687a122b323798d4674666443f6d78e9475a05"
+    },
+    {
+      "file": "backbone/model.safetensors",
+      "bytes": 3763685328,
+      "sha256": "168f7f90bf410130da92fd5270305286d53d5cbb68be02f1c611c40cb9243fe2"
+    },
+    {
+      "file": "code/decision_model.py",
+      "bytes": 10114,
+      "sha256": "d3e28489c09f3bd7130e2d43d92e0b5c4a08e25b09b21303904defb0ff1c3646"
+    },
+    {
+      "file": "code/predict.py",
+      "bytes": 3164,
+      "sha256": "02352e8385ab47157b6459910da54d962e5da4bb4940d571faedc86bc5da9aee"
+    },
+    {
+      "file": "decision_config.json",
+      "bytes": 753,
+      "sha256": "9fa5b80f0df63964410b8498b5473d8cba4aa82e506aae83619d5423c0c0e67a"
+    },
+    {
+      "file": "decision_head.safetensors",
+      "bytes": 8424272,
+      "sha256": "924a43288d48804c22e7e2573898d694303879ab9bb819d1b20eadc2ab32f00d"
+    },
+    {
+      "file": "temperature.json",
+      "bytes": 13733,
+      "sha256": "63af0fbd1c817c4bc354aab81df66626bcd9908ac7a71bcbb921e63ba7361d44"
+    },
+    {
+      "file": "tokenizer.json",
+      "bytes": 19989325,
+      "sha256": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523"
+    },
+    {
+      "file": "tokenizer_config.json",
+      "bytes": 1123,
+      "sha256": "bee8eba30f0eb4af73c0fe2cd06d0f89b657d7819941c438157ec42f7c80ea87"
+    }
+  ],
+  "training_selection_calibration_unchanged": true,
+  "recomputed_or_recalibrated": false,
+  "private_compiled_cache_required": false
+}

SOURCE_BUNDLE_MANIFEST.json

@@ -0,0 +1,129 @@
+{
+  "format": "research-pointer-bundle-v1",
+  "status": "candidate-export-awaiting-independent-reload-and-quality-gates",
+  "files": [
+    {
+      "file": "backbone/config.json",
+      "bytes": 1789,
+      "sha256": "e7bed6f1a1a4d5f029a28040e8687a122b323798d4674666443f6d78e9475a05"
+    },
+    {
+      "file": "backbone/model.safetensors",
+      "bytes": 3763685328,
+      "sha256": "168f7f90bf410130da92fd5270305286d53d5cbb68be02f1c611c40cb9243fe2"
+    },
+    {
+      "file": "chat_template.jinja",
+      "bytes": 7755,
+      "sha256": "273d8e0e683b885071fb17e08d71e5f2a5ddfb5309756181681de4f5a1822d80"
+    },
+    {
+      "file": "code/decision_api.py",
+      "bytes": 6724,
+      "sha256": "147b2fec32cbbbbb1b92cf2a19bb887f9d945e4974e141ca7ea93b8c542f5b21"
+    },
+    {
+      "file": "code/decision_model.py",
+      "bytes": 10114,
+      "sha256": "d3e28489c09f3bd7130e2d43d92e0b5c4a08e25b09b21303904defb0ff1c3646"
+    },
+    {
+      "file": "code/predict.py",
+      "bytes": 3164,
+      "sha256": "02352e8385ab47157b6459910da54d962e5da4bb4940d571faedc86bc5da9aee"
+    },
+    {
+      "file": "decision_config.json",
+      "bytes": 753,
+      "sha256": "9fa5b80f0df63964410b8498b5473d8cba4aa82e506aae83619d5423c0c0e67a"
+    },
+    {
+      "file": "decision_head.safetensors",
+      "bytes": 8424272,
+      "sha256": "924a43288d48804c22e7e2573898d694303879ab9bb819d1b20eadc2ab32f00d"
+    },
+    {
+      "file": "runtime.json",
+      "bytes": 378,
+      "sha256": "c5d3521358b2817f4e56ea8150c5c612139b4e2b2c5bb09f512d9a7c0b5298b9"
+    },
+    {
+      "file": "temperature.json",
+      "bytes": 13733,
+      "sha256": "63af0fbd1c817c4bc354aab81df66626bcd9908ac7a71bcbb921e63ba7361d44"
+    },
+    {
+      "file": "tokenizer.json",
+      "bytes": 19989325,
+      "sha256": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523"
+    },
+    {
+      "file": "tokenizer_config.json",
+      "bytes": 1123,
+      "sha256": "bee8eba30f0eb4af73c0fe2cd06d0f89b657d7819941c438157ec42f7c80ea87"
+    }
+  ],
+  "tensors": [
+    {
+      "file": "backbone/model.safetensors",
+      "elements": 1881825088,
+      "elements_by_dtype": {
+        "BF16": 1881825088
+      }
+    },
+    {
+      "file": "decision_head.safetensors",
+      "elements": 2105856,
+      "elements_by_dtype": {
+        "F32": 2105856
+      }
+    }
+  ],
+  "source_checkpoint_files": [
+    {
+      "file": "backbone/config.json",
+      "bytes": 1788,
+      "sha256": "071e97d8291168ba712237744c9a60e733acce554e6164322d22550dc96de19b"
+    },
+    {
+      "file": "backbone/model-00001-of-00002.safetensors",
+      "bytes": 3999855496,
+      "sha256": "93059b3982827486760a708d4f10aa6ae61de94ce90607b080021529bd1562ab"
+    },
+    {
+      "file": "backbone/model-00002-of-00002.safetensors",
+      "bytes": 3527479552,
+      "sha256": "2ef40a6f2cd250aa0c08a6ced808d1aa025bff3dd6720857651c8f8943cf5cc9"
+    },
+    {
+      "file": "decision_config.json",
+      "bytes": 1119,
+      "sha256": "1e6d926e1593ba243e58448c3d3cca39c6731faa55e5acec369832e4efab074d"
+    },
+    {
+      "file": "decision_head.safetensors",
+      "bytes": 8424272,
+      "sha256": "924a43288d48804c22e7e2573898d694303879ab9bb819d1b20eadc2ab32f00d"
+    },
+    {
+      "file": "tokenizer.json",
+      "bytes": 19989325,
+      "sha256": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523"
+    },
+    {
+      "file": "tokenizer_config.json",
+      "bytes": 1123,
+      "sha256": "bee8eba30f0eb4af73c0fe2cd06d0f89b657d7819941c438157ec42f7c80ea87"
+    }
+  ],
+  "source_model_code_sha256": "d3e28489c09f3bd7130e2d43d92e0b5c4a08e25b09b21303904defb0ff1c3646",
+  "source_api_code_sha256": "147b2fec32cbbbbb1b92cf2a19bb887f9d945e4974e141ca7ea93b8c542f5b21",
+  "dev_sha256": "5ba3bd7527da831d77cec8c6a2f260a71e10a57742e356510c719f8f30d61a70",
+  "production_predictions_sha256": "b9e2cf2c466e6f9422e627905625400d683c6ce0cecb015c16e5784b0c3c577c",
+  "temperature_sha256": "63af0fbd1c817c4bc354aab81df66626bcd9908ac7a71bcbb921e63ba7361d44",
+  "runtime_sha256": "c5d3521358b2817f4e56ea8150c5c612139b4e2b2c5bb09f512d9a7c0b5298b9",
+  "production_batch_size": 8,
+  "input_length_limit": 16384,
+  "original_checkpoint_name": "checkpoint-001563",
+  "no_publication_performed": true
+}

TIMING.md

@@ -1,164 +0,0 @@
-# Local request efficiency
-
-One AMD gfx942 accelerator, 261824 MiB visible device memory; the same physical device for every model.
-
-Synchronized local API request latency including input rendering, tokenization, transfers, forward passes, and answer assembly. Excludes model loading, eligibility diagnostics, warmup, network, and post-response shape validation.
-
-Three outer blocks with independently randomized model order. Each model is reloaded per panel/block, with three warmups per case followed by ten timed repetitions. Case order is shuffled within each block. 30 timings per eligible cell.
-
-No other training or inference GPU process during measurement; each job has two idle-process preflight checks. ROCm telemetry is sampled during execution. CPU affinity, host services, and clocks are not locked; this is a GPU-quiet reproducible request benchmark, not a noise-free hardware microbenchmark.
-
-One API request at a time; Q1/Q8/Q32 are questions within one request, not concurrent client requests. Native adapters retain their question packing/batch strategy. No cross-request cache.
-
-Qwen/Decider backbone BF16, Sol/Nox decision head FP32; Laya preserves upstream FP32 parameters with native BF16 autocast. GDN-capable Qwen paths use FLA 0.5.2; Laya architecture has no GDN.
-
-The timing panel contains English states and questions. Routed Laya selects its English branch here; these speeds do not characterize its multilingual branch.
-
-p50/p95 are empirical percentiles of 30 observations, not confidence intervals. Questions/s equals 1000 times total questions divided by total measured milliseconds; it is not maximum concurrent-serving capacity.
-
-Peak PyTorch allocated memory, including resident model parameters, in GiB. Excludes allocator-reserved memory, non-PyTorch allocations, and device-driver overhead.
-
-All models receive identical byte-level states and question specifications. Token counts differ by native tokenizer/template. Truncated/unsupported cells are explicitly excluded; no short-input substitution. These repeated timing prompts do not measure semantic accuracy.
-
-Every timing model_runtime_key is identical to that model’s reported final-quality key. No reference-quality/optimized-speed mixture.
-
-After 25 successful jobs, preflight detected a transient ROCm process before starting job 26. The process exited; an explicit resume executed only the remaining 11 jobs after idle checks. No timing job was repeated.
-
-No comparable local GPU timing is available for hosted Jev. Existing API wall times are a different workload/network measure and are not in the local latency ranking.
-
-## choice
-
-| Model | Case | Q / K | Tokens per row | p50 / p95 ms | Questions/s | Peak GiB |
-|---|---|---:|---:|---:|---:|---:|
-| Sol 2B | short-q1 | 1 / 4 | 309–309 | 21.04 / 21.41 | 47.6 | 3.61 |
-| Sol 2B | medium-q1 | 1 / 4 | 2243–2243 | 35.88 / 36.10 | 27.9 | 3.76 |
-| Sol 2B | long-q1 | 1 / 4 | 8863–8863 | 134.31 / 135.52 | 7.4 | 4.33 |
-| Sol 2B | short-q8 | 8 / 4 | 309–309 | 37.29 / 37.51 | 214.3 | 3.78 |
-| Sol 2B | medium-q8 | 8 / 4 | 2243–2243 | 216.20 / 216.66 | 37.0 | 5.00 |
-| Sol 2B | short-q32 | 32 / 4 | 309–309 | 147.50 / 147.91 | 216.8 | 3.78 |
-| Sol 2B | short-k255 | 1 / 255 | 7870–7870 | 118.33 / 119.12 | 8.4 | 4.24 |
-| Nox 4B | short-q1 | 1 / 4 | 309–309 | 27.37 / 28.95 | 36.0 | 8.06 |
-| Nox 4B | medium-q1 | 1 / 4 | 2243–2243 | 68.72 / 78.11 | 14.3 | 8.32 |
-| Nox 4B | long-q1 | 1 / 4 | 8863–8863 | 278.96 / 296.37 | 3.6 | 9.25 |
-| Nox 4B | short-q8 | 8 / 4 | 309–309 | 69.25 / 69.84 | 115.4 | 8.35 |
-| Nox 4B | medium-q8 | 8 / 4 | 2243–2243 | 442.80 / 445.90 | 18.0 | 10.43 |
-| Nox 4B | short-q32 | 32 / 4 | 309–309 | 275.65 / 278.26 | 115.8 | 8.35 |
-| Nox 4B | short-k255 | 1 / 255 | 7870–7870 | 248.31 / 249.58 | 4.0 | 9.10 |
-| Laya routed | short-q1 | 1 / 4 | 228–228 | 11.54 / 12.21 | 85.8 | 3.56 |
-| Laya routed | medium-q1 | 1 / 4 | — | input_truncated | — | — |
-| Laya routed | long-q1 | 1 / 4 | — | input_truncated | — | — |
-| Laya routed | short-q8 | 8 / 4 | 228–228 | 14.56 / 15.02 | 548.5 | 3.61 |
-| Laya routed | medium-q8 | 8 / 4 | — | input_truncated | — | — |
-| Laya routed | short-q32 | 32 / 4 | 228–228 | 37.89 / 38.30 | 843.8 | 3.78 |
-| Laya routed | short-k255 | 1 / 255 | — | unsupported | — | — |
-| Decider 2B | short-q1 | 1 / 4 | 250–250 | 23.35 / 23.91 | 42.9 | 3.62 |
-| Decider 2B | medium-q1 | 1 / 4 | 2184–2184 | 35.24 / 35.41 | 28.4 | 3.80 |
-| Decider 2B | long-q1 | 1 / 4 | 8804–8804 | 131.47 / 131.99 | 7.6 | 4.42 |
-| Decider 2B | short-q8 | 8 / 4 | 250–250 | 31.59 / 31.77 | 253.4 | 3.90 |
-| Decider 2B | medium-q8 | 8 / 4 | 2184–2184 | 209.86 / 210.64 | 38.1 | 5.28 |
-| Decider 2B | short-q32 | 32 / 4 | 250–250 | 96.67 / 97.02 | 330.9 | 4.87 |
-| Decider 2B | short-k255 | 1 / 255 | 5556–5556 | 78.83 / 79.27 | 12.7 | 4.10 |
-| Qwen3.5 2B LM-head | short-q1 | 1 / 4 | 296–296 | 20.51 / 20.71 | 49.2 | 3.60 |
-| Qwen3.5 2B LM-head | medium-q1 | 1 / 4 | 2230–2230 | 35.22 / 35.53 | 28.4 | 3.74 |
-| Qwen3.5 2B LM-head | long-q1 | 1 / 4 | 8850–8850 | 129.02 / 131.25 | 7.7 | 4.21 |
-| Qwen3.5 2B LM-head | short-q8 | 8 / 4 | 296–296 | 37.32 / 37.61 | 214.5 | 3.75 |
-| Qwen3.5 2B LM-head | medium-q8 | 8 / 4 | 2230–2230 | 202.06 / 202.75 | 39.6 | 4.86 |
-| Qwen3.5 2B LM-head | short-q32 | 32 / 4 | 296–296 | 146.88 / 147.16 | 218.3 | 3.75 |
-| Qwen3.5 2B LM-head | short-k255 | 1 / 255 | — | unsupported | — | — |
-| Qwen3.5 4B LM-head | short-q1 | 1 / 4 | 296–296 | 27.47 / 28.76 | 36.3 | 8.04 |
-| Qwen3.5 4B LM-head | medium-q1 | 1 / 4 | 2230–2230 | 66.97 / 67.66 | 14.9 | 8.29 |
-| Qwen3.5 4B LM-head | long-q1 | 1 / 4 | 8850–8850 | 257.47 / 258.85 | 3.9 | 9.12 |
-| Qwen3.5 4B LM-head | short-q8 | 8 / 4 | 296–296 | 69.03 / 69.46 | 115.8 | 8.31 |
-| Qwen3.5 4B LM-head | medium-q8 | 8 / 4 | 2230–2230 | 414.85 / 415.72 | 19.3 | 10.25 |
-| Qwen3.5 4B LM-head | short-q32 | 32 / 4 | 296–296 | 274.69 / 275.19 | 116.5 | 8.31 |
-| Qwen3.5 4B LM-head | short-k255 | 1 / 255 | — | unsupported | — | — |
-
-## noul-score
-
-| Model | Case | Q / K | Tokens per row | p50 / p95 ms | Questions/s | Peak GiB |
-|---|---|---:|---:|---:|---:|---:|
-| Sol 2B | noul-short-q1 | 1 / 2 | 262–262 | 20.84 / 21.38 | 47.8 | 3.61 |
-| Sol 2B | noul-medium-q1 | 1 / 2 | 2196–2196 | 34.86 / 35.12 | 28.7 | 3.76 |
-| Sol 2B | noul-long-q1 | 1 / 2 | 8816–8816 | 131.69 / 132.53 | 7.6 | 4.33 |
-| Sol 2B | noul-short-q8 | 8 / 2 | 262–262 | 33.20 / 33.91 | 239.9 | 3.76 |
-| Sol 2B | noul-medium-q8 | 8 / 2 | 2196–2196 | 207.14 / 208.02 | 38.6 | 4.96 |
-| Sol 2B | noul-short-q32 | 32 / 2 | 262–262 | 131.43 / 132.38 | 243.5 | 3.77 |
-| Sol 2B | score-short-q1 | 1 / 4 | 307–307 | 20.91 / 21.50 | 47.6 | 3.61 |
-| Sol 2B | score-medium-q1 | 1 / 4 | 2241–2241 | 35.80 / 36.03 | 27.9 | 3.76 |
-| Sol 2B | score-long-q1 | 1 / 4 | 8861–8861 | 134.45 / 135.55 | 7.4 | 4.33 |
-| Sol 2B | score-short-q8 | 8 / 4 | 307–307 | 37.37 / 37.68 | 214.1 | 3.78 |
-| Sol 2B | score-medium-q8 | 8 / 4 | 2241–2241 | 215.72 / 216.75 | 37.1 | 5.00 |
-| Sol 2B | score-short-q32 | 32 / 4 | 307–307 | 146.54 / 147.00 | 218.4 | 3.78 |
-| Sol 2B | score-short-q1-levels2 | 1 / 2 | 274–274 | 20.99 / 22.65 | 47.2 | 3.61 |
-| Sol 2B | score-short-q1-levels10 | 1 / 10 | 490–490 | 21.10 / 21.56 | 47.3 | 3.63 |
-| Nox 4B | noul-short-q1 | 1 / 2 | 262–262 | 27.44 / 29.16 | 35.8 | 8.05 |
-| Nox 4B | noul-medium-q1 | 1 / 2 | 2196–2196 | 67.56 / 67.75 | 14.8 | 8.31 |
-| Nox 4B | noul-long-q1 | 1 / 2 | 8816–8816 | 273.90 / 275.65 | 3.6 | 9.25 |
-| Nox 4B | noul-short-q8 | 8 / 2 | 262–262 | 64.61 / 64.87 | 124.2 | 8.32 |
-| Nox 4B | noul-medium-q8 | 8 / 2 | 2196–2196 | 427.53 / 429.06 | 18.7 | 10.36 |
-| Nox 4B | noul-short-q32 | 32 / 2 | 262–262 | 256.45 / 257.09 | 124.8 | 8.32 |
-| Nox 4B | score-short-q1 | 1 / 4 | 307–307 | 27.41 / 29.51 | 35.7 | 8.06 |
-| Nox 4B | score-medium-q1 | 1 / 4 | 2241–2241 | 68.76 / 69.12 | 14.5 | 8.32 |
-| Nox 4B | score-long-q1 | 1 / 4 | 8861–8861 | 279.43 / 280.17 | 3.6 | 9.25 |
-| Nox 4B | score-short-q8 | 8 / 4 | 307–307 | 69.56 / 69.74 | 115.4 | 8.35 |
-| Nox 4B | score-medium-q8 | 8 / 4 | 2241–2241 | 442.15 / 443.60 | 18.1 | 10.43 |
-| Nox 4B | score-short-q32 | 32 / 4 | 307–307 | 277.89 / 278.51 | 115.1 | 8.35 |
-| Nox 4B | score-short-q1-levels2 | 1 / 2 | 274–274 | 27.76 / 28.89 | 35.8 | 8.05 |
-| Nox 4B | score-short-q1-levels10 | 1 / 10 | 490–490 | 27.75 / 29.39 | 35.4 | 8.08 |
-| Laya routed | noul-short-q1 | 1 / 2 | 202–202 | 12.05 / 12.24 | 84.2 | 3.56 |
-| Laya routed | noul-medium-q1 | 1 / 2 | — | input_truncated | — | — |
-| Laya routed | noul-long-q1 | 1 / 2 | — | input_truncated | — | — |
-| Laya routed | noul-short-q8 | 8 / 2 | 202–202 | 13.83 / 14.32 | 577.7 | 3.60 |
-| Laya routed | noul-medium-q8 | 8 / 2 | — | input_truncated | — | — |
-| Laya routed | noul-short-q32 | 32 / 2 | 202–202 | 34.38 / 34.68 | 929.9 | 3.75 |
-| Laya routed | score-short-q1 | 1 / 4 | 231–231 | 12.17 / 12.59 | 83.3 | 3.56 |
-| Laya routed | score-medium-q1 | 1 / 4 | — | input_truncated | — | — |
-| Laya routed | score-long-q1 | 1 / 4 | — | input_truncated | — | — |
-| Laya routed | score-short-q8 | 8 / 4 | 231–231 | 16.92 / 17.22 | 471.6 | 3.61 |
-| Laya routed | score-medium-q8 | 8 / 4 | — | input_truncated | — | — |
-| Laya routed | score-short-q32 | 32 / 4 | 231–231 | 38.64 / 38.95 | 827.9 | 3.79 |
-| Laya routed | score-short-q1-levels2 | 1 / 2 | 216–216 | 11.73 / 11.98 | 86.4 | 3.56 |
-| Laya routed | score-short-q1-levels10 | 1 / 10 | 328–328 | 11.91 / 12.54 | 84.4 | 3.56 |
-| Decider 2B | noul-short-q1 | 1 / 2 | 204–204 | 23.29 / 23.83 | 42.9 | 3.62 |
-| Decider 2B | noul-medium-q1 | 1 / 2 | 2138–2138 | 34.91 / 35.29 | 28.6 | 3.79 |
-| Decider 2B | noul-long-q1 | 1 / 2 | 8758–8758 | 132.39 / 133.57 | 7.5 | 4.41 |
-| Decider 2B | noul-short-q8 | 8 / 2 | 204–204 | 31.21 / 31.46 | 256.5 | 3.90 |
-| Decider 2B | noul-medium-q8 | 8 / 2 | 2138–2138 | 204.89 / 205.29 | 39.1 | 5.24 |
-| Decider 2B | noul-short-q32 | 32 / 2 | 204–204 | 95.38 / 95.82 | 335.8 | 4.87 |
-| Decider 2B | score-short-q1 | 1 / 4 | 222–225 | 24.42 / 24.66 | 41.2 | 3.74 |
-| Decider 2B | score-medium-q1 | 1 / 4 | 2156–2159 | 107.51 / 108.72 | 9.3 | 4.41 |
-| Decider 2B | score-long-q1 | 1 / 4 | 8776–8779 | 469.64 / 472.21 | 2.1 | 6.93 |
-| Decider 2B | score-short-q8 | 8 / 4 | 222–225 | 96.03 / 96.75 | 83.2 | 4.87 |
-| Decider 2B | score-medium-q8 | 8 / 4 | 2156–2159 | 773.01 / 774.95 | 10.4 | 9.79 |
-| Decider 2B | score-short-q32 | 32 / 4 | 222–225 | 353.70 / 355.63 | 90.4 | 8.71 |
-| Decider 2B | score-short-q1-levels2 | 1 / 2 | 234–234 | 23.60 / 24.24 | 42.3 | 3.66 |
-| Decider 2B | score-short-q1-levels10 | 1 / 10 | 234–234 | 37.79 / 38.16 | 26.4 | 3.98 |
-| Qwen3.5 2B LM-head | noul-short-q1 | 1 / 2 | 263–263 | 19.87 / 20.91 | 49.8 | 3.60 |
-| Qwen3.5 2B LM-head | noul-medium-q1 | 1 / 2 | 2197–2197 | 34.91 / 35.11 | 28.6 | 3.74 |
-| Qwen3.5 2B LM-head | noul-long-q1 | 1 / 2 | 8817–8817 | 127.50 / 128.67 | 7.8 | 4.21 |
-| Qwen3.5 2B LM-head | noul-short-q8 | 8 / 2 | 263–263 | 34.26 / 34.58 | 233.3 | 3.73 |
-| Qwen3.5 2B LM-head | noul-medium-q8 | 8 / 2 | 2197–2197 | 199.54 / 200.27 | 40.1 | 4.83 |
-| Qwen3.5 2B LM-head | noul-short-q32 | 32 / 2 | 263–263 | 134.40 / 136.62 | 237.4 | 3.74 |
-| Qwen3.5 2B LM-head | score-short-q1 | 1 / 4 | 303–303 | 19.90 / 20.24 | 50.2 | 3.60 |
-| Qwen3.5 2B LM-head | score-medium-q1 | 1 / 4 | 2237–2237 | 35.52 / 35.90 | 28.1 | 3.74 |
-| Qwen3.5 2B LM-head | score-long-q1 | 1 / 4 | 8857–8857 | 128.28 / 129.84 | 7.8 | 4.21 |
-| Qwen3.5 2B LM-head | score-short-q8 | 8 / 4 | 303–303 | 37.81 / 37.98 | 211.6 | 3.75 |
-| Qwen3.5 2B LM-head | score-medium-q8 | 8 / 4 | 2237–2237 | 202.06 / 202.79 | 39.6 | 4.86 |
-| Qwen3.5 2B LM-head | score-short-q32 | 32 / 4 | 303–303 | 148.68 / 149.07 | 215.8 | 3.76 |
-| Qwen3.5 2B LM-head | score-short-q1-levels2 | 1 / 2 | 286–286 | 19.93 / 20.44 | 50.0 | 3.60 |
-| Qwen3.5 2B LM-head | score-short-q1-levels10 | 1 / 10 | 438–438 | 20.13 / 20.45 | 49.6 | 3.61 |
-| Qwen3.5 4B LM-head | noul-short-q1 | 1 / 2 | 263–263 | 26.61 / 26.89 | 37.6 | 8.04 |
-| Qwen3.5 4B LM-head | noul-medium-q1 | 1 / 2 | 2197–2197 | 65.78 / 65.96 | 15.2 | 8.28 |
-| Qwen3.5 4B LM-head | noul-long-q1 | 1 / 2 | 8817–8817 | 256.56 / 258.56 | 3.9 | 9.11 |
-| Qwen3.5 4B LM-head | noul-short-q8 | 8 / 2 | 263–263 | 64.86 / 64.99 | 123.3 | 8.27 |
-| Qwen3.5 4B LM-head | noul-medium-q8 | 8 / 2 | 2197–2197 | 410.51 / 411.79 | 19.5 | 10.22 |
-| Qwen3.5 4B LM-head | noul-short-q32 | 32 / 2 | 263–263 | 257.86 / 258.60 | 124.4 | 8.28 |
-| Qwen3.5 4B LM-head | score-short-q1 | 1 / 4 | 303–303 | 26.60 / 26.74 | 37.6 | 8.04 |
-| Qwen3.5 4B LM-head | score-medium-q1 | 1 / 4 | 2237–2237 | 66.41 / 80.27 | 14.7 | 8.29 |
-| Qwen3.5 4B LM-head | score-long-q1 | 1 / 4 | 8857–8857 | 257.84 / 258.14 | 3.9 | 9.12 |
-| Qwen3.5 4B LM-head | score-short-q8 | 8 / 4 | 303–303 | 69.96 / 70.42 | 114.3 | 8.31 |
-| Qwen3.5 4B LM-head | score-medium-q8 | 8 / 4 | 2237–2237 | 416.17 / 446.59 | 19.0 | 10.26 |
-| Qwen3.5 4B LM-head | score-short-q32 | 32 / 4 | 303–303 | 278.88 / 280.80 | 114.7 | 8.31 |
-| Qwen3.5 4B LM-head | score-short-q1-levels2 | 1 / 2 | 286–286 | 26.65 / 30.12 | 36.4 | 8.04 |
-| Qwen3.5 4B LM-head | score-short-q1-levels10 | 1 / 10 | 438–438 | 26.82 / 27.25 | 37.2 | 8.06 |
-

USAGE.md

@@ -64,3 +64,7 @@ The bound is **16,384 tokens per complete question**, including its state, instr
 The qualified device is an AMD ROCm GPU accessed as `cuda:0` in PyTorch. Backbone parameters remain BF16, the candidate head FP32, with FLA Gated DeltaNet and SDPA. CPU and MPS inference are not implemented by the current engine and are rejected explicitly. Other GPU/runtime combinations require validation. The loader checks manifest hashes and the recorded runtime before loading. `allow_unvalidated_runtime=True` downgrades runtime differences to an explicit warning for experiments; it does not silently substitute a claimed validated environment or bypass unsupported CPU execution.
 
 This package is a loading wrapper, not a trainer, HTTP service or new model selection policy. No prediction logs or private local paths are required by the bundle.
+
+For the Sol v1.1 bundle, the validated normalization profile loads automatically. Use one model profile per fresh Python process. See [normalization runtime](NORMALIZATION_RUNTIME.md) for the supported shape and hardware contract.
+
+For this profiled Sol release, `allow_unvalidated_runtime=True` does not bypass the normalization profile version, GPU architecture or supported-shape checks.

assets/decision-capabilities.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1314aae927ea0aa598e62a80138e2ea9bfd8e851bc6faa91504d5aa691426e24
-size 45170
+oid sha256:41f1b6e93b41c7f2057ad9c0743ddf0fc643e0c48ba0de58d8a3b58dbc897f1e
+size 51635

backbone/model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cab4b9b9dfd490781acf489b78b6d3faa1d56a1be8fab5d8baf6c464e4c0ea1d
+oid sha256:168f7f90bf410130da92fd5270305286d53d5cbb68be02f1c611c40cb9243fe2
 size 3763685328

bundle-manifest.json

@@ -1,7 +1,22 @@
 {
   "format": "research-pointer-bundle-v1",
-  "status": "candidate-export-awaiting-independent-reload-and-quality-gates",
+  "status": "profile-bound-candidate-awaiting-default-public-entrypoint-offline-proof",
   "files": [
+    {
+      "file": "NORMALIZATION_RUNTIME.md",
+      "bytes": 1441,
+      "sha256": "9938bbe9ffdb890490c1db08903659edde6d7c5a5b9ce0487949e7a4dc2d38f6"
+    },
+    {
+      "file": "RUNTIME_BINDING.json",
+      "bytes": 1764,
+      "sha256": "c050b5a43d9ddf98e07924878a4e72cfb1c4746131a3045c31557c262791f922"
+    },
+    {
+      "file": "SOURCE_BUNDLE_MANIFEST.json",
+      "bytes": 4206,
+      "sha256": "9242b7fd4ce226f638108810ec5156dca6f270cce5a16efcf5ef091180ad75be"
+    },
     {
       "file": "backbone/config.json",
       "bytes": 1789,
@@ -10,7 +25,7 @@
     {
       "file": "backbone/model.safetensors",
       "bytes": 3763685328,
-      "sha256": "cab4b9b9dfd490781acf489b78b6d3faa1d56a1be8fab5d8baf6c464e4c0ea1d"
+      "sha256": "168f7f90bf410130da92fd5270305286d53d5cbb68be02f1c611c40cb9243fe2"
     },
     {
       "file": "chat_template.jinja",
@@ -19,8 +34,8 @@
     },
     {
       "file": "code/decision_api.py",
-      "bytes": 6724,
-      "sha256": "147b2fec32cbbbbb1b92cf2a19bb887f9d945e4974e141ca7ea93b8c542f5b21"
+      "bytes": 8251,
+      "sha256": "1b068eccdffd3c3b67bfa52f8f526e6b482d92551c927668ad28a794767f8a40"
     },
     {
       "file": "code/decision_model.py",
@@ -32,6 +47,16 @@
       "bytes": 3164,
       "sha256": "02352e8385ab47157b6459910da54d962e5da4bb4940d571faedc86bc5da9aee"
     },
+    {
+      "file": "code/profile_guard.py",
+      "bytes": 6131,
+      "sha256": "1603c39038ff783b9d5a5f69110d1695bbb7e28accffd0c03525854e8f258452"
+    },
+    {
+      "file": "code/runtime_profile.py",
+      "bytes": 2857,
+      "sha256": "afb59dda5e4c3890ee5c971549799e69f2e4fa2eafe5893de2168299fd368bda"
+    },
     {
       "file": "decision_config.json",
       "bytes": 753,
@@ -40,17 +65,77 @@
     {
       "file": "decision_head.safetensors",
       "bytes": 8424272,
-      "sha256": "308bc01085a2bcee89a6668cb06ca4b771e01338040689527127bdefd3cc7415"
+      "sha256": "924a43288d48804c22e7e2573898d694303879ab9bb819d1b20eadc2ab32f00d"
+    },
+    {
+      "file": "pyproject.toml",
+      "bytes": 436,
+      "sha256": "135a9516e87ca2fb41ef54a974e29132256b6c2d528a1cbdea1001e28f306946"
+    },
+    {
+      "file": "runtime-profile/l2norm_fwd_kernel.json",
+      "bytes": 13210,
+      "sha256": "a67f3b4624edc07e2c3f9f1d05553c654a3ff005c6b96a1831d62e323dea9edc"
+    },
+    {
+      "file": "runtime-profile/profile.json",
+      "bytes": 18101,
+      "sha256": "6b03450d42dbb68f0ffe14945ffcf3e6ea043e1033a819fb7211a8176a51722f"
     },
     {
       "file": "runtime.json",
-      "bytes": 952,
-      "sha256": "7ed527476826756d6076f23d0eaa8b4b53ff34b9f8bd57ae8e65b50e4d28cfa6"
+      "bytes": 1133,
+      "sha256": "9fe0a2bd626a8b974c85e4925cdc79dca320c9116c6d0196c12a87b9aeae8d2a"
+    },
+    {
+      "file": "src/decision/__init__.py",
+      "bytes": 167,
+      "sha256": "70de37df98b6fc8e3b9f9d43935ba31a32350496214c11adf8c5fba72c433313"
+    },
+    {
+      "file": "src/decision/example.py",
+      "bytes": 3581,
+      "sha256": "a54dec885f92c2d38d07ef2333dff51965be52bc119f772cc74200ed314e69b6"
+    },
+    {
+      "file": "src/decision/model.py",
+      "bytes": 9415,
+      "sha256": "ba240d7493fc29203fe036966f0ab911200cd4a9252b50b05423409977639ee0"
+    },
+    {
+      "file": "src/decision_local.egg-info/PKG-INFO",
+      "bytes": 225,
+      "sha256": "cc5266721b2e02c5c963b59856d979d619f7d9d687b60bda037c7bad2f22cd1d"
+    },
+    {
+      "file": "src/decision_local.egg-info/SOURCES.txt",
+      "bytes": 361,
+      "sha256": "741f317b6e9b27453e8f2fbeb76c0d1da326e89de7978fe87ce906ed1e2a68a9"
+    },
+    {
+      "file": "src/decision_local.egg-info/dependency_links.txt",
+      "bytes": 1,
+      "sha256": "01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b"
+    },
+    {
+      "file": "src/decision_local.egg-info/entry_points.txt",
+      "bytes": 59,
+      "sha256": "5400ff8d993d37849bc01b702e4674b8d9c4122af514101ace0bcd691eab39fb"
+    },
+    {
+      "file": "src/decision_local.egg-info/requires.txt",
+      "bytes": 31,
+      "sha256": "b8bf334329a333c3bba85388fbf41377b17305ee71b402c431c1af7445eadc08"
+    },
+    {
+      "file": "src/decision_local.egg-info/top_level.txt",
+      "bytes": 9,
+      "sha256": "834608781b00d5df58e3ae55a6b15c201aeebacc26ec99a1035fd75b83e16760"
     },
     {
       "file": "temperature.json",
-      "bytes": 667,
-      "sha256": "159d308a6b970c425b8938b6c50c4fff6a10bb40ca82b119e43386a855d609ab"
+      "bytes": 13733,
+      "sha256": "63af0fbd1c817c4bc354aab81df66626bcd9908ac7a71bcbb921e63ba7361d44"
     },
     {
       "file": "tokenizer.json",
@@ -88,22 +173,22 @@
     {
       "file": "backbone/model-00001-of-00002.safetensors",
       "bytes": 3999855496,
-      "sha256": "0cdb50da983e4d803d0a853bb6b1ad03634a118f734cfe5ecc8439090bd0e8dd"
+      "sha256": "93059b3982827486760a708d4f10aa6ae61de94ce90607b080021529bd1562ab"
     },
     {
       "file": "backbone/model-00002-of-00002.safetensors",
       "bytes": 3527479552,
-      "sha256": "9147e23c767af4994fa55ef1208fe27764eb3043e31c54585f509206bf053f7d"
+      "sha256": "2ef40a6f2cd250aa0c08a6ced808d1aa025bff3dd6720857651c8f8943cf5cc9"
     },
     {
       "file": "decision_config.json",
-      "bytes": 1109,
-      "sha256": "b68c7406a7dd5da3f6558d6404aef3aee81f1168a5fc56a009fe918056628202"
+      "bytes": 1119,
+      "sha256": "1e6d926e1593ba243e58448c3d3cca39c6731faa55e5acec369832e4efab074d"
     },
     {
       "file": "decision_head.safetensors",
       "bytes": 8424272,
-      "sha256": "308bc01085a2bcee89a6668cb06ca4b771e01338040689527127bdefd3cc7415"
+      "sha256": "924a43288d48804c22e7e2573898d694303879ab9bb819d1b20eadc2ab32f00d"
     },
     {
       "file": "tokenizer.json",
@@ -117,13 +202,16 @@
     }
   ],
   "source_model_code_sha256": "d3e28489c09f3bd7130e2d43d92e0b5c4a08e25b09b21303904defb0ff1c3646",
-  "source_api_code_sha256": "147b2fec32cbbbbb1b92cf2a19bb887f9d945e4974e141ca7ea93b8c542f5b21",
-  "dev_sha256": "a38f9be168553d5a82a91991aca86aaa602e1951957bac7c475abcec8dcc232d",
-  "production_predictions_sha256": "38df5ace521ee8ec7decc752ab7a49e5e7b2cda751c7d7d6a4fdf00ad8af4202",
-  "temperature_sha256": "159d308a6b970c425b8938b6c50c4fff6a10bb40ca82b119e43386a855d609ab",
-  "runtime_sha256": "7ed527476826756d6076f23d0eaa8b4b53ff34b9f8bd57ae8e65b50e4d28cfa6",
+  "source_api_code_sha256": "1b068eccdffd3c3b67bfa52f8f526e6b482d92551c927668ad28a794767f8a40",
+  "dev_sha256": "5ba3bd7527da831d77cec8c6a2f260a71e10a57742e356510c719f8f30d61a70",
+  "production_predictions_sha256": "b9e2cf2c466e6f9422e627905625400d683c6ce0cecb015c16e5784b0c3c577c",
+  "temperature_sha256": "63af0fbd1c817c4bc354aab81df66626bcd9908ac7a71bcbb921e63ba7361d44",
+  "runtime_sha256": "9fe0a2bd626a8b974c85e4925cdc79dca320c9116c6d0196c12a87b9aeae8d2a",
   "production_batch_size": 8,
   "input_length_limit": 16384,
-  "original_checkpoint_name": "checkpoint-000200",
-  "no_publication_performed": true
+  "original_checkpoint_name": "checkpoint-001563",
+  "no_publication_performed": true,
+  "source_bundle_manifest_sha256": "9242b7fd4ce226f638108810ec5156dca6f270cce5a16efcf5ef091180ad75be",
+  "normalization_profile_sha256": "6b03450d42dbb68f0ffe14945ffcf3e6ea043e1033a819fb7211a8176a51722f",
+  "public_wrapper_included": true
 }

code/decision_api.py

@@ -10,6 +10,32 @@ import math
 from pathlib import Path
 
 
+def prepare_runtime_profile(checkpoint, device='cuda:0'):
+    # This profile is verified before any dependency import can choose kernels.
+    import hashlib, json, sys
+    root=Path(checkpoint);runtime=json.loads((root/'runtime.json').read_text())
+    spec=runtime.get('normalization_profile')
+    if spec is None:
+        if '_decision_process_normalization_profile_v1' in sys.modules:
+            raise RuntimeError('Use separate processes for profiled and unprofiled models')
+        return None
+    import torch
+    target=torch.device(device)
+    if target.type!='cuda' or not torch.cuda.is_available():
+        raise RuntimeError('The bound profile requires a ROCm CUDA device')
+    arch=getattr(torch.cuda.get_device_properties(target),'gcnArchName','').split(':')[0]
+    if arch!=spec['validated_arch']:
+        raise RuntimeError('Target GPU architecture does not match the bound profile: '+arch)
+    relative=Path(spec['loader_file'])
+    if relative.is_absolute() or '..' in relative.parts:raise ValueError('Unsafe profile loader path')
+    path=root/relative
+    if hashlib.sha256(path.read_bytes()).hexdigest()!=spec['loader_sha256']:
+        raise ValueError('Bound runtime profile loader changed')
+    definition=importlib.util.spec_from_file_location('decision_bundle_runtime_profile',path)
+    module=importlib.util.module_from_spec(definition);definition.loader.exec_module(module)
+    return module.ensure_profile(root)
+
+
 def question_row(state, name, question):
     kind=question.get('type')
     if kind not in {'choice','noul','score'}:raise ValueError('Unknown question type')
@@ -61,6 +87,7 @@ def typed_answer(row, probabilities):
 class DecisionEngine:
     def __init__(self, checkpoint, model_code, *, device='cuda:0', max_length=16384,
                  batch_size=8, temperatures=None, model_name='local-decision-research'):
+        self.normalization_profile=prepare_runtime_profile(checkpoint, device=device)
         import torch
         path=Path(model_code)/'decision_model.py'
         spec=importlib.util.spec_from_file_location('research_decision_runtime',path)

code/profile_guard.py

@@ -0,0 +1,82 @@
+"""Fail-closed official FLA strict-config setup for one isolated process.
+
+No model/prompt/head changes. The guard prevents FLA STRICT's ordinary missing
+key fallback and records actual configured calls. This is a diagnostic module,
+not an installed change to the published wrapper or dependency environment.
+"""
+import hashlib,importlib,json,os,sys
+from pathlib import Path
+
+def sha(p):return hashlib.sha256(Path(p).read_bytes()).hexdigest()
+def serialized(key):return json.dumps(key,separators=(',',':'),sort_keys=True)
+def config_fields(config):
+ if isinstance(config,dict):return {k:config.get(k) for k in ['kwargs','num_warps','num_stages','num_ctas','maxnreg','ir_override']}
+ return {k:getattr(config,k,None) for k in ['kwargs','num_warps','num_stages','num_ctas','maxnreg','ir_override']}
+def validate_profile(path,expected_sha):
+ path=Path(path).resolve()
+ if sha(path)!=expected_sha:raise ValueError('Profile hash changed')
+ profile=json.loads(path.read_text())
+ if profile['format']!='decision-fla-l2norm-profile-v1' or profile['cache_mode']!='strict':raise ValueError('Profile format/mode unsupported')
+ if len(profile['files'])!=1 or profile['files'][0]['file']!='l2norm_fwd_kernel.json':raise ValueError('Unexpected profile file set')
+ f=path.parent/'l2norm_fwd_kernel.json'
+ if sha(f)!=profile['files'][0]['sha256']:raise ValueError('Explicit kernel config changed')
+ data=json.loads(f.read_text());entries={}
+ if data.get('default_config') is not None:raise ValueError('Implicit fallback defaults forbidden')
+ for h,item in data['autotune_entries'].items():
+  key=item['autotune_key'];encoded=serialized(key)
+  if hashlib.md5(encoded.encode()).hexdigest()!=h or encoded in entries:raise ValueError('Invalid/duplicate key')
+  if len(key)!=5 or key[0]!=128 or type(key[1]) is not int or not 1<=key[1]<=32 or key[2:]!=['torch.bfloat16','torch.bfloat16','torch.float32']:raise ValueError('Unsupported numerical key')
+  c=item['config']
+  if c['kwargs'].keys()!={'BT'} or c['kwargs']['BT'] not in [8,16,32] or c['num_warps'] not in [1,2,4,8,16] or c['num_stages']!=3 or c['num_ctas']!=1 or any(c.get(x) is not None for x in ['maxnreg','pre_hook','ir_override']):raise ValueError('Unexpected launch configuration')
+  entries[encoded]=c
+ if {json.loads(k)[1] for k in entries}!=set(range(1,33)):raise ValueError('Incomplete legal NB coverage')
+ return path,profile,entries
+
+def attach_guard(kernel,cache_module,entries,telemetry):
+ original=kernel.run
+ def guarded(*args,**kwargs):
+  if cache_module.FLA_CACHE_MODE is not cache_module.FlaCacheMode.STRICT:raise RuntimeError('FLA strict mode changed')
+  key=cache_module.AutotuneKey.build(kernel.arg_names,kernel.keys,args,kwargs);encoded=serialized(list(key.autotune_key))
+  if encoded not in entries:raise RuntimeError('Uncontracted FLA l2norm key: '+encoded)
+  expected=entries[encoded];loaded=cache_module.load_cached_config(kernel.kernel_name,key)
+  if config_fields(loaded)!=config_fields(expected):raise RuntimeError('FLA exact config lookup mismatch')
+  if key.autotune_key in kernel.cache and config_fields(kernel.cache[key.autotune_key])!=config_fields(expected):raise RuntimeError('A conflicting in-process kernel cache exists')
+  # Explicit official configuration load guarantees that the following original
+  # run finds this exact cache entry and cannot perform timing-based autotune.
+  kernel.maybe_load_cached_config(key)
+  if key.autotune_key not in kernel.cache or config_fields(kernel.cache[key.autotune_key])!=config_fields(expected):raise RuntimeError('Official strict config did not load')
+  result=original(*args,**kwargs)
+  if config_fields(kernel.cache[key.autotune_key])!=config_fields(expected):raise RuntimeError('Kernel config changed during call')
+  telemetry['calls']+=1;telemetry['keys'][encoded]=telemetry['keys'].get(encoded,0)+1
+  return result
+ kernel.run=guarded
+ return original
+
+def install(profile_path,expected_sha):
+ path,profile,entries=validate_profile(profile_path,expected_sha)
+ if any(n=='fla' or n.startswith('fla.') for n in sys.modules):raise RuntimeError('Install profile before importing FLA; use a fresh isolated process')
+ for name,wanted in {'FLA_CACHE_MODE':'strict','FLA_CONFIG_DIR':str(path.parent)}.items():
+  actual=os.environ.get(name)
+  if actual is not None and actual!=wanted:raise RuntimeError('Conflicting '+name)
+  os.environ[name]=wanted
+ import torch,triton,fla
+ actual={'torch':str(torch.__version__),'hip':torch.version.hip,'triton':triton.__version__,'fla':fla.__version__}
+ for name,value in actual.items():
+  if value!=profile['runtime'][name]:raise RuntimeError('Runtime mismatch: '+name)
+ if not torch.cuda.is_available():raise RuntimeError('Profile is only qualified for the specified ROCm GPU')
+ arch=torch.cuda.get_device_properties(0).gcnArchName.split(':')[0]
+ if arch!=profile['runtime']['gpu_arch']:raise RuntimeError('Unsupported GPU architecture '+arch)
+ module=importlib.import_module('fla.modules.l2norm');cache_module=importlib.import_module('fla.ops.utils.cache');root=Path(fla.__file__).parent
+ for name,value in profile['fla_source_sha256'].items():
+  if sha(root/name)!=value:raise RuntimeError('Pinned FLA source changed: '+name)
+ kernel=module.l2norm_fwd_kernel
+ if kernel.kernel_name!='l2norm_fwd_kernel' or kernel.keys!=['D','NB'] or kernel.cache:raise RuntimeError('Kernel identity or fresh-cache precondition failed')
+ telemetry={'profile_sha256':expected_sha,'status':'installed','calls':0,'keys':{},'strict_guard':True,'unknown_keys':'raise','autotune_fallback_permitted':False,'runtime':actual,'gpu_arch':arch,'process_scope':'one explicitly profiled Sol model; other model loading in this process is not supported'}
+ attach_guard(kernel,cache_module,entries,telemetry)
+ # The validated inference path only uses the vectorized D128 forward kernel.
+ # Other dimensions/backward must not silently enter a different autotuner.
+ for name in ['l2norm_fwd_kernel1','l2norm_bwd_kernel','l2norm_bwd_kernel1']:
+  other=getattr(module,name)
+  def reject(*args,_name=name,**kwargs):raise RuntimeError('Uncontracted normalization kernel: '+_name)
+  other.run=reject
+ return telemetry

code/runtime_profile.py

@@ -0,0 +1,36 @@
+"""Bundle-local automatic launch-profile binding, before importing FLA.
+
+One profile is active per Python process. Repeated loading of the same verified
+profile is allowed; mixing with an unprofiled/different-profile model is not.
+"""
+import hashlib,importlib.util,json,os,sys,types
+from pathlib import Path
+STATE='_decision_process_normalization_profile_v1'
+def sha(p):return hashlib.sha256(Path(p).read_bytes()).hexdigest()
+def safe(root,relative):
+ p=Path(relative)
+ if p.is_absolute() or '..' in p.parts:raise ValueError('Unsafe runtime-profile path')
+ return root/p
+
+def ensure_profile(bundle):
+ bundle=Path(bundle).resolve();runtime=json.loads((bundle/'runtime.json').read_text());spec=runtime.get('normalization_profile');active=sys.modules.get(STATE)
+ if spec is None:
+  if active is not None:raise RuntimeError('Load unprofiled and profiled Decision models in separate processes')
+  return None
+ if spec.get('kind')!='decision-fla-l2norm-profile-v1' or spec.get('validated_arch')!='gfx942':raise ValueError('Unknown normalization profile contract')
+ profile=safe(bundle,spec['profile_file']);guard=safe(bundle,spec['guard_file'])
+ if sha(profile)!=spec['profile_sha256'] or sha(guard)!=spec['guard_sha256']:raise ValueError('Bound runtime profile bytes changed')
+ if json.loads((bundle/'decision_config.json').read_text()).get('base_model')!='Qwen/Qwen3.5-2B':raise ValueError('This bundle profile is bound to the validated Sol family only')
+ if active is not None:
+  if active.profile_sha256!=spec['profile_sha256'] or active.guard_sha256!=spec['guard_sha256']:raise RuntimeError('Different Decision normalization profile is already active; use a separate process')
+  active.guard.validate_profile(profile,spec['profile_sha256'])
+  if os.environ.get('FLA_CACHE_MODE')!='strict' or os.environ.get('FLA_CONFIG_DIR')!=active.profile_dir:raise RuntimeError('Active FLA profile environment changed')
+  return {'profile_sha256':active.profile_sha256,'guard_sha256':active.guard_sha256,'validated_arch':'gfx942','automatic_bundle_binding':True,'scope':'single profile per process'}
+ module_spec=importlib.util.spec_from_file_location('decision_profile_guard_'+spec['guard_sha256'][:16],guard);module=importlib.util.module_from_spec(module_spec);module_spec.loader.exec_module(module)
+ telemetry=module.install(profile,spec['profile_sha256'])
+ state=types.ModuleType(STATE);state.profile_sha256=spec['profile_sha256'];state.guard_sha256=spec['guard_sha256'];state.profile_dir=str(profile.parent);state.guard=module;state.telemetry=telemetry;sys.modules[STATE]=state
+ return {'profile_sha256':state.profile_sha256,'guard_sha256':state.guard_sha256,'validated_arch':'gfx942','automatic_bundle_binding':True,'scope':'single profile per process'}
+
+def active_telemetry():
+ state=sys.modules.get(STATE)
+ return None if state is None else dict(state.telemetry)

decision_head.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:308bc01085a2bcee89a6668cb06ca4b771e01338040689527127bdefd3cc7415
+oid sha256:924a43288d48804c22e7e2573898d694303879ab9bb819d1b20eadc2ab32f00d
 size 8424272

metrics/asset-hashes.json

@@ -1,17 +0,0 @@
-{
-  "readout.png": "fb6b3d1cf149a1a3ccb584eb1c213df73c07bcd328732b4238928e2a2a4af085",
-  "readout.pdf": "d489b0d1814e37495a8612c12fabd3bf4a6e73d11c7e356bf1ce0cedded8a745",
-  "decision-family-header.png": "213511289ce8df038d938ac470e803c427ed57f0f85cc397dd4d79964b866541",
-  "readout.svg": "701084bf7b0858a3adb247ae446df737b77d4b75004341dfd5f956b933f3d253",
-  "architecture.svg": "215d9af6f94cc24847fc1d23a0b2a287b36afdc1e10ea4f5c82895b2fe103002",
-  "decision-quality.png": "50af70f9120e1f13024a18b264d9471c0f441464f33188306e5f5e0dc0cc4955",
-  "decision-capabilities.png": "ec86958b569c84c9ed8fcd8180095f88bbc3fe8f8ded70f685b08bd6f59f2ef6",
-  "decision-capabilities.pdf": "8cd467055c05d5c7616e776174ed0e3e953e71672558ea346e347e8663b5e5ce",
-  "decision-quality.pdf": "403a02af4ba6292d27fe2863dc568a026278002fbb58082befed37ac55f31ce4",
-  "architecture-atlas.pdf": "d1582ac115aa6a40129b6221e15a197e33d20847e03d69396f02fdab82771376",
-  "decision-mark.png": "d83cf7878c3839f3085feb8c02334217ad2fef0a615b805e1726b292552c46e8",
-  "decision-capabilities.svg": "e3b03c292b420b039b972756ae40e93e73c0610672a9bca5d795736faeb7f697",
-  "decision-quality.svg": "b430df49b3c74fa0cd8b7e39ab1a7f6768295feaa423213816fa6ce8f2fde8a2",
-  "architecture.pdf": "ee87f997a0aa4343c2ec66038b7526361e0a3c0ae02b383f8928aac9dd9bd74c",
-  "architecture.png": "8fcfa10b3941ee4aec1f5fba20a25e382a7d5de83074071c9ef68e627996de40"
-}

metrics/natural-confirmation.json

@@ -0,0 +1,877 @@
+{
+  "version": "sol-v4-observed-reading-public-1",
+  "requested_examples": 480,
+  "families": {
+    "boolq": {
+      "requested": 160,
+      "false": 80,
+      "true": 80,
+      "weight": "1/2"
+    },
+    "belebele_en": {
+      "requested": 160,
+      "weight": "1/4"
+    },
+    "belebele_zh": {
+      "requested": 160,
+      "weight": "1/4"
+    }
+  },
+  "score_definition": "0.5 * BoolQ accuracy + 0.25 * Belebele English accuracy + 0.25 * Belebele Chinese accuracy; all 480 requested examples retained.",
+  "models": {
+    "jev-1.13.0": {
+      "label": "Jev \u00b7 1.13.0",
+      "weighted_accuracy": 0.9453125,
+      "weighted_accuracy_exact": {
+        "numerator": 121,
+        "denominator": 128
+      },
+      "weighted_accuracy_ci95": [
+        0.9200550899983875,
+        0.9676459580838324
+      ],
+      "families": {
+        "boolq": {
+          "correct": 148,
+          "requested": 160,
+          "accuracy": 0.925,
+          "ci95": [
+            0.88125,
+            0.9625
+          ]
+        },
+        "belebele_en": {
+          "correct": 155,
+          "requested": 160,
+          "accuracy": 0.96875,
+          "ci95": [
+            0.9386503067484663,
+            0.9936708860759493
+          ]
+        },
+        "belebele_zh": {
+          "correct": 154,
+          "requested": 160,
+          "accuracy": 0.9625,
+          "ci95": [
+            0.9308176100628931,
+            0.9877300613496932
+          ]
+        }
+      },
+      "unweighted": {
+        "correct": 457,
+        "requested": 480,
+        "accuracy": 0.9520833333333333
+      },
+      "coverage": {
+        "requested": 480,
+        "accepted": 480,
+        "finite_valid_outputs": 480,
+        "full_input_verified": null,
+        "truncated": null,
+        "retention": "Unknown inside the closed API."
+      },
+      "typed_probability_metrics": {
+        "noul": {
+          "n": 160,
+          "valid_probability_coverage": 160,
+          "missing_or_invalid": 0,
+          "accuracy": 0.925,
+          "brier": 0.10445375000000001,
+          "nll": 0.19516064889282234,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null,
+          "class_support": {
+            "0": 80,
+            "1": 80
+          },
+          "recall_false": 0.925,
+          "recall_true": 0.925,
+          "balanced_accuracy": 0.925,
+          "recall_denominator": "All gold rows, including invalid/missing predictions as wrong"
+        },
+        "choice": {
+          "n": 320,
+          "valid_probability_coverage": 320,
+          "missing_or_invalid": 0,
+          "accuracy": 0.965625,
+          "brier": 0.040956875,
+          "nll": 0.07843730477267183,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null
+        }
+      },
+      "natural_choice": {
+        "correct": 309,
+        "requested": 320,
+        "accuracy": 0.965625
+      }
+    },
+    "decider": {
+      "label": "Decider \u00b7 2B",
+      "weighted_accuracy": 0.9203125,
+      "weighted_accuracy_exact": {
+        "numerator": 589,
+        "denominator": 640
+      },
+      "weighted_accuracy_ci95": [
+        0.8916901676829269,
+        0.9464989496201366
+      ],
+      "families": {
+        "boolq": {
+          "correct": 147,
+          "requested": 160,
+          "accuracy": 0.91875,
+          "ci95": [
+            0.875,
+            0.95625
+          ]
+        },
+        "belebele_en": {
+          "correct": 149,
+          "requested": 160,
+          "accuracy": 0.93125,
+          "ci95": [
+            0.8902439024390244,
+            0.967948717948718
+          ]
+        },
+        "belebele_zh": {
+          "correct": 146,
+          "requested": 160,
+          "accuracy": 0.9125,
+          "ci95": [
+            0.8630864845938376,
+            0.9559748427672956
+          ]
+        }
+      },
+      "unweighted": {
+        "correct": 442,
+        "requested": 480,
+        "accuracy": 0.9208333333333333
+      },
+      "coverage": {
+        "requested": 480,
+        "accepted": 480,
+        "finite_valid_outputs": 480,
+        "full_input_verified": 480,
+        "truncated": 0,
+        "retention": "Complete frozen inputs verified."
+      },
+      "typed_probability_metrics": {
+        "noul": {
+          "n": 160,
+          "valid_probability_coverage": 160,
+          "missing_or_invalid": 0,
+          "accuracy": 0.91875,
+          "brier": 0.14787474037499998,
+          "nll": 0.2631311328165121,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null,
+          "class_support": {
+            "0": 80,
+            "1": 80
+          },
+          "recall_false": 0.9375,
+          "recall_true": 0.9,
+          "balanced_accuracy": 0.91875,
+          "recall_denominator": "All gold rows, including invalid/missing predictions as wrong"
+        },
+        "choice": {
+          "n": 320,
+          "valid_probability_coverage": 320,
+          "missing_or_invalid": 0,
+          "accuracy": 0.921875,
+          "brier": 0.12688092915001592,
+          "nll": 0.24728184456069316,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null
+        }
+      },
+      "natural_choice": {
+        "correct": 295,
+        "requested": 320,
+        "accuracy": 0.921875
+      }
+    },
+    "nox-v1.1": {
+      "label": "Nox \u00b7 4B \u00b7 v1.1",
+      "weighted_accuracy": 0.784375,
+      "weighted_accuracy_exact": {
+        "numerator": 251,
+        "denominator": 320
+      },
+      "weighted_accuracy_ci95": [
+        0.7399293154761905,
+        0.8265123552522746
+      ],
+      "families": {
+        "boolq": {
+          "correct": 140,
+          "requested": 160,
+          "accuracy": 0.875,
+          "ci95": [
+            0.81875,
+            0.925
+          ]
+        },
+        "belebele_en": {
+          "correct": 114,
+          "requested": 160,
+          "accuracy": 0.7125,
+          "ci95": [
+            0.6380368098159509,
+            0.782608695652174
+          ]
+        },
+        "belebele_zh": {
+          "correct": 108,
+          "requested": 160,
+          "accuracy": 0.675,
+          "ci95": [
+            0.5975609756097561,
+            0.7469135802469136
+          ]
+        }
+      },
+      "unweighted": {
+        "correct": 362,
+        "requested": 480,
+        "accuracy": 0.7541666666666667
+      },
+      "coverage": {
+        "requested": 480,
+        "accepted": 480,
+        "finite_valid_outputs": 480,
+        "full_input_verified": 480,
+        "truncated": 0,
+        "retention": "Complete frozen inputs verified."
+      },
+      "typed_probability_metrics": {
+        "noul": {
+          "n": 160,
+          "valid_probability_coverage": 160,
+          "missing_or_invalid": 0,
+          "accuracy": 0.875,
+          "brier": 0.21024758152305395,
+          "nll": 0.4175424597184884,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null,
+          "class_support": {
+            "0": 80,
+            "1": 80
+          },
+          "recall_false": 0.875,
+          "recall_true": 0.875,
+          "balanced_accuracy": 0.875,
+          "recall_denominator": "All gold rows, including invalid/missing predictions as wrong"
+        },
+        "choice": {
+          "n": 320,
+          "valid_probability_coverage": 320,
+          "missing_or_invalid": 0,
+          "accuracy": 0.69375,
+          "brier": 0.4329685852043793,
+          "nll": 1.0936456725285504,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null
+        }
+      },
+      "natural_choice": {
+        "correct": 222,
+        "requested": 320,
+        "accuracy": 0.6937500000000001
+      }
+    },
+    "sol-v1.1": {
+      "label": "Sol \u00b7 2B \u00b7 v1.1",
+      "weighted_accuracy": 0.734375,
+      "weighted_accuracy_exact": {
+        "numerator": 47,
+        "denominator": 64
+      },
+      "weighted_accuracy_ci95": [
+        0.6869478485838779,
+        0.7795678401898735
+      ],
+      "families": {
+        "boolq": {
+          "correct": 127,
+          "requested": 160,
+          "accuracy": 0.79375,
+          "ci95": [
+            0.73125,
+            0.85625
+          ]
+        },
+        "belebele_en": {
+          "correct": 109,
+          "requested": 160,
+          "accuracy": 0.68125,
+          "ci95": [
+            0.60625,
+            0.7532467532467533
+          ]
+        },
+        "belebele_zh": {
+          "correct": 107,
+          "requested": 160,
+          "accuracy": 0.66875,
+          "ci95": [
+            0.5925925925925926,
+            0.7419354838709677
+          ]
+        }
+      },
+      "unweighted": {
+        "correct": 343,
+        "requested": 480,
+        "accuracy": 0.7145833333333333
+      },
+      "coverage": {
+        "requested": 480,
+        "accepted": 480,
+        "finite_valid_outputs": 480,
+        "full_input_verified": 480,
+        "truncated": 0,
+        "retention": "Complete frozen inputs verified."
+      },
+      "typed_probability_metrics": {
+        "noul": {
+          "n": 160,
+          "valid_probability_coverage": 160,
+          "missing_or_invalid": 0,
+          "accuracy": 0.79375,
+          "brier": 0.3233785546556422,
+          "nll": 0.6047157313227172,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null,
+          "class_support": {
+            "0": 80,
+            "1": 80
+          },
+          "recall_false": 0.7375,
+          "recall_true": 0.85,
+          "balanced_accuracy": 0.79375,
+          "recall_denominator": "All gold rows, including invalid/missing predictions as wrong"
+        },
+        "choice": {
+          "n": 320,
+          "valid_probability_coverage": 320,
+          "missing_or_invalid": 0,
+          "accuracy": 0.675,
+          "brier": 0.4428554242484616,
+          "nll": 1.0762990015746408,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null
+        }
+      },
+      "natural_choice": {
+        "correct": 216,
+        "requested": 320,
+        "accuracy": 0.675
+      }
+    },
+    "qwen35-4b-lm-head": {
+      "label": "Qwen3.5 \u00b7 4B \u00b7 untuned",
+      "weighted_accuracy": 0.8796875,
+      "weighted_accuracy_exact": {
+        "numerator": 563,
+        "denominator": 640
+      },
+      "weighted_accuracy_ci95": [
+        0.8453125000000001,
+        0.9125
+      ],
+      "families": {
+        "boolq": {
+          "correct": 134,
+          "requested": 160,
+          "accuracy": 0.8375,
+          "ci95": [
+            0.78125,
+            0.89375
+          ]
+        },
+        "belebele_en": {
+          "correct": 149,
+          "requested": 160,
+          "accuracy": 0.93125,
+          "ci95": [
+            0.89171974522293,
+            0.967948717948718
+          ]
+        },
+        "belebele_zh": {
+          "correct": 146,
+          "requested": 160,
+          "accuracy": 0.9125,
+          "ci95": [
+            0.8647011385199241,
+            0.9559748427672956
+          ]
+        }
+      },
+      "unweighted": {
+        "correct": 429,
+        "requested": 480,
+        "accuracy": 0.89375
+      },
+      "coverage": {
+        "requested": 480,
+        "accepted": 480,
+        "finite_valid_outputs": 480,
+        "full_input_verified": 480,
+        "truncated": 0,
+        "retention": "Complete frozen inputs verified."
+      },
+      "typed_probability_metrics": {
+        "noul": {
+          "n": 160,
+          "valid_probability_coverage": 160,
+          "missing_or_invalid": 0,
+          "accuracy": 0.8375,
+          "brier": 0.2685250413432339,
+          "nll": 0.43323629871265307,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null,
+          "class_support": {
+            "0": 80,
+            "1": 80
+          },
+          "recall_false": 0.7875,
+          "recall_true": 0.8875,
+          "balanced_accuracy": 0.8374999999999999,
+          "recall_denominator": "All gold rows, including invalid/missing predictions as wrong"
+        },
+        "choice": {
+          "n": 320,
+          "valid_probability_coverage": 320,
+          "missing_or_invalid": 0,
+          "accuracy": 0.921875,
+          "brier": 0.12427284941036536,
+          "nll": 0.26369763063759766,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null
+        }
+      },
+      "natural_choice": {
+        "correct": 295,
+        "requested": 320,
+        "accuracy": 0.921875
+      }
+    },
+    "qwen35-2b-lm-head": {
+      "label": "Qwen3.5 \u00b7 2B \u00b7 untuned",
+      "weighted_accuracy": 0.7375,
+      "weighted_accuracy_exact": {
+        "numerator": 59,
+        "denominator": 80
+      },
+      "weighted_accuracy_ci95": [
+        0.6917801490514904,
+        0.7823309748427674
+      ],
+      "families": {
+        "boolq": {
+          "correct": 104,
+          "requested": 160,
+          "accuracy": 0.65,
+          "ci95": [
+            0.575,
+            0.725
+          ]
+        },
+        "belebele_en": {
+          "correct": 134,
+          "requested": 160,
+          "accuracy": 0.8375,
+          "ci95": [
+            0.7784810126582279,
+            0.89171974522293
+          ]
+        },
+        "belebele_zh": {
+          "correct": 130,
+          "requested": 160,
+          "accuracy": 0.8125,
+          "ci95": [
+            0.746986092518221,
+            0.8727272727272727
+          ]
+        }
+      },
+      "unweighted": {
+        "correct": 368,
+        "requested": 480,
+        "accuracy": 0.7666666666666667
+      },
+      "coverage": {
+        "requested": 480,
+        "accepted": 480,
+        "finite_valid_outputs": 480,
+        "full_input_verified": 480,
+        "truncated": 0,
+        "retention": "Complete frozen inputs verified."
+      },
+      "typed_probability_metrics": {
+        "noul": {
+          "n": 160,
+          "valid_probability_coverage": 160,
+          "missing_or_invalid": 0,
+          "accuracy": 0.65,
+          "brier": 0.4013130395712821,
+          "nll": 0.5740625822372918,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null,
+          "class_support": {
+            "0": 80,
+            "1": 80
+          },
+          "recall_false": 0.9375,
+          "recall_true": 0.3625,
+          "balanced_accuracy": 0.65,
+          "recall_denominator": "All gold rows, including invalid/missing predictions as wrong"
+        },
+        "choice": {
+          "n": 320,
+          "valid_probability_coverage": 320,
+          "missing_or_invalid": 0,
+          "accuracy": 0.825,
+          "brier": 0.23224826689575004,
+          "nll": 0.47036782610781563,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null
+        }
+      },
+      "natural_choice": {
+        "correct": 264,
+        "requested": 320,
+        "accuracy": 0.825
+      }
+    },
+    "laya-routed": {
+      "label": "Laya \u00b7 EN/ML",
+      "weighted_accuracy": 0.5125,
+      "weighted_accuracy_exact": {
+        "numerator": 41,
+        "denominator": 80
+      },
+      "weighted_accuracy_ci95": [
+        0.465988068600994,
+        0.5583870648734177
+      ],
+      "families": {
+        "boolq": {
+          "correct": 111,
+          "requested": 160,
+          "accuracy": 0.69375,
+          "ci95": [
+            0.61875,
+            0.7625
+          ]
+        },
+        "belebele_en": {
+          "correct": 61,
+          "requested": 160,
+          "accuracy": 0.38125,
+          "ci95": [
+            0.30538922155688625,
+            0.4591194968553459
+          ]
+        },
+        "belebele_zh": {
+          "correct": 45,
+          "requested": 160,
+          "accuracy": 0.28125,
+          "ci95": [
+            0.2147239263803681,
+            0.3515151515151515
+          ]
+        }
+      },
+      "unweighted": {
+        "correct": 217,
+        "requested": 480,
+        "accuracy": 0.45208333333333334
+      },
+      "coverage": {
+        "requested": 480,
+        "accepted": 480,
+        "finite_valid_outputs": 480,
+        "full_input_verified": 480,
+        "truncated": 0,
+        "retention": "Complete frozen inputs verified."
+      },
+      "typed_probability_metrics": {
+        "noul": {
+          "n": 160,
+          "valid_probability_coverage": 160,
+          "missing_or_invalid": 0,
+          "accuracy": 0.69375,
+          "brier": 0.41904120474999995,
+          "nll": 0.6464549663099783,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null,
+          "class_support": {
+            "0": 80,
+            "1": 80
+          },
+          "recall_false": 0.525,
+          "recall_true": 0.8625,
+          "balanced_accuracy": 0.6937500000000001,
+          "recall_denominator": "All gold rows, including invalid/missing predictions as wrong"
+        },
+        "choice": {
+          "n": 320,
+          "valid_probability_coverage": 320,
+          "missing_or_invalid": 0,
+          "accuracy": 0.33125,
+          "brier": 0.8108916245962444,
+          "nll": 1.5869034433374543,
+          "zero_true_probability_count": 0,
+          "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+          "soft_target_coverage": 0,
+          "soft_target_nll": null
+        }
+      },
+      "natural_choice": {
+        "correct": 106,
+        "requested": 320,
+        "accuracy": 0.33125
+      }
+    }
+  },
+  "previous_sol_reference": {
+    "label": "Sol \u00b7 2B \u00b7 v1.0",
+    "weighted_accuracy": 0.69375,
+    "weighted_accuracy_exact": {
+      "numerator": 111,
+      "denominator": 160
+    },
+    "weighted_accuracy_ci95": [
+      0.6437499999999999,
+      0.7417484177215189
+    ],
+    "families": {
+      "boolq": {
+        "correct": 119,
+        "requested": 160,
+        "accuracy": 0.74375,
+        "ci95": [
+          0.675,
+          0.8125
+        ]
+      },
+      "belebele_en": {
+        "correct": 106,
+        "requested": 160,
+        "accuracy": 0.6625,
+        "ci95": [
+          0.5870967741935483,
+          0.7349397590361446
+        ]
+      },
+      "belebele_zh": {
+        "correct": 100,
+        "requested": 160,
+        "accuracy": 0.625,
+        "ci95": [
+          0.546583850931677,
+          0.7005988023952096
+        ]
+      }
+    },
+    "unweighted": {
+      "correct": 325,
+      "requested": 480,
+      "accuracy": 0.6770833333333334
+    },
+    "coverage": {
+      "requested": 480,
+      "accepted": 480,
+      "finite_valid_outputs": 480,
+      "full_input_verified": 480,
+      "truncated": 0,
+      "retention": "Complete frozen inputs verified."
+    },
+    "typed_probability_metrics": {
+      "noul": {
+        "n": 160,
+        "valid_probability_coverage": 160,
+        "missing_or_invalid": 0,
+        "accuracy": 0.74375,
+        "brier": 0.43895629019878823,
+        "nll": 0.9002462702649545,
+        "zero_true_probability_count": 0,
+        "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+        "soft_target_coverage": 0,
+        "soft_target_nll": null,
+        "class_support": {
+          "0": 80,
+          "1": 80
+        },
+        "recall_false": 0.575,
+        "recall_true": 0.9125,
+        "balanced_accuracy": 0.7437499999999999,
+        "recall_denominator": "All gold rows, including invalid/missing predictions as wrong"
+      },
+      "choice": {
+        "n": 320,
+        "valid_probability_coverage": 320,
+        "missing_or_invalid": 0,
+        "accuracy": 0.64375,
+        "brier": 0.5085778581327408,
+        "nll": 1.2810086045362705,
+        "zero_true_probability_count": 0,
+        "nll_policy": "Natural log; no epsilon clipping; infinity is explicit JSON string; proper scores conditional on valid coverage",
+        "soft_target_coverage": 0,
+        "soft_target_nll": null
+      }
+    }
+  },
+  "sol_update_comparison": {
+    "point_delta": 0.040625,
+    "exact_fraction": "13/320",
+    "paired_ci95": [
+      0.010993975903614502,
+      0.0703125
+    ]
+  },
+  "bootstrap": {
+    "resamples": 10000,
+    "seed": 202609215404,
+    "weights": {
+      "boolq": "1/2",
+      "belebele_en": "1/4",
+      "belebele_zh": "1/4"
+    },
+    "clusters": "BoolQ complete normalized passage; Belebele source link with all question IDs and both translations coupled. Identical random draws shared across models.",
+    "point_denominator": "All requested rows; invalid/unsupported count as incorrect.",
+    "readiness": "Descriptive confirmation only; no change to active V3 update policy."
+  },
+  "sources": [
+    {
+      "name": "BoolQ validation",
+      "revision": "35b264d03638db9f4ce671b711558bf7ff0f80d5",
+      "license": "CC BY-SA 3.0",
+      "url": "https://huggingface.co/datasets/google/boolq/tree/35b264d03638db9f4ce671b711558bf7ff0f80d5"
+    },
+    {
+      "name": "Belebele English/Chinese test",
+      "revision": "7899cdfa4e1e0d733fd77c848e2c273cb1d32be2",
+      "license": "CC BY-SA 4.0",
+      "url": "https://huggingface.co/datasets/facebook/belebele/tree/7899cdfa4e1e0d733fd77c848e2c273cb1d32be2"
+    }
+  ],
+  "scope": "Evidence yes/no and four-choice reading comprehension; no ordinal Score or all-capability claim.",
+  "changes_release_gate": false,
+  "candidate_bundle_manifest_sha256": "009d4892300c0a98f6a9b61706cdb6cc7ce2f6883c25c120477b98fe7fe82d0f",
+  "candidate_normalization_profile_sha256": "6b03450d42dbb68f0ffe14945ffcf3e6ea043e1033a819fb7211a8176a51722f",
+  "nox_peer_bundle_manifest_sha256": "a8cfd42bc908fc93d5a2a20be527d09704b6e3f6c1217f27307c32cc2d822971",
+  "nox_peer_published_revision": "5bee3061a7abf675813963e4e05267cb1a28f92f",
+  "sol_v1_revision": "b6a4ce881ca625dcf19663a277d27577cfe4540a",
+  "independence": {
+    "excluded_from_custom_train_select_cal": true,
+    "original_human_annotations": true,
+    "new_human_adjudication": false,
+    "upstream_pretraining_exposure": "Unknown",
+    "first_global_unseal_at": "2026-09-21T16:40:29.779741+00:00",
+    "candidate_frozen_before_first_unseal": false,
+    "subsequent_use": "Observed regression for candidates not frozen before first exposure.",
+    "source_parent_clusters": {
+      "boolq_passages": 160,
+      "belebele_shared_passages": 144
+    },
+    "overlap_audit": "No exact or character-5-gram Jaccard >=0.5 overlap across 51 custom-data and earlier-panel files; this is not proof of semantic or upstream-pretraining independence.",
+    "boolq_grouping_limit": "Passage grouping only; this distribution lacks article titles, so article-level independence is not established.",
+    "candidate_exposure": {
+      "global_first_unseal": {
+        "policy_sha256": "259b907f3634471f7c077acbe6cb4eba6ebb85d24ba305c86ea6a679349491c3",
+        "first_global_unseal_at": "2026-09-21T16:40:29.779741+00:00",
+        "first_candidate": "Decision-1.0-Nox-stage4-selected",
+        "first_candidate_bundle_manifest_sha256": "a8cfd42bc908fc93d5a2a20be527d09704b6e3f6c1217f27307c32cc2d822971",
+        "first_candidate_dispatch_sha256": "c04b96e7beb312cfd2196299aaaef23c8bd5a0acccc01c7d83002da76830652a",
+        "meaning": "V4 natural confirmation is observed regression for all subsequent unfrozen candidates. This first exposure record is never replaced."
+      },
+      "this_evaluation_at": "2026-09-21T18:05:29.511389+00:00",
+      "candidate_frozen_at": "2026-09-21T17:53:15.626713+00:00",
+      "candidate_frozen_before_first_global_unseal": false,
+      "candidate_is_first_unseal_identity": false,
+      "confirmation_interpretation": "Observed regression for this candidate; do not relabel it as fresh confirmation.",
+      "historical_panel_key": "v4_natural_confirmation",
+      "historical_key_is_not_a_new_freshness_claim": true
+    }
+  },
+  "provenance": {
+    "sol_report": {
+      "aggregate": "bb5fab144cb57513bbb7ce22ce2c8af38a5a3d28709df362e230bfa2995454a9",
+      "statistics": "fa3fe24079d7221b19fa522b38da7d9b4c03687130749c46e77c029d0f7bf951",
+      "receipt": "968ab2a8c6228bd61444001d19529598c90f69f89b097b1269e132eb86b35878"
+    },
+    "nox_peer_report": {
+      "aggregate": "c55a218a533751cf6fd8e783182bb2edfa1731f1a7aa3446fbed4575d6291dcb",
+      "statistics": "23105cc6fbffb15b1fac390b64bf31b70403091b5f0d4a3de06a78a27b3c5474",
+      "receipt": "e3220736d2da9c87d7b08808fa8903af0ce1c90961f88e7a54a8348abca29794"
+    },
+    "nox_publication_receipt_sha256": "5e0a3b7ba2637c5689e225b850441d532816f41b9f9240cce73a3ec9b66c0ea5",
+    "policy_sha256": "259b907f3634471f7c077acbe6cb4eba6ebb85d24ba305c86ea6a679349491c3",
+    "suite_manifest_sha256": "969a92c4c5a2f395a3ba55d3a4aff28be060d89327c3b09de3e1f5188cf74830",
+    "prediction_sha256": {
+      "jev-1.13.0": {
+        "predictions_sha256": "d6583f903e974d70747fae2944014828954dc4d7344ceddcbb2cd1b06fbfc799"
+      },
+      "decider": {
+        "predictions_sha256": "626ef366ae471ec10dfb89ef2ff2f7b29a6c879976d3ff14fcbb2bafd7d9b045"
+      },
+      "nox-v1.1": {
+        "predictions_sha256": "dedd6747282d694f2e5c7a79ee4133225f01cb55b1c7d8fd6882d3dddce6c723",
+        "raw_T1_sha256": "08e3cd068b1ccbbffc216e53d6fb9deb64c5a66b2ce6a5b0cbe9aa2667304ea4"
+      },
+      "sol-v1.1": {
+        "predictions_sha256": "0c7156e7d8709175ecf352c8aa5e8511f1eb0669c965ea5311992efb3d479dc3",
+        "raw_T1_sha256": "c7d581453f2c0f51f89051bcd8574fc6333c39475b20fb7dcb82e757580f24bc"
+      },
+      "qwen35-4b-lm-head": {
+        "predictions_sha256": "7a6d252ccaa72bf54b94445190bbad023fb465a3eaf4cf657532e2cdd4f750f5"
+      },
+      "qwen35-2b-lm-head": {
+        "predictions_sha256": "93d84cdbe607691e2cef35ece3aa2dd5900f6410b0d72796b86d74c9156f3cdd"
+      },
+      "laya-routed": {
+        "predictions_sha256": "a10922609e2223ca0a818667bc1425d476aec58a4d881002e0e322596c1d0b8d"
+      }
+    }
+  },
+  "quality_only": true,
+  "timing_claims": false
+}

model-card-example.json

@@ -31,25 +31,25 @@
       "destination": {
         "type": "choice",
         "probabilities": {
-          "billing": 0.9999959977119728,
-          "technical": 4.00228802725166e-06
+          "billing": 0.9998550772527804,
+          "technical": 0.00014492274721960988
         },
-        "confidence": 0.9999919954239456,
+        "confidence": 0.9997101545055609,
         "choice": "billing"
       },
       "refund_requested": {
         "type": "noul",
-        "noul": 0.9998253639445287
+        "noul": 0.9984572750397842
       },
       "urgency": {
         "type": "score",
         "probabilities": {
-          "0": 0.02944745914852826,
-          "1": 0.9703999910699086,
-          "2": 0.00015254978156315938
+          "0": 0.13640706014048662,
+          "1": 0.8623479461162635,
+          "2": 0.0012449937432499015
         },
-        "confidence": 0.955599986604863,
-        "score": 0.9707050906330349,
+        "confidence": 0.7935219191743954,
+        "score": 0.8648379336027633,
         "legend": {
           "0": "Routine information request with no payment problem or outage",
           "1": "A payment or billing problem, with no product outage",
@@ -64,6 +64,50 @@
   },
   "direct_engine_exact_response": true,
   "overflow_rejected": true,
-  "bundle_manifest_sha256": "076558961011e924bf17084f940401150245bae6e135d4b1b0a6ee16be01f13c",
-  "example_source_sha256": "d52c20b602bd1049aecda55244e71c27d8ce44bd1aa8b1941db62ae41d31ba12"
+  "overflow_message": "check: 40083 tokens exceeds max_length=16384; no truncation allowed",
+  "bundle_manifest_sha256": "009d4892300c0a98f6a9b61706cdb6cc7ce2f6883c25c120477b98fe7fe82d0f",
+  "runtime": {
+    "actual": {
+      "torch": "2.12.0+git6bbd260",
+      "hip": "7.2.53211",
+      "transformers": "5.17.0",
+      "fla": "0.5.2",
+      "tokenizers": "0.23.2",
+      "safetensors": "0.8.0",
+      "triton": "3.7.1",
+      "gated_delta": "fla.ops.gated_delta_rule.chunk"
+    },
+    "differences": {},
+    "matches_validated_runtime": true,
+    "normalization_profile": {
+      "profile_sha256": "6b03450d42dbb68f0ffe14945ffcf3e6ea043e1033a819fb7211a8176a51722f",
+      "guard_sha256": "1603c39038ff783b9d5a5f69110d1695bbb7e28accffd0c03525854e8f258452",
+      "validated_arch": "gfx942",
+      "automatic_bundle_binding": true,
+      "scope": "single profile per process"
+    }
+  },
+  "revision": null,
+  "device": "cuda:0",
+  "model_name": "Decision-1.0-Sol",
+  "example_source_sha256": "a54dec885f92c2d38d07ef2333dff51965be52bc119f772cc74200ed314e69b6",
+  "normalization_telemetry": {
+    "profile_sha256": "6b03450d42dbb68f0ffe14945ffcf3e6ea043e1033a819fb7211a8176a51722f",
+    "status": "installed",
+    "calls": 72,
+    "keys": {
+      "[128,1,\"torch.bfloat16\",\"torch.bfloat16\",\"torch.float32\"]": 72
+    },
+    "strict_guard": true,
+    "unknown_keys": "raise",
+    "autotune_fallback_permitted": false,
+    "runtime": {
+      "torch": "2.12.0+git6bbd260",
+      "hip": "7.2.53211",
+      "triton": "3.7.1",
+      "fla": "0.5.2"
+    },
+    "gpu_arch": "gfx942",
+    "process_scope": "one explicitly profiled Sol model; other model loading in this process is not supported"
+  }
 }

peer-projection-provenance.json

@@ -0,0 +1,36 @@
+{
+  "format": "decision-public-peer-projection-v1",
+  "target_release": "Sol v1.1",
+  "peer_release": "Nox v1.1",
+  "peer_repository": "llm-semantic-router/Decision-1.0-Nox",
+  "peer_revision": "5bee3061a7abf675813963e4e05267cb1a28f92f",
+  "peer_bundle_manifest_sha256": "a8cfd42bc908fc93d5a2a20be527d09704b6e3f6c1217f27307c32cc2d822971",
+  "target_bundle_manifest_sha256": "009d4892300c0a98f6a9b61706cdb6cc7ce2f6883c25c120477b98fe7fe82d0f",
+  "target_report_sha256": {
+    "public": "d6712e27f1f0aed164404ad14dab4c920f024566d46dd42971fa385ce78b9c25",
+    "statistics": "af1eb2ea01eb096e7c41febb95dddeb76e76da0693fa40c3df25f39212dcca57",
+    "qualification": "0ac43b7803eedaf3c5c3cc450e9daddad5f9335b3ea82dc9401693a8cfc9e454",
+    "receipt": "bca099653d5b92cf282a5a82c19998c65dae0ac89d6fd039123b60b3680e85b3"
+  },
+  "peer_report_sha256": {
+    "public": "68adfdde02ec23a9947a1df8d44d3837bc66a23681a364bca8986f2644120b8b",
+    "statistics": "0f55bc193c0f7f4359dac177df1d03ef641ea4c09ca71e0bc7eec84452559f94",
+    "qualification": "469defd8f05af52eea4b7d2eec856484be694e82d6f7f4c930c71d8fc7028975",
+    "receipt": "927c92121c2df534f2ba12523e466ecebd8bd8936f949c7792de3e747c92a47c"
+  },
+  "peer_publication_receipt_sha256": "5e0a3b7ba2637c5689e225b850441d532816f41b9f9240cce73a3ec9b66c0ea5",
+  "common_policy_sha256": "359de2b6b81be91a66c6c538b3655bed06f4728357328cd318b1a765e64565cd",
+  "same_frozen_panels_families_baseline_predictions_and_statistical_sources_verified": true,
+  "peer_prediction_sha256": {
+    "old_core": "b706110fba0da5a7e00c122586094a3947077ba9142374207ecea56717acccb9",
+    "fresh_core": "a565bd5f362d31e51d631605bc3852700cca094c8047e0fec1672a3e4860a381",
+    "old_native": "c9304a3153f8dbe2598d2678b955c886b48276db44c4bdb6988ddaf81ec484e3",
+    "fresh_native": "0a969ae31845f3522651c32d7d269b81c9984bed352040a00e20bc7be097c330",
+    "old_core_raw_T1": "1c17319811808e45f7596f756cb07fd6f28b8e9e6187e9f55b9ebe11ae06ea31",
+    "fresh_core_raw_T1": "a370f801490f2d253dd96d40481be45ccf1338c5345e6ea2b008cf47e71db679"
+  },
+  "own_v1_0_qualification_unchanged": true,
+  "no_new_pairwise_Sol_Nox_test_claim": true,
+  "no_examples_or_predictions_read": true,
+  "exposure": "Observed regression for Sol; Nox first-exposure provenance preserved in its source receipt."
+}

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "decision-local"
-version = "1.0.0"
+version = "1.1.0"
 description = "Local typed inference for exported Decision decoder models"
 requires-python = ">=3.10"
 dependencies = []

release-manifest.json

@@ -1,18 +1,18 @@
 {
   "format": "decision-public-release-v1",
-  "status": "documentation-refresh",
-  "bundle_manifest_sha256": "076558961011e924bf17084f940401150245bae6e135d4b1b0a6ee16be01f13c",
-  "readiness_sha256": "46888ec7eef0c62fb477bbd2c1ca5368ab40c151221d3051debfcc9d6fcd8cc1",
-  "model_card_sha256": "9eed30c689e947d913204fcfd839fb296d9a2e5921e9c6b67bd34791d2e6a09c",
+  "status": "assembled-not-published",
+  "bundle_manifest_sha256": "009d4892300c0a98f6a9b61706cdb6cc7ce2f6883c25c120477b98fe7fe82d0f",
+  "readiness_sha256": "80c971663626c56275585c7e854e06b138e1459b28ef9d76293cb45d4fc11b68",
+  "model_card_sha256": "05cacb9febff3519c7ff3daad2ac2d872a06067019214c2d3cf8b1e466e26401",
   "repo_id": "llm-semantic-router/Decision-1.0-Sol",
-  "assembly_script_sha256": "0ca32612da99c16b9801149d69f375b157e51b6e426b0a74efa3efb6bcb9f792",
+  "assembly_script_sha256": "7831ef5952b8256e585d4efa91cac4ce87738f4eb93c84b1cf1e7250def1f430",
   "original_bundle_manifest_preserved": true,
   "files_exclude_this_manifest": true,
   "files": [
     {
       "file": "ATTRIBUTIONS.md",
-      "bytes": 2610,
-      "sha256": "47c35fdcde567fb511f31d1d7d0f2e2ba21daef0beb31217c025132aafe0d2e9"
+      "bytes": 4182,
+      "sha256": "76ff86ae499bdabb46ef75f974dfbc48eec794b8651bd9671780b55a1fc108e3"
     },
     {
       "file": "Dockerfile.runtime",
@@ -21,48 +21,58 @@
     },
     {
       "file": "EVALUATION.md",
-      "bytes": 16973,
-      "sha256": "1f4f7bc151e757da8e35db2bf690cd13839ab67de130655004400c2bd425ffb3"
-    },
-    {
-      "file": "FIGURE-NOTICES.md",
-      "bytes": 1041,
-      "sha256": "d70794c70cc8811a74ef19e9dbeb7368d3ff89b71e0117bff40149dc9b72055d"
+      "bytes": 11526,
+      "sha256": "3d5568d6a8e97556c33877145327036c2b43737064d692f0991aac13946ce8bf"
     },
     {
       "file": "LICENSE",
       "bytes": 11544,
       "sha256": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a"
     },
+    {
+      "file": "NORMALIZATION_RUNTIME.md",
+      "bytes": 1441,
+      "sha256": "9938bbe9ffdb890490c1db08903659edde6d7c5a5b9ce0487949e7a4dc2d38f6"
+    },
     {
       "file": "QUESTION-SCALING.md",
-      "bytes": 3608,
-      "sha256": "ad72203f319d9344ffa917b6d12ae34fd976c71c39fc715fa862dd23685bbbf1"
+      "bytes": 3763,
+      "sha256": "7dde01b31f9292116e715b501f78bf7577061f979ece9e95f7e3739ff2ead141"
     },
     {
       "file": "QWEN-LICENSE",
       "bytes": 11544,
       "sha256": "bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a"
     },
+    {
+      "file": "READING-SUPPLEMENT.md",
+      "bytes": 4359,
+      "sha256": "674a9efdbc28387f4526e1794890a0377db268b8fa89915601d3ce11a46d272a"
+    },
     {
       "file": "README.md",
-      "bytes": 4480,
-      "sha256": "9eed30c689e947d913204fcfd839fb296d9a2e5921e9c6b67bd34791d2e6a09c"
+      "bytes": 5434,
+      "sha256": "05cacb9febff3519c7ff3daad2ac2d872a06067019214c2d3cf8b1e466e26401"
     },
     {
       "file": "RUNTIME.md",
-      "bytes": 5205,
-      "sha256": "be35b2f8a9c977f38be9b5eb892fdc809db5167c327575c9311355aae3e781c7"
+      "bytes": 6018,
+      "sha256": "4eb7f04296218bb859f223f27d000bdd4415bfbc0326a4e98358a66e67c73273"
+    },
+    {
+      "file": "RUNTIME_BINDING.json",
+      "bytes": 1764,
+      "sha256": "c050b5a43d9ddf98e07924878a4e72cfb1c4746131a3045c31557c262791f922"
     },
     {
-      "file": "TIMING.md",
-      "bytes": 13432,
-      "sha256": "bdf371d3e93056125840829105b7f5a91abce343889530c8f7cb339f6612f9a6"
+      "file": "SOURCE_BUNDLE_MANIFEST.json",
+      "bytes": 4206,
+      "sha256": "9242b7fd4ce226f638108810ec5156dca6f270cce5a16efcf5ef091180ad75be"
     },
     {
       "file": "USAGE.md",
-      "bytes": 5435,
-      "sha256": "46ec1078d2bb8cb54a1850809a101f07a77bb54c18b92276cb89d40245dbbc8f"
+      "bytes": 5827,
+      "sha256": "819bf0195a511d149a15e415261307b75e0b9052e1aa828ada9b9b5d3ca9fe8c"
     },
     {
       "file": "assets/architecture-atlas.pdf",
@@ -86,18 +96,18 @@
     },
     {
       "file": "assets/decision-capabilities.pdf",
-      "bytes": 45170,
-      "sha256": "1314aae927ea0aa598e62a80138e2ea9bfd8e851bc6faa91504d5aa691426e24"
+      "bytes": 51635,
+      "sha256": "41f1b6e93b41c7f2057ad9c0743ddf0fc643e0c48ba0de58d8a3b58dbc897f1e"
     },
     {
       "file": "assets/decision-capabilities.png",
-      "bytes": 276453,
-      "sha256": "ed88c09a355c8583ac4d7b31bd4ff93b7be8725fde436f3d76e1754ae92a5962"
+      "bytes": 445740,
+      "sha256": "81a551e4b353bf2934d722f6f6038972bb8b4cd68ffd22ed546506a8303daebe"
     },
     {
       "file": "assets/decision-capabilities.svg",
-      "bytes": 69169,
-      "sha256": "cae4bf2c1aa13e8d6dc94c435eed1fd8bde1e523fbfa83a0b7dde5ba37650022"
+      "bytes": 114828,
+      "sha256": "5d7c39e232094f74fbbb10915ec9dc9c5a030159a2e7497dee4180b46251d643"
     },
     {
       "file": "assets/decision-family-header.png",
@@ -111,18 +121,18 @@
     },
     {
       "file": "assets/decision-quality.pdf",
-      "bytes": 37484,
-      "sha256": "aec3769c803f3b87668332e2a6d036cab1bd65b21c8ebbd919fbb79f1637290f"
+      "bytes": 37119,
+      "sha256": "c85f236e2d74ae5eb87018277bdac22ec8b4132c5fb10d1ea20231c593c33b41"
     },
     {
       "file": "assets/decision-quality.png",
-      "bytes": 175813,
-      "sha256": "dbf0e32d3e7846f94b8d6fd224eaf54677cd1ecb140f1671fb1b80f2a6d25935"
+      "bytes": 170682,
+      "sha256": "2176d6ffb40424f9121989ec5a63e51789f39e79d086ba0a27735aa37938df01"
     },
     {
       "file": "assets/decision-quality.svg",
-      "bytes": 37293,
-      "sha256": "e324b1446ac18beeefd6da1d3d4c2ec4db579b7d989390663c1ab5e43bf76841"
+      "bytes": 36923,
+      "sha256": "367cb8096c8441fce35259973198a4eb6fa79424531884b066cfc502634c3cec"
     },
     {
       "file": "assets/decision-question-scaling.pdf",
@@ -162,12 +172,12 @@
     {
       "file": "backbone/model.safetensors",
       "bytes": 3763685328,
-      "sha256": "cab4b9b9dfd490781acf489b78b6d3faa1d56a1be8fab5d8baf6c464e4c0ea1d"
+      "sha256": "168f7f90bf410130da92fd5270305286d53d5cbb68be02f1c611c40cb9243fe2"
     },
     {
       "file": "bundle-manifest.json",
-      "bytes": 4204,
-      "sha256": "076558961011e924bf17084f940401150245bae6e135d4b1b0a6ee16be01f13c"
+      "bytes": 7251,
+      "sha256": "009d4892300c0a98f6a9b61706cdb6cc7ce2f6883c25c120477b98fe7fe82d0f"
     },
     {
       "file": "chat_template.jinja",
@@ -176,8 +186,8 @@
     },
     {
       "file": "code/decision_api.py",
-      "bytes": 6724,
-      "sha256": "147b2fec32cbbbbb1b92cf2a19bb887f9d945e4974e141ca7ea93b8c542f5b21"
+      "bytes": 8251,
+      "sha256": "1b068eccdffd3c3b67bfa52f8f526e6b482d92551c927668ad28a794767f8a40"
     },
     {
       "file": "code/decision_model.py",
@@ -189,6 +199,16 @@
       "bytes": 3164,
       "sha256": "02352e8385ab47157b6459910da54d962e5da4bb4940d571faedc86bc5da9aee"
     },
+    {
+      "file": "code/profile_guard.py",
+      "bytes": 6131,
+      "sha256": "1603c39038ff783b9d5a5f69110d1695bbb7e28accffd0c03525854e8f258452"
+    },
+    {
+      "file": "code/runtime_profile.py",
+      "bytes": 2857,
+      "sha256": "afb59dda5e4c3890ee5c971549799e69f2e4fa2eafe5893de2168299fd368bda"
+    },
     {
       "file": "decision_config.json",
       "bytes": 753,
@@ -197,17 +217,12 @@
     {
       "file": "decision_head.safetensors",
       "bytes": 8424272,
-      "sha256": "308bc01085a2bcee89a6668cb06ca4b771e01338040689527127bdefd3cc7415"
-    },
-    {
-      "file": "metrics/asset-hashes.json",
-      "bytes": 1394,
-      "sha256": "b4aec27dd4a4032c37ef013cda0cb72161b6c1de25161d61864a5f1971af952a"
+      "sha256": "924a43288d48804c22e7e2573898d694303879ab9bb819d1b20eadc2ab32f00d"
     },
     {
-      "file": "metrics/quality-aggregate.json",
-      "bytes": 1214948,
-      "sha256": "b2181b7636f0ddace6af33a45010bfb4e62fb32f987f4abe2d2a4d58b3a007e4"
+      "file": "metrics/natural-confirmation.json",
+      "bytes": 27933,
+      "sha256": "64863075f92fd7e9df489fa37b8ddaedf29e2ee53da3da2a6009d1ee5625268e"
     },
     {
       "file": "metrics/question-scaling.json",
@@ -215,30 +230,45 @@
       "sha256": "4708635c6f26cd0209dafb4e45af6921ed22091e1996bbeb5861c98214ff5886"
     },
     {
-      "file": "metrics/timing-aggregate.json",
-      "bytes": 148812,
-      "sha256": "a0e5f1c94277b31a922ae1e1a708ce70a2f03b5e49996a688112fc1e9697615a"
+      "file": "model-card-example.json",
+      "bytes": 3774,
+      "sha256": "2e4b902caef5ffbc935f20f2812aa1799d83e5c4b4561cf79746cbbf847faa96"
     },
     {
-      "file": "model-card-example.json",
-      "bytes": 2271,
-      "sha256": "77c4450389fe872d7e61ee4d02f0c89b3a0126913379255abc7f7f03ab5b8723"
+      "file": "peer-projection-provenance.json",
+      "bytes": 2281,
+      "sha256": "0dfae5adc5d3c028d6fd0f59ac8b3660b55f2e8b10787b5d840b1f5437d4f301"
     },
     {
       "file": "pyproject.toml",
       "bytes": 436,
-      "sha256": "6a557fbe472027af103ca2cfc07e977caab3e257de930aa57cf9fea8cf90f8a0"
+      "sha256": "135a9516e87ca2fb41ef54a974e29132256b6c2d528a1cbdea1001e28f306946"
+    },
+    {
+      "file": "quality-metrics.json",
+      "bytes": 1379273,
+      "sha256": "f2e1e71b391ccbac4ceeacbbd6a2cf128ace27594203a548d71c11aa8952f9f8"
     },
     {
       "file": "runtime-build-provenance.json",
-      "bytes": 2894,
-      "sha256": "61d064b0db35d0ebd04a038e376896d2aab8b69e6525a8ec1f317f73b5284345"
+      "bytes": 3033,
+      "sha256": "8a22aed74b1210c44c918c1dc3f8ac9778b55ebcc70a9a805eb2f2de76a24870"
     },
     {
       "file": "runtime-fla-requirements.lock",
       "bytes": 654,
       "sha256": "35e1fedff9ca49092a7ada277bbd794abe7e8474e14a678a7e4bcec6406d0eb5"
     },
+    {
+      "file": "runtime-profile/l2norm_fwd_kernel.json",
+      "bytes": 13210,
+      "sha256": "a67f3b4624edc07e2c3f9f1d05553c654a3ff005c6b96a1831d62e323dea9edc"
+    },
+    {
+      "file": "runtime-profile/profile.json",
+      "bytes": 18101,
+      "sha256": "6b03450d42dbb68f0ffe14945ffcf3e6ea043e1033a819fb7211a8176a51722f"
+    },
     {
       "file": "runtime-provenance.json",
       "bytes": 5805,
@@ -246,8 +276,8 @@
     },
     {
       "file": "runtime.json",
-      "bytes": 952,
-      "sha256": "7ed527476826756d6076f23d0eaa8b4b53ff34b9f8bd57ae8e65b50e4d28cfa6"
+      "bytes": 1133,
+      "sha256": "9fe0a2bd626a8b974c85e4925cdc79dca320c9116c6d0196c12a87b9aeae8d2a"
     },
     {
       "file": "src/decision/__init__.py",
@@ -256,18 +286,48 @@
     },
     {
       "file": "src/decision/example.py",
-      "bytes": 3381,
-      "sha256": "d52c20b602bd1049aecda55244e71c27d8ce44bd1aa8b1941db62ae41d31ba12"
+      "bytes": 3581,
+      "sha256": "a54dec885f92c2d38d07ef2333dff51965be52bc119f772cc74200ed314e69b6"
     },
     {
       "file": "src/decision/model.py",
-      "bytes": 8790,
-      "sha256": "ab428234d508a1245d3723c5a5bb0c796685abebd0447f77c7fbed9113954b21"
+      "bytes": 9415,
+      "sha256": "ba240d7493fc29203fe036966f0ab911200cd4a9252b50b05423409977639ee0"
+    },
+    {
+      "file": "src/decision_local.egg-info/PKG-INFO",
+      "bytes": 225,
+      "sha256": "cc5266721b2e02c5c963b59856d979d619f7d9d687b60bda037c7bad2f22cd1d"
+    },
+    {
+      "file": "src/decision_local.egg-info/SOURCES.txt",
+      "bytes": 361,
+      "sha256": "741f317b6e9b27453e8f2fbeb76c0d1da326e89de7978fe87ce906ed1e2a68a9"
+    },
+    {
+      "file": "src/decision_local.egg-info/dependency_links.txt",
+      "bytes": 1,
+      "sha256": "01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b"
+    },
+    {
+      "file": "src/decision_local.egg-info/entry_points.txt",
+      "bytes": 59,
+      "sha256": "5400ff8d993d37849bc01b702e4674b8d9c4122af514101ace0bcd691eab39fb"
+    },
+    {
+      "file": "src/decision_local.egg-info/requires.txt",
+      "bytes": 31,
+      "sha256": "b8bf334329a333c3bba85388fbf41377b17305ee71b402c431c1af7445eadc08"
+    },
+    {
+      "file": "src/decision_local.egg-info/top_level.txt",
+      "bytes": 9,
+      "sha256": "834608781b00d5df58e3ae55a6b15c201aeebacc26ec99a1035fd75b83e16760"
     },
     {
       "file": "temperature.json",
-      "bytes": 667,
-      "sha256": "159d308a6b970c425b8938b6c50c4fff6a10bb40ca82b119e43386a855d609ab"
+      "bytes": 13733,
+      "sha256": "63af0fbd1c817c4bc354aab81df66626bcd9908ac7a71bcbb921e63ba7361d44"
     },
     {
       "file": "tokenizer.json",
@@ -284,25 +344,31 @@
     "ATTRIBUTIONS.md": "copy",
     "Dockerfile.runtime": "copy",
     "EVALUATION.md": "copy",
-    "FIGURE-NOTICES.md": "copy",
     "LICENSE": "copy",
+    "NORMALIZATION_RUNTIME.md": "copy",
+    "QUESTION-SCALING.md": "copy",
     "QWEN-LICENSE": "copy",
-    "README.md": "documentation-copy",
+    "READING-SUPPLEMENT.md": "copy",
+    "README.md": "copy",
     "RUNTIME.md": "copy",
-    "TIMING.md": "copy",
+    "RUNTIME_BINDING.json": "copy",
+    "SOURCE_BUNDLE_MANIFEST.json": "copy",
     "USAGE.md": "copy",
     "assets/architecture-atlas.pdf": "copy",
     "assets/architecture.pdf": "copy",
     "assets/architecture.png": "copy",
     "assets/architecture.svg": "copy",
-    "assets/decision-capabilities.pdf": "documentation-copy",
-    "assets/decision-capabilities.png": "documentation-copy",
-    "assets/decision-capabilities.svg": "documentation-copy",
+    "assets/decision-capabilities.pdf": "copy",
+    "assets/decision-capabilities.png": "copy",
+    "assets/decision-capabilities.svg": "copy",
     "assets/decision-family-header.png": "copy",
     "assets/decision-mark.png": "copy",
-    "assets/decision-quality.pdf": "documentation-copy",
-    "assets/decision-quality.png": "documentation-copy",
-    "assets/decision-quality.svg": "documentation-copy",
+    "assets/decision-quality.pdf": "copy",
+    "assets/decision-quality.png": "copy",
+    "assets/decision-quality.svg": "copy",
+    "assets/decision-question-scaling.pdf": "copy",
+    "assets/decision-question-scaling.png": "copy",
+    "assets/decision-question-scaling.svg": "copy",
     "assets/readout.pdf": "copy",
     "assets/readout.png": "copy",
     "assets/readout.svg": "copy",
@@ -313,31 +379,34 @@
     "code/decision_api.py": "copy",
     "code/decision_model.py": "copy",
     "code/predict.py": "copy",
+    "code/profile_guard.py": "copy",
+    "code/runtime_profile.py": "copy",
     "decision_config.json": "copy",
     "decision_head.safetensors": "hardlink",
-    "metrics/asset-hashes.json": "copy",
-    "metrics/quality-aggregate.json": "copy",
-    "metrics/timing-aggregate.json": "copy",
+    "metrics/natural-confirmation.json": "copy",
+    "metrics/question-scaling.json": "copy",
     "model-card-example.json": "copy",
+    "peer-projection-provenance.json": "copy",
     "pyproject.toml": "copy",
+    "quality-metrics.json": "copy",
     "runtime-build-provenance.json": "copy",
     "runtime-fla-requirements.lock": "copy",
+    "runtime-profile/l2norm_fwd_kernel.json": "copy",
+    "runtime-profile/profile.json": "copy",
     "runtime-provenance.json": "copy",
     "runtime.json": "copy",
     "src/decision/__init__.py": "copy",
     "src/decision/example.py": "copy",
     "src/decision/model.py": "copy",
+    "src/decision_local.egg-info/PKG-INFO": "copy",
+    "src/decision_local.egg-info/SOURCES.txt": "copy",
+    "src/decision_local.egg-info/dependency_links.txt": "copy",
+    "src/decision_local.egg-info/entry_points.txt": "copy",
+    "src/decision_local.egg-info/requires.txt": "copy",
+    "src/decision_local.egg-info/top_level.txt": "copy",
     "temperature.json": "copy",
     "tokenizer.json": "copy",
-    "tokenizer_config.json": "copy",
-    "QUESTION-SCALING.md": "documentation-copy",
-    "metrics/question-scaling.json": "documentation-copy",
-    "assets/decision-question-scaling.pdf": "documentation-copy",
-    "assets/decision-question-scaling.png": "documentation-copy",
-    "assets/decision-question-scaling.svg": "documentation-copy"
+    "tokenizer_config.json": "copy"
   },
-  "scope": "Inference weights, original numerical code, calibrated runtime metadata, wrapper, model card, license, attribution and approved assets only.",
-  "previous_release_manifest_sha256": "996b5b0244bf7d75b7330f93560000701b1cc12bdd33fa50a09565d2eaee9d15",
-  "documentation_refresh_script_sha256": "faf40a7508c73266e2337acac18d05d5fe1a119cee13768d41947b4369a44afc",
-  "inference_files_unchanged": true
+  "scope": "Inference weights, original numerical code, calibrated runtime metadata, wrapper, model card, license, attribution and approved assets only."
 }

runtime-build-provenance.json

@@ -25,7 +25,7 @@
   "build_exit_code": 0,
   "gpu_validation": {
     "status": "passed",
-    "scope": "Real packaged three-question example and explicit complete-input overflow rejection for each bundle. Exact response equality against qualified runtime on this request; not a replacement for full benchmark rerun.",
+    "scope": "Real packaged three-question example and explicit complete-input overflow rejection for each bundle. Exact response equality against qualified runtime on this request; not a replacement for full benchmark rerun. These entries identify the initial v1.0 weights; they are not new measurements of a later weight update.",
     "models": {
       "Decision-1.0-Sol": {
         "bundle_manifest_sha256": "076558961011e924bf17084f940401150245bae6e135d4b1b0a6ee16be01f13c",
@@ -51,7 +51,8 @@
         "overflow_rejected": true,
         "board_product_name_verified": false
       }
-    }
+    },
+    "validation_release": "v1.0"
   },
   "built_utc": "2026-09-21T12:57:54.147566+00:00",
   "files": {

runtime-profile/l2norm_fwd_kernel.json

@@ -0,0 +1,646 @@
+{
+  "kernel_name": "l2norm_fwd_kernel",
+  "triton_version": "3.7.1",
+  "autotune_entries": {
+    "878a9638751a7375a0fb15d3a7767659": {
+      "autotune_key": [
+        128,
+        1,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "eca743e0790be43b6867d648e9430f67": {
+      "autotune_key": [
+        128,
+        2,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 16
+        },
+        "num_warps": 1,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "7e2d02ce45e4f072828b4c73e16fe0af": {
+      "autotune_key": [
+        128,
+        3,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 16
+        },
+        "num_warps": 2,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "424d41ef00a3161badd7f960c34e6ece": {
+      "autotune_key": [
+        128,
+        4,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "6bc2b57d9035425140554c225e667aad": {
+      "autotune_key": [
+        128,
+        5,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "91b9289c43a2d1b21042009d44c57b8a": {
+      "autotune_key": [
+        128,
+        6,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "f44702dfc86a96df1de10439f5c5575a": {
+      "autotune_key": [
+        128,
+        7,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "87c993ac4858bff6042f319bec1647f7": {
+      "autotune_key": [
+        128,
+        8,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "b35c6bfa2cd85845ba58306ee63e28b1": {
+      "autotune_key": [
+        128,
+        9,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 4,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "3527d130e8f6f1415c6d24df9821e638": {
+      "autotune_key": [
+        128,
+        10,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 2,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "fe2e43a7c21c6a4814d2636f3004d7dc": {
+      "autotune_key": [
+        128,
+        11,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 4,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "c39c159634377c459eccde8a272f74fe": {
+      "autotune_key": [
+        128,
+        12,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "a10453829fb545fd2e01acfb4a809d17": {
+      "autotune_key": [
+        128,
+        13,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 4,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "bd8734e29b9704051fe467ede8e8cc03": {
+      "autotune_key": [
+        128,
+        14,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "d5ffd2f70a70e13ac03901bd22a150a7": {
+      "autotune_key": [
+        128,
+        15,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "d522f744092358110c9ad2cc5f54b501": {
+      "autotune_key": [
+        128,
+        16,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "6003bee20af205e16410f1983eadd67c": {
+      "autotune_key": [
+        128,
+        17,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "89d2cb2e8a818ce3ecb80bd340ba42b6": {
+      "autotune_key": [
+        128,
+        18,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "157376310735556f015b8ac2dc6e9e2e": {
+      "autotune_key": [
+        128,
+        19,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "e19965e082b861f637bf4bab6abe025b": {
+      "autotune_key": [
+        128,
+        20,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "49feda728414835e54048468abd808ab": {
+      "autotune_key": [
+        128,
+        21,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "6b419788330d3468ef22dc741ec9d7f7": {
+      "autotune_key": [
+        128,
+        22,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "8827081d0dda52c9376b3a1d97a82a8e": {
+      "autotune_key": [
+        128,
+        23,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "49e16fa1a715a8913c54c4ff9ce61aa9": {
+      "autotune_key": [
+        128,
+        24,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "7bf76d12f90b95bbddff57fbc8fe3c39": {
+      "autotune_key": [
+        128,
+        25,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "3f2cd7a3a429540e08f0e4dd8ac9afc5": {
+      "autotune_key": [
+        128,
+        26,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "837afa436bfab6256a535e9360403f34": {
+      "autotune_key": [
+        128,
+        27,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "d4e411f29afbb396dcfe1455427e5178": {
+      "autotune_key": [
+        128,
+        28,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "1b96f1802f4f63f5c872de3d1972ad8a": {
+      "autotune_key": [
+        128,
+        29,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "aec76b66508d5e884ffcc743e8f55a92": {
+      "autotune_key": [
+        128,
+        30,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "439d493be8fe1159bd65368245a4aa60": {
+      "autotune_key": [
+        128,
+        31,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    "801422dba59f6c5e144328aabdd135a3": {
+      "autotune_key": [
+        128,
+        32,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    }
+  }
+}

runtime-profile/profile.json

@@ -0,0 +1,822 @@
+{
+  "format": "decision-fla-l2norm-profile-v1",
+  "status": "diagnostic-only-awaiting-clean-room-proof",
+  "target": "Decision Sol Qwen3.5-2B pointer Stage4; one profile per process",
+  "model_changes": false,
+  "numerical_tolerance_unchanged": 1e-06,
+  "cache_mode": "strict",
+  "supported": {
+    "head_dimension": 128,
+    "key_heads": 16,
+    "body_dtype": "bfloat16",
+    "output_dtype": "bfloat16",
+    "rstd_dtype": "float32",
+    "batch_size_min": 1,
+    "batch_size_max": 8,
+    "padded_tokens_max": 16384,
+    "NB_min": 1,
+    "NB_max": 32,
+    "NB_formula": "ceil(B*padded_tokens*16/65536)",
+    "unknown_key": "raise before kernel/autotune"
+  },
+  "runtime": {
+    "torch": "2.12.0+git6bbd260",
+    "hip": "7.2.53211",
+    "triton": "3.7.1",
+    "fla": "0.5.2",
+    "gpu_arch": "gfx942"
+  },
+  "files": [
+    {
+      "file": "l2norm_fwd_kernel.json",
+      "sha256": "a67f3b4624edc07e2c3f9f1d05553c654a3ff005c6b96a1831d62e323dea9edc"
+    }
+  ],
+  "fla_source_sha256": {
+    "modules/l2norm.py": "30f7feebdbfa87b8e90c143ec62a7fa0781bfd736c856c0ec69452f0c778be3f",
+    "ops/utils/cache.py": "a64b09ffd3f51aed547ad72559dee94cb9aeafd2a5f02c6c8e60a6e020147a9c"
+  },
+  "source_evidence": [
+    {
+      "key": [
+        128,
+        1,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "030be4b902759f8a02c66edfe9724c8a04d320c32eb8298ef42da42f8ff02a54",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        7,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "b868fb0bcb06cfdc5b4432392e35ade0daeeb40374645dbfd65476996362dc48",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        8,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "ee3b7b3a9facff1f50feacbc32427f005e6d344b1e452e9daabc68b19a9fed0d",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        9,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "fd0b189927eb2475fd73e9a71899a6655ae4d73a010c618b00316f9eb15e429e",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 4,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        5,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "a8ad1053d222fb9d5751a6ebe11fc58679a9e44366be3dc9a2c214042485778c",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        2,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "00499403ea55da8fc3e1110dfb41a94204f19813ac8302dc2cd0894b2e1f51a9",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 16
+        },
+        "num_warps": 1,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        12,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "bad146284201145ccafd77500eba52f4b20483596065a3da91e6b1bccb0548e3",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        10,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "cb8bd713aee627cb50fcca3ed750e7ac9976ee56f9d6d3d44ef469325049822c",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 2,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        13,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "f2c9fe94fbf5ff3f48b4ec875cbfb8016148c772fe085b69d6e42dd8053f9549",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 4,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        11,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "0d2cc4c185b6c73918b1397271e98276a6ee8aa510d345506b8fd30343f4e38d",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 4,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        3,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "a426f2ff58b25d5f0e7b23876254d3f4fc255ac596a902eeb14e47917688f508",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 16
+        },
+        "num_warps": 2,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        4,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "615e132647ea6735a52e431485cd2fd4706dd754e7d4aec356d0a006e18d9166",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        18,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "ef64a3156f4dd428d7b9f80f9435c25eabbcfc192882c199c6fe5fc1c2316be8",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "key": [
+        128,
+        6,
+        "torch.bfloat16",
+        "torch.bfloat16",
+        "torch.float32"
+      ],
+      "source_autotune_sha256": "adca9d81bb4df7c1a64dc7e525544b05c362d943c95446f4871bf92f09fb603b",
+      "chosen_config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    }
+  ],
+  "per_NB_policy": [
+    {
+      "NB": 1,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 2,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 16
+        },
+        "num_warps": 1,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 3,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 16
+        },
+        "num_warps": 2,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 4,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 5,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 6,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 7,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 8,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 16,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 9,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 4,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 10,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 2,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 11,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 4,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 12,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 13,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 8
+        },
+        "num_warps": 4,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 14,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 15,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 16,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 17,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 18,
+      "basis": "original-production-chosen",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 19,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 20,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 21,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 22,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 23,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 24,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 25,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 26,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 27,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 28,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 29,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 30,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 31,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    },
+    {
+      "NB": 32,
+      "basis": "prospective-fixed-original-NB18-setting",
+      "config": {
+        "kwargs": {
+          "BT": 32
+        },
+        "num_warps": 8,
+        "num_ctas": 1,
+        "num_stages": 3,
+        "maxnreg": null,
+        "pre_hook": null,
+        "ir_override": null
+      }
+    }
+  ],
+  "unseen_NB_claim": "Deterministic prospective settings, not proof of matching an undefined prior autotune choice. No speed claim.",
+  "other_kernels": "Unmodified FLA/Transformers dispatch; this profile changes only l2norm_fwd_kernel launch settings.",
+  "compiled_binaries_included": false,
+  "private_paths_included": false,
+  "builder_sha256": "833d5dcfad43a94c239ea2f5ad68d69b0c501e82d9a3b3ef530b5b85b72db379"
+}

runtime.json

@@ -2,22 +2,24 @@
   "torch": "2.12.0+git6bbd260",
   "torch_git": "6bbd26020da1c6dc198625dfcdd968b1e4e6b1c5",
   "hip": "7.2.53211",
+  "transformers": "5.17.0",
   "triton": "3.7.1",
   "fla": "0.5.2",
-  "image_digest": "sha256:670f9f4cced18cccfb196188a3d42ff81ab617f17c447bfe033e150824d3448b",
-  "wheels": {
-    "fla_core-0.5.2-py3-none-any.whl": "5e830c85bad3d0d34677f98ac7074d08687a3756f0f0499d95ceb96eb6920761",
-    "flash_linear_attention-0.5.2-py3-none-any.whl": "dcf405d81f5426393b59037097aa700d0f4a841465d5028d5aa543f4502f2400"
-  },
-  "gated_delta": "fla.ops.gated_delta_rule.chunk",
-  "causal_conv": "transformers-reference-PyTorch",
-  "full_attention": "sdpa",
-  "runtime_installation": "isolated PYTHONPATH, no shared image mutation",
-  "warm_start": "successful pointer head warmup checkpoint100; no failed smoke weights reused",
-  "transformers": "5.17.0",
+  "gdn_implementation": "fla.ops.gated_delta_rule.chunk",
+  "gdn_new_implementation": true,
   "tokenizers": "0.23.2",
   "safetensors": "0.8.0",
-  "huggingface-hub": "1.31.0",
-  "accelerate": "1.15.0",
-  "numpy": "2.3.5"
+  "gated_delta": "fla.ops.gated_delta_rule.chunk",
+  "normalization_profile": {
+    "kind": "decision-fla-l2norm-profile-v1",
+    "profile_file": "runtime-profile/profile.json",
+    "profile_sha256": "6b03450d42dbb68f0ffe14945ffcf3e6ea043e1033a819fb7211a8176a51722f",
+    "guard_file": "code/profile_guard.py",
+    "guard_sha256": "1603c39038ff783b9d5a5f69110d1695bbb7e28accffd0c03525854e8f258452",
+    "loader_file": "code/runtime_profile.py",
+    "loader_sha256": "afb59dda5e4c3890ee5c971549799e69f2e4fa2eafe5893de2168299fd368bda",
+    "validated_arch": "gfx942",
+    "activation": "automatic before FLA import by public from_pretrained and bundle DecisionEngine",
+    "process_scope": "one profile per fresh Python process; same-profile loads allowed; unprofiled/different-profile mixing rejected"
+  }
 }

src/decision/example.py

@@ -52,6 +52,9 @@ def main():
               'bundle_manifest_sha256': hashlib.sha256((model.bundle_path / 'bundle-manifest.json').read_bytes()).hexdigest(),
               'runtime': model.runtime, 'revision': args.revision, 'device': args.device,
               'model_name': response['model'], 'example_source_sha256': hashlib.sha256(Path(__file__).read_bytes()).hexdigest()}
+    if model.runtime.get('normalization_profile') is not None:
+        import sys
+        record['normalization_telemetry'] = dict(sys.modules['_decision_process_normalization_profile_v1'].telemetry)
     args.output.parent.mkdir(parents=True, exist_ok=True)
     args.output.write_text(json.dumps(record, ensure_ascii=False, indent=2) + '\n')
     print(json.dumps(response, ensure_ascii=False, indent=2))

src/decision/model.py

@@ -131,10 +131,21 @@ class DecisionModel:
                 isinstance(v, bool) or not isinstance(v, (int, float)) or not math.isfinite(v) or v <= 0
                 for v in temperatures.values()):
             raise ValueError('Bundle must contain finite positive temperatures for all three types')
-        runtime = _runtime_report(json.loads((path / 'runtime.json').read_text()), device, allow_unvalidated_runtime)
+        api = _load_api(path)
+        expected_runtime = json.loads((path / 'runtime.json').read_text())
+        if expected_runtime.get('normalization_profile') is not None:
+            if not hasattr(api, 'prepare_runtime_profile'):
+                raise RuntimeError('Profiled bundle omits its automatic runtime entrypoint')
+            profile = api.prepare_runtime_profile(path, device=device)
+        else:
+            import sys
+            if '_decision_process_normalization_profile_v1' in sys.modules:
+                raise RuntimeError('Use separate processes for profiled and unprofiled models')
+            profile = None
+        runtime = _runtime_report(expected_runtime, device, allow_unvalidated_runtime)
+        runtime['normalization_profile'] = profile
         names = {'Qwen/Qwen3.5-2B': 'Decision-1.0-Sol', 'Qwen/Qwen3.5-4B': 'Decision-1.0-Nox'}
         name = model_name or config.get('model_name') or names.get(config.get('base_model'), path.name)
-        api = _load_api(path)
         engine = api.DecisionEngine(path, path / 'code', device=device, max_length=maximum,
                                     batch_size=manifest['production_batch_size'], temperatures=temperatures,
                                     model_name=name)

src/decision_local.egg-info/PKG-INFO

@@ -0,0 +1,7 @@
+Metadata-Version: 2.4
+Name: decision-local
+Version: 1.0.0
+Summary: Local typed inference for exported Decision decoder models
+Requires-Python: >=3.10
+Provides-Extra: hub
+Requires-Dist: huggingface-hub==1.31.0; extra == "hub"

src/decision_local.egg-info/SOURCES.txt

@@ -0,0 +1,11 @@
+pyproject.toml
+src/decision/__init__.py
+src/decision/example.py
+src/decision/model.py
+src/decision_local.egg-info/PKG-INFO
+src/decision_local.egg-info/SOURCES.txt
+src/decision_local.egg-info/dependency_links.txt
+src/decision_local.egg-info/entry_points.txt
+src/decision_local.egg-info/requires.txt
+src/decision_local.egg-info/top_level.txt
+tests/test_wrapper.py

src/decision_local.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

src/decision_local.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+decision-example = decision.example:main

src/decision_local.egg-info/requires.txt

@@ -0,0 +1,3 @@
+
+[hub]
+huggingface-hub==1.31.0

src/decision_local.egg-info/top_level.txt

@@ -0,0 +1 @@
+decision

temperature.json

@@ -1,22 +1,420 @@
 {
-  "temperature": 0.7033302993804421,
+  "version": "stage4-nine-family-macro-nll-beta-bisection-v1",
+  "objective": "equal-family macro NLL",
+  "input_sha256": {
+    "cal": "5ba3bd7527da831d77cec8c6a2f260a71e10a57742e356510c719f8f30d61a70",
+    "predictions": "b9e2cf2c466e6f9422e627905625400d683c6ce0cecb015c16e5784b0c3c577c",
+    "inference-summary": "8e537d9729e9f56636eef327063e356e9f32bc96b46edd252263bb880cb450b1",
+    "inference-binding": "8d7d7a5cbb58104b5bf2037088b770d1173e719a1701d56001efc35c7d8b7224",
+    "selection": "b1d31cb6eea09b75471f256845530217cb0fd19d668c524c6e4e092267c8e9cd",
+    "model-manifest": "55fd2542828d86625123f5b0e59b2f8e967ec9db38ec6d3f67b64c1f079408b6",
+    "plan": "e5086ac34dbf190bd1f1ce6f0a426a37710160878b67120aab9bdbfed49e8348"
+  },
+  "temperature": 0.9982273489600023,
   "temperatures": {
-    "choice": 0.7033302993804421,
-    "noul": 0.7033302993804421,
-    "score": 0.7033302993804421
+    "choice": 0.9982273489600023,
+    "noul": 0.9982273489600023,
+    "score": 0.9982273489600023
   },
-  "method": "one-global-positive-temperature; family-macro-development-Brier",
-  "bounds": [
-    0.25,
-    4.0
-  ],
-  "dev_examples": 256,
-  "families": 8,
-  "raw_brier": 0.045776584869665994,
-  "calibrated_dev_brier": 0.04377952758544348,
-  "boundary_solution": false,
   "ranking_unchanged": true,
+  "ranking_claim_scope": "Positive global scaling preserves mathematical logit ordering; finite-precision inference parity is audited separately",
+  "source_sha256": "80e008d3a0eab5604de4b15bc897ddfa26c6761430753478ad5972a6cf08e5df",
+  "family_rows": {
+    "stage4_dense_table": 100,
+    "stage4_scope": 100,
+    "stage4_arithmetic": 100,
+    "stage4_ordinal": 100,
+    "stage4_natural_nli": 200,
+    "stage4_automaton": 100,
+    "stage4_boolean": 100,
+    "stage4_relations": 100,
+    "stage4_registers": 100
+  },
+  "selected_checkpoint": "checkpoint-001563",
+  "selected_arm": "ce",
+  "fit": {
+    "temperature": 0.9982273489600023,
+    "inverse_temperature": 1.0017757989117855,
+    "status": "relative-bracket-converged",
+    "boundary_hit": false,
+    "iterations": 38,
+    "beta_final_bracket": [
+      1.0017757988754965,
+      1.0017757989480742
+    ],
+    "objective_at_T1": 0.6400866469651806,
+    "objective_at_selected_T": 0.6400860719946871,
+    "derivative_at_selected_beta": 5.17162410925694e-12,
+    "initial_boundary_derivatives": {
+      "beta_0.05": -2.9645769693802926,
+      "beta_20": 0.25425204205854235
+    },
+    "derivative_at_beta1": -0.0006479866804508699,
+    "initial_boundary_objectives": [
+      1.299894618307853,
+      5.139564840374214
+    ]
+  },
+  "raw": {
+    "temperature": 1.0,
+    "family_macro_nll": 0.6400866469651806,
+    "family_macro_brier": 0.3141964724320727,
+    "family_macro_native_accuracy": 0.7488888888888889,
+    "by_family": {
+      "stage4_dense_table": {
+        "nll": 0.44557900925725874,
+        "brier": 0.21564659042540416,
+        "native_accuracy": 0.88,
+        "rows": 100
+      },
+      "stage4_scope": {
+        "nll": 0.0498655140340889,
+        "brier": 0.028826434428688895,
+        "native_accuracy": 0.98,
+        "rows": 100
+      },
+      "stage4_arithmetic": {
+        "nll": 1.0564804127990468,
+        "brier": 0.3403965985820037,
+        "native_accuracy": 0.74,
+        "rows": 100
+      },
+      "stage4_ordinal": {
+        "nll": 0.4764526628150301,
+        "brier": 0.25096833388492373,
+        "native_accuracy": 0.79,
+        "rows": 100
+      },
+      "stage4_natural_nli": {
+        "nll": 0.5360083749004084,
+        "brier": 0.3208918487912171,
+        "native_accuracy": 0.76,
+        "rows": 200
+      },
+      "stage4_automaton": {
+        "nll": 1.3367082473383853,
+        "brier": 0.6598012810538971,
+        "native_accuracy": 0.41,
+        "rows": 100
+      },
+      "stage4_boolean": {
+        "nll": 0.45878482504605034,
+        "brier": 0.3056423821560782,
+        "native_accuracy": 0.75,
+        "rows": 100
+      },
+      "stage4_relations": {
+        "nll": 0.8561881989781089,
+        "brier": 0.39314463269485195,
+        "native_accuracy": 0.69,
+        "rows": 100
+      },
+      "stage4_registers": {
+        "nll": 0.5447125775182473,
+        "brier": 0.31245014987158926,
+        "native_accuracy": 0.74,
+        "rows": 100
+      }
+    },
+    "by_task_type": {
+      "score": {
+        "family_macro_nll": 0.48249895787513475,
+        "family_macro_brier": 0.24589504270711923,
+        "family_macro_native_accuracy": 0.8325,
+        "by_family": {
+          "stage4_dense_table": {
+            "nll": 0.4885452529352393,
+            "brier": 0.24082175152931473,
+            "native_accuracy": 0.875,
+            "rows": 16
+          },
+          "stage4_ordinal": {
+            "nll": 0.4764526628150301,
+            "brier": 0.25096833388492373,
+            "native_accuracy": 0.79,
+            "rows": 100
+          }
+        }
+      },
+      "choice": {
+        "family_macro_nll": 0.7578712143392436,
+        "family_macro_brier": 0.3425364771873384,
+        "family_macro_native_accuracy": 0.7228736900165471,
+        "by_family": {
+          "stage4_scope": {
+            "nll": 0.06248730610766334,
+            "brier": 0.03778805355766209,
+            "native_accuracy": 0.972972972972973,
+            "rows": 74
+          },
+          "stage4_arithmetic": {
+            "nll": 1.0564804127990468,
+            "brier": 0.3403965985820037,
+            "native_accuracy": 0.74,
+            "rows": 100
+          },
+          "stage4_natural_nli": {
+            "nll": 0.5360083749004084,
+            "brier": 0.3208918487912171,
+            "native_accuracy": 0.76,
+            "rows": 200
+          },
+          "stage4_automaton": {
+            "nll": 1.6447244044728033,
+            "brier": 0.7579729767796513,
+            "native_accuracy": 0.3142857142857143,
+            "rows": 70
+          },
+          "stage4_relations": {
+            "nll": 1.3089671086886498,
+            "brier": 0.5900762617472654,
+            "native_accuracy": 0.5166666666666667,
+            "rows": 60
+          },
+          "stage4_registers": {
+            "nll": 0.4960359448890203,
+            "brier": 0.26223171850149674,
+            "native_accuracy": 0.8133333333333334,
+            "rows": 75
+          },
+          "stage4_dense_table": {
+            "nll": 0.20039494851711318,
+            "brier": 0.08839788235207212,
+            "native_accuracy": 0.9428571428571428,
+            "rows": 35
+          }
+        }
+      },
+      "noul": {
+        "family_macro_nll": 0.427528942482448,
+        "family_macro_brier": 0.2664778929921975,
+        "family_macro_native_accuracy": 0.7816780045351474,
+        "by_family": {
+          "stage4_scope": {
+            "nll": 0.013941951978530863,
+            "brier": 0.003320287676995936,
+            "native_accuracy": 1.0,
+            "rows": 26
+          },
+          "stage4_boolean": {
+            "nll": 0.45878482504605034,
+            "brier": 0.3056423821560782,
+            "native_accuracy": 0.75,
+            "rows": 100
+          },
+          "stage4_relations": {
+            "nll": 0.17701983441229752,
+            "brier": 0.09774718911623191,
+            "native_accuracy": 0.95,
+            "rows": 40
+          },
+          "stage4_dense_table": {
+            "nll": 0.6066806873604711,
+            "brier": 0.29831806399487465,
+            "native_accuracy": 0.8367346938775511,
+            "rows": 49
+          },
+          "stage4_registers": {
+            "nll": 0.690742475405928,
+            "brier": 0.463105443981867,
+            "native_accuracy": 0.52,
+            "rows": 25
+          },
+          "stage4_automaton": {
+            "nll": 0.6180038806914102,
+            "brier": 0.4307339910271372,
+            "native_accuracy": 0.6333333333333333,
+            "rows": 30
+          }
+        }
+      }
+    },
+    "proper_score_convention": "multiclass sum-squared Brier, including two entries for noul; NLL in nats; equal family weight within each reported task type"
+  },
+  "calibrated": {
+    "temperature": 0.9982273489600023,
+    "family_macro_nll": 0.6400860719946871,
+    "family_macro_brier": 0.31418442074316877,
+    "family_macro_native_accuracy": 0.7488888888888889,
+    "by_family": {
+      "stage4_dense_table": {
+        "nll": 0.445938296164393,
+        "brier": 0.2156572009593377,
+        "native_accuracy": 0.88,
+        "rows": 100
+      },
+      "stage4_scope": {
+        "nll": 0.04986573972928388,
+        "brier": 0.02882890500497602,
+        "native_accuracy": 0.98,
+        "rows": 100
+      },
+      "stage4_arithmetic": {
+        "nll": 1.056168898096333,
+        "brier": 0.3402942706104855,
+        "native_accuracy": 0.74,
+        "rows": 100
+      },
+      "stage4_ordinal": {
+        "nll": 0.4763746819749599,
+        "brier": 0.2509243877991786,
+        "native_accuracy": 0.79,
+        "rows": 100
+      },
+      "stage4_natural_nli": {
+        "nll": 0.5358254005746533,
+        "brier": 0.3208298384181523,
+        "native_accuracy": 0.76,
+        "rows": 200
+      },
+      "stage4_automaton": {
+        "nll": 1.3368995049684889,
+        "brier": 0.6598855736978168,
+        "native_accuracy": 0.41,
+        "rows": 100
+      },
+      "stage4_boolean": {
+        "nll": 0.45874553976333154,
+        "brier": 0.30562878190540066,
+        "native_accuracy": 0.75,
+        "rows": 100
+      },
+      "stage4_relations": {
+        "nll": 0.8561430453929648,
+        "brier": 0.3931506584302739,
+        "native_accuracy": 0.69,
+        "rows": 100
+      },
+      "stage4_registers": {
+        "nll": 0.5448135412877755,
+        "brier": 0.31246016986289765,
+        "native_accuracy": 0.74,
+        "rows": 100
+      }
+    },
+    "by_task_type": {
+      "score": {
+        "family_macro_nll": 0.4824392478193705,
+        "family_macro_brier": 0.24583611006762263,
+        "family_macro_native_accuracy": 0.8325,
+        "by_family": {
+          "stage4_dense_table": {
+            "nll": 0.48850381366378104,
+            "brier": 0.24074783233606664,
+            "native_accuracy": 0.875,
+            "rows": 16
+          },
+          "stage4_ordinal": {
+            "nll": 0.4763746819749599,
+            "brier": 0.2509243877991786,
+            "native_accuracy": 0.79,
+            "rows": 100
+          }
+        }
+      },
+      "choice": {
+        "family_macro_nll": 0.757831004980382,
+        "family_macro_brier": 0.34252562919037327,
+        "family_macro_native_accuracy": 0.7228736900165471,
+        "by_family": {
+          "stage4_scope": {
+            "nll": 0.06250739149911984,
+            "brier": 0.03779678126229391,
+            "native_accuracy": 0.972972972972973,
+            "rows": 74
+          },
+          "stage4_arithmetic": {
+            "nll": 1.056168898096333,
+            "brier": 0.3402942706104855,
+            "native_accuracy": 0.74,
+            "rows": 100
+          },
+          "stage4_natural_nli": {
+            "nll": 0.5358254005746533,
+            "brier": 0.3208298384181523,
+            "native_accuracy": 0.76,
+            "rows": 200
+          },
+          "stage4_automaton": {
+            "nll": 1.6449019417622601,
+            "brier": 0.758039538459674,
+            "native_accuracy": 0.3142857142857143,
+            "rows": 70
+          },
+          "stage4_relations": {
+            "nll": 1.3088570377916526,
+            "brier": 0.5900881012282487,
+            "native_accuracy": 0.5166666666666667,
+            "rows": 60
+          },
+          "stage4_registers": {
+            "nll": 0.4961240001814749,
+            "brier": 0.2622496602881456,
+            "native_accuracy": 0.8133333333333334,
+            "rows": 75
+          },
+          "stage4_dense_table": {
+            "nll": 0.20043236495718011,
+            "brier": 0.0883812140656126,
+            "native_accuracy": 0.9428571428571428,
+            "rows": 35
+          }
+        }
+      },
+      "noul": {
+        "family_macro_nll": 0.42770221644099665,
+        "family_macro_brier": 0.26650089018224704,
+        "family_macro_native_accuracy": 0.7816780045351474,
+        "by_family": {
+          "stage4_scope": {
+            "nll": 0.013885653922827704,
+            "brier": 0.0033049495033789502,
+            "native_accuracy": 1.0,
+            "rows": 26
+          },
+          "stage4_boolean": {
+            "nll": 0.45874553976333154,
+            "brier": 0.30562878190540066,
+            "native_accuracy": 0.75,
+            "rows": 100
+          },
+          "stage4_relations": {
+            "nll": 0.17707205679493318,
+            "brier": 0.09774449423331166,
+            "native_accuracy": 0.95,
+            "rows": 40
+          },
+          "stage4_dense_table": {
+            "nll": 0.6074007311085204,
+            "brier": 0.298375760944087,
+            "native_accuracy": 0.8367346938775511,
+            "rows": 49
+          },
+          "stage4_registers": {
+            "nll": 0.6908821646066775,
+            "brier": 0.46309169858715404,
+            "native_accuracy": 0.52,
+            "rows": 25
+          },
+          "stage4_automaton": {
+            "nll": 0.6182271524496894,
+            "brier": 0.4308596559201498,
+            "native_accuracy": 0.6333333333333333,
+            "rows": 30
+          }
+        }
+      }
+    },
+    "proper_score_convention": "multiclass sum-squared Brier, including two entries for noul; NLL in nats; equal family weight within each reported task type"
+  },
   "heldout_used": false,
-  "dev_sha256": "a38f9be168553d5a82a91991aca86aaa602e1951957bac7c475abcec8dcc232d",
-  "raw_predictions_sha256": "38df5ace521ee8ec7decc752ab7a49e5e7b2cda751c7d7d6a4fdf00ad8af4202"
+  "SELECT_used_for_temperature": false,
+  "generalization_improvement_claim": false,
+  "raw_predictions_sha256": "b9e2cf2c466e6f9422e627905625400d683c6ce0cecb015c16e5784b0c3c577c",
+  "dev_sha256": "5ba3bd7527da831d77cec8c6a2f260a71e10a57742e356510c719f8f30d61a70",
+  "cal_sha256": "5ba3bd7527da831d77cec8c6a2f260a71e10a57742e356510c719f8f30d61a70",
+  "calibration_dataset_role": "CAL; dev_sha256 is a legacy portable-bundle schema alias, not SELECT",
+  "source_calibration_artifact_sha256": "f6be19545ba1f81456b201b5cf83360e9b80e722080b0ce6c997daa4f228d11c",
+  "export_metadata_bridge_sha256": "e4cd2ceaa9bd9697f5c47f5a4504526612e6f23fd87a095fe9c92fd0fae9ae31",
+  "temperature_refit_performed_by_bridge": false
 }