Publish approved Winnow card, usage and evidence

.gitattributes

@@ -37,3 +37,4 @@ gguf/Winnow-E2B-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
 gguf/mmproj-Winnow-E2B.gguf filter=lfs diff=lfs merge=lfs -text
 gguf/Winnow-E2B-BF16.gguf filter=lfs diff=lfs merge=lfs -text
 gguf/Gemma-4-E2B-IT-Assistant-BF16.gguf filter=lfs diff=lfs merge=lfs -text
+assets/e2b-e4b-adaptive-portrait.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -13,47 +13,51 @@ tags:
 
 # Winnow-E2B
 
-**Download the Q8 target for the reported results; a matching BF16 target is also available for higher-memory setups.** Add the projector for vision or the E2B assistant for optional MTP. Verify each downloaded file against [SHA256SUMS](SHA256SUMS). Generic Hub model snippets do not provide Winnow's `/v1/systemone` decision API.
+Winnow-E2B is an EldanRing decision fine-tune of [Gemma 4 E2B IT](https://huggingface.co/google/gemma-4-E2B-it). It scores supplied answer options through Winnow's `/v1/systemone` API and supports ordinary text chat. Image input uses the matching vision projector.
 
-Winnow-E2B is an EldanRing decision fine-tune of [Google's Gemma 4 E2B IT](https://huggingface.co/google/gemma-4-E2B-it). It scores the candidate answers supplied for each question through Winnow's native `/v1/systemone` API. It also supports ordinary text chat; image input requires the matching vision projector. Candidate probabilities are conditional on the supplied options, not a general guarantee of correctness.
+The Q8 and BF16 GGUFs contain the same merged fine-tune. No separate adapter or base-model download is needed.
 
-The Q8 and BF16 GGUFs contain the same fine-tune merged into the language-model weights; no separate adapter or base-model download is needed. The source base revision is [`3e22461f65e89153144f8adb70e3b8c2cc9845a7`](https://huggingface.co/google/gemma-4-E2B-it/tree/3e22461f65e89153144f8adb70e3b8c2cc9845a7). The projector is the corresponding F16 vision export. Audio support has not been established for this package.
+[Inference code](https://github.com/EldanRing/winnow-inference) · [Evaluation details](docs/EVALUATION.md) · [Downloads](#downloads)
 
-## Exact model files
+## Decision results
 
-| Role | File | Bytes | SHA-256 |
-|---|---|---:|---|
-| Q8 target | [gguf/Winnow-E2B-Q8_0.gguf](https://huggingface.co/EldanRing/Winnow-E2B/resolve/main/gguf/Winnow-E2B-Q8_0.gguf?download=true) | 4,954,594,336 | `cb37414b46b5c4ac54ea9a44f71bb6147991c34e6be414a8cdd359cc10f0dfb1` |
-| BF16 target | [gguf/Winnow-E2B-BF16.gguf](https://huggingface.co/EldanRing/Winnow-E2B/resolve/main/gguf/Winnow-E2B-BF16.gguf?download=true) | 9,311,304,736 | `bcebcf44720a5409bed1960fc7caf511b9506b33b232ef46072f3f1713f61f7f` |
-| F16 vision projector | [gguf/mmproj-Winnow-E2B.gguf](https://huggingface.co/EldanRing/Winnow-E2B/resolve/main/gguf/mmproj-Winnow-E2B.gguf?download=true) | 985,653,600 | `7faa5282cff7250380c191bdea4e13a7b397ad851e2d883873cd9c323a5e192f` |
-| Optional BF16 MTP assistant | [gguf/Gemma-4-E2B-IT-Assistant-BF16.gguf](https://huggingface.co/EldanRing/Winnow-E2B/resolve/main/gguf/Gemma-4-E2B-IT-Assistant-BF16.gguf?download=true) | 170,194,016 | `0772dcc50761a47a2da87151d3105ad2b82994d9dd1dae3ad3c5c0cd26086b16` |
+Direct decisions score the supplied candidates. Adaptive reasoning generates context for selected questions, then blends direct and augmented probabilities. The results below use the Q8 target on an RTX 5070 Ti: F16 target K/V cache, an 8K text context, MTP4, a raw maximum-probability gate below 0.99, and a 50:50 blend.
 
-The Q8 target, projector, and assistant hashes identify the **original files used in the E2B tests**. The BF16 target is a separate export of the same merged checkpoint; the results below do **not** evaluate or calibrate it. The shorter target and projector filenames name byte-identical copies. See the small [checksum file](SHA256SUMS) and [release manifest](release-manifest.json). Check each downloaded file; do not substitute an E4B or 12B model, projector, or assistant.
+| Panel | Direct | Adaptive reasoning |
+|---|---:|---:|
+| JevBench public, verified labels | 175/231 (75.76%) | 202/231 (87.45%) |
+| Kev v9 clean, verified labels | 728/1,046 (69.60%) | 851/1,046 (81.36%) |
+| Typed, synthetic teacher agreement | 1,237/2,000 (61.85%) | 1,366/2,000 (68.30%) |
 
-The optional MTP draft is a CPU BF16 GGUF conversion of Google's [official E2B IT assistant](https://huggingface.co/google/gemma-4-E2B-it-assistant/tree/2d874ef7d29f9a30599a1e4b3c1cbc9595f005df), revision `2d874ef7d29f9a30599a1e4b3c1cbc9595f005df`. Its pinned source `model.safetensors` has SHA-256 `93682eb1c97639d18f007704dc880bd74cbe530adaf7b1bb561213863fdad2a6`. For MTP, download the bundled assistant alongside the Q8 target:
+All 3,277 cases remain in the totals, including two Kev context-limit attempts that fell back to direct scoring. These frozen panels were used during development. Jev and Kev use verified labels; Typed measures agreement with synthetic teacher labels.
 
-```sh
-huggingface-cli download EldanRing/Winnow-E2B gguf/Winnow-E2B-Q8_0.gguf gguf/Gemma-4-E2B-IT-Assistant-BF16.gguf --local-dir Winnow-E2B
-```
+![E2B and E4B direct and adaptive reasoning scores on the same frozen cases. Lighter extensions show gains; E4B Typed declines by 0.35 percentage points.](assets/e2b-e4b-adaptive-portrait.png)
 
-The evaluated GGUF was produced with the pinned [`llama.cpp` converter revision `911f6cdc8ab8a530b2bee09ee61471a6f3178eeb`](https://github.com/ggml-org/llama.cpp/tree/911f6cdc8ab8a530b2bee09ee61471a6f3178eeb), using BF16 output. Verify its hash before using MTP; other assistants are not interchangeable. Direct decisions and ordinary chat do not require the assistant. [Assistant attribution and conversion provenance](docs/assistants/README.md) are included.
+The comparison uses Q8 targets, F16 target K/V cache, and MTP4. Reasoning prompts, policies, samplers, and batch sizes differ by model. E4B's seven-agreement Typed decline is shown to scale with a downward marker. See [evaluation methods](docs/EVALUATION.md#comparison-with-e4b) for exact settings.
 
-## Decision results
+Mean serial adaptive reasoning time was **1.356 s** on Jev, **1.212 s** on Kev, and **1.951 s** on Typed. Typed timing excludes 16 cases affected by GPU overlap; all quality results remain included. See [timing definitions](docs/EVALUATION.md#latency) for the measurement boundaries.
 
-The table evaluates the downloadable Q8 target, with its matching projector resident, on an RTX 5070 Ti. Direct is native candidate scoring. The experimental selective path uses raw direct maximum probability `<0.99` to trigger generated reasoning, then blends direct and augmented probabilities 50:50. It used F16 target KV, Q8 draft KV, the official BF16 assistant with MTP draft length 4, backend temperature sampling, one chat slot, and an 8,192-token context.
+## Downloads
 
-| Panel | Direct | Selective reasoning |
-|---|---:|---:|
-| JevBench public, verified labels | 175/231 (75.8%) | 203/231 (87.9%) |
-| Kev v9 clean, verified labels | 728/1,046 (69.6%) | 851/1,046 (81.4%) |
-| Typed decisions, synthetic hard-teacher agreement | 1,237/2,000 (61.9%) | 1,365/2,000 (68.3%) |
+| File | Role | Size |
+|---|---|---:|
+| [Winnow-E2B-Q8_0.gguf](https://huggingface.co/EldanRing/Winnow-E2B/resolve/main/gguf/Winnow-E2B-Q8_0.gguf?download=true) | Q8 target used for the reported results | 4.95 GB |
+| [Winnow-E2B-BF16.gguf](https://huggingface.co/EldanRing/Winnow-E2B/resolve/main/gguf/Winnow-E2B-BF16.gguf?download=true) | Higher-precision target | 9.31 GB |
+| [mmproj-Winnow-E2B.gguf](https://huggingface.co/EldanRing/Winnow-E2B/resolve/main/gguf/mmproj-Winnow-E2B.gguf?download=true) | F16 vision projector | 986 MB |
+| [Gemma-4-E2B-IT-Assistant-BF16.gguf](https://huggingface.co/EldanRing/Winnow-E2B/resolve/main/gguf/Gemma-4-E2B-IT-Assistant-BF16.gguf?download=true) | Optional MTP draft assistant | 170 MB |
+
+Text-only use needs one target; images also need the matching E2B projector. Download the exact filenames and verify them against [SHA256SUMS](SHA256SUMS).
+
+## Running the model
+
+The [quickstart](https://github.com/EldanRing/winnow-inference/blob/main/docs/QUICKSTART.md) covers installation and launch commands. Use `/v1/systemone` for decisions and `/v1/chat/completions` for ordinary chat and images.
 
-All three full panels were previously exposed during development, so these are regression and transfer results, not independent generalization estimates. In a separate 288-case disjoint holdout, the selective path gained only 3 agreements over the earlier transferred rule; its paired 95% interval included zero. Typed agreement is not independently verified correctness.
+**F16 is the default and recommended target K/V cache.** Cache precision is separate from the GGUF weight format; use `--cache q8_0` for an explicit Q8 override. The pinned MTP assistant uses the target's shared K/V cache.
 
-For the three full panels in table order, mean serial API time was **0.051 → 1.368 s**, **0.023 → 1.233 s**, and **0.029 → 1.981 s** from direct to selective reasoning. These selective costs are paired-output replay estimates, not deployed-service timing or concurrent throughput. Generation length and fallback costs are included. The current integration's request-to-prompt mapping differs from the benchmark serialization, so these figures do not certify exact production-path equivalence.
+MTP uses the bundled BF16 conversion of Google's official E2B IT assistant. Follow the [assistant download instructions](docs/ASSISTANT.md); no conversion is needed. The assistant is a separate draft model, distinct from the BF16 target. Direct decisions and ordinary chat need only the target. Exact provenance is in the [release manifest](release-manifest.json).
 
-Image evidence is limited to a narrow 180-decision synthetic procedural panel. It does not establish photo, screenshot, or document quality. A 64K-context launch and small text/image requests were checked operationally; full-length 64K quality was not measured.
+Candidate probabilities are normalized over the supplied options. Confidence describes concentration among those options, not a guarantee of correctness.
 
 ## License and attribution
 
-Gemma 4 E2B IT and the separate official E2B IT assistant are by Google DeepMind under [Apache License 2.0](https://ai.google.dev/gemma/apache_2). Winnow modifications are by EldanRing. The [license](LICENSE) and [notice](NOTICE) apply to this model package; the bundled converted assistant also has [upstream attribution](docs/assistants/README.md). The Winnow inference code is separately licensed; no inference binary or source is part of this model repository. No Google endorsement is implied.
+Gemma 4 E2B IT and the official E2B IT assistant are by Google DeepMind under [Apache License 2.0](https://ai.google.dev/gemma/apache_2). Winnow modifications are by EldanRing. See [LICENSE](LICENSE) and [NOTICE](NOTICE). The bundled assistant includes [upstream attribution](docs/assistants/README.md). Inference code is separately licensed and is not included in this model package. No Google endorsement is implied.

docs/ASSISTANT.md

@@ -0,0 +1,29 @@
+# Optional E2B MTP assistant
+
+Direct decisions and ordinary chat need only the Winnow-E2B target. MTP additionally uses the bundled BF16 GGUF conversion of Google's official [Gemma 4 E2B IT assistant](https://huggingface.co/google/gemma-4-E2B-it-assistant/tree/2d874ef7d29f9a30599a1e4b3c1cbc9595f005df). No conversion is required to use this download.
+
+## Download the evaluated assistant
+
+Download the Q8 target and its matching assistant from the model repository:
+
+```sh
+huggingface-cli download EldanRing/Winnow-E2B \
+  gguf/Winnow-E2B-Q8_0.gguf \
+  gguf/Gemma-4-E2B-IT-Assistant-BF16.gguf \
+  --local-dir Winnow-E2B
+```
+
+Verify both files against [SHA256SUMS](../SHA256SUMS). An E4B or 12B assistant is not interchangeable with the E2B assistant. Add `gguf/mmproj-Winnow-E2B.gguf` for image input.
+
+The assistant's BF16 precision is distinct from the optional `Winnow-E2B-BF16.gguf` target. Reported MTP4 results use Q8 target weights with BF16 assistant weights and shared F16 target K/V cache.
+
+## Provenance
+
+The bundled assistant is the evaluated CPU conversion from upstream revision `2d874ef7d29f9a30599a1e4b3c1cbc9595f005df`, made with `convert_hf_to_gguf.py` at pinned [llama.cpp revision](https://github.com/ggml-org/llama.cpp/tree/911f6cdc8ab8a530b2bee09ee61471a6f3178eeb) and BF16 output. Conversion changes serialization and performs no training. A different conversion does not inherit these measurements.
+
+| Artifact | Bytes | SHA-256 |
+|---|---:|---|
+| Upstream `model.safetensors` | 157,565,344 | `93682eb1c97639d18f007704dc880bd74cbe530adaf7b1bb561213863fdad2a6` |
+| Evaluated BF16 GGUF | 170,194,016 | `0772dcc50761a47a2da87151d3105ad2b82994d9dd1dae3ad3c5c0cd26086b16` |
+
+The assistant is by Google DeepMind under Apache License 2.0. The [assistant attribution directory](assistants/README.md) includes the license, notice, upstream model card, and conversion provenance. Target, projector, base revision, and assistant identities are recorded in the [release manifest](../release-manifest.json).

docs/EVALUATION.md

@@ -0,0 +1,69 @@
+# Winnow-E2B evaluation
+
+## Scope
+
+The completed October 6, 2026 evaluation uses the Q8 target and official BF16 MTP assistant identified in the [release manifest](../release-manifest.json), on an RTX 5070 Ti. It evaluates the adaptive reasoning path at source revision `ee6bd37d34ae35d2e69ebb0c4b0957a10c727357`.
+
+JevBench and Kev use verified source labels. Typed measures synthetic hard-teacher agreement: 2,000 decisions from 400 five-question groups. All three panels were previously exposed during development. They are regression and transfer evidence, not independent estimates of generalization. JevBench means public-subset accuracy, not its official composite leaderboard score. No tuning on these labels occurred in this campaign.
+
+## E2B method
+
+The frozen adaptive rule routes when raw direct maximum probability is below 0.99, scores at temperature 1, and combines direct and augmented probabilities 50:50. Only completed natural-EOS generations are rescored. A failed reasoning attempt retains direct probabilities and stays in the denominator.
+
+The text-only run used the `e2b-q8-text8k-mtp` preset: an 8,192-position context, F16 target KV shared with the pinned MTP assistant, MTP draft length 4, uncapped generation at temperature 0, and seed 314159. The retained draft-cache launch flags were Q8_0, but the assistant attention cache shares the target's F16 K/V tensors; this is not a separately measured Q8 attention-cache pool. The reasoning prompt serializes options with native labels while preserving input escaping and structured-state ownership. This is an actual integrated-path run, not the earlier selective replay.
+
+All 3,277 cases completed with zero request errors and exact direct-vector parity against frozen references (maximum difference 0; tolerance 1e-10). Routing selected 131 Jev, 676 Kev and 1,777 Typed cases. Every selected generation completed except two Kev attempts that reached the 8K context limit; those cases retained direct scoring. Partial text was not rescored or blended.
+
+## Comparison with E4B
+
+The chart uses the same frozen cases, source labels and answer order. Shared settings are Q8 targets, F16 target KV shared with the pinned MTP assistant, 8K context, one chat slot and MTP4. Model-specific prompts, samplers, augmentation paths and adaptive policies differ. Winnow-12B's published campaigns remain separate and are not included in this matched comparison.
+
+| Panel | E2B direct → adaptive reasoning | E2B change | E4B direct → adaptive reasoning | E4B change |
+|---|---:|---:|---:|---:|
+| JevBench, 231 | 175 → 202 | +11.69 pp | 183 → 203 | +8.66 pp |
+| Kev v9 clean, 1,046 | 728 → 851 | +11.76 pp | 762 → 842 | +7.65 pp |
+| Typed teacher agreement, 2,000 | 1,237 → 1,366 | +6.45 pp | 1,446 → 1,439 | −0.35 pp |
+
+Changes are adaptive minus the same model's direct result, in percentage points. Chart percentages are computed from exact counts, with two decimal places. The card and graph both round percentages to two decimals.
+
+E4B used the live `e4b-calibrated75-g95-v1` path: raw maxP below 0.95, direct/augmented temperatures 1.2041180007310734 / 3.4209273427377678, and 25:75 direct/augmented weighting. Its brief reasoning prompt, released sampler, 75-second deadline, and batch 2048 differ from E2B's native-label prompt, backend temperature sampler, 180-second deadline and batch 1024. Both use ubatch 1024. F16 KV was outside E4B's measured release policy profile; calibration was held fixed, not revalidated here. The comparison does not isolate model size or the effect of reasoning alone.
+
+E4B corrected/regressed 22/2 Jev answers, 114/34 Kev answers, and 129/136 Typed agreements. Its Typed NLL, Brier, and rating MAE also worsened slightly. The graph uses a downward cap for the seven-agreement loss. These fresh E4B counts differ from its older public Q8-KV comparisons; the historical release tables remain separate.
+
+## Latency
+
+| Panel | Clean cases | Mean | Median | p95 | p99 |
+|---|---:|---:|---:|---:|---:|
+| JevBench | 231 | 1.356 s | 0.887 s | 4.099 s | 6.057 s |
+| Kev v9 clean | 1,046 | 1.212 s | 1.029 s | 3.522 s | 6.895 s |
+| Typed | 1,984 | 1.951 s | 2.095 s | 3.005 s | 3.744 s |
+
+These are serial adaptive reasoning pipeline times, including failed reasoning attempts. Sixteen Typed cases marked by the GPU-overlap evidence are excluded from timing only; all 2,000 remain in quality scores. Scored-case pipeline cost across all cases was 5,482.2 seconds. Startup, diagnostic replays, controller overhead and idle recovery gaps are outside those times. The full launch-to-cleanup span was 169.3 minutes and must not be presented as summed benchmark latency.
+
+The earlier E4B live comparison used different output lengths, prompts and timing boundaries. It does not establish intrinsic relative speed, equal-output MTP speedup or concurrent service throughput. Native backend allocation figures are not continuously measured peak GPU VRAM.
+
+## Earlier paths
+
+| Panel | Earlier numbered options | Historical F16/backend replay | Current adaptive reasoning path |
+|---|---:|---:|---:|
+| JevBench, 231 | 195 | 203 | 202 |
+| Kev v9 clean, 1,046 | 845 | 851 | 851 |
+| Typed teacher agreement, 2,000 | 1,342 | 1,365 | 1,366 |
+
+The current path restores most historical reasoning behavior but is not identical: it loses one Jev answer, matches Kev's winners, and gains one Typed agreement relative to the historical path. Native labels, safe escaping and state augmentation can change outcomes. These older paths are historical comparisons, not current release scores or independent calibration studies.
+
+## Disjoint holdout
+
+An earlier separately frozen 288-case holdout contained 96 LogiQA2 choices, 96 PAWS Boolean decisions, and 96 HelpSteer2 consensus ratings. The 0.99/50% rule scored 158/288 versus 155/288 for the transferred 0.80/50% rule: nine fixes and six regressions. Its paired 95% change interval was −1.74 to +3.82 percentage points; NLL and Brier change intervals also included zero. Mean serial replay time rose from 1.066 to 1.677 seconds.
+
+This holdout was disjoint from selection, but was not wholly independent external validation: 74 HelpSteer2 cases had previously been scored by other models, and public sources may have appeared in base-model pretraining. The small inconclusive gain does not establish a general benefit. It is not a holdout result for the current path.
+
+## Vision
+
+This campaign is text only and adds no new vision or always-mode evidence. Vision evidence is limited to an earlier 180-decision procedural-image panel, with four context-limit fallbacks. The full text comparison above does not establish photo, screenshot, document, or image-conditioned reasoning quality.
+
+## Evidence identity
+
+The [machine-readable comparison](comparison-summary.json) contains exact chart counts and receipt hashes. The retained [analysis](evidence/e2b-v3-analysis.json), [freeze](evidence/e2b-v3-freeze.json), and [completion receipt](evidence/e2b-v3-completion.json) identify the evaluated source, artifacts, fallbacks and timing exclusions. Complete case logs remain in the private review evidence bundle, separate from the model package.
+
+The final result log SHA-256 is `db3d7cb81972ad2909a75aca659b97015d9e96a770f04620275b79b9fbdbd80d`. The chart builder verifies the analysis, freeze and completion identities alongside the retained E4B receipt. Producing these figures performed no additional inference or fitting.

docs/comparison-summary.json

@@ -0,0 +1,113 @@
+{
+  "title": "E2B / E4B direct and adaptive comparison",
+  "date": "2026-10-06",
+  "metric": "Verified-label accuracy for Jev/Kev; synthetic teacher-label agreement for Typed",
+  "scope": "Previously exposed panels; same frozen cases; E2B integrated v3 and retained E4B comparison; model-specific policies and generation paths",
+  "common_settings": {
+    "target_weights": "Q8",
+    "target_kv": "F16",
+    "context": 8192,
+    "mtp_draft_length": 4,
+    "chat_slots": 1,
+    "draft_cache_flags": "Q8_0 (retained launch flags)",
+    "assistant_attention_kv": "F16 (shared target K/V tensors)"
+  },
+  "e2b": {
+    "gate": "raw maxP < 0.99",
+    "direct_augmented_weights": [
+      0.5,
+      0.5
+    ],
+    "sampler": "backend temperature sampling",
+    "batch": 1024,
+    "ubatch": 1024,
+    "deadline_seconds": 180,
+    "adaptive_execution": "actual integrated v3 run",
+    "source_commit": "ee6bd37d34ae35d2e69ebb0c4b0957a10c727357",
+    "policy": "e2b-raw99-blend50-v3",
+    "prompt": "native labels with safe input escaping and structured-state ownership"
+  },
+  "e4b": {
+    "policy": "e4b-calibrated75-g95-v1",
+    "gate": "raw maxP < 0.95",
+    "direct_augmented_weights": [
+      0.25,
+      0.75
+    ],
+    "sampler": "released sampler; no E2B backend-temperature flag",
+    "batch": 2048,
+    "ubatch": 1024,
+    "deadline_seconds": 75,
+    "adaptive_execution": "live selective client",
+    "calibration_scope": "F16 KV outside measured release profile; temperatures held fixed"
+  },
+  "receipt_sha256": {
+    "e4b-e2b-comparison-analysis.json": "b19713c86fc58ab837137009bb8796a6fb82c717465f9c84cfc63cafe5f8d19f",
+    "e4b-e2b-comparison-freeze.json": "6ffc48de7155b73d69659ee7757b122354813b0984b02712a84b89854c40eacd",
+    "e2b-v3-analysis.json": "1cd872865228fb33adaf12dd48d3656b98fd6f1f499d8adf364e04de4f1ddb27",
+    "e2b-v3-freeze.json": "eb163f63b0eb3a706adaaf18d20ebe3acaef4ddb0690005149eef2cbc750d2a6",
+    "e2b-v3-completion.json": "6e8eb1e7f38bb725321b5bc24683ec8ccfb24b34a92d5de15e94f776d30e26b0"
+  },
+  "panels": [
+    {
+      "key": "jevbench-public",
+      "title": "JevBench",
+      "n": 231,
+      "source_sha256": "bd304bc512d70bcbd9ff0b7d039d241b86aa0072ae5533c3c8d493c170aeb4d0",
+      "E2B": {
+        "direct": 175,
+        "adaptive": 202,
+        "direct_pct": "75.76",
+        "adaptive_pct": "87.45",
+        "change_pp": "+11.69"
+      },
+      "E4B": {
+        "direct": 183,
+        "adaptive": 203,
+        "direct_pct": "79.22",
+        "adaptive_pct": "87.88",
+        "change_pp": "+8.66"
+      }
+    },
+    {
+      "key": "kev-v9-clean",
+      "title": "Kev v9 clean",
+      "n": 1046,
+      "source_sha256": "4ca4b28a171d6130f00e078316e9cdfcdedd0933a1b5fb967517c2e20244ad90",
+      "E2B": {
+        "direct": 728,
+        "adaptive": 851,
+        "direct_pct": "69.60",
+        "adaptive_pct": "81.36",
+        "change_pp": "+11.76"
+      },
+      "E4B": {
+        "direct": 762,
+        "adaptive": 842,
+        "direct_pct": "72.85",
+        "adaptive_pct": "80.50",
+        "change_pp": "+7.65"
+      }
+    },
+    {
+      "key": "typed-decisions",
+      "title": "Typed",
+      "n": 2000,
+      "source_sha256": "143541319ddc4445ba67098c0af0480646e8124d682365d649b8e5947298cc0d",
+      "E2B": {
+        "direct": 1237,
+        "adaptive": 1366,
+        "direct_pct": "61.85",
+        "adaptive_pct": "68.30",
+        "change_pp": "+6.45"
+      },
+      "E4B": {
+        "direct": 1446,
+        "adaptive": 1439,
+        "direct_pct": "72.30",
+        "adaptive_pct": "71.95",
+        "change_pp": "\u22120.35"
+      }
+    }
+  ]
+}

docs/evidence/e2b-v3-analysis.json

@@ -0,0 +1,1534 @@
+{
+  "results_sha256": "db3d7cb81972ad2909a75aca659b97015d9e96a770f04620275b79b9fbdbd80d",
+  "source_commit": "ee6bd37d34ae35d2e69ebb0c4b0957a10c727357",
+  "freeze_sha256": "eb163f63b0eb3a706adaaf18d20ebe3acaef4ddb0690005149eef2cbc750d2a6",
+  "observed": 3277,
+  "full_panel_complete": true,
+  "timing_exclusions": {
+    "latency_exclusion_only": true,
+    "quality_rows_retained": true,
+    "prior_uncertain_ranges": {
+      "1491": [
+        1496,
+        "first foreign-GPU watchdog stop; per-case timestamps unavailable"
+      ],
+      "2651": [
+        2660,
+        "RustDesk watchdog stop; per-case timestamps unavailable"
+      ]
+    },
+    "new_overlap_samples": 0,
+    "excluded_cases": [
+      {
+        "index": 1491,
+        "panel": "typed-decisions",
+        "id": "agent_trace_observability_000042/risk",
+        "reasons": [
+          "first foreign-GPU watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 1492,
+        "panel": "typed-decisions",
+        "id": "agent_trace_observability_000042/urgency",
+        "reasons": [
+          "first foreign-GPU watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 1493,
+        "panel": "typed-decisions",
+        "id": "agent_trace_observability_000043/action",
+        "reasons": [
+          "first foreign-GPU watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 1494,
+        "panel": "typed-decisions",
+        "id": "agent_trace_observability_000043/needs_review",
+        "reasons": [
+          "first foreign-GPU watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 1495,
+        "panel": "typed-decisions",
+        "id": "agent_trace_observability_000043/outcome",
+        "reasons": [
+          "first foreign-GPU watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 1496,
+        "panel": "typed-decisions",
+        "id": "agent_trace_observability_000043/risk",
+        "reasons": [
+          "first foreign-GPU watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2651,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000074/matches_order",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2652,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000074/urgency",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2653,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000075/discrepancy_severity",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2654,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000075/disposition",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2655,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000075/duplicate",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2656,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000075/matches_order",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2657,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000075/urgency",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2658,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000076/discrepancy_severity",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2659,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000076/disposition",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      },
+      {
+        "index": 2660,
+        "panel": "typed-decisions",
+        "id": "invoice_processing_000076/duplicate",
+        "reasons": [
+          "RustDesk watchdog stop; per-case timestamps unavailable"
+        ]
+      }
+    ]
+  },
+  "panels": {
+    "jevbench-public": {
+      "observed": 231,
+      "expected": 231,
+      "complete": true,
+      "valid": 231,
+      "semantics": "verified label accuracy",
+      "errors": [],
+      "fallbacks": [],
+      "direct": {
+        "n": 231,
+        "top1": 175,
+        "nll": 0.6548821957788397,
+        "brier": 0.3384088069425174,
+        "score_expected_level_mae": 0.18163427398620083
+      },
+      "v2": {
+        "n": 231,
+        "top1": 195,
+        "nll": 0.5102332515038959,
+        "brier": 0.24715639752392515,
+        "score_expected_level_mae": 0.1169276614338399
+      },
+      "v3": {
+        "n": 231,
+        "top1": 202,
+        "nll": 0.4439350676441406,
+        "brier": 0.21173286180768633,
+        "score_expected_level_mae": 0.08733389277501005
+      },
+      "historical": {
+        "n": 231,
+        "top1": 203,
+        "nll": 0.4433104727261114,
+        "brier": 0.211353951167008,
+        "score_expected_level_mae": 0.08733389277501005
+      },
+      "direct_to_v3": {
+        "n": 231,
+        "fixes": 35,
+        "regressions": 8,
+        "winner_changes": 54,
+        "max_probability_delta": 0.4938297051869622
+      },
+      "v2_to_v3": {
+        "n": 231,
+        "fixes": 15,
+        "regressions": 8,
+        "winner_changes": 26,
+        "max_probability_delta": 0.49954774239515043
+      },
+      "historical_to_v3": {
+        "n": 231,
+        "fixes": 0,
+        "regressions": 1,
+        "winner_changes": 1,
+        "max_probability_delta": 0.07599484280024105
+      },
+      "routed": 131,
+      "completed_blends": 131,
+      "max_direct_delta": 0.0,
+      "historical_text_equal": 126,
+      "historical_text_compared": 131,
+      "generation_output_tokens": {
+        "sum": 123713,
+        "mean": 944.3740458015267,
+        "max": 2952,
+        "distribution": {
+          "n": 131,
+          "sum": 123713,
+          "mean": 944.3740458015267,
+          "median": 856,
+          "p95": 1936.5,
+          "p99": 2826.3999999999987,
+          "max": 2952
+        }
+      },
+      "pipeline_seconds_all_cost": {
+        "n": 231,
+        "sum": 313.31913769757375,
+        "mean": 1.3563599034527003,
+        "median": 0.8873364869505167,
+        "p95": 4.099002962524537,
+        "p99": 6.056786250858566,
+        "max": 7.182334668934345
+      },
+      "pipeline_seconds_clean": {
+        "n": 231,
+        "sum": 313.31913769757375,
+        "mean": 1.3563599034527003,
+        "median": 0.8873364869505167,
+        "p95": 4.099002962524537,
+        "p99": 6.056786250858566,
+        "max": 7.182334668934345
+      },
+      "timing_excluded_count": 0,
+      "v2_selective_seconds_clean_matched": {
+        "n": 231,
+        "sum": 314.0389325831784,
+        "mean": 1.3594758986284778,
+        "median": 0.8249348889803514,
+        "p95": 4.52698246500222,
+        "p99": 5.536772494891189,
+        "max": 5.958748230012134
+      },
+      "max_native_backend_allocated_bytes": 6601572352,
+      "natural_eos_output_tokens": {
+        "n": 131,
+        "sum": 123713,
+        "mean": 944.3740458015267,
+        "median": 856,
+        "p95": 1936.5,
+        "p99": 2826.3999999999987,
+        "max": 2952
+      },
+      "clean_phase_seconds": {
+        "/v1/chat/completions": {
+          "n": 231,
+          "sum": 283.491011307342,
+          "mean": 1.2272338151832987,
+          "median": 0.8251801070291549,
+          "p95": 3.6716601609950885,
+          "p99": 5.589920414669901,
+          "max": 6.879954903037287
+        },
+        "/v1/systemone": {
+          "n": 231,
+          "sum": 28.06992222147528,
+          "mean": 0.1215148148115813,
+          "median": 0.05774561199359596,
+          "p95": 0.4264363810652867,
+          "p99": 0.5270983364898711,
+          "max": 0.5750212969724089
+        },
+        "/v1/winnow/inspect": {
+          "n": 231,
+          "sum": 1.6974414952564985,
+          "mean": 0.007348231581196963,
+          "median": 0.004081002902239561,
+          "p95": 0.02844280900899321,
+          "p99": 0.0330876518972218,
+          "max": 0.039218239951878786
+        }
+      },
+      "clean_chat_metrics": {
+        "n": 131,
+        "output_tokens": 123713,
+        "decode_seconds": 272.717043,
+        "aggregate_decode_tokens_per_second": 453.63134859158765,
+        "draft_proposed": 152252,
+        "draft_accepted": 85751,
+        "draft_acceptance": 0.563217560360455,
+        "prefill_ms": {
+          "n": 131,
+          "sum": 10241.078,
+          "mean": 78.1761679389313,
+          "median": 46.572,
+          "p95": 210.508,
+          "p99": 227.93679999999998,
+          "max": 248.51
+        },
+        "decode_ms": {
+          "n": 131,
+          "sum": 272717.043,
+          "mean": 2081.8094885496184,
+          "median": 1831.607,
+          "p95": 4055.5175,
+          "p99": 5768.822199999997,
+          "max": 6808.545
+        }
+      },
+      "clean_chat_metrics_natural_eos": {
+        "n": 131,
+        "output_tokens": 123713,
+        "decode_seconds": 272.717043,
+        "aggregate_decode_tokens_per_second": 453.63134859158765,
+        "draft_proposed": 152252,
+        "draft_accepted": 85751,
+        "draft_acceptance": 0.563217560360455,
+        "prefill_ms": {
+          "n": 131,
+          "sum": 10241.078,
+          "mean": 78.1761679389313,
+          "median": 46.572,
+          "p95": 210.508,
+          "p99": 227.93679999999998,
+          "max": 248.51
+        },
+        "decode_ms": {
+          "n": 131,
+          "sum": 272717.043,
+          "mean": 2081.8094885496184,
+          "median": 1831.607,
+          "p95": 4055.5175,
+          "p99": 5768.822199999997,
+          "max": 6808.545
+        }
+      },
+      "by_state": {
+        "dict": {
+          "direct": {
+            "n": 35,
+            "top1": 24,
+            "nll": 0.8098638610329222,
+            "brier": 0.4492148897620551,
+            "score_expected_level_mae": null
+          },
+          "v2": {
+            "n": 35,
+            "top1": 30,
+            "nll": 0.5730443227047298,
+            "brier": 0.2778388271622036,
+            "score_expected_level_mae": null
+          },
+          "v3": {
+            "n": 35,
+            "top1": 28,
+            "nll": 0.5986073814983259,
+            "brier": 0.2874113653226452,
+            "score_expected_level_mae": null
+          },
+          "historical": {
+            "n": 35,
+            "top1": 29,
+            "nll": 0.5952986408668535,
+            "brier": 0.2851932314045603,
+            "score_expected_level_mae": null
+          },
+          "direct_to_v3": {
+            "n": 35,
+            "fixes": 7,
+            "regressions": 3,
+            "winner_changes": 12,
+            "max_probability_delta": 0.4847284009193595
+          },
+          "v2_to_v3": {
+            "n": 35,
+            "fixes": 1,
+            "regressions": 3,
+            "winner_changes": 5,
+            "max_probability_delta": 0.4927516672740554
+          },
+          "historical_to_v3": {
+            "n": 35,
+            "fixes": 0,
+            "regressions": 1,
+            "winner_changes": 1,
+            "max_probability_delta": 0.07599484280024105
+          }
+        },
+        "str": {
+          "direct": {
+            "n": 196,
+            "top1": 151,
+            "nll": 0.6272068984120392,
+            "brier": 0.31862200643902855,
+            "score_expected_level_mae": 0.18163427398620083
+          },
+          "v2": {
+            "n": 196,
+            "top1": 165,
+            "nll": 0.49901698878946127,
+            "brier": 0.24167739223137544,
+            "score_expected_level_mae": 0.1169276614338399
+          },
+          "v3": {
+            "n": 196,
+            "top1": 174,
+            "nll": 0.41631501159875034,
+            "brier": 0.19821884332287223,
+            "score_expected_level_mae": 0.08733389277501005
+          },
+          "historical": {
+            "n": 196,
+            "top1": 174,
+            "nll": 0.4161697284152646,
+            "brier": 0.19816836541030225,
+            "score_expected_level_mae": 0.08733389277501005
+          },
+          "direct_to_v3": {
+            "n": 196,
+            "fixes": 28,
+            "regressions": 5,
+            "winner_changes": 42,
+            "max_probability_delta": 0.4938297051869622
+          },
+          "v2_to_v3": {
+            "n": 196,
+            "fixes": 14,
+            "regressions": 5,
+            "winner_changes": 21,
+            "max_probability_delta": 0.49954774239515043
+          },
+          "historical_to_v3": {
+            "n": 196,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.010968767130365098
+          }
+        }
+      },
+      "by_kind": {
+        "choice": {
+          "direct": {
+            "n": 139,
+            "top1": 100,
+            "nll": 0.8075315805861729,
+            "brier": 0.39522492072947685,
+            "score_expected_level_mae": null
+          },
+          "v2": {
+            "n": 139,
+            "top1": 115,
+            "nll": 0.6044044303940057,
+            "brier": 0.2790112001195667,
+            "score_expected_level_mae": null
+          },
+          "v3": {
+            "n": 139,
+            "top1": 119,
+            "nll": 0.5397349946166345,
+            "brier": 0.24901431746883707,
+            "score_expected_level_mae": null
+          },
+          "historical": {
+            "n": 139,
+            "top1": 119,
+            "nll": 0.5396015160770918,
+            "brier": 0.24895584375003005,
+            "score_expected_level_mae": null
+          },
+          "direct_to_v3": {
+            "n": 139,
+            "fixes": 24,
+            "regressions": 5,
+            "winner_changes": 39,
+            "max_probability_delta": 0.4938297051869622
+          },
+          "v2_to_v3": {
+            "n": 139,
+            "fixes": 9,
+            "regressions": 5,
+            "winner_changes": 17,
+            "max_probability_delta": 0.49954774239515043
+          },
+          "historical_to_v3": {
+            "n": 139,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.010968767130365098
+          }
+        },
+        "noul": {
+          "direct": {
+            "n": 74,
+            "top1": 61,
+            "nll": 0.43519757844670776,
+            "brier": 0.26491784130498336,
+            "score_expected_level_mae": null
+          },
+          "v2": {
+            "n": 74,
+            "top1": 64,
+            "nll": 0.3773254061123309,
+            "brier": 0.207808720831387,
+            "score_expected_level_mae": null
+          },
+          "v3": {
+            "n": 74,
+            "top1": 66,
+            "nll": 0.3184973321031578,
+            "brier": 0.16806665325527254,
+            "score_expected_level_mae": null
+          },
+          "historical": {
+            "n": 74,
+            "top1": 67,
+            "nll": 0.31679830630493755,
+            "brier": 0.1669936733757791,
+            "score_expected_level_mae": null
+          },
+          "direct_to_v3": {
+            "n": 74,
+            "fixes": 8,
+            "regressions": 3,
+            "winner_changes": 11,
+            "max_probability_delta": 0.4496872566776117
+          },
+          "v2_to_v3": {
+            "n": 74,
+            "fixes": 5,
+            "regressions": 3,
+            "winner_changes": 8,
+            "max_probability_delta": 0.4990190336545404
+          },
+          "historical_to_v3": {
+            "n": 74,
+            "fixes": 0,
+            "regressions": 1,
+            "winner_changes": 1,
+            "max_probability_delta": 0.07599484280024105
+          }
+        },
+        "score": {
+          "direct": {
+            "n": 18,
+            "top1": 14,
+            "nll": 0.3792375954654199,
+            "brier": 0.20179167587530403,
+            "score_expected_level_mae": 0.18163427398620083
+          },
+          "v2": {
+            "n": 18,
+            "top1": 16,
+            "nll": 0.32942140112892654,
+            "brier": 0.16292920388246132,
+            "score_expected_level_mae": 0.1169276614338399
+          },
+          "v3": {
+            "n": 18,
+            "top1": 17,
+            "nll": 0.21983521102503334,
+            "brier": 0.10335381158427896,
+            "score_expected_level_mae": 0.08733389277501005
+          },
+          "historical": {
+            "n": 18,
+            "top1": 17,
+            "nll": 0.21983521102503334,
+            "brier": 0.10335381158427896,
+            "score_expected_level_mae": 0.08733389277501005
+          },
+          "direct_to_v3": {
+            "n": 18,
+            "fixes": 3,
+            "regressions": 0,
+            "winner_changes": 4,
+            "max_probability_delta": 0.4173095607519327
+          },
+          "v2_to_v3": {
+            "n": 18,
+            "fixes": 1,
+            "regressions": 0,
+            "winner_changes": 1,
+            "max_probability_delta": 0.49626072038671637
+          },
+          "historical_to_v3": {
+            "n": 18,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.0
+          }
+        }
+      }
+    },
+    "kev-v9-clean": {
+      "observed": 1046,
+      "expected": 1046,
+      "complete": true,
+      "valid": 1046,
+      "semantics": "verified label accuracy",
+      "errors": [],
+      "fallbacks": [
+        {
+          "id": "mmlu_pro/test/9247/answer",
+          "reason": "generation_context_limit_or_unverified"
+        },
+        {
+          "id": "unknowable_control/warranty_claim/v5-test-20260920-warranty_claim-0004/a/decision",
+          "reason": "generation_context_limit_or_unverified"
+        }
+      ],
+      "direct": {
+        "n": 1046,
+        "top1": 728,
+        "nll": 0.9841937846927966,
+        "brier": 0.44164586585817817,
+        "score_expected_level_mae": 0.29737852879492405
+      },
+      "v2": {
+        "n": 1046,
+        "top1": 845,
+        "nll": 0.7616620243944804,
+        "brier": 0.31723865346147595,
+        "score_expected_level_mae": 0.16328748801957052
+      },
+      "v3": {
+        "n": 1046,
+        "top1": 851,
+        "nll": 0.743563618944474,
+        "brier": 0.3110523200923067,
+        "score_expected_level_mae": 0.16352370646939757
+      },
+      "historical": {
+        "n": 1046,
+        "top1": 851,
+        "nll": 0.7439440288222531,
+        "brier": 0.3113648971845476,
+        "score_expected_level_mae": 0.1636530071709598
+      },
+      "direct_to_v3": {
+        "n": 1046,
+        "fixes": 173,
+        "regressions": 50,
+        "winner_changes": 264,
+        "max_probability_delta": 0.49877661297158915
+      },
+      "v2_to_v3": {
+        "n": 1046,
+        "fixes": 38,
+        "regressions": 32,
+        "winner_changes": 100,
+        "max_probability_delta": 0.4998410532110379
+      },
+      "historical_to_v3": {
+        "n": 1046,
+        "fixes": 0,
+        "regressions": 0,
+        "winner_changes": 2,
+        "max_probability_delta": 0.49223936331112117
+      },
+      "routed": 676,
+      "completed_blends": 674,
+      "max_direct_delta": 0.0,
+      "historical_text_equal": 674,
+      "historical_text_compared": 676,
+      "generation_output_tokens": {
+        "sum": 508239,
+        "mean": 751.8328402366864,
+        "max": 8012,
+        "distribution": {
+          "n": 676,
+          "sum": 508239,
+          "mean": 751.8328402366864,
+          "median": 565.5,
+          "p95": 1889.75,
+          "p99": 3618.25,
+          "max": 8012
+        }
+      },
+      "pipeline_seconds_all_cost": {
+        "n": 1046,
+        "sum": 1267.75761802739,
+        "mean": 1.212005370963088,
+        "median": 1.028721651993692,
+        "p95": 3.5224499475152697,
+        "p99": 6.895163247355956,
+        "max": 12.624659975990653
+      },
+      "pipeline_seconds_clean": {
+        "n": 1046,
+        "sum": 1267.75761802739,
+        "mean": 1.212005370963088,
+        "median": 1.028721651993692,
+        "p95": 3.5224499475152697,
+        "p99": 6.895163247355956,
+        "max": 12.624659975990653
+      },
+      "timing_excluded_count": 0,
+      "v2_selective_seconds_clean_matched": {
+        "n": 1046,
+        "sum": 1300.9493182514561,
+        "mean": 1.2437373979459427,
+        "median": 0.9990504610468633,
+        "p95": 3.6775082352978643,
+        "p99": 6.746468279679535,
+        "max": 9.355459655052982
+      },
+      "max_native_backend_allocated_bytes": 6603538432,
+      "natural_eos_output_tokens": {
+        "n": 674,
+        "sum": 492665,
+        "mean": 730.9569732937686,
+        "median": 563.5,
+        "p95": 1851.4500000000003,
+        "p99": 3412.4799999999996,
+        "max": 5334
+      },
+      "clean_phase_seconds": {
+        "/v1/chat/completions": {
+          "n": 1046,
+          "sum": 1196.6142937389668,
+          "mean": 1.1439907205917466,
+          "median": 0.9535917384782806,
+          "p95": 3.3786256242892705,
+          "p99": 6.638901050179261,
+          "max": 12.581271308939904
+        },
+        "/v1/systemone": {
+          "n": 1046,
+          "sum": 66.71202236169484,
+          "mean": 0.06377822405515758,
+          "median": 0.06213975552236661,
+          "p95": 0.13457796376314946,
+          "p99": 0.23082962960470413,
+          "max": 0.39296685194130987
+        },
+        "/v1/winnow/inspect": {
+          "n": 1046,
+          "sum": 4.167626591864973,
+          "mean": 0.003984346646142422,
+          "median": 0.004179205046966672,
+          "p95": 0.008609114302089438,
+          "p99": 0.010857289913110432,
+          "max": 0.01358645805157721
+        }
+      },
+      "clean_chat_metrics": {
+        "n": 676,
+        "output_tokens": 508239,
+        "decode_seconds": 1172.2648219999999,
+        "aggregate_decode_tokens_per_second": 433.55305939565255,
+        "draft_proposed": 666445,
+        "draft_accepted": 342220,
+        "draft_acceptance": 0.513500738995716,
+        "prefill_ms": {
+          "n": 676,
+          "sum": 22825.388,
+          "mean": 33.76536686390533,
+          "median": 32.602500000000006,
+          "p95": 41.06625,
+          "p99": 48.360749999999996,
+          "max": 80.819
+        },
+        "decode_ms": {
+          "n": 676,
+          "sum": 1172264.822,
+          "mean": 1734.1195591715975,
+          "median": 1317.3195,
+          "p95": 4071.9035000000003,
+          "p99": 8340.009,
+          "max": 12524.258
+        }
+      },
+      "clean_chat_metrics_natural_eos": {
+        "n": 674,
+        "output_tokens": 492665,
+        "decode_seconds": 1147.630398,
+        "aggregate_decode_tokens_per_second": 429.28890769935845,
+        "draft_proposed": 653084,
+        "draft_accepted": 329989,
+        "draft_acceptance": 0.5052780346785406,
+        "prefill_ms": {
+          "n": 674,
+          "sum": 22739.867,
+          "mean": 33.73867507418397,
+          "median": 32.602500000000006,
+          "p95": 41.03915,
+          "p99": 46.812639999999995,
+          "max": 80.819
+        },
+        "decode_ms": {
+          "n": 674,
+          "sum": 1147630.398,
+          "mean": 1702.7157240356082,
+          "median": 1315.1734999999999,
+          "p95": 4007.4803000000006,
+          "p99": 7377.133189999987,
+          "max": 11452.211
+        }
+      },
+      "by_state": {
+        "dict": {
+          "direct": {
+            "n": 757,
+            "top1": 506,
+            "nll": 1.0263358490810022,
+            "brier": 0.4660603526231764,
+            "score_expected_level_mae": 0.29737852879492405
+          },
+          "v2": {
+            "n": 757,
+            "top1": 621,
+            "nll": 0.7171781019173493,
+            "brier": 0.3007661792699979,
+            "score_expected_level_mae": 0.16328748801957052
+          },
+          "v3": {
+            "n": 757,
+            "top1": 619,
+            "nll": 0.7081713179511532,
+            "brier": 0.30210716310944896,
+            "score_expected_level_mae": 0.16352370646939757
+          },
+          "historical": {
+            "n": 757,
+            "top1": 619,
+            "nll": 0.7087038104127751,
+            "brier": 0.3025430916237622,
+            "score_expected_level_mae": 0.1636530071709598
+          },
+          "direct_to_v3": {
+            "n": 757,
+            "fixes": 150,
+            "regressions": 37,
+            "winner_changes": 225,
+            "max_probability_delta": 0.49877661297158915
+          },
+          "v2_to_v3": {
+            "n": 757,
+            "fixes": 24,
+            "regressions": 26,
+            "winner_changes": 78,
+            "max_probability_delta": 0.4998410532110379
+          },
+          "historical_to_v3": {
+            "n": 757,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 2,
+            "max_probability_delta": 0.49223936331112117
+          }
+        },
+        "list": {
+          "direct": {
+            "n": 11,
+            "top1": 9,
+            "nll": 0.45089530046520027,
+            "brier": 0.2974764690463965,
+            "score_expected_level_mae": null
+          },
+          "v2": {
+            "n": 11,
+            "top1": 10,
+            "nll": 0.1983185854747978,
+            "brier": 0.12111648754883751,
+            "score_expected_level_mae": null
+          },
+          "v3": {
+            "n": 11,
+            "top1": 11,
+            "nll": 0.1368364542171407,
+            "brier": 0.07533518108769578,
+            "score_expected_level_mae": null
+          },
+          "historical": {
+            "n": 11,
+            "top1": 11,
+            "nll": 0.13636481228161182,
+            "brier": 0.07505861319212954,
+            "score_expected_level_mae": null
+          },
+          "direct_to_v3": {
+            "n": 11,
+            "fixes": 2,
+            "regressions": 0,
+            "winner_changes": 2,
+            "max_probability_delta": 0.4603535862018596
+          },
+          "v2_to_v3": {
+            "n": 11,
+            "fixes": 1,
+            "regressions": 0,
+            "winner_changes": 1,
+            "max_probability_delta": 0.48803378909777917
+          },
+          "historical_to_v3": {
+            "n": 11,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.0013950736901103267
+          }
+        },
+        "str": {
+          "direct": {
+            "n": 278,
+            "top1": 213,
+            "nll": 0.8905417724072999,
+            "brier": 0.38086923594388294,
+            "score_expected_level_mae": null
+          },
+          "v2": {
+            "n": 278,
+            "top1": 214,
+            "nll": 0.9050832731114038,
+            "brier": 0.3698537857923678,
+            "score_expected_level_mae": null
+          },
+          "v3": {
+            "n": 278,
+            "top1": 221,
+            "nll": 0.8639448083831234,
+            "brier": 0.34473711277242924,
+            "score_expected_level_mae": null
+          },
+          "historical": {
+            "n": 278,
+            "top1": 221,
+            "nll": 0.8639448083831234,
+            "brier": 0.34473711277242924,
+            "score_expected_level_mae": null
+          },
+          "direct_to_v3": {
+            "n": 278,
+            "fixes": 21,
+            "regressions": 13,
+            "winner_changes": 37,
+            "max_probability_delta": 0.4906821240628119
+          },
+          "v2_to_v3": {
+            "n": 278,
+            "fixes": 13,
+            "regressions": 6,
+            "winner_changes": 21,
+            "max_probability_delta": 0.498794986167734
+          },
+          "historical_to_v3": {
+            "n": 278,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.0
+          }
+        }
+      },
+      "by_kind": {
+        "choice": {
+          "direct": {
+            "n": 576,
+            "top1": 343,
+            "nll": 1.3449669496263348,
+            "brier": 0.5692832548094673,
+            "score_expected_level_mae": null
+          },
+          "v2": {
+            "n": 576,
+            "top1": 430,
+            "nll": 1.0454033035132229,
+            "brier": 0.4124806597366923,
+            "score_expected_level_mae": null
+          },
+          "v3": {
+            "n": 576,
+            "top1": 433,
+            "nll": 1.0170622984634021,
+            "brier": 0.40463624557307726,
+            "score_expected_level_mae": null
+          },
+          "historical": {
+            "n": 576,
+            "top1": 433,
+            "nll": 1.018041954435723,
+            "brier": 0.405391147268223,
+            "score_expected_level_mae": null
+          },
+          "direct_to_v3": {
+            "n": 576,
+            "fixes": 120,
+            "regressions": 30,
+            "winner_changes": 191,
+            "max_probability_delta": 0.49877661297158915
+          },
+          "v2_to_v3": {
+            "n": 576,
+            "fixes": 27,
+            "regressions": 24,
+            "winner_changes": 81,
+            "max_probability_delta": 0.4998410532110379
+          },
+          "historical_to_v3": {
+            "n": 576,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 2,
+            "max_probability_delta": 0.49223936331112117
+          }
+        },
+        "noul": {
+          "direct": {
+            "n": 370,
+            "top1": 311,
+            "nll": 0.5029575433062489,
+            "brier": 0.2629504138741171,
+            "score_expected_level_mae": null
+          },
+          "v2": {
+            "n": 370,
+            "top1": 318,
+            "nll": 0.45439529360118935,
+            "brier": 0.22222817979439785,
+            "score_expected_level_mae": null
+          },
+          "v3": {
+            "n": 370,
+            "top1": 321,
+            "nll": 0.4465732399814046,
+            "brier": 0.2164619524787992,
+            "score_expected_level_mae": null
+          },
+          "historical": {
+            "n": 370,
+            "top1": 321,
+            "nll": 0.44608433300775163,
+            "brier": 0.21615934481882615,
+            "score_expected_level_mae": null
+          },
+          "direct_to_v3": {
+            "n": 370,
+            "fixes": 28,
+            "regressions": 18,
+            "winner_changes": 46,
+            "max_probability_delta": 0.4901768441417294
+          },
+          "v2_to_v3": {
+            "n": 370,
+            "fixes": 11,
+            "regressions": 8,
+            "winner_changes": 19,
+            "max_probability_delta": 0.49871298353542987
+          },
+          "historical_to_v3": {
+            "n": 370,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.028342231557091452
+          }
+        },
+        "score": {
+          "direct": {
+            "n": 100,
+            "top1": 74,
+            "nll": 0.6867144478058433,
+            "brier": 0.36762767783977907,
+            "score_expected_level_mae": 0.29737852879492405
+          },
+          "v2": {
+            "n": 100,
+            "top1": 97,
+            "nll": 0.2641991606057006,
+            "brier": 0.12018344988441894,
+            "score_expected_level_mae": 0.16328748801957052
+          },
+          "v3": {
+            "n": 100,
+            "top1": 97,
+            "nll": 0.26707562707880517,
+            "brier": 0.12199326949304623,
+            "score_expected_level_mae": 0.16352370646939757
+          },
+          "historical": {
+            "n": 100,
+            "top1": 97,
+            "nll": 0.2672208518023219,
+            "brier": 0.12203424045574633,
+            "score_expected_level_mae": 0.1636530071709598
+          },
+          "direct_to_v3": {
+            "n": 100,
+            "fixes": 25,
+            "regressions": 2,
+            "winner_changes": 27,
+            "max_probability_delta": 0.4908280003843208
+          },
+          "v2_to_v3": {
+            "n": 100,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.2012603587773023
+          },
+          "historical_to_v3": {
+            "n": 100,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.0027665416307858237
+          }
+        }
+      }
+    },
+    "typed-decisions": {
+      "observed": 2000,
+      "expected": 2000,
+      "complete": true,
+      "valid": 2000,
+      "semantics": "synthetic teacher agreement",
+      "errors": [],
+      "fallbacks": [],
+      "direct": {
+        "n": 2000,
+        "top1": 1237,
+        "nll": 0.98012896186682,
+        "brier": 0.5366594455578602,
+        "score_expected_level_mae": 0.5718142114825938
+      },
+      "v2": {
+        "n": 2000,
+        "top1": 1342,
+        "nll": 0.9213980450857948,
+        "brier": 0.48476748880056186,
+        "score_expected_level_mae": 0.5018408163629516
+      },
+      "v3": {
+        "n": 2000,
+        "top1": 1366,
+        "nll": 0.8821571197982664,
+        "brier": 0.4700707747544247,
+        "score_expected_level_mae": 0.48623367185770394
+      },
+      "historical": {
+        "n": 2000,
+        "top1": 1365,
+        "nll": 0.8821582687252656,
+        "brier": 0.4699970802405581,
+        "score_expected_level_mae": 0.48607930491573853
+      },
+      "direct_to_v3": {
+        "n": 2000,
+        "fixes": 334,
+        "regressions": 205,
+        "winner_changes": 632,
+        "max_probability_delta": 0.49734083274590535
+      },
+      "v2_to_v3": {
+        "n": 2000,
+        "fixes": 173,
+        "regressions": 149,
+        "winner_changes": 388,
+        "max_probability_delta": 0.49990235636218977
+      },
+      "historical_to_v3": {
+        "n": 2000,
+        "fixes": 1,
+        "regressions": 0,
+        "winner_changes": 1,
+        "max_probability_delta": 0.08040670116881293
+      },
+      "routed": 1777,
+      "completed_blends": 1777,
+      "max_direct_delta": 0.0,
+      "historical_text_equal": 1777,
+      "historical_text_compared": 1777,
+      "generation_output_tokens": {
+        "sum": 1367542,
+        "mean": 769.5790658413056,
+        "max": 2712,
+        "distribution": {
+          "n": 1777,
+          "sum": 1367542,
+          "mean": 769.5790658413056,
+          "median": 762,
+          "p95": 1051.1999999999998,
+          "p99": 1254.3600000000001,
+          "max": 2712
+        }
+      },
+      "pipeline_seconds_all_cost": {
+        "n": 2000,
+        "sum": 3901.123203223222,
+        "mean": 1.950561601611611,
+        "median": 2.093935512995813,
+        "p95": 3.0059627732262015,
+        "p99": 3.7421011833066586,
+        "max": 8.422256662975997
+      },
+      "pipeline_seconds_clean": {
+        "n": 1984,
+        "sum": 3871.4878648194717,
+        "mean": 1.9513547705743306,
+        "median": 2.0951888249837793,
+        "p95": 3.0047559435595756,
+        "p99": 3.743849764257903,
+        "max": 8.422256662975997
+      },
+      "timing_excluded_count": 16,
+      "v2_selective_seconds_clean_matched": {
+        "n": 1984,
+        "sum": 4021.9517563040135,
+        "mean": 2.0271934255564585,
+        "median": 2.178467755962629,
+        "p95": 3.0571437860082367,
+        "p99": 4.1093565620703165,
+        "max": 8.320580558967777
+      },
+      "max_native_backend_allocated_bytes": 6753550336,
+      "natural_eos_output_tokens": {
+        "n": 1777,
+        "sum": 1367542,
+        "mean": 769.5790658413056,
+        "median": 762,
+        "p95": 1051.1999999999998,
+        "p99": 1254.3600000000001,
+        "max": 2712
+      },
+      "clean_phase_seconds": {
+        "/v1/chat/completions": {
+          "n": 1984,
+          "sum": 3670.5011939010583,
+          "mean": 1.8500510049904528,
+          "median": 1.9909222425194457,
+          "p95": 2.8783525532402563,
+          "p99": 3.5913489963649803,
+          "max": 8.18496916803997
+        },
+        "/v1/systemone": {
+          "n": 1984,
+          "sum": 187.3238072542008,
+          "mean": 0.09441724155957702,
+          "median": 0.09884031204273924,
+          "p95": 0.12675707904854788,
+          "p99": 0.14250179579830732,
+          "max": 0.2218943398911506
+        },
+        "/v1/winnow/inspect": {
+          "n": 1984,
+          "sum": 12.908032948966138,
+          "mean": 0.006506064994438577,
+          "median": 0.006854985433164984,
+          "p95": 0.008897835278185084,
+          "p99": 0.01095620173728094,
+          "max": 0.016994947916828096
+        }
+      },
+      "clean_chat_metrics": {
+        "n": 1763,
+        "output_tokens": 1356905,
+        "decode_seconds": 3597.4189079999996,
+        "aggregate_decode_tokens_per_second": 377.18848838607374,
+        "draft_proposed": 2043980,
+        "draft_accepted": 847677,
+        "draft_acceptance": 0.4147188328652922,
+        "prefill_ms": {
+          "n": 1763,
+          "sum": 68639.242,
+          "mean": 38.933205899035734,
+          "median": 39.707,
+          "p95": 45.585,
+          "p99": 51.141939999999984,
+          "max": 55.325
+        },
+        "decode_ms": {
+          "n": 1763,
+          "sum": 3597418.908,
+          "mean": 2040.5098740782757,
+          "median": 2028.45,
+          "p95": 2874.6335999999988,
+          "p99": 3605.586639999997,
+          "max": 8142.273
+        }
+      },
+      "clean_chat_metrics_natural_eos": {
+        "n": 1763,
+        "output_tokens": 1356905,
+        "decode_seconds": 3597.4189079999996,
+        "aggregate_decode_tokens_per_second": 377.18848838607374,
+        "draft_proposed": 2043980,
+        "draft_accepted": 847677,
+        "draft_acceptance": 0.4147188328652922,
+        "prefill_ms": {
+          "n": 1763,
+          "sum": 68639.242,
+          "mean": 38.933205899035734,
+          "median": 39.707,
+          "p95": 45.585,
+          "p99": 51.141939999999984,
+          "max": 55.325
+        },
+        "decode_ms": {
+          "n": 1763,
+          "sum": 3597418.908,
+          "mean": 2040.5098740782757,
+          "median": 2028.45,
+          "p95": 2874.6335999999988,
+          "p99": 3605.586639999997,
+          "max": 8142.273
+        }
+      },
+      "by_state": {
+        "dict": {
+          "direct": {
+            "n": 2000,
+            "top1": 1237,
+            "nll": 0.98012896186682,
+            "brier": 0.5366594455578602,
+            "score_expected_level_mae": 0.5718142114825938
+          },
+          "v2": {
+            "n": 2000,
+            "top1": 1342,
+            "nll": 0.9213980450857948,
+            "brier": 0.48476748880056186,
+            "score_expected_level_mae": 0.5018408163629516
+          },
+          "v3": {
+            "n": 2000,
+            "top1": 1366,
+            "nll": 0.8821571197982664,
+            "brier": 0.4700707747544247,
+            "score_expected_level_mae": 0.48623367185770394
+          },
+          "historical": {
+            "n": 2000,
+            "top1": 1365,
+            "nll": 0.8821582687252656,
+            "brier": 0.4699970802405581,
+            "score_expected_level_mae": 0.48607930491573853
+          },
+          "direct_to_v3": {
+            "n": 2000,
+            "fixes": 334,
+            "regressions": 205,
+            "winner_changes": 632,
+            "max_probability_delta": 0.49734083274590535
+          },
+          "v2_to_v3": {
+            "n": 2000,
+            "fixes": 173,
+            "regressions": 149,
+            "winner_changes": 388,
+            "max_probability_delta": 0.49990235636218977
+          },
+          "historical_to_v3": {
+            "n": 2000,
+            "fixes": 1,
+            "regressions": 0,
+            "winner_changes": 1,
+            "max_probability_delta": 0.08040670116881293
+          }
+        }
+      },
+      "by_kind": {
+        "choice": {
+          "direct": {
+            "n": 600,
+            "top1": 358,
+            "nll": 1.0383454926500515,
+            "brier": 0.5461734111343014,
+            "score_expected_level_mae": null
+          },
+          "v2": {
+            "n": 600,
+            "top1": 390,
+            "nll": 0.9958149854968003,
+            "brier": 0.5063973319925225,
+            "score_expected_level_mae": null
+          },
+          "v3": {
+            "n": 600,
+            "top1": 393,
+            "nll": 0.9563012200742745,
+            "brier": 0.4976482188673111,
+            "score_expected_level_mae": null
+          },
+          "historical": {
+            "n": 600,
+            "top1": 393,
+            "nll": 0.9558431778730259,
+            "brier": 0.4973719628971465,
+            "score_expected_level_mae": null
+          },
+          "direct_to_v3": {
+            "n": 600,
+            "fixes": 93,
+            "regressions": 58,
+            "winner_changes": 190,
+            "max_probability_delta": 0.49734083274590535
+          },
+          "v2_to_v3": {
+            "n": 600,
+            "fixes": 50,
+            "regressions": 47,
+            "winner_changes": 123,
+            "max_probability_delta": 0.49990235636218977
+          },
+          "historical_to_v3": {
+            "n": 600,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.02832484742590488
+          }
+        },
+        "noul": {
+          "direct": {
+            "n": 600,
+            "top1": 438,
+            "nll": 0.781763055105711,
+            "brier": 0.4553453509151658,
+            "score_expected_level_mae": null
+          },
+          "v2": {
+            "n": 600,
+            "top1": 470,
+            "nll": 0.7073079185383874,
+            "brier": 0.3751118454262236,
+            "score_expected_level_mae": null
+          },
+          "v3": {
+            "n": 600,
+            "top1": 477,
+            "nll": 0.6776685742578799,
+            "brier": 0.36021695730742676,
+            "score_expected_level_mae": null
+          },
+          "historical": {
+            "n": 600,
+            "top1": 476,
+            "nll": 0.6782093598018603,
+            "brier": 0.360367007607702,
+            "score_expected_level_mae": null
+          },
+          "direct_to_v3": {
+            "n": 600,
+            "fixes": 73,
+            "regressions": 34,
+            "winner_changes": 107,
+            "max_probability_delta": 0.49051290118046414
+          },
+          "v2_to_v3": {
+            "n": 600,
+            "fixes": 31,
+            "regressions": 24,
+            "winner_changes": 55,
+            "max_probability_delta": 0.49906662495041354
+          },
+          "historical_to_v3": {
+            "n": 600,
+            "fixes": 1,
+            "regressions": 0,
+            "winner_changes": 1,
+            "max_probability_delta": 0.010967217955198949
+          }
+        },
+        "score": {
+          "direct": {
+            "n": 800,
+            "top1": 441,
+            "nll": 1.085240993850228,
+            "brier": 0.59050954235755,
+            "score_expected_level_mae": 0.5718142114825938
+          },
+          "v2": {
+            "n": 800,
+            "top1": 482,
+            "nll": 1.0261529346880962,
+            "brier": 0.550786838937345,
+            "score_expected_level_mae": 0.5018408163629516
+          },
+          "v3": {
+            "n": 800,
+            "top1": 496,
+            "nll": 0.9799154537465503,
+            "brier": 0.5317780547550084,
+            "score_expected_level_mae": 0.48623367185770394
+          },
+          "historical": {
+            "n": 800,
+            "top1": 496,
+            "nll": 0.9798562685569995,
+            "brier": 0.5316884727227588,
+            "score_expected_level_mae": 0.48607930491573853
+          },
+          "direct_to_v3": {
+            "n": 800,
+            "fixes": 168,
+            "regressions": 113,
+            "winner_changes": 335,
+            "max_probability_delta": 0.49632280685398744
+          },
+          "v2_to_v3": {
+            "n": 800,
+            "fixes": 92,
+            "regressions": 78,
+            "winner_changes": 210,
+            "max_probability_delta": 0.49985733751079203
+          },
+          "historical_to_v3": {
+            "n": 800,
+            "fixes": 0,
+            "regressions": 0,
+            "winner_changes": 0,
+            "max_probability_delta": 0.08040670116881293
+          }
+        }
+      }
+    }
+  }
+}

docs/evidence/e2b-v3-completion.json

@@ -0,0 +1,24 @@
+{
+  "done": 3277,
+  "total": 3277,
+  "panels": {
+    "jevbench-public": 231,
+    "kev-v9-clean": 1046,
+    "typed-decisions": 2000
+  },
+  "routed": 2584,
+  "completed_blends": 2582,
+  "fallbacks": 2,
+  "errors": 0,
+  "max_direct_delta": 0.0,
+  "recovered_diagnosed_cases": [
+    1001,
+    1236
+  ],
+  "recovery_elapsed_seconds": 1326.1061786840437,
+  "time_utc": "2026-10-06T16:38:00Z",
+  "complete": true,
+  "source_commit": "ee6bd37d34ae35d2e69ebb0c4b0957a10c727357",
+  "freeze_sha256": "eb163f63b0eb3a706adaaf18d20ebe3acaef4ddb0690005149eef2cbc750d2a6",
+  "output_sha256": "db3d7cb81972ad2909a75aca659b97015d9e96a770f04620275b79b9fbdbd80d"
+}

docs/evidence/e2b-v3-freeze.json

@@ -0,0 +1,142 @@
+{
+  "source_commit": "ee6bd37d34ae35d2e69ebb0c4b0957a10c727357",
+  "parent_commit": "947bf3ad350e180e5fb126ee80c921fab19465b5",
+  "source_archive_sha256": "999179036fa60ab7524706bf78050cda8f70d130fbfd756f6ab41d5325f1ed71",
+  "policy_id": "e2b-raw99-blend50-v3",
+  "runtime_profile": "e2b-q8-text8k-mtp",
+  "panels": {
+    "jevbench-public": 231,
+    "kev-v9-clean": 1046,
+    "typed-decisions": 2000
+  },
+  "baseline_files": {
+    "jevbench-public": "8384afe4d943ce1ddfb9e86fde9599d773a0023229094c9c2ef5126e56a00de9",
+    "kev-v9-clean": "0702129a098dfb2dcc33dc8c64fa67180c829532c5fa470bc2e279d42ec8bf86",
+    "typed-decisions": "5b02533afcd7ff0845bc755d80d70e35584fce3748b4cd5c787f9e7bca5c3aa7"
+  },
+  "direct_tolerance": 1e-10,
+  "wall_cap_seconds": 10800,
+  "files": {
+    "FULL-VALIDATION-PROTOCOL.md": "c249f73a9499a21a1bc65c5220691219d3f49a84b6357bc1ce2c8132e0f55b01",
+    "direct-reference.jsonl": "98914abd055858cc9a9dbc435397059c60c1e97497ba02dee244e28882114865",
+    "e2b-v3-source.tar.gz": "999179036fa60ab7524706bf78050cda8f70d130fbfd756f6ab41d5325f1ed71",
+    "frozen-requests.jsonl": "bedf68610bc24dbf660bca54fd6d5a47dabb9e46404d059e2439121e8025c9ca",
+    "full_controller.py": "28041d4dde21a34471230276dce525d1a5f3b4e6a0d1957e932f32525c336091",
+    "full_probe.py": "f595767c69b2a9c8df7a340cc8572c33c235ed4abd4f658d0ee0dff0c67e607e"
+  },
+  "source_files": {
+    ".clang-format": "96cd8537a8a322552fd6d23a9e3470e6e748233116f05093e10c334cdf4eb787",
+    ".dockerignore": "9712f1e0aa72bf285399c336dc20937acb3d7171a8e1eb47db6707ed0cb02466",
+    ".gitattributes": "0bcc1cafcbe5d86f8f6f245e0db28ae06eb9b403dba8821d72fa7c51d96d249d",
+    ".gitignore": "d6497c2e377dc24fa49e418c95bd3136ee22b15e2b632897e7ea2c9821dba2f4",
+    "CMakeLists.txt": "5b797d4ca694e875803b02b7979cfa89c3d8fc64cccb2a0d61ccbdb2466ed01d",
+    "Dockerfile": "9da5f2a7b799c8ea28d9963c61e52ff4b0e7937631d2363775e632d026340034",
+    "LICENSE": "6cc43947bc50233b132fe3cf248c86aa70ed38c7d7b5253e01ea5b3f65fe88e0",
+    "README.md": "914df5cb8e8b010ad6d5850085ce813ca4fcc56e3c9dc9f66585020712e0d64b",
+    "docs/ADAPTIVE.md": "d84abd5a17ccaf3c4457a8edc3cd4b70c2a65a5b3015c99383a6a648a0c52ae1",
+    "docs/API.md": "aa93d4be062cdceb42e8ad9a0618cf742e7986bb5eff2d82e23491778ea2ecda",
+    "docs/BENCHMARKS.md": "319118495adde81bda28bec6d25886c3edd787a7e41d2ba83305c881efec96d5",
+    "docs/E2B.md": "fe0e75c6dba1730d7a948481c2b5d441c79e0301d912398ae1b97cb21aeb4715",
+    "docs/EVALUATION.md": "60bc73264c9373aa2526016c6f52a312210510a58a476419d33caae51f90fb29",
+    "docs/IMAGE-REASONING.md": "ddc2a4efa91c411c2c6a5733cbf686d25a0d7b1f3d97c6fb16bf2092ca026d99",
+    "docs/IMPLEMENTATION.md": "8093f0dbd9e766fdfc994016e1487889d09057910d28fc50e366ba30ea428414",
+    "docs/INSTALL.md": "439b55cdbbe45d4cce4b3d36aa7cdbe4e8f05ca2f1657f71b196099f27259cf1",
+    "docs/PRESETS-AND-ASSETS.md": "afa29722eb43cc75cfec2fc3278dd6421cb2ba90732c4a25a2ba353b52ecaba7",
+    "docs/QUICKSTART.md": "deedf4fb4bb8587b3d50c6d715ccbea5fb7851847bd5e973c7f75a8c2d83ac61",
+    "docs/REASONING-RESULTS.md": "fa12e0a4f95845aeec5b219f25a5c309e677a2c3a8fed2f3f5441c1c79edef50",
+    "docs/RELEASE.md": "521b0dea553bec89dd3b8f936afa28c9b48168a40489c64976079b33a7e2eb97",
+    "docs/VALIDATION.md": "2e4a2aeff052f4fb6ec76c3ac4673572ba070befe0c30c77e5a3a3dc56508a76",
+    "docs/assets/01-5070ti-capability.png": "ae66377e15a8f72f1d9013dc6287b7f8ee0183fca982075a9d2abe7393631b62",
+    "docs/assets/01-5070ti-capability.svg": "3edf0643e3523530b1f583bad8a0fa6cb91af4dba4f7508dee331df8c0e47040",
+    "docs/assets/02-decision-quality.png": "6b8fc874fbebe0dae497ebb485b373db674461f88d1544b1d465213fb1df1bc9",
+    "docs/assets/02-decision-quality.svg": "e618d9a18241fce10fe53bfb1bcbc3a3806993a7ad640fb2f161dd23f70968cf",
+    "docs/assets/03-deployment-memory.png": "59ae7f71e9fd54a8ecc5d7540812a044802d8e0ef95508ba54b9a892be3d821c",
+    "docs/assets/03-deployment-memory.svg": "2134d134df1a7f64f98105b5b8df8ba7fb84839e125df190e8cc194aafc34034",
+    "docs/assets/04-inference-speed.png": "e0620bb64850df6cc63304ddbf56aefb77cb8d5d36a640479bfe4aeb81fb7773",
+    "docs/assets/04-inference-speed.svg": "08699674712723922413c7b89a3480117b26024d7a91a21da616bf5a762b263e",
+    "docs/assets/12b-03-q8-reasoning.png": "937894a60a0fa571116e98a5250e47a3e7e3a66f6782199d8ab1263e7a490f84",
+    "docs/assets/12b-03-q8-reasoning.svg": "a3511ade45f22cb18dcac6648d522af6c7132b7d1b8cebf8c8e247254a923449",
+    "docs/assets/IBM-Plex-Mono-LICENSE.txt": "7e6b2818edbd8f6a01ae80641cc8f16a51080d08fb4e532be3a0b6f74adb07da",
+    "docs/assets/Instrument-Sans-LICENSE.txt": "9e27a72ed30eb49a08678f6a5d6ed98ec7ba5368f541637ee0683ec9134ef966",
+    "docs/assets/e4b-01-reasoning-outcomes.png": "37746e0cc052931e02d707072ee4a2424a6a7e8386121c687a042960c1610527",
+    "docs/assets/e4b-01-reasoning-outcomes.svg": "f86516098d157457b5f1df7e98c77f4356f01636dc40230911ae12077ba4f046",
+    "docs/assets/e4b-02-reasoning-latency.png": "e8ea198a76a638d41dc3eeeba1511f23434b2dcc0321f63b105e254485f3d0ea",
+    "docs/assets/e4b-02-reasoning-latency.svg": "565c71da5306159f455140fe7160dc20fc8513a225abc68a3d607ac9432e1501",
+    "docs/benchmarks.json": "84a7e835d1dce3d079c4d5206f8cd2ba8cff6dca3aac1600de20f140cd82513c",
+    "docs/e2b-evidence.json": "4a28f9f39f5de4015591675c6e4d5e52ebe78c9932f9367c6c2e61e7bd922e25",
+    "docs/reasoning-results.json": "d3a0c5c9dd2ff856eaa330b2002da11cbdd40c38aa0297ff19cf05703c3816f8",
+    "examples/adaptive-decision.json": "01a243f2dff4097c0a61c888706047d1951dbb539c5592e84d0b905f740abff6",
+    "examples/client.mjs": "e1d4d697b21f518acedf117940fbea2bc42c0489e3a1f72cad1b8444452b493a",
+    "examples/client.py": "a2ebcda4327947bf35703012440f43714da8c5cc4eb9bc131567d2ab4b218b24",
+    "examples/decisions.json": "f30cd9b930f62b12c4a4c13a94e029b33a818cd4d87a883f44e0df833b4135d3",
+    "manifests/adaptive-v1.json": "9a38af7981f2c035ee11f29cce02751dcef798c5149d19c376f34bfbdb125eb9",
+    "manifests/assistants-v1.json": "4e1e7a63b4a5676c26323574a7b8a8577cb25c986f89c4f9163ee16fcb26360a",
+    "manifests/models.json": "62d53fdfd1b68def79f68e2ab957e7d2e2ce5ec26d09902676248a8bc3cd72a0",
+    "manifests/release-assets-v1.json": "23bd549612c31ee9f373ca64b825757094b62e16801d9ead6d308cd428acfc25",
+    "manifests/runtime-presets-v1.json": "4fb87a1ae9ffd4e00c95e465b9440c292b1d8791fd711098f06f6ed7bf9dae79",
+    "native/bridge.cpp": "c8815cfc741238011612e49cca6112b79289acff7a9ba74b7abcab873fdced98",
+    "native/bridge.h": "6db293f54c7f9411d5e9b99e1d7b1f5ed066d515066324b772d5566f7a49af9f",
+    "native/engine.h": "9a27e5ad9644b69424aacba0f54a36e8c489733aed01c3918c6cf50012d629ac",
+    "native/images.h": "5200be571801ec4ab107c06f2364dfa136a71b177ccc1c8ec5db927fa3630a34",
+    "native/planner.h": "37cac384797573c55c8d95684dad3f149aa69d2341982f763332e5166822fe86",
+    "native/protocol.h": "bc9918e5fe0c072c5a5dc54e4e6a9c73adeb30e08f7dc7a0abdbfc1d11f3b2f5",
+    "patches/llama-bounded-swa-fork.patch": "79f095b43ea7e072aaf51ab5f3a3edd290d2c598fadd8450f0ab508306715561",
+    "patches/llama-classifier-head.patch": "0c381d753cdf5705cd55e8b72d07904241ae45e3d6005010545031774594237d",
+    "patches/llama-e2b-mtp-shape.patch": "47e85e9bff800e1c46934f70bc65571c372ea6d170b06a9a3fad57ac9a955bb0",
+    "patches/llama-gemma4-tied-embedding.patch": "caeaa51bec0c2dc5c8e48ccea2f5fce227333f60013e749f92152b87f8250b32",
+    "patches/llama-managed-config.patch": "ab58bd6ff5ea296dedb5871f18aa51f11f8de18be56b7bcee52b43f1c38b5a95",
+    "patches/llama-resident-mtp-opt-in.patch": "f3c66d57b2202c4b2e9ef3fdd366c3685c045c22f9c0a9dc5c70c376e41d4d45",
+    "patches/llama-vision-mtp-state-guard.patch": "3c24df4e4d234824603254ad8a0ceb7ba95ea6d2bae6e3a973aaa24c947a46c5",
+    "patches/llama-winnow-server.patch": "608c8642d164a5676cb817ae5d7e851148dd6bd008459cc2b3080a1652b39423",
+    "ruff.toml": "21ec6966cdc163503d041d2e0b844799f43b1cd9dbe915fe82263e90219536d5",
+    "runtime.lock.json": "9b20caf5c6032f6274350c8d20b1d90e5cbc5b7dfa468ca85c77e198e24845df",
+    "scripts/adaptive_policy.py": "43c8652373e949778ab9e099b6a52c338f94d71724ba10e70f6729b1f9f858c8",
+    "scripts/assets.py": "23eec166f908fe411835bb6acd130da16c57061a60c1667ebd39643526b4f5b0",
+    "scripts/bench.py": "204822b2c58370791dcf2acf090b9e35c13e32e6b0b3061d909cedc4adb42145",
+    "scripts/build.py": "045df273eb93bc918feeedf37286c2eecaf6a0cebc93c6a9d61b1de60db5bc75",
+    "scripts/chat_capacity.py": "8aac9044024a1a2278083c407f57c113ead60d4d4b270c599088b91b55ece4fb",
+    "scripts/check.py": "af58b3082893e68052ecac98bb3bbf4db0a7e676fda57fe11f2ae4767c196562",
+    "scripts/check_http.py": "f3b8a26f4176c50675f73960f24f2f934336e5c568bdbf68a085d842794bff5e",
+    "scripts/check_package.py": "329e6d3c9e2e64052985d87c7ff2c40c7bb9e00f71a1d8b6e16cae2778fa39ea",
+    "scripts/decision_client.py": "ceaff3cdb7ab505e0427fca4e96daeed041a6110918498f7282ca5ad92bc2c86",
+    "scripts/download.py": "c385532c1b2e5586af0ffcac68ac3f05b0f8d8d0f71105e8c7961dbb569e5e0b",
+    "scripts/http_client.py": "b4d267d4f21a978915c7fc62a33e794387bcd626eafcc6cb6c6415e182e8d08d",
+    "scripts/install_assistants.py": "fa5c75fe8292cfdee5977e67c4c94fade941a0dd729599cae20cb2edc0255256",
+    "scripts/launch_options.py": "77ced9e55dd028065f133e62c06427889da55803880544ab1fdf41a0570c6567",
+    "scripts/lifecycle.py": "701438434c3752fa9fc17df46f74860e04500d947853f3915b9dbecca60bdab9",
+    "scripts/package_local_candidate.py": "5d66f7f7022e54d358846ee855e0c5fc378822cd747d7ff76e93a1df2275221c",
+    "scripts/package_macos.py": "12050d90958e7bdabb0b0b3275cfd719f401726cc51f1ecb55f28e82788ee3c6",
+    "scripts/package_source.py": "f2a207e1771dc2f721b96f774bce0a1b4c89fe1bc4566d40a672f8a7cf38dee2",
+    "scripts/parity.py": "a943df2ca7c27bd85435992dea76d3eb3de76ea59391fc19174d4981a6b33d21",
+    "scripts/profiles.py": "dd4dd47c36bb9f5f5496988ad6192b5306eacf941c73324f18c88b954b319821",
+    "scripts/reasoning_contract.py": "ec71323a25965411f71b23adeb7c278b9942928c4471bc238771e3fb8806442d",
+    "scripts/release_check.py": "22f1c53ca42a53909292bb867a3dc3a3b8e51dfc003db11eca84fba002344ef2",
+    "scripts/serve.py": "67d748639c7691c6a350b44a29dd0ffbf7a783bdca6fadc0d1519e0f53581e73",
+    "scripts/setup.py": "1a963c2118e1ac14fdab1cc94976b11535be034574c7cfbdb160ac721845f84a",
+    "scripts/verify_model.py": "31e8d97669cadedf71c219c1e51ff1a794444c3ad91a85bc50079821ef794fe9",
+    "scripts/winnow.py": "2d4c540800a0bc89c8966db479d7188e0571b64a3a961a2b000852be1c5671d8",
+    "tests/parity-requests.json": "f0316bfba49d8bae27f7ec06cdc5dc165e4c57623363287ddb6b49f42df2e768",
+    "tests/test_adaptive_pipeline.py": "33ae167dc2ef2ed35a89be8b965cd17247d1177a787da3a430fae706e8d5b5da",
+    "tests/test_assistant_install.py": "1ef64a0176bebbf9a29911ec3008106121e064425627cf8c32f291937e211a93",
+    "tests/test_e2b_native_labels.py": "be93515a0235a7c7c0ae03d666e567b130205db100e59f36e709f8e8573ceda2",
+    "tests/test_e2b_release.py": "855e6328d9ac5816bd8e6ace07ecd8fd26e521b33041a85e8e7b4e4cde4f7ea8",
+    "tests/test_image_reasoning.py": "73c3e8466c570babbfa1981705fec6f88b53adaa5613ae7f1fce5c370f661daf",
+    "tests/test_onboarding.py": "915677b83349f2c5f01a63af4bf48b4a6c0d8dced3259029c103df534daa64c1",
+    "tests/test_public_reasoning_parity.py": "835c0662ef551081b1a9a129ab8852686effd0cfbdfe45e9269787edae379047",
+    "tests/test_reasoning_modes.py": "ecd9b40e7d5a209ccafdbc425b98acf8bf73af8fb14a799b70361992eec7e03c",
+    "tests/test_release_tools.py": "0c568317107231ee6ec3f01d8d02ed998333ca2d7c97df2eee79333868f9c0d1",
+    "tests/test_release_usability.py": "6550dee609d693a859639a273b50b9bfff49499f0c5435ddb63163547ab77182",
+    "tests/test_runtime_package.py": "de41b66660bbb2a45a0eab3b9930377fad6c55545b8bbfada266d07441b58de3",
+    "tests/test_runtime_patch_lock.py": "af9d28cd7ff4e62960b431f90875ed59a6ce27f674d1215f5849fb79616f4d35",
+    "tests/test_runtime_presets.py": "e91aee90d2983277d84f4b8c6c67ac524fe02755fd0f83df556d6e95e136bf80",
+    "tests/test_transport_cancellation.py": "1204773b4940ee856b577cdc4f244ebe84c49a30057cadad7251bb2f49081780",
+    "tests/unit.cpp": "1c8d8e3023f9ff7d4c4e9e5f2e200e7410f193f2ece7efdf7518a49201240ec4",
+    "third_party/README.md": "52d1407fd8944dd816d0f2ec078242882b3f6958edd1ed609785fa88ba0f6b8c",
+    "third_party/cpp-httplib-LICENSE.txt": "4b45cbe16d7b71b89ae6127e26e0d90a029198ca5e958ad8e3d0b8bbed364d8b",
+    "third_party/gemma-assistants-LICENSE-APACHE-2.0.txt": "cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30",
+    "third_party/gemma-assistants-NOTICE.txt": "f54573f51292300e9aa87df0d2351260e78c2bd220f9249fcc9eded1e448b7df",
+    "third_party/llama.cpp-LICENSE": "94f29bbed6a22c35b992c5c6ebf0e7c92f13b836b90f36f461c9cf2f0f1d010d",
+    "third_party/nlohmann-json-LICENSE.txt": "c0d068392ea65358b798b8c165103560f06e9e3b38c4ab4e2d8810a7b931af86",
+    "third_party/rotate-bits-LICENSE.txt": "ca0b44aec101afb6b46f4c39b23369fa06a64fb9c0add37c2689d702156fee55"
+  }
+}

release-manifest.json

@@ -22,7 +22,10 @@
       "converter_sha256": "e9a1da876330bbce9687541ab31736542a01b4ac43c6686126514a50f122fb7f",
       "merge_receipt_sha256": "36e83f26a69bc797cef332c2b4c93aadb41dcd3860dbac8446410da872ad05a3",
       "gguf_file_type": 32,
-      "tensor_types": {"bf16": 318, "f32": 283},
+      "tensor_types": {
+        "bf16": 318,
+        "f32": 283
+      },
       "tensor_layout_sha256": "05a32f8c9728812a06d6c7075be4e5daaf8b678bc62c1af92136b618987d3227",
       "evaluation_scope": "Not evaluated or calibrated in the reported Q8 tests."
     },
@@ -57,6 +60,7 @@
     "raw_max_probability_gate": 0.99,
     "direct_probability_weight": 0.5,
     "augmented_probability_weight": 0.5,
-    "evidence_scope": "Previously exposed full text panels; current v2 integration prompt mapping differs from historical benchmark serialization."
+    "evidence_scope": "Previously exposed full text panels; current v2 integration prompt mapping differs from historical benchmark serialization.",
+    "draft_cache_note": "The draft_kv value records retained launch flags. The pinned Gemma 4 assistant shares target K/V tensors and uses F16 attention cache in this evaluated profile; it has no separately measured Q8 attention-cache pool."
   }
 }