Publish corrected qualification card and evidence

.gitattributes

@@ -36,3 +36,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
 assets/context-frontier-dark.png filter=lfs diff=lfs merge=lfs -text
 assets/context-frontier-light.png filter=lfs diff=lfs merge=lfs -text
+assets/context-per-card-dark.png filter=lfs diff=lfs merge=lfs -text
+assets/context-per-card-light.png filter=lfs diff=lfs merge=lfs -text
+assets/fidelity-vs-size-dark.png filter=lfs diff=lfs merge=lfs -text
+assets/fidelity-vs-size-light.png filter=lfs diff=lfs merge=lfs -text

DOCS-SHA256SUMS

@@ -1,2 +1,15 @@
-34448b82c17d60fec9b65b1f093c115ddbaadc04beb1b0140b6bfed2e012a930  .gitattributes
-8791630cada33f122210d4af4e571fef6a2caf315c8bfdc5e964bacd305759ef  README.md
+34448b82c17d60fec9b65b1f093c115ddbaadc04beb1b0140b6bfed2e012a930 .gitattributes
+264d838d7f2fef99be8e12c1166e80273f531f94419622efd940c438cd30fc5c README.md
+d71eb992953ffd25e07d268a268dbf4953cae40eb14e35ef0d3d7e12f32709c2 assets/context-frontier-dark.png
+df4d8cfdb5ded9e330edbc16223d83ca524f3b67e7bcea910c5a7b4530f11b1b assets/context-frontier-dark.svg
+bafed8e1de217f9311dbf604ab69950adb00e8f3d23a9c2a3ea12184cb47ac8a assets/context-frontier-light.png
+b2b897b7429e517aa0e91ca12a63285672e23eeb8db5ffd4effbb45e50327c62 assets/context-frontier-light.svg
+6339e328f0b865cf03a7d1c608a8e0da555af0b2c30e618733db2d9c8a8ce369 assets/context-per-card-dark.png
+0c34a0701922ddcffdc80f7a3c1105154dd3c9cd773d26fbedc8a9660d14265b assets/context-per-card-dark.svg
+605d513b38a88f16d42c437ebe2c827cb6e3c2973ea7e6c3af6ff1f7df5b3073 assets/context-per-card-light.png
+0df334eec3108065d383674d50789d4d0956e655954b2cf6c55ecdcf46493f65 assets/context-per-card-light.svg
+f8c77a2822223f012b48c6d52aa19caef5b317b515a1ba49ab5d406d963a2ab9 assets/fidelity-vs-size-dark.png
+c01fd0ae8986a68081413ae0d63b4d20853788f07eb0c9816c34655e6cd62b2c assets/fidelity-vs-size-dark.svg
+eeb68bfb3cdeef240bc4a19368cdb7f0cb29493bc3570dfddda519af0a41e1ea assets/fidelity-vs-size-light.png
+61ef12b7b07ae57914ba73199614341fa68fce0149d74851e14b4a878aba03df assets/fidelity-vs-size-light.svg
+0602ce8b2ba9e95b84adc5933ab1138faec605a6c0b12af935bbcd91d94d6aff release-evidence.json

README.md

@@ -35,8 +35,8 @@ every cold start. Measured consequences:
 | first load, cold | 957 s (encodes 208 projections) | **178 s** (5.4x faster) |
 | restart with a warm cache | 173 s | 178 s (same) |
 | persistent encode cache | required, and must be writable | **not used** |
-| mean KL divergence from BF16 (body-only) | 0.008157 | **0.007406** (−9.2 %, but see the floor caveat) |
-| mean KL divergence, as served | 0.008284 | **0.007532** (measured, not estimated) |
+| mean KL divergence from BF16 (body-only) | 0.007945 | **0.007172** (−9.7 %, but see the floor caveat) |
+| mean KL divergence, as served | 0.008078 | **0.007300** (measured, not estimated) |
 | attention width | K6 / K5 / K4, chosen at launch | fixed at K6 |
 
 **Choose this build** for the smaller download, the best fidelity in the family, and a start
@@ -48,24 +48,26 @@ fidelity for KV room, and on a 32 GB card that difference matters (see
 
 ## Which of the four builds
 
-Same architecture and tokenizer; they differ in where the bits go. Contexts are what the engine
-serves on a 32 GB card with MTP-3 and vision enabled, at utilisation 0.97
+Same architecture and tokenizer; the KLD column is the overlap-corrected 127-context v3
+subset. Capacity uses each card's documented profile: hydrated, online and K4 are real RTX
+5090 MTP-3 tests; context is MTP-3 with an 8.4 MP cap on a budget-capped RTX PRO 6000.
+These profiles are not interchangeable
 ([collection](https://huggingface.co/collections/qwen38-27b-mixed-precision-exl3-measured-6a7fe0cb27817c23e4a57025)).
 
 <picture>
   <source media="(prefers-color-scheme: dark)" srcset="assets/context-frontier-dark.svg">
-  <img alt="Mean KL divergence versus context served on a 32 GB card: hydrated 0.007406, online K6 0.008157, context edition 0.009673 with verified needle retrieval, online K5 0.012135, K4 0.030736; official FP8 0.013126." src="assets/context-frontier-light.svg">
+  <img alt="Overlap-corrected v3 mean KL divergence versus demonstrated or configured context. Circles are real RTX 5090 MTP-3 results: hydrated and online K6 at 185,600, K4 at 262,144. Stars have generation proof: online K5 at 206,400 on the 5090, and the context edition at 262,144 with MTP-3 and an 8.4 MP image cap under a 30.24 GiB engine budget; the latter's hard-limit 5090 rerun is pending." src="assets/context-frontier-light.svg">
 </picture>
 
-| build | download | resident | mean KLD | context on 32 GB | pick it when |
+| build | download | resident | corrected v3 mean KLD | context profile | pick it when |
 |---|---:|---:|---:|---:|---|
-| [-hydrated](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated) | 21.61 GB | 20.31 GiB | **0.007406** | ~180k | fidelity first, smallest download |
-| [-EXL3-K5K6](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6) | 30.57 GB | 20.32 GiB | 0.008157 | ~180k | you want the attention width knob at launch |
-| [-context](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-context) | 20.70 GB | **18.13 GiB** | 0.009738 | **262,144 native** | you need native context on 32 GB — int8 embeddings, retrieval verified at 227,334 tokens |
-| [-K4](https://huggingface.co/malaiwah/Qwen3.8-27B-K4) | 28.31 GB | 17.89 GiB | 0.030736 | 262,144 | smallest footprint, native context without any overlay |
+| [-hydrated](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated) | 21.61 GB | 20.31 GiB | **0.007172** | ~180k | fidelity first, smallest download |
+| [-EXL3-K5K6](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6) | 30.57 GB | 20.32 GiB | 0.007945 | ~180k | you want the attention width knob at launch |
+| [-context](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-context) | 20.70 GB | **18.41 GiB** | 0.009459 | **262,144, MTP-3, 8.4 MP cap** | native window; hard-limit RTX 5090 check pending |
+| [-K4](https://huggingface.co/malaiwah/Qwen3.8-27B-K4) | 28.31 GB | 17.89 GiB | 0.029679 | 262,144 | smallest footprint, native context without any overlay |
 
-Official `Qwen/Qwen3.8-27B-FP8` is 28.51 GiB resident at 0.013126 on the same suite and runs on
-stock vLLM, which none of these do.
+Official `Qwen/Qwen3.8-27B-FP8` is 28.51 GiB resident at 0.012798 on the same subset and runs
+on stock vLLM, which none of these do.
 
 ## Recipe
 
@@ -115,6 +117,11 @@ operands, source-cluster bootstrap. Same suite, reference and head as every comp
 | `Qwen/Qwen3.8-27B-FP8` | 28.51 GiB | 0.013126 | [0.00981, 0.01709] | 0.002343 | 96.22 % |
 | `unsloth/Qwen3.8-27B-NVFP4` | 21.34 GiB | 0.094978 | [0.06858, 0.12688] | 0.012911 | 90.53 % |
 
+**Overlap-corrected subset:** a later all-position 12-token scan found exact calibration
+overlap in 2/41 source documents that the original fixed-stride scan missed. Conservatively
+removing their nine contexts gives this build **0.007172**, the online-K6 sibling **0.007945**,
+official FP8 **0.012798**, and NVFP4 **0.092727** over 127 contexts. No ordering changes.
+
 Paired on identical contexts:
 
 - versus the BF16-attention sibling: **−0.000751**, 95 % CI [−0.000977, −0.000572],
@@ -131,7 +138,10 @@ head, so no candidate's own head quantization is counted — that is what makes
 since official FP8 serves a BF16 head. Measured directly on **this** checkpoint with asymmetric
 heads (reference through the true BF16 head, candidate through this build's dequantized K6
 head): the head costs **+0.000125** (95 % CI [+0.000107, +0.000144], 9/136 contexts favour it),
-so **as served this build is 0.007532** with 97.08 % top-1. Still 1.74x better than FP8's
+so **as served this build is 0.007532** with 97.08 % top-1. On the overlap-corrected
+127-context subset, the measured result is **0.007300** with 97.05 % top-1; the head increment
+is +0.000128 (95 % CI [+0.000110, +0.000146], 8/127 contexts favour it). The original
+as-served result is still 1.74x better than FP8's
 body-only 0.013126.
 
 **Weakest control:** live-versus-replayed logit qualification is 6.54e-04 on this harness,
@@ -140,9 +150,9 @@ so differences below ~1e-3 are not resolvable. The −0.000751 offline-versus-on
 direction, but treat the magnitude as a point estimate. The FP8 gap, at 7.6x the floor, is
 not in question.
 
-**Development-set caveat:** the recipe was chosen with these 136 contexts visible, and the
-suite's source clusters are shared with the partition that guided selection. A
-source-disjoint qualification suite is being built to test the frozen recipe once.
+**Development-set caveat:** the recipe was chosen with the 136-context development suite
+visible. The source-disjoint qualification below is the post-selection test; its later lexical
+overlap correction is reported rather than hidden.
 
 ## Post-selection qualification
 
@@ -151,20 +161,33 @@ not: **160 new contexts from 100 documents with zero intersection with the devel
 (context token hashes 0/160, document names 0/100, content hashes 0/100), partitioned by whole
 source cluster, run **once**, with no recipe changed afterwards.
 
+The original 42-context table used a fixed-stride character overlap scan. A later,
+offset-independent scan found exact 12-token calibration overlap in four qualification source
+documents. Applying the same conservative rule to every candidate — exclude every context from
+any source document with even one hit — leaves **36 contexts / 24 clusters**:
+
 | candidate | mean KLD | 95 % CI | top-1 | paired vs FP8 |
 |---|---:|---|---:|---|
-| [hydrated](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated) | **0.003029** | [0.002572, 0.003536] | 97.68 % | −0.002691, **42/42** |
-| [K5/K6 online K6](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6) | 0.003395 | [0.002925, 0.003910] | 97.58 % | −0.002325, **42/42** |
-| [context edition](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-context) | 0.003900 | [0.003275, 0.004588] | 97.43 % | −0.001820, **42/42** |
-| `Qwen/Qwen3.8-27B-FP8` | 0.005720 | [0.004849, 0.006680] | 96.82 % | — |
-
-The ranking is preserved and the advantage over FP8 is **larger** on unseen sources (47 / 41 /
-32 % lower) than on the development suite (44 / 38 / 26 %), which is the opposite of what
-selection bias would produce.
-
-**Absolute magnitudes are suite-specific:** every candidate, FP8 included, measures ~2.4x lower
-on this corpus than on the development one. Quote the suite with the number.
-Details in [docs/31](https://github.com/malaiwah/qwen38-27b-exl3/blob/main/docs/31-frozen-qualification.md).
+| [hydrated](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated) | **0.003093** | [0.002577, 0.003684] | 97.63 % | −0.002798, **36/36** |
+| [K5/K6 online K6](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6) | 0.003455 | [0.002916, 0.004060] | 97.50 % | −0.002436, **36/36** |
+| [context edition](https://huggingface.co/malaiwah/Qwen3.8-27B-EXL3-K5K6-context) | 0.003990 | [0.003268, 0.004797] | 97.36 % | −0.001901, **36/36** |
+| `Qwen/Qwen3.8-27B-FP8` | 0.005891 | [0.004901, 0.006985] | 96.72 % | — |
+
+The correction changes no ordering or paired win: the three EXL3 builds remain 47 / 41 / 32 %
+below FP8. The original 42-context figures and the candidate-independent correction are both
+preserved in [docs/31](https://github.com/malaiwah/qwen38-27b-exl3/blob/main/docs/31-frozen-qualification.md).
+Absolute magnitudes remain suite-specific.
+
+## Downstream task retention
+
+On 40 deterministic generated tasks (10 each arithmetic, executable builtins-only code,
+exact-list instruction following and tool-call schema), BF16 and every comparator scored
+40/40. This build had **zero regressions** and matched BF16's exact final-answer text on
+**35/40**; all five differing answers still passed their contracts. Wilson 95 % lower bound
+is 91.2 %. This is a transparent smoke suite, not a public leaderboard; full responses are in
+the [run receipt](https://github.com/malaiwah/qwen38-27b-exl3/blob/main/receipts/tasks-v2-hydrated.json);
+its extracted-value agreement field is superseded by the
+[strict rescore](https://github.com/malaiwah/qwen38-27b-exl3/blob/main/receipts/task-retention-v2-strict-rescore.json).
 
 ## Context capacity
 
@@ -178,13 +201,15 @@ graphs, vision enabled):
 | K6 attention, seqs 8, util 0.95 | 187,050 tok / 6.71 GiB | **185,600 run** | stable, multimodal-safe |
 | K6 attention, seqs 8, util 0.98 | 202,185 tok / 7.55 GiB | — | text fine, **a 3,264-token image OOMed** with 33 MiB free |
 
-**Native 262,144 does not fit any 20 GiB-class build on a 32 GB card.** With MTP-3 the engine
-needs **9.13 GiB** of KV for native context (37.4 KB/token: 16 full-attention layers, 4 KV
-heads, head_dim 256; the other 48 layers are Gated DeltaNet and hold per-sequence state).
-Without MTP that falls to 33.5 KB/token and buys ~11 % more length. For native context on
-32 GB, the smaller [`malaiwah/Qwen3.8-27B-K4`](https://huggingface.co/malaiwah/Qwen3.8-27B-K4)
-build is the only one in this family that fits (289,577-token capacity measured), at 3.8x the
-divergence. On 48 GB and larger, native context fits here at the best fidelity.
+**Native 262,144 does not fit this checkpoint with MTP-3 on a 32 GB card.** The engine needs
+**9.13 GiB** of KV (37.4 KB/token: 16 full-attention layers, 4 KV heads, head_dim 256; the
+other 48 layers are Gated DeltaNet and hold per-sequence state). Without MTP that falls to
+33.5 KB/token and buys ~11 % more length. The smaller
+[`malaiwah/Qwen3.8-27B-K4`](https://huggingface.co/malaiwah/Qwen3.8-27B-K4) is the only family
+member hardware-qualified at native length on a real 32 GB card (289,577-token capacity).
+The context edition starts natively with MTP-3 and an 8.4 MP image ceiling under a capped
+32 GB-equivalent engine budget, but its physical RTX 5090 rerun remains pending. On 48 GB and
+larger, native context fits here at the best fidelity.
 
 Keep utilisation at **0.95** if you serve images: 0.98 leaves no vision headroom.
 
@@ -196,18 +221,22 @@ overstated usable VRAM, which produced an overclaim that the tester correctly re
 ## Serving
 
 ```bash
-docker run --rm --gpus '"device=0"' --ipc host -p 8000:8000 \
+docker run --rm --gpus '"device=0"' --ipc host -p 127.0.0.1:8000:8000 \
   -v /models:/models:ro \
   --entrypoint /opt/venv/bin/vllm \
   voipmonitor/vllm@sha256:820181fbbc975cd5291c411cda9771d58fecee1636d916f508f47230df20592b \
   serve /models/Qwen3.8-27B-EXL3-K5K6-hydrated \
     --served-model-name qwen38 --quantization exl3 --enforce-eager \
-    --quantization-config '{"linear":{"weight":"mxfp8"},"ignore":["re:.*visual\\..*","re:.*in_proj_a$","re:.*in_proj_b$","re:.*in_proj_ba$","lm_head"]}' \
+    --quantization-config '{"linear":{"weight":"mxfp8"},"ignore":["re:.*visual\\..*","re:.*in_proj_a$","re:.*in_proj_b$","re:.*in_proj_ba$","re:.*mtp\\..*","lm_head"]}' \
     --mm-processor-kwargs '{"truncation":false}' \
     --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder \
     --max-model-len 8192 --gpu-memory-utilization 0.95 --max-num-seqs 8 \
     --host 0.0.0.0 --port 8000
 ```
+The container listens on all interfaces internally, but Docker publishes the port to host
+loopback only. For remote clients, keep that binding and put an authenticated TLS proxy in
+front; do not expose this unauthenticated generation endpoint directly.
+
 
 No `VLLM_EXL3_ONLINE_TRELLIS_BITS`, no `VLLM_EXL3_ONLINE_CACHE_DIR`: nothing is encoded at
 load, which is the point of this build. Three notes carry over from the sibling:
@@ -217,32 +246,33 @@ load, which is the point of this build. Three notes carry over from the sibling:
    `model.`, so `re:.*visual\..*` matches while `re:.*\.visual\..*` silently does not, and the
    wrong pattern **crashes** startup
    ([#311](https://github.com/local-inference-lab/vllm/issues/311), fixed by
-   [PR #312](https://github.com/local-inference-lab/vllm/pull/312)). `mtp.*` must **not** be
-   ignored: the draft head is quantized and belongs to the EXL3 loader.
+   [PR #312](https://github.com/local-inference-lab/vllm/pull/312)). The tested config ignores
+   `mtp.*` from the generic BF16 online overlay; EXL3-owned draft projections are selected
+   before that ignore check and still load from their serialized quantized tensors.
 3. **`--mm-processor-kwargs '{"truncation":false}'` is required for images** whose expanded
    token sequence exceeds 2,048, or requests fail with HTTP 400
    ([#313](https://github.com/local-inference-lab/vllm/issues/313)).
 
 The command above is what the **pinned image runs unmodified**, so it is eager-only. CUDA-graph
-decode (+46-50 %) and the prefill dispatch (+113 %) need patches still open upstream
-([#314](https://github.com/local-inference-lab/vllm/pull/314),
-[#316](https://github.com/local-inference-lab/vllm/pull/316)); the sibling card carries the
-exact patch recipe with its sha256. There is no published image digest containing them.
+decode (+46-50 %) and the measured reconstructed-prefill path (+113 %) need patches still
+open upstream ([#314](https://github.com/local-inference-lab/vllm/pull/314),
+[#316](https://github.com/local-inference-lab/vllm/pull/316), and
+[#318](https://github.com/local-inference-lab/vllm/pull/318)); the sibling card carries the
+exact one-module patch recipe and current sha256. No published image digest contains it.
 
 ## What is not verified
 
-No downstream task benchmarks, no OCR/chart/document/video evaluation, no long-context
-retrieval or perplexity **for this build**, no native-262K generation, no YaRN-1M run, no
-multi-GPU or TP>1, no non-SM120 GPU, and no quant-specific safety regression testing.
+The deterministic downstream smoke passes 40/40 with zero BF16 regressions, but no public
+task benchmark has run. Also unverified: real OCR/chart/document/video quality, long-context
+retrieval or perplexity **for this build**, native-262K or YaRN-1M generation, multi-GPU or
+TP>1, non-SM120 hardware, and quant-specific safety regression testing.
 Throughput was not re-measured here: this build shares the sibling's kernels and resident
 footprint, so its decode and prefill figures should carry over — that is an inference, not a
 measurement.
 
-Evaluation captures for this build are being published in the
-[fidelity dataset](https://huggingface.co/datasets/malaiwah/qwen38-27b-fidelity-suite-v3);
-until they land, the
-[receipts](https://github.com/malaiwah/qwen38-27b-exl3/tree/main/receipts) let a third party
-check the arithmetic but not recompute the row independently.
+Evaluation captures and reports for this build are published in the
+[fidelity dataset](https://huggingface.co/datasets/malaiwah/qwen38-27b-fidelity-suite-v3),
+so the row is independently recomputable without this checkpoint or a GPU.
 
 ## Machine-readable evidence
 

release-evidence.json

@@ -1,6 +1,7 @@
 {
   "schema": "qwen38-release-evidence/1",
   "generated_utc": "2026-08-15T03:42:51Z",
+  "amended_utc": "2026-08-15T12:27:32Z",
   "artifact": {
     "hf_repo": "malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated",
     "label": "K5/K6, attention serialized offline at K6",
@@ -45,7 +46,7 @@
     "suite_token_sha256": "3f9d17f1b55f64872ad3ac19c8711654e09ba70b7ca14b0851525088fe735691",
     "partition": "analysis",
     "contexts": 136,
-    "scored_positions": null,
+    "scored_positions": 278392,
     "shared_comparator_head": "BF16, identical for every candidate (body-only)",
     "mean_kld": 0.007406422525617622,
     "ci95": [
@@ -83,24 +84,59 @@
         "bf16_control": "24/32, 0.0128"
       }
     },
-    "development_set_caveat": "Recipe selection saw this partition, and its source clusters are shared with the qualification partition, so these are development-set numbers. A source-disjoint suite is being built.",
-    "captures_published": false,
+    "development_set_caveat": "Recipe selection saw v3; source-disjoint v4 qualification is published. The conservative overlap-corrected v3 subset and v4 results are linked by amendments below.",
+    "captures_published": true,
     "as_served": {
-      "mean_kld": 0.007532,
-      "top1_agreement": 0.9708,
-      "head_cost": 0.00012516804670361007,
-      "head_cost_ci95": [
-        0.00010680418990943814,
-        0.0001438008910763626
-      ],
-      "method": "asymmetric heads: reference through the true BF16 head, candidate through this checkpoint's dequantized K6 head; measured, not inferred from the sibling"
+      "original": {
+        "contexts": 136,
+        "scored_positions": 278392,
+        "mean_kld": 0.007531590572321231,
+        "top1_agreement": 0.9707822063852409,
+        "head_delta_mean_kld": 0.00012516804670361007,
+        "head_delta_ci95": [
+          0.00010680418990943814,
+          0.0001438008910763626
+        ]
+      },
+      "overlap_corrected": {
+        "contexts": 127,
+        "scored_positions": 259969,
+        "mean_kld": 0.007300129113882429,
+        "top1_agreement": 0.9704657093730407,
+        "head_delta_mean_kld": 0.00012809525810428388,
+        "head_delta_ci95": [
+          0.00010955942052458132,
+          0.0001463054626847932
+        ],
+        "contexts_favouring_quantized_head": 8
+      },
+      "method": "asymmetric heads: BF16 reference head versus this checkpoint's dequantized K6 head"
     }
   },
   "serving": {
     "note": "No online encoding: no VLLM_EXL3_ONLINE_TRELLIS_BITS, no cache directory. Eager on the pinned image."
   },
+  "amendments": {
+    "fidelity_dataset_revision": "73252e77e96bbf5596484e29dd8041f9f38c95f1",
+    "overlap_corrected_v3": {
+      "path": "receipts/analysis-v3-contamination-corrected.json",
+      "sha256": "026fc453787eccef15879f29266489ba9ec04bbff59ab7e35827e92394480417"
+    },
+    "source_disjoint_v4": {
+      "path": "receipts/qualification-v4-contamination-corrected.json",
+      "sha256": "17148ed02713f426dcb4b7b69c8072a5eb8a9abd49bcbfcea585ca8e5c2f5ab3"
+    },
+    "task_run": {
+      "path": "receipts/tasks-v2-hydrated.json",
+      "sha256": "704b9c51f5e902a680f087e2eed00d51c6166d35461e667f0d6d7c5ce6d3aec4"
+    },
+    "task_strict_rescore": {
+      "path": "receipts/task-retention-v2-strict-rescore.json",
+      "sha256": "b1cddff1a9ae012a48ec0258096c3a258e652aef887fc9128d8665e62c867080"
+    }
+  },
   "not_verified": [
-    "downstream task retention (code, reasoning, instruction following, tool calling)",
+    "public downstream benchmark beyond the deterministic paired smoke",
     "multimodal quality beyond single image smoke tests",
     "long-context accuracy beyond an external 205,021-token retrieval check",
     "native 262,144 generation on this box",