Q3 exact-canvas gate report
Post-hoc data notice (2026-08-27): the upstream LoRA was trained with held-out-slide leakage. Canvas mechanics remain engineering evidence, but no biological/category conclusion or output is submission-eligible. See
reports/SLIDE_SPLIT_AUDIT_REPORT.md.
Outcome
FAIL for candidate production. Run 20260826T134924Z-canvas-gate completed successfully and passed every preregistered mechanical, seam, and privacy check. Full-resolution engineering review nevertheless rejected the four images. The two positive-category fields contain procedurally rendered H. pylori as isolated one- to two-pixel dark hairlines rather than convincing basophilic curved bacilli, and all four fields remain too soft at the challenge canvas scale. The normal field also has substantial vacuolated/goblet-like morphology, so category separation is not reliable enough to justify a candidate pool.
No Q3 v1 image is submission-eligible.
Frozen inputs
- Deployed code:
218a3051b9be98aa4d4557648c0e703bb3734e37. - Generator:
Qwen/Qwen-Image-2512only, with DiffSynth-Studio3f3df523ed6273ee4c9dae245de2c7abd8fd4b5f. - LoRA: caption-neutral epoch 3, SHA-256
ab11358b0b54a7a2c461f93fdc2a2fcccc15da4944dfbe5d6c6340b3d4df6d24. - Geometry: 1024x512 global scaffold, eight overlapping 1280x1152 detail tiles, 128-pixel cosine feather, final 4096x2048 RGB canvas.
- Sampling: 35 steps, denoising strength 0.62, one fixed seed per category.
- Microdetail: deterministic
q3-surface-lumen-v1, 28 organisms for H. pylori-gastritis, 18 for mixed, and zero for both controls.
Integrity and mechanical results
- Worker interval:
2026-08-26T13:49:24Zto2026-08-26T14:45:13Z; exit code 0 andQWEN_CANVAS_GATE_COMPLETEpresent. - All 71 files listed in
ALL_SHA256SUMSpassed independent local SHA-256 verification. - Verification report SHA-256:
b2655196a12068ed2b6eb2c097667888512e3a149e97507e4e16c7bb48be3d9f. - Exactly four candidates are decodable 4096x2048 RGB PNGs.
- Maximum seam ratios were 1.244324 normal, 1.160639 H. pylori-gastritis, 1.187302 intestinal metaplasia, and 1.303911 mixed, all below the frozen 1.75 limit; uncovered pixels were zero.
- Recorded microdetail counts were exactly 28 and 18 in the two positive fields, with declared lengths 10–20 pixels and zero organisms in the controls.
Candidate SHA-256 values:
| Category | SHA-256 |
|---|---|
| normal | ea15d72dbecaccaa774d52f3db8dd4c819bd849b37269e4c8adb21a84eed475a |
| H. pylori-gastritis | a946ec874e3963df6d9799ca50e48df4407a533037793054339ff6dae7f4d927 |
| intestinal metaplasia | 4cd6160442dd8007ada5fd31335151936bfd5bcda1f316702476f89fc9131611 |
| mixed | 72a44a1ab2bd8cca92430cb125f57f542ced43391f52499871eeb345ff62fde4 |
Privacy result
Privacy triage passed across 32 non-overlapping 1024-pixel candidate tiles against all 128 caption-neutral training fields:
- zero exact matches;
- maximum DINOv2 cosine 0.953180, below the 0.995 failure threshold;
- minimum dHash distance 19, above the distance-at-most-4 failure threshold.
Privacy report SHA-256: 01386726f59905e96fedb81c86b96cd6263704381967073e3c8f71369e391420.
Full-resolution engineering review
- Normal: continuous gland-like tissue and no seam failure, but extensive clear/vacuolated cytoplasm resembles goblet-cell or metaplastic morphology. It is not a reliable normal control.
- H. pylori-gastritis: inflammatory tissue signal is present, but the added organisms resolve as detached dark hairlines against pale lumen or blurred tissue. Their measured scale alone does not make their appearance clinically credible.
- Intestinal metaplasia: the strongest category signal of the four, with multiple clear-cell/goblet-like structures, but the field is still visibly soft and several cellular boundaries are generative approximations.
- Mixed: metaplastic/vacuolated tissue and some inflammation are present, but large empty regions and soft cellular detail weaken realism. The added organisms repeat the hairline failure.
- The seam metric correctly rules out obvious tile boundaries but does not measure cellular sharpness, category fidelity, or organism appearance. It therefore cannot override the visual gate.
Runtime, cost, and cleanup
- Q3 worker runtime was 3,349 seconds. A100 instance uptime including startup and automatic shutdown was 3,873.580 seconds.
- At the recorded conservative on-demand list rate of US$3.673385/hour, A100 compute is at most US$3.95 before credits and tax; the instance used Spot capacity.
- The bounded CPU recovery worker ran 1,045.666 seconds, about US$0.04 at US$0.134/hour. Compute plus temporary recovery storage is conservatively below US$4.15.
- The A100 and recovery CPU are terminated. The exact temporary recovery instance, disk, and snapshot were deleted after local verification; the original stopped Qwen disk is retained.
Scientific conclusion
The exact-canvas and Qwen image-conditioned tiling implementation are operational, reproducible, seam-safe, and privacy-safe. The experiment falsifies the claim that those properties are sufficient for submission quality. Q3 v1 fails on two independent clinical-appearance axes: high-frequency tissue fidelity and H. pylori rendering.
The next experiment must remain bounded and causal. First replace only the failed organism renderer with an antialiased, physically scaled, stain-matched surface cluster and test it CPU-only on the frozen no-organism canvases. Separately preregister a one-tile Qwen detail ablation that measures whether denoising strength can materially improve cellular sharpness without changing the accepted checkpoint, prompt, seed, or geometry. Do not generate a candidate pool unless both corrections pass full-resolution engineering review; reopen the backbone comparison if Qwen cannot clear the tissue-detail gate.