# Q3 corrected exact-canvas gate report

> **Post-hoc data notice (2026-08-27):** the upstream LoRA was trained with held-out-slide leakage. This run was already rejected; its mechanics remain historical evidence only. See `reports/SLIDE_SPLIT_AUDIT_REPORT.md`.

## Outcome

**FAIL; interrupted and mechanically ineligible.** Spot A100 run `20260826T165518Z-canvas-gate` used the preregistered corrected Q3 configuration on deployed commit `14e22d7b72593afda2b5e40be4561bfe2a760987`. It completed normal, H. pylori-gastritis, and intestinal-metaplasia fields, then GCP preempted the worker while it was generating the eighth mixed-category detail tile. The run never wrote `QWEN_CANVAS_GATE_COMPLETE`, `ALL_SHA256SUMS`, `canvas_gate_verification.json`, or `privacy_screen.json`.

The shell EXIT trap recorded exit code 0 during termination. That value is not a success result: the missing terminal marker and GCP operation `systemevent-1787766526972-659f6d4c81380-877dc487-b87d418a` establish an interrupted run. The worker is corrected so TERM, HUP, and INT now produce nonzero codes, `RUN_FAILED`, and an explicit signal record.

No output from this attempt is submission-eligible.

## Frozen inputs

- Generator: Qwen/Qwen-Image-2512 with caption-neutral epoch-3 LoRA SHA-256 `ab11358b0b54a7a2c461f93fdc2a2fcccc15da4944dfbe5d6c6340b3d4df6d24`.
- Gate configuration SHA-256: `890744c320718f8b6c3a12caf23035255f43210dd4902f3e8a9a182dd2570d13`.
- Prompt override SHA-256: `4cb227a987db2790f563f5cf9ac5c549c3997238918757b741d02c89640e0a98`.
- Correction relative to Q3 v1: denoising strength 0.95 for every detail tile and antialiased `q3-surface-lumen-antialiased-v1` microdetail for the two positive categories. All other checkpoint, prompt, seed, sampling, geometry, and assembly variables remained frozen.

## Recovered evidence

- Worker interval: `2026-08-26T16:55:18Z` to termination handling at `2026-08-26T17:48:34Z`, 3,196 seconds.
- GCP records `compute.instances.preempted` at `2026-08-26T17:48:46.972Z`; the instance reached TERMINATED at `2026-08-26T17:49:02.245Z`.
- Three categories contain `CANVAS_COMPLETE`, generation manifests, all eight detail tiles, exact 4096x2048 RGB canvases, and matching final hashes. Mixed contains only seven of eight detail tiles and no candidate or completion marker.
- The locally generated recovery manifest contains 64 recovered source files and has SHA-256 `503eef671e5106322bb148218409783f7b279eb78a9f00d94043f9d87d2cc941`. It is explicitly labelled local recovery evidence, not a remote completion manifest.

| Completed category | Candidate SHA-256 | Seam ratio | No-organism mean absolute gradient | Frozen floor |
| --- | --- | ---: | ---: | ---: |
| normal | `47627c87ddd82a50db2f4379485ed17a9bb6f6b1d074061e7347710e2dad7c5c` | 1.105794 | 2.982779 | 3.0 |
| H. pylori-gastritis | `77d4e0326b23a6e29caaf711836e08300b8e8242ff49669aacdc456ebecf66fa` | 1.124654 | 3.018435 | 3.0 |
| intestinal metaplasia | `d2677a3e8df206c06a8e5327df21b954bc230816d171869c3a0fc9c96eb99b97` | 1.094998 | 3.065252 | 3.0 |

Even ignoring the interruption, the normal field misses the preregistered sharpness floor by 0.017221, or 0.57%. The whole four-field gate therefore cannot pass. Privacy was not run because the candidate set is incomplete.

## Full-resolution engineering review

- The 0.95 detail setting is materially sharper than Q3 v1 and produces coherent continuous glandular tissue without visible tile seams.
- The normal field still contains extensive clear/vacuolated cells and is not a reliable normal control.
- The H. pylori-gastritis and intestinal-metaplasia fields remain too visually similar to the normal field. This confirms that the shared four-category LoRA does not provide reliable category separation at final scale.
- Antialiased positive-field organisms are less hairline-like than Q3 v1 but remain an engineering rendering pending clinical review.

## Runtime, cost, and cleanup

- A100 uptime was 3,420.715 seconds. At the conservative on-demand list rate of US$3.673385/hour, compute is at most US$3.49 before credits and tax; the interrupted instance used Spot capacity.
- The bounded recovery CPU ran 748.769 seconds, approximately US$0.03 at US$0.134/hour.
- The A100 and recovery CPU are terminated. The exact temporary recovery instance, disk, and snapshot were deleted after local retrieval; the original stopped Qwen disk remains.

## Scientific conclusion

The corrected tile denoising materially improves texture, but the experiment fails both the immutable whole-run contract and category fidelity. The next experiment must not tune the observed sharpness threshold or another nearby denoising value. Reopen the model path as preregistered: compare an eligible category-specialist Qwen LoRA strategy against the shared checkpoint, while excluding pathology-native challengers whose non-commercial/no-derivatives terms are incompatible with direct company submission and downstream challenge rights.