# Q2 tissue-prompt ablation gate

> **Post-hoc invalidation (2026-08-27):** the upstream LoRA was trained with held-out-slide leakage. This remains an exploratory prompt-control record only; it cannot support generalisation or checkpoint selection. See `reports/SLIDE_SPLIT_AUDIT_REPORT.md`.

Date: 2026-08-26 UTC

Run: `20260826T095638Z-prompt-ablation`

Code: `2ed42eb97905706ce7e7b5a97482d85b225c3762`

Result: **FAIL for tissue-scaffold selection — the curved-object failure is suppressed, but positive-category morphology collapses toward nonspecific tissue**

## Frozen comparison

This ablation loaded the no-overlay epoch-3 checkpoint with SHA-256 `5fb8d73646228a2eb047593113e6c64490c2fe2c6c9c71b803c04b64718e9de7`. Width, height, steps, guidance, model, checkpoint, seeds, and normal/metaplasia positive prompts remained fixed. The two H. pylori-positive prompts were changed to request lesion tissue without visible organisms, and the negative prompt was strengthened against worm-like or thick curved objects. Prompt-override SHA-256 was `091d247a1badd4e4e72b0535609bb65eb6322f824ebf4eaa320d9947c99b86a7`.

## Mechanical, runtime, and privacy results

- The run completed from `09:56:38Z` through `10:05:31Z`, or 533 seconds, with `QWEN_PROMPT_ABLATION_COMPLETE` and exit code 0.
- All fourteen files in `ALL_SHA256SUMS` matched after retrieval. All four outputs decode as 1024x1024 RGB PNGs, and every generation-manifest record points to the frozen LoRA hash.
- A100 uptime including startup and automatic shutdown was 735.278 seconds, at most about US$0.75 using the conservative US$3.673385/hour on-demand rate; the instance was Spot. The bounded CPU recovery worker ran for 543.165 seconds, about US$0.02 at US$0.134/hour. Both workers are terminated and all temporary recovery resources were deleted.
- Privacy screening against the 128 no-overlay training fields passed: zero exact matches, maximum DINO cosine 0.951791 against the 0.995 threshold, and minimum dHash distance 17 against the distance-at-most-4 threshold. Privacy report SHA-256 is `a01b1afe62742a163fdd5483fc671dcbdfb05891d6b8a238635aed93bd5f9505`.

## Morphology finding

The held-out/Q0/Q2-v1/no-overlay/prompt-ablation sheet SHA-256 is `fe84106049f0e4cacffdaf68d038486c0384cb7e8079a9d981266359cfca45b3`.

- Removing positive organism wording suppresses the gross dark curved rods in both positive cells. This confirms that inference wording is a material cause of the shortcut.
- The H. pylori-gastritis result becomes pale, diffuse, and poorly glandular, without convincing chronic-active inflammatory architecture.
- The mixed result loses convincing goblet-cell/metaplastic and gastritis coexistence. It resembles nonspecific pale tissue rather than a mixed lesion.
- Mean nearest-held-out DINO cosine is 0.919272, but every cell's nearest reference tile is from a normal or gastritis reference rather than preserving its requested category. The high aggregate value therefore reflects tissue texture, not category fidelity.
- No organism-neutral tissue cell is submission-eligible, and the four-cell set fails the whole-checkpoint morphology gate.

Held-out diagnostic SHA-256 is `26fa8a10847aa6f2bed075c9885da791453a70c5c44965fa53c2f71558efbd03`.

## Gate decision

The experiment identifies a two-part failure: explicit organism wording drives oversized curves, while the existing checkpoint does not retain enough positive-category tissue signal when that wording is removed. The next bounded Q2 correction changes only the 64 positive-category training captions to organism-neutral gastritis and mixed-lesion descriptions. All 128 training pixels, slides, coordinates, provenance, split, model, optimiser, rank, learning rate, epochs, validation seeds, and normal/metaplasia captions remain fixed. Q3 remains blocked.