# Q3 corrected normal specialist

## Result

**FAIL for candidate production.** Run `20260826T212418Z-lora-train` completed successfully on deployed commit `03b77c8455f7300c7e6e18b03c1e8b6720d60f82`, but no whole checkpoint met the preregistered normal-morphology gate. The run is retained as valid negative evidence; the corrected intestinal-metaplasia specialist must not start from this result.

## Integrity and split

- Exit code is 0 and `QWEN_LORA_COMPLETE` is present. Every entry in `ALL_SHA256SUMS` passes locally.
- The immutable corpus gate records 32 exact 1024×1024 RGB fields from nine non-held-out slides at 0.25 µm/pixel, with zero records from the three exact M1 held-out UUIDs.
- Corpus metadata SHA-256 is `024754db101c6371b6e67f4f1247df921ab63b6ba618afcb2aa5c07578a2bf9a`; provenance SHA-256 is `4d66a9d56dd9c9beef75fedebd3f69c7b04b8c4411a90f546048f585efab2831`; held-out manifest SHA-256 is `18377920a2dcca040e05ce875e4c7ecfe7a3f09d512539bf69d2eb681a26d0af`.
- Four whole rank-32 checkpoints and four fixed-seed 1024×1024 RGB validation images were recovered. Checkpoint hashes are `75eba9e8da3ed393a6146c6744c749b85d0d30b57ace5709da80ab8d98626183`, `a28a99120c61a14c69154b571ca59df5e402388263310e9637494023f86a8b0a`, `191097e97019394b367d2addfa7b7a8714f7f043c2b26d7948f5ff61cc3ef987`, and `d0770dcf2c5ed7bc2dd96b75758249794b8af38fda663bde0f7641b87c353d9f`.

## Privacy

Both preregistered screens pass. Against the corrected 32-field normal corpus, maximum DINO cosine is 0.818119 and minimum dHash distance is 23. Against the full corrected 128-field caption-neutral corpus, maximum DINO cosine is 0.838639 and minimum dHash distance is 20. Neither screen found an exact pixel match. These are privacy/nearest-neighbour checks, not clinical-quality evidence.

## Whole-checkpoint engineering review

| Epoch | Mean absolute gradient | Sydney-QA glands/mm² | Sydney-QA mucin/vacuole candidates/mm² | Full-resolution result |
| --- | ---: | ---: | ---: | --- |
| 0 | 4.518327 | 0.0 | 0.0 | Diffuse pale texture; no plausible gastric glands; reject |
| 1 | 7.352303 | 184.5 | 5,182.1 | Sharper but dominated by diffuse fibrous/vacuolated texture; reject |
| 2 | 6.037955 | 30.6 | 704.3 | Diffuse wavy tissue with almost no coherent gland architecture; reject |
| 3 | 5.925170 | 229.6 | 4,362.3 | Most organised checkpoint, but repetitive elongated columns and excessive clear/vacuolated morphology remain unlike the untouched real normal field; reject |

The untouched M1 real normal reference measures 7.682061 mean absolute gradient, 340.9 gland/lumen candidates/mm², 2,258.0 mucin/vacuole candidates/mm², and a lumen-to-tissue ratio of 0.0810 under the same classical extractor. These single-field comparisons are descriptive only. The untrained Q0 control is also visibly non-histologic; its high 18.323101 gradient demonstrates why sharpness alone cannot qualify an image.

Sydney-QA extracted interpretable features and audit masks for all four checkpoints, Q0, and the real reference. Every synthetic and real record remains `CALIBRATION_REQUIRED`: no GI-pathologist-annotated, patient/slide-disjoint reference envelope exists. Sydney-QA does not establish clinical correctness, diagnosis, or OLGA/OLGIM and does not replace pathologist review.

## Runtime, cost, and cleanup

The model run lasted 966 seconds. Measured A100 instance uptime was 1,420.175 seconds, approximately US$1.45 at the conservative US$3.673385/hour list-price assumption. Recovery CPU uptime was 2,051.721 seconds, approximately US$0.08 at US$0.134/hour. Total conservative compute is below US$1.53 before temporary storage, credits, and tax; the actual A100 was Spot.

The A100 and recovery CPU are terminated. After local checksum verification, the temporary recovery instance, disk, and snapshot were deleted; the original stopped Qwen disk is retained.

## Decision

No epoch advances, and the corrected intestinal-metaplasia specialist is not run. A normal-only adapter did not solve the generator's tissue-topology failure, so adding the paired category would spend GPU budget without satisfying the decisive first half of the preregistered gate. Reopen the pathology-native backbone or pathology-super-resolution route under compatible submission rights, while preserving Qwen-Image-2512 as the measured baseline.