# Slide-split audit report

Date: 2026-08-27

## Outcome

**FAIL — the Q2 HiESD corpus was not slide-disjoint from the four M1 held-out reference fields.** The three `excluded_slides` values committed in the Q2 configurations were malformed 32-character identifiers rather than the exact hyphenated HiESD UUIDs recorded in `manifests/m1_source_fields.csv`. Both corpus construction and verification compared those strings literally, so the intended exclusions did not match any source slide and the gate passed incorrectly.

The exact held-out manifest SHA-256 is `18377920a2dcca040e05ce875e4c7ecfe7a3f09d512539bf69d2eb681a26d0af`. Its three unique slide UUIDs are:

- `0d55cc29-bb23-4d0d-8b79-93f70be520b8`
- `0dddc3e2-b189-43fb-81c1-d0cac38f3cbd`
- `1a203f16-4d3c-440a-a6f1-82619763a54c`

## Leakage count

The 128-image no-overlay/caption-neutral corpus contains 40 fields from held-out slides:

| Category | Leaked fields | Category total |
| --- | ---: | ---: |
| Normal | 8 | 32 |
| H. pylori-gastritis | 12 | 32 |
| Intestinal metaplasia | 12 | 32 |
| Mixed | 8 | 32 |
| **Total** | **40** | **128** |

The invalid normal-specialist corpus contains eight held-out-slide fields among its 32 normal fields. None of the 40 training crops has the exact M1 reference coordinate, but slide-level separation—not coordinate inequality—was the preregistered boundary, so this does not rescue the experiments.

## Scientific impact

- Q0 and Q1 infrastructure findings remain valid because Q0 is untrained and Q1 used a separate GAGL smoke corpus.
- Q2 v1, no-overlay, prompt-ablation, and caption-neutral artifact-integrity and execution records remain reproducible, but every held-out similarity/generalisation result and checkpoint-selection conclusion is invalid.
- Caption-neutral epoch 3 is no longer an accepted tissue scaffold. It is an exploratory artifact only until a corrected slide-disjoint experiment independently reproduces the result.
- Q3 operational findings about exact-canvas assembly, seams, denoising, shutdown, and failure handling remain engineering evidence. Their biological/category conclusions cannot qualify any checkpoint for submission because the upstream LoRA used the invalid corpus.
- Normal-specialist run `20260826T184640Z-lora-train` completed with valid hashes but is scientifically rejected. No checkpoint or validation image may be selected from it.

## Root cause and correction

The builder and verifier accepted arbitrary exclusion strings and had no cryptographic binding to the held-out manifest. The correction:

1. replaces every Q2 and specialist exclusion with the exact canonical UUID;
2. requires canonical hyphenated UUIDs in corpus construction and verification;
3. requires the held-out manifest as an input, verifies its SHA-256 through the preregistered config, and requires the configured exclusion set to equal the manifest's unique slide set;
4. proves the old corpus now fails with `Held-out slide entered corpus`;
5. keeps all GPU compute stopped until a new IOFlood-built corpus contains zero records from the three held-out UUIDs and passes full mechanical and visual review.

The corrected corpus receives new metadata/provenance hashes and a new protocol version. Historical artifacts and reports are preserved rather than overwritten.

## Corrected rebuild v2 gate

The first genuinely slide-disjoint rebuild completed on IOFlood and passed independent remote and local mechanical verification:

- 128 images, exactly 32 per category;
- 13 source slides;
- zero records from the three M1 held-out slide UUIDs;
- metadata SHA-256 `ab54ffb3b4f4e22a81595319e31481f23b43950ef4054ee76711b3c5c2a95e29`;
- provenance SHA-256 `8fc600045163585961cb4557fb832c89e788c9a0e8572bf43b74758866b9426e`;
- local verification SHA-256 `4bd619e53f7c120931b236b8458af3a645b2846302a359389db3c1c0d270802f`.

It nevertheless **fails the engineering visual input gate**. Three normal crops from slide `6c6bcc80-c875-471e-87e9-1aef6725d80e` (`0009`, `0011`, and `0012`) are dominated by blank space and torn/pale tissue rather than a continuous gastric field. Their fractions of pixels below mean RGB intensity 230 are 0.275, 0.254, and 0.321. No specialist run may use this corpus.

The preregistered v3 correction applies one category-independent real-input rule before selection: at least 50% of crop pixels must have mean RGB intensity below 230. It evaluates up to twelve non-overlapping candidates per slide/category but retains at most four, preserving the prior slide-diversity cap while allowing inadequate fields to be replaced. The rule is applied uniformly to all real candidate categories and recorded in provenance. It does not use synthetic outputs, Sydney labels, or generator artifacts. The v3 rebuild must again pass the split, hash, dimension, full contact-sheet, and full-resolution visual gates before GPU work resumes.

## Corrected rebuild v3 result

The v3 rebuild passes the bounded engineering input gate:

- 128 exact 1024×1024 RGB PNGs, 32 per category, from 13 non-held-out slides;
- zero provenance records from the three M1 held-out slide UUIDs;
- at most four selected fields per slide/category;
- minimum stained-pixel fractions of 0.537 normal, 0.509 H. pylori-gastritis, 0.506 intestinal metaplasia, and 0.578 mixed;
- metadata SHA-256 `ab54ffb3b4f4e22a81595319e31481f23b43950ef4054ee76711b3c5c2a95e29`;
- provenance SHA-256 `59c826ee7693826cef2eef55f1f7c931b88a270026e8bfee33435e20ddd2398e`;
- independent local-verification SHA-256 `2b2ac5ddc09643159c8853253f084760e15acb45ad2e122d39cb9cd5fd98effe`;
- deterministic 32-field review-sheet SHA-256 `0208eb502f83565bcc70c062061c57a6a7d4fc1bc4906a3759ce8864cf941867`.

The review sheet and the lowest-coverage field in every category were inspected at full resolution. All contain interpretable gastric tissue rather than the blank/torn-field failure seen in v2. This is an engineering input-adequacy decision only; it does not validate the source category annotations or replace pathologist review.

The first orchestration wrapper returned exit code 1 after image construction because it supplied the caption-neutral prompt override while verifying the base-prompt corpus. Re-running the verifier with the corpus's actual pinned prompt set passes. The failed wrapper status is retained as procedural evidence; the corpus was not rebuilt or altered after verification.

The hash-preserving caption-neutral view changes only the 64 positive captions and has metadata SHA-256 `0fa3b61a6b76e311b52af163f9ea621efbd00a868769046d7c3954f6d1d11ac9`. Its corrected specialist views are:

- normal: metadata `024754db101c6371b6e67f4f1247df921ab63b6ba618afcb2aa5c07578a2bf9a`, provenance `4d66a9d56dd9c9beef75fedebd3f69c7b04b8c4411a90f546048f585efab2831`, 32 fields from nine slides;
- intestinal metaplasia: metadata `5ba91f805d795680790a87c2fa3f94afabc17aa95968a3cf6729f7188533fc5d`, provenance `7449e3d94e0ac90ddcaf3ed5aeb83a688bd441e095c50c889ce4fdd10d022bdb`, 32 fields from eight slides.

Both specialist views independently pass the strengthened held-out, hash, scale, dimension, and prompt gate. They restore eligibility for only the preregistered normal-versus-intestinal-metaplasia specialist pair; they do not restore any conclusion from the invalid historical Q2/Q3 runs.