Dated experiment record. Statements describe this run, not current submission readiness. Historical H. pylori category names do not establish infection; copy screens are bounded tests. Claims that a source does not exist mean none was identified in that recorded search, not proof of absence. Licensing interpretations in the notes remain unconfirmed. Local access details have been omitted. Current limitations and remaining work.

Q0: pinned Qwen-Image-2512 zero-shot baseline

Date: 2026-08-25 UTC
Run: 20260825T201315Z-smoke
Code: 81b6503a4418b9a199938733e405790439be539a
Result: mechanical pass; clinical-plausibility control failed

Frozen configuration

Transformer: Qwen/Qwen-Image-2512 at 25468b98e3276ca6700de15c6628e51b7de54a26.
Shared components: Qwen/Qwen-Image at 75e0b4be04f60ec59a75f475837eced720f823b6.
DiffSynth-Studio: 3f3df523ed6273ee4c9dae245de2c7abd8fd4b5f.
Hardware: one NVIDIA A100-SXM4 40 GB.
Four fixed prompts and seeds, 1024×1024 RGB PNG, 30 steps, CFG 4.0.
No LoRA, no image conditioning, no retries, and no candidate selection.

Mechanical results

All four pre-registered cells completed and passed decoding, dimensions, RGB mode, non-blankness, manifest, and SHA-256 checks.

Category	Seconds	Peak allocated VRAM	SHA-256 prefix
Normal	61.710	19.499 GiB	`1ca48092bfeb`
H. pylori with gastritis	59.024	19.499 GiB	`a4520c8f6925`
Intestinal metaplasia	56.953	19.498 GiB	`ffaa006d5c06`
Mixed	56.911	19.498 GiB	`10aeeea92281`

The run lasted 9 minutes 42 seconds including pinned dependency/model preparation. Conservative chargeable A100 time including artifact retrieval was approximately 15.8 minutes, or about US$0.97 at the recorded US$3.673385/hour list rate, before disk, credits, and tax.

Visual audit

The hash-verified review sheet compares each output with a held-out HiESD field. Its SHA-256 is 807664988f6dc0a5c98e90244be51553f6a533faca8e0aa809b9b9b7abe66354.

This is an engineering visual audit, not pathologist sign-off:

The normal output is dominated by folded, fibrous, wood-grain-like texture rather than orderly gastric foveolar and glandular architecture.
The H. pylori/gastritis output contains curved dark objects that are tens to hundreds of pixels long. At the requested physical scale, organisms should be approximately 10–20 pixels long. The output therefore fails scale and localisation.
The intestinal-metaplasia output contains some gland-like forms but does not reliably reproduce the held-out intestinal-type gland and goblet-cell morphology.
The mixed output repeats the oversized curved-object failure and does not establish a credible mixed lesion.
Stain colours are superficially H&E-like, showing that colour/style language alone is not sufficient for clinical fidelity.

No Q0 image is eligible for the challenge package.

Decision

Qwen-Image-2512 is operationally viable on the existing A100 40 GB, but zero-shot prompting is not clinically viable. Advance to Q1 solely to validate the pinned NF4 rank-32 LoRA path, then use slide-disjoint HiESD fields at 0.25 µm/pixel for morphology training. Q1's eight-image GAGL corpus is an infrastructure test, not evidence of model quality.

Download public Markdown export