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.

Q3 detail denoising ablation

Post-hoc data notice (2026-08-27): the upstream LoRA was trained with held-out-slide leakage. The denoising comparison remains operational evidence only and cannot qualify the checkpoint for submission. See reports/SLIDE_SPLIT_AUDIT_REPORT.md.

Result

Run 20260826T160906Z-detail-ablation completed successfully on deployed commit 63126a9ee6e719c265897a87639a5fd741cf5aaa. All fifteen recorded artifact hashes pass locally. The exact caption-neutral epoch-3 checkpoint, positive-category prompt file, source scaffold, crop, seed, 35 steps, CFG 4.0, and 1280x1152 geometry match the preregistration.

The ablation selects denoising strength 0.95 for the next bounded four-field gate. It is a controlled engineering selection, not a submission or clinical pass.

Fixed comparison

Denoising strength	Mean absolute gradient	Change from 0.62 control	Full-resolution result
0.62 control	1.651991	—	Soft cellular detail; rejected
0.75	1.628052	-1.4%	No meaningful improvement; rejected
0.85	1.846297	+11.8%	Still soft; rejected
0.95	3.349442	+102.8%	Coherent gastric glands, materially sharper nuclei and epithelium, no gross repetition or pasted structures; selected

The real-reference preregistration range is 5.45–8.64, so 0.95 remains 38.5% below its lower bound. Selection is permitted by the recorded alternative criterion of material improvement with coherent morphology. The next four-field run must therefore recheck assembled-canvas sharpness and cannot treat this one-tile result as final quality acceptance.

Integrity, runtime, and cost

Output hashes: ee2e8c3a619eec206ccb31a580d88f1c88aa2eae2e4ab9ff5a07fec6e4d3fc60 at 0.75, 8fa8b0e55b0de6efcadc4fb7adf142f1adb1437bc103e65a202c34327d385dd5 at 0.85, and 3d9e540f84528445dc11787fba5424de20910cab98d5fb51debcd338ed41b36e at 0.95.
Worker runtime was 535 seconds; the measured A100 instance uptime was 784.437 seconds.
Conservative A100 compute cost is US$0.81 at US$3.673385/hour. The bounded recovery CPU ran 166.436 seconds, approximately US$0.01 at US$0.134/hour. Total compute is below US$0.85 before temporary storage, credits, and tax; the actual A100 was Spot.
The A100 and recovery CPU are terminated. Temporary recovery resources are deleted after local verification; the original stopped Qwen disk is retained.

Gate decision

Advance only to a preregistered four-field correction using denoising strength 0.95 for every detail tile and the already tested antialiased microdetail renderer. Keep every other Q3 variable frozen. Candidate pooling remains prohibited until all four exact canvases pass dimensions, seams, tissue topology, category fidelity, organism rendering, privacy, and full-resolution review together.

Download public Markdown export