Q2 v1: Qwen-Image-2512 morphology gate

Post-hoc invalidation (2026-08-27): the training corpus contained fields from all three M1 held-out slides because malformed exclusion identifiers bypassed literal matching. Held-out metrics and checkpoint-selection conclusions in this historical report are invalid. See reports/SLIDE_SPLIT_AUDIT_REPORT.md.

Date: 2026-08-26 UTC Run: 20260826T003100Z-lora-train Code: 4d54800ba827ae7744c460166b7d105ed9e12562 Result: FAIL — partial tissue improvement and privacy pass, but no checkpoint passes aggregate clinical-fidelity review

Purpose and frozen protocol

Q2 v1 tested whether the accepted, slide-disjoint HiESD corpus v2 could move the pinned Qwen-Image-2512 prior toward gastric H&E morphology. The preregistered run used 128 balanced 1024×1024 fields at 0.25 µm/pixel, 32 per category, fourteen non-held-out source slides, NF4 transformer training, rank-32 LoRA, learning rate 1e-4, one dataset repeat, and four epochs. Every epoch checkpoint was evaluated using the same four prompts and seeds. Checkpoints were accepted or rejected as whole units; cells were never mixed across epochs.

The corpus metadata SHA-256 was 4e475c7d260b023d9b25cc0b6f17d8899fe753ac776d4416e667d979b4cbf8c7; the v2 provenance SHA-256 was 835f4401d31f6f5cde98c3bc452482b76dc273fb13b45338c4ff975f3a44fce4. Corpus v1 remained prohibited and was not present in the run.

Mechanical and reproducibility results

• The immutable remote input gate passed: 128 images, fourteen permitted slides, 32 images per category, exact pinned metadata and provenance hashes, and all held-out slide IDs absent.
• Split preprocessing created 128 cached feature files. Transformer-only NF4 training completed all 512 exposure steps and saved four checkpoints.
• The worker ran from 2026-08-26T00:31:00Z to 01:13:48Z, or 2,568 seconds.
• The A100 instance was active for 3,365.826 seconds including startup and automatic shutdown. At the conservative previously recorded on-demand list rate of US$3.673385/hour, that is at most US$3.44 before credits and tax; the actual worker was Spot. The bounded CPU recovery worker ran for 2,044.018 seconds. Including short-lived recovery compute and storage, Q2 v1 is conservatively below US$3.60.
• QWEN_LORA_COMPLETE is present and exit_code.txt is 0.
• All 166 files listed by ALL_SHA256SUMS independently matched after retrieval.
• All four 472,047,184-byte LoRA checkpoints matched lora_checkpoints.sha256:

• All sixteen validation outputs decoded as 1024×1024, 8-bit RGB, non-interlaced PNGs and matched the per-run hashes.
• The A100 and bounded CPU recovery worker are stopped. The original stopped Qwen disk remains the reproducible model/cache source.

Privacy triage

All sixteen validation images were screened against all 128 training fields using exact RGB equality, 64-bit difference hash, and the pinned non-generative DINOv2-base evaluator at revision f9e44c814b77203eaa57a6bdbbd535f21ede1415.

• Exact matches: 0.
• Maximum DINO cosine similarity: 0.943365.
• Minimum dHash Hamming distance: 19.
• Fail thresholds: cosine at least 0.995 or dHash distance at most 4.
• Result: PASS, with no flagged candidate.

The privacy report SHA-256 is 34d3881247dd44ace7d6f471a379394f7d5482879944d21650eaaba3a4975eb0.

Morphology audit

The held-out-reference/Q0/Q2 sheet has SHA-256 474c036bd21b891c1f7e0390231ad696bba5b579fad139aaf3496095e13b8a74. Engineering review found a real and useful shift away from Q0's wood-grain texture:

• Epochs 1–3 produced more gland-like gastric tissue in the normal and intestinal-metaplasia cells.
• A DINOv2 feature comparison against the four held-out fields rose from a Q0 mean nearest-reference cosine of 0.7602 to 0.9170 at epoch 1. This is an out-of-domain engineering diagnostic, not a clinical metric.
• The H. pylori-gastritis and mixed cells retained conspicuous dark curved rods that are grossly oversized and poorly localised relative to the expected 10–20-pixel organism length at 0.25 µm/pixel.
• Epoch 2 suppressed some of the largest curves but did not establish credible organism morphology or convincing aggregate category fidelity.
• No single epoch simultaneously passed normal architecture, gastritis, metaplasia, mixed-lesion, and organism-scale review. Selecting different cells from different epochs would violate the preregistered aggregate checkpoint rule.

The held-out feature diagnostic SHA-256 is 93fe1178ca353bd115d6efdfb8eaf51e931e3817a53b64e4955aeb7533f789d0.

Gate decision

Q2 v1 passes training, artifact-integrity, mechanical-output, and privacy gates, and it demonstrates that the LoRA can materially alter the tissue prior. It fails the aggregate morphology gate because the positive categories learn a visually dominant curved-organism shortcut rather than clinically plausible H. pylori scale and localisation. Every Q2 v1 checkpoint is rejected for candidate production. Do not begin Q3 or assemble submission images from this run.

The next technical work remains inside Q2: remove the failed organism shortcut from the training design, preregister a bounded correction or ablation, and require another equal-cell aggregate gate. Clinical sign-off remains mandatory before any final set can be frozen.

Download public Markdown export