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.

Q2 caption-neutral correction: tissue-scaffold gate

Post-hoc invalidation (2026-08-27): 40 of 128 training fields came from M1 held-out slides. Epoch 3 is revoked as an accepted scaffold and retained only as an exploratory artifact. See reports/SLIDE_SPLIT_AUDIT_REPORT.md.

Date: 2026-08-26 UTC

Run: 20260826T111002Z-lora-train

Code: 603c4284f89919fc90a8be9394b104cfb524baf0

Result: PASS for the tissue-scaffold checkpoint only — epoch 3 removes the oversized-organism shortcut and preserves aggregate gastric category signal; no validation image is a submission candidate

Controlled change

This run changed only the 64 positive-category captions from the accepted no-overlay corpus. All 128 training images, source slides, coordinates, categories, split, dimensions, physical scale, seed, model, optimiser, rank, learning rate, repeats, epochs, and validation seeds remained frozen. The corpus metadata SHA-256 was 3c84066676775f38d696f3103a7345210f2cb769cbdf2fbc73e1b11003d8768e; provenance SHA-256 remained 86cbc623a28e6a22807f3bdb082c58f8ed4a1e86ba6e7a37d513ea3b7177134c; prompt-override SHA-256 was 4cb227a987db2790f563f5cf9ac5c549c3997238918757b741d02c89640e0a98.

Mechanical and cost results

The immutable remote input gate passed: 128 pixel-identical no-overlay fields, 32 per category, fourteen non-held-out slides, and exact pinned corpus/config hashes.
Training completed all 512 exposure steps, saved four 472,047,184-byte rank-32 LoRA checkpoints, generated sixteen fixed validation images, and ended with exit_code.txt equal to 0 plus QWEN_LORA_COMPLETE.
Worker time was 2,586 seconds, from 11:10:02Z through 11:53:08Z.
A100 instance uptime including upload, preflight, startup, and automatic shutdown was approximately 3,499 seconds. At the conservative recorded on-demand list rate of US$3.673385/hour, this is at most US$3.57 before credits and tax; the instance actually used Spot capacity. The bounded CPU recovery worker ran approximately 1,449 seconds, about US$0.05 at US$0.134/hour. Total compute plus temporary storage is conservatively below US$3.75.
Every entry in ALL_SHA256SUMS and all four checkpoint hashes matched locally. All sixteen validation outputs decode as 1024x1024, 8-bit RGB PNGs; all four generation manifests and verification records are present.
Checkpoint SHA-256 values: epoch 0 c939528b6d578fedf3de7c05b3f09fae76560ade996651963de57677231b9785; epoch 1 5d2358252765303de14b2cc467caf9525324707d29a25185b9e9420a72903793; epoch 2 0107997ad381c9d20b45a6d59e855990801e7dc2d9546ee7efae52a43eafab75; epoch 3 ab11358b0b54a7a2c461f93fdc2a2fcccc15da4944dfbe5d6c6340b3d4df6d24.
The A100 and CPU workers are terminated. The exact temporary recovery instance, disk, and snapshot were deleted after local verification; the original stopped Qwen disk remains.

Privacy triage

All sixteen outputs were screened against all 128 caption-neutral training fields using exact RGB equality, 64-bit difference hash, and pinned DINOv2-base revision f9e44c814b77203eaa57a6bdbbd535f21ede1415.

Exact matches: 0.
Maximum DINO cosine: 0.959761 against the 0.995 failure threshold.
Minimum dHash Hamming distance: 18 against the distance-at-most-4 failure threshold.
Result: PASS. Privacy report SHA-256: 02758c807ea4ea6c97db8e67f684f4ab994aef0488cff7994574e7163982e439.

Whole-checkpoint morphology review

The held-out/Q0/prior-control/caption-neutral sheet SHA-256 is 2d50f30eed40a1bc8cabfa9e449b0d79a0c9634b8bcfbbe4d3b97545b6548987. Every caption-neutral cell was inspected at full 1024x1024 resolution.

Mean nearest-held-out DINO cosine rises by epoch: 0.891029, 0.907996, 0.924516, and 0.940878. This is an out-of-domain engineering diagnostic, not a clinical metric.
Epochs 0 and 1 retain excessive vacuolar/honeycomb texture and weaker category separation. Epoch 2 improves glandular structure but normal remains vacuolated.
Epoch 3 provides the strongest whole-checkpoint balance: coherent glandular tissue, inflammatory signal in the gastritis cell, goblet-like/metaplastic morphology in the intestinal-metaplasia cell, and combined gland/inflammatory structure in mixed. No positive cell contains the prior worm-like or grossly oversized curved rods.
The caption-neutral checkpoint intentionally does not establish visible H. pylori organisms. At 0.25 micrometres per pixel, organisms must be validated separately at the final 4096x2048 scale as approximately 10–20-pixel, surface-localised microdetails.

Held-out diagnostic SHA-256: a66e9015775ce5b9cd2857080d52bce95ec42417e424775b131c6cdc80fc0f6f.

Gate decision

Accept epoch 3 as the pinned Q3 tissue-scaffold checkpoint. Reject epochs 0–2 and prohibit mixing cells across epochs. None of the sixteen validation images is submission-eligible. Q3 may begin only as a bounded one-field-per-category final-resolution experiment using the epoch-3 checkpoint for the global Qwen tissue canvas and a separately preregistered, physically scaled, epithelial-surface-localised H. pylori microdetail stage. Advance to a candidate pool only if exact dimensions, seams, category morphology, organism scale/localisation, and privacy all pass; pathologist review remains mandatory before package freeze.

Download public Markdown export