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 no-overlay ablation: morphology gate

Post-hoc invalidation (2026-08-27): this corpus contains held-out-slide fields. Its execution record remains reproducible, but held-out metrics and checkpoint-selection conclusions are invalid. See reports/SLIDE_SPLIT_AUDIT_REPORT.md.

Date: 2026-08-26 UTC

Run: 20260826T075929Z-lora-train

Code: efd0de00a317896c91f28c0e4d2f267cc6bdd694

Result: FAIL for candidate production — privacy and integrity pass, tissue similarity improves, but no whole checkpoint removes the positive-category curved-object failure

Purpose and causal control

This run changed one training intervention from Q2 v1: all simulated H. pylori overlays were removed. The 128 source fields, fourteen permitted slides, coordinates, categories, dimensions, physical scale, split, seed, model, optimiser, rank, learning rate, repeats, epochs, validation prompts, and validation seeds remained frozen. The corpus metadata SHA-256 was 4e475c7d260b023d9b25cc0b6f17d8899fe753ac776d4416e667d979b4cbf8c7; no-overlay provenance SHA-256 was 86cbc623a28e6a22807f3bdb082c58f8ed4a1e86ba6e7a37d513ea3b7177134c.

Mechanical and cost results

The immutable remote input gate passed: 128 1024x1024 RGB fields at 0.25 micrometres per pixel, exactly 32 per category, fourteen non-held-out slides, and explicit ablation-none-v1 provenance with zero simulated organisms.
Training completed all 512 exposure steps and saved four 472,047,184-byte rank-32 LoRA checkpoints. QWEN_LORA_COMPLETE is present and exit_code.txt is 0.
Worker time was 2,586 seconds, from 07:59:29Z through 08:42:35Z.
A100 instance uptime including startup and automatic shutdown was 4,094.897 seconds. At the conservative recorded on-demand list rate of US$3.673385/hour, this is at most US$4.18 before credits and tax; the instance actually used Spot capacity. The bounded CPU recovery worker ran for 1,695.464 seconds, conservatively about US$0.06 at US$0.134/hour. Total compute plus temporary storage remains below US$4.35.
All 170 files in ALL_SHA256SUMS and all checkpoint hashes matched after retrieval. All sixteen validation outputs decode as 1024x1024 RGB PNGs.
Checkpoint SHA-256 values: epoch 0 c5b70ade0706177ab414388b03c79470fa92ba77f2c89e84b0142de80c15af2f; epoch 1 23dad231886c9a3fbe3892cd906421979d66435b19012521802701c35ea95f1a; epoch 2 f96671a1ff40814770e6732c01ace93065b0fa73480268b1b7fec7e24feef7fd; epoch 3 5fb8d73646228a2eb047593113e6c64490c2fe2c6c9c71b803c04b64718e9de7.
The A100 and recovery CPU are terminated. The 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 no-overlay training fields with exact RGB equality, 64-bit difference hash, and pinned DINOv2-base revision f9e44c814b77203eaa57a6bdbbd535f21ede1415.

Exact matches: 0.
Maximum DINO cosine: 0.937824 against the 0.995 failure threshold.
Minimum dHash Hamming distance: 19 against the distance-at-most-4 failure threshold.
Result: PASS. Privacy report SHA-256: fba6378a00472a02e3d25fb44c26dd355d169e82d13a1ea77292d5103e6a9a18.

Morphology and causal finding

The Q0/Q2-v1/no-overlay comparison sheet SHA-256 is ec7e57682848ba39789f86ae6988fdfae86636842aaf4198160cdc79452c22b5; the 1024-pixel audit montage SHA-256 is d2fce1e991a51a01ce80eb794f46536a51147c039d7a3d1edcae64d08801fed9.

Mean nearest-held-out DINO cosine was 0.760188 for Q0 and 0.754595, 0.862106, 0.739393, and 0.907731 for no-overlay epochs 0 through 3. This is an out-of-domain engineering diagnostic, not a clinical metric.
Epoch 3 gives the strongest aggregate tissue-similarity diagnostic and more coherent gastric/metaplastic structures than Q0.
Despite removing every simulated organism from training pixels, H. pylori-gastritis and mixed prompts still produce conspicuous dark curved objects in epochs 0 through 2, with residual curved forms in epoch 3. Several exceed the expected 10–20-pixel organism length at 0.25 micrometres per pixel.
Epoch 3 suppresses the worst rods but loses or weakens some category-specific gland and inflammatory architecture. No whole checkpoint simultaneously passes all four category cells.

The ablation falsifies the narrow claim that the pixel overlays alone caused the failure. They amplified it, but inference wording and/or the general model's small-object prior also contributes. Selecting individual cells across epochs remains prohibited.

Gate decision

All four checkpoints are rejected for Q3 and submission. The next bounded test remains inside Q2 and changes inference wording only: use epoch 3 with identical seeds and sampler settings, remove requests for visible organisms from the two positive tissue prompts, and strengthen the negative prompt against worm-like or thick curved objects. If that yields acceptable tissue scaffolds, physically scaled 10–20-pixel organisms must be added through a separately preregistered final-resolution method and reviewed by a gastrointestinal pathologist. If it fails, rebuild a caption-neutral tissue LoRA before any canvas production.

Held-out diagnostic SHA-256: 9fa46c8bf149e47f75c915887faf90d4eb649c15afa41ea3e3d3bb2aa18938e9.

Download public Markdown export