# M1 report — foundations and zero-shot baseline

Status: complete; founder review required before M2  
Date: 2026-08-25  

## Outcome

M1 produced and checksummed four real HiESD reference fields plus all eight pre-registered zero-shot model/category cells. All images are decodable RGB PNGs at their required M1 dimensions. The A100 worker is stopped.

Z-Image Base is the recommended backbone for the first controlled LoRA pilot. This is a relative engineering choice, not an assertion that its zero-shot images are clinically acceptable. None of the generated M1 images is submission-ready.

## Evidence

- Real references: four 4096×2048 fields extracted at 0.25 µm/pixel from source slides reporting 0.2458 µm/pixel and approximately 40× objective.
- Generated baseline: four categories × two models at 1024×1024, 50 steps, guidance 4.0, one fixed seed per category.
- FLUX.2 Klein Base 4B: 18.61 GB peak allocated GPU memory; 21.8–24.7 seconds per image.
- Z-Image Base: 23.27 GB peak allocated GPU memory; 42.6–43.3 seconds per image.
- SHA-256 verification passed for every generated PNG and both generated contact sheets.
- Conservative A100 uptime reconstructed from GCP audit events: 2,056 seconds, or 34.3 minutes, including failed preflights and artifact-retrieval boots.

At the official `a2-highgpu-1g` on-demand rate of US$3.673385/hour in `us-central1`, compute is approximately US$2.10. Short-lived 200 GB balanced-disk storage adds well under US$0.10; the M1 estimate is therefore under US$2.25 before credits and taxes. Source: https://cloud.google.com/products/compute/pricing/accelerator-optimized

## Visual screening

This screening was performed by the engineering agent and is not a pathologist review.

### Z-Image Base

- **Normal:** recognisable mucosal/glandular organization and the strongest tissue-scale structure of the two models, although framing and some gland geometry do not match the real 40× field closely.
- **Intestinal metaplasia:** produces organised glands but does not reliably express convincing goblet-cell and gastric-metaplasia morphology.
- **H. pylori/gastritis:** renders abundant dark curved objects at a biologically impossible size and largely loses gastric tissue architecture.
- **Mixed:** separates surface organisms and tissue into implausible layers; organism scale is again grossly wrong.

### FLUX.2 Klein Base 4B

- **Normal:** washed-out, repetitive gland-like texture with poor cellular detail.
- **Intestinal metaplasia:** repetitive ring/bubble texture rather than coherent epithelium and goblet cells.
- **H. pylori/gastritis:** diffuse inflammatory/spindle-like texture without convincing gastric glands or correctly scaled organisms.
- **Mixed:** repetitive circular structures without biologically plausible mixed gastric pathology.

## Decision

Proceed with Z-Image Base in M2, subject to founder approval, because it preserves more tissue-scale organisation and stronger category separation. The next experiment must teach domain morphology rather than merely upscale these zero-shot outputs. H. pylori requires an explicit scale-aware training/evaluation strategy; a generic super-resolution pass would sharpen the current oversized hallucinations rather than correct their biology.

M2 should remain a small LoRA pilot with slide-level train/validation separation and fixed before/after seeds. It must stop again for review before any multiscale or organism-detail system is built.

## Known limitations

- Only one zero-shot seed per category/model was generated; no statistical quality claim is made.
- HiESD does not annotate individual H. pylori organisms.
- The real references and generated square baselines have different field aspect ratios; the review sheet uses labelled centre crops of the real fields only for side-by-side viewing.
- M1 includes no clinical scoring, privacy/memorization test, LoRA training, 4096×2048 generation, or submission candidate.