# Z-Image exact-caption alignment

## Result

**FAIL; close the Z-Image LoRA path.** Inference-only run `20260827T024006Z-zimage-prompt-alignment` completed successfully on deployed commit `b4de1124989b695911ea4dc89db865b79a2b18e9`, but every whole-checkpoint cell is mechanically disqualified by generated text and fails the full-resolution normal-gastric topology gate.

## Frozen intervention and integrity

- The intervention changed only the positive prompt to the exact byte-identical caption present in all 32 corrected normal training records. Z-Image revision, the four LoRA checkpoint hashes, negative prompt, seed 41001, 1024x1024 RGB geometry, 50 steps, CFG 4.0, and CFG-normalization setting remained frozen.
- Config SHA-256 is `e75dbd19631a958d11938e80035e9c1a506e9e4672d28e2f2263be2dfc5e134f`. Remote immutable preflight reverified the source run, deployed source commit, all four checkpoint hashes, and exact prompt equality across all 32 metadata rows.
- Exit code is 0, `ZIMAGE_PROMPT_ALIGNMENT_COMPLETE` is present, and every entry in the retrieved `ALL_SHA256SUMS` passes locally.
- All four outputs are 1024x1024 RGB PNGs. Step/output SHA-256 values are: 32 `7edfc4a9a7ca65b3f28c5f5db577c82c5be80857fba839a6fe028eee9c7f3c42`; 64 `4760c171d97a966cda84544aa41b80bf1ced5f1972b031d83d5dc362a9f604f7`; 96 `8c922ce392a991d2e357324e132596d12dd8a8a65dc2701c0e47b9ad8a6649d2`; 128 `dd5e61780be5cb57c4d5a1481c6bc006bac10a6b6ca79a18d918f5751c89545b`.

## Privacy and Sydney-QA

Privacy passes against both corrected corpora. Against the 32-field normal specialist, maximum DINO cosine is 0.900356 and minimum dHash distance is 22. Against the full corrected 128-field corpus, maximum DINO cosine is 0.900797 and minimum dHash distance is 20. There are no exact pixel matches. These are privacy checks, not clinical-quality evidence.

Sydney-QA completed on all four cells and remains `CALIBRATION_REQUIRED`. No GI-pathologist-annotated patient/slide-disjoint real envelope exists, so its measurements cannot establish clinical correctness, diagnosis, Sydney grade, or OLGA/OLGIM and do not replace pathologist review.

## Whole-checkpoint engineering review

| Step | Mean absolute gradient | Sydney-QA glands/mm2 | Mucin/vacuole candidates/mm2 | Block max cosine | Result |
| --- | ---: | ---: | ---: | ---: | --- |
| 32 | 13.139663 | 99.0 | 742.7 | 0.9993 | Generated pseudo-label text; mirrored/stacked tissue bands and papillary projections; reject |
| 64 | 11.913954 | 165.1 | 924.6 | 1.0000 | Generated pseudo-label text; repeated rosettes and papillary structures with large blank separation; reject |
| 96 | 9.660728 | 115.1 | 954.1 | 0.9994 | Generated pseudo-label text; repeated villiform/papillary structures and non-gastric composition; reject |
| 128 | 8.030359 | 131.6 | 855.4 | 0.9986 | Generated pseudo-label text; simplified repetitive papillary topology and poor cellular detail; reject |

The exact caption made the physical-scale phrase visibly render as text despite the frozen negative prompt prohibiting text and annotations. All four cells are therefore automatic mechanical failures before morphology selection. Their stacked/mirrored tissue organization is also materially worse than the original attempt-4 cells and untouched real normal reference. No whole checkpoint passes.

## Runtime, cost, and cleanup

The worker ran from `2026-08-27T02:40:06Z` to `2026-08-27T02:46:12Z`, or 366 seconds. Conservative A100 instance uptime is below 564 seconds, approximately US$0.58 at US$3.673385/hour. Bounded recovery-CPU uptime is below 278 seconds, approximately US$0.01 at US$0.134/hour. Total conservative compute is below US$0.59 before temporary storage, credits, and tax; the A100 itself was Spot.

Both workers are terminated. After local hash, dimension, privacy, and Sydney-QA extraction, the exact temporary recovery instance, disk, and snapshot were deleted. The original stopped disk remains retained.

## Decision

The preregistered stop rule fires: close the Z-Image LoRA path. Do not run Z-Image IM, gastritis, mixed, canvas, or candidate-pool jobs. Prompt tuning would now select against synthetic failures and would not address the model's repeated inability to reproduce local normal gastric topology.

The next eligible challenger must be separately preregistered. The strongest already-measured remaining open-weight candidate is FLUX.2 Klein Base 4B at the immutable M1 revision: its zero-shot normal was weaker than Z-Image, but its Apache-2.0 4B base now has a first-party-supported official Diffusers LoRA path. Any run remains normal-only and must use the same corrected corpus, untouched reference, privacy thresholds, and Sydney-QA uncertainty before another category can open.