# Z-Image split-corrected normal challenger

## Result

**FAIL for candidate production.** Run `20260827T013450Z-zimage-normal` completed successfully on deployed commit `5ad0fefce453778933b4f2393df4b1b70af256f9`, but no whole checkpoint met the frozen normal-gastric morphology gate. The three earlier launches remain infrastructure-only failures; this fourth launch is the first valid scientific Z-Image result.

## Integrity, rights evidence, and split

- Exit code is 0, `ZIMAGE_LORA_COMPLETE` is present, and every entry in `ALL_SHA256SUMS` passes after local retrieval.
- The exact model is `Tongyi-MAI/Z-Image` revision `04cc4abb7c5069926f75c9bfde9ef43d49423021`; publisher code is `26f23eda626ffadda020b04ff79488e1d72004cd`; the official Diffusers trainer is `d57cecde92a6d396845ab35425aa27469dff8173`. The archived model card and publisher/Diffusers licence files record Apache-2.0.
- Immutable config SHA-256 is `ce6319bc8fc2e21dba72e63401bd48c21d6c892c648755e19871e62ea4a581d3`. The matched CUDA 12.9 runtime lock SHA-256 is `e5359bf567d77ad8c502696cf8e3caf6605471d77ee398e913024ec9cf2bd082`.
- The corpus gate passed 32 exact 1024x1024 RGB normal fields from nine non-held-out slides at 0.25 micrometres/pixel. Metadata SHA-256 is `024754db101c6371b6e67f4f1247df921ab63b6ba618afcb2aa5c07578a2bf9a`, provenance SHA-256 is `4d66a9d56dd9c9beef75fedebd3f69c7b04b8c4411a90f546048f585efab2831`, and held-out manifest SHA-256 is `18377920a2dcca040e05ce875e4c7ecfe7a3f09d512539bf69d2eb681a26d0af`.
- Four whole rank-32 LoRAs were recovered at steps 32, 64, 96, and 128. Each is exactly 66,882,272 bytes. Their SHA-256 values are `bcbf92c3aad802a6b473616d457c907eb168991b2e58254e7fd2cfcb711f480b`, `88574a26284810c5e17a87a8a68830255e03dcc8e1160dc11dee9b2a2e770c8b`, `fc6546cf629ca30670a5c67c39320af72588469a0b3b511f325552072f98c887`, and `b00b18230f4db052cb0a6499a9479e28c516dddc32f9234fa51cb93be29f127c`.
- Four fixed-seed validation images are exact 1024x1024 RGB PNGs. The local comparison sheet SHA-256 is `8319c8e07232c958ab0f3aa39cc7f0b77e5bfaa9ef59bdfaea5a62f8ead6e2db`.

## Privacy

Both preregistered screens pass. Against the corrected 32-field normal corpus, maximum DINO cosine is 0.894674 and minimum dHash distance is 23. Against the full corrected 128-field caption-neutral corpus, maximum DINO cosine is 0.897912 and minimum dHash distance is 15. Neither screen found an exact pixel match. DINO and dHash are privacy/nearest-neighbour checks, not clinical-quality evidence.

## Whole-checkpoint engineering review

| Step | Mean absolute gradient | Sydney-QA glands/mm2 | Sydney-QA mucin/vacuole candidates/mm2 | Lumen/tissue | Full-resolution result |
| --- | ---: | ---: | ---: | ---: | --- |
| 32 | 14.020908 | 442.1 | 933.3 | 0.1230 | Nearly preserves the zero-shot composition: an implausibly regular surface-to-deep column with oversized, dotted lumina rather than the local normal gastric topology in the untouched reference; reject |
| 64 | 11.742981 | 518.1 | 1,219.9 | 0.1267 | Adds dense, diffuse surface cellularity while retaining elongated repetitive lumina; reject |
| 96 | 7.178376 | 215.9 | 847.0 | 0.0456 | Strong pink shift, simplified parallel pits, oversized isolated nuclei, and loss of realistic gland diversity; reject |
| 128 | 5.120024 | 318.4 | 939.3 | 0.1796 | Strongest block repetition, excessive open lumina, simplified cellular detail, and no convincing improvement over the reference; reject |

The zero-shot control measures 14.328430 mean absolute gradient and the untouched M1 real normal field 7.682061. Step 32 differs from the zero-shot pixels by mean absolute error 5.51844 but does not materially correct its composition. Sydney-QA block-feature maximum cosine rises from 0.9945 at step 32 to 0.9997 at step 128, consistent with the visible increase in repetition. These measurements are descriptive; sharpness and an uncalibrated classical feature count cannot qualify an image.

Sydney-QA extracted raw/normalized stain, nuclei, gland/lumen, mucin/vacuole, inflammatory-surrogate, bacillus-surrogate, and repetition features plus audit masks for all four cells. Every record remains `CALIBRATION_REQUIRED`. No GI-pathologist-annotated patient/slide-disjoint real envelope exists, so Sydney-QA cannot establish clinical correctness, diagnosis, Sydney grade, or OLGA/OLGIM and does not replace pathologist review.

## Runtime, cost, and cleanup

The worker ran for 607 seconds. Conservative A100 instance uptime is 800 seconds, approximately US$0.82 at the project's US$3.673385/hour list-price assumption. The bounded `e2-standard-4` recovery worker was chargeable for at most 816 seconds, approximately US$0.03 at US$0.134/hour. Total conservative compute is below US$0.85 before temporary storage, credits, and tax; the A100 itself was Spot.

The A100 and recovery CPU are terminated. After complete local verification, the temporary recovery instance, read-only disk, and snapshot were deleted. The original stopped Z-Image/Qwen disk remains retained.

## Decision and smallest correction

No checkpoint advances to another category, canvas, or candidate pool. The failure is specifically validation composition and topology, not integrity, privacy, rights evidence, or runtime.

The smallest causal follow-up is inference-only: validate the same four immutable whole checkpoints with the exact normal caption used for all 32 training fields. The original validation prompt was semantically similar but not byte-identical and omitted the frozen 0.25-micrometre/pixel scale language. `config/zimage_normal_prompt_alignment_v1.json` preregisters one exact training-caption prompt, the unchanged negative prompt, seed, geometry, 50 steps, CFG 4.0, model, and checkpoint hashes. It changes no weight or training variable. If no whole checkpoint corrects local gastric topology under that single aligned prompt, close this Z-Image LoRA path rather than opening a category or pool.