# FLUX.2 Klein split-corrected normal challenger

## Result

**FAIL for candidate production.** Run `20260827T034257Z-flux2-klein-normal` completed successfully on deployed commit `6db7fb3c93242e76202c6d85c3e9d9055b23c536`, but no whole checkpoint met the frozen normal-gastric morphology gate. The preregistered stop rule closes the FLUX.2 Klein LoRA path: no second category, final canvas, or candidate pool may use these checkpoints.

## Integrity, rights evidence, and split

- Exit code is 0, `FLUX2_KLEIN_LORA_COMPLETE` is present, and all 57 entries in `ALL_SHA256SUMS` pass after local retrieval.
- The exact model is `black-forest-labs/FLUX.2-klein-base-4B` revision `a3b4f4849157f664bdbc776fd7453c2783562f4d`; publisher code is `50fe5162777813d869182b139e83b10743caef15`; official Diffusers is `ec94eec6cabd536c44c77647038303e96ab9355b`. Archived model, publisher, and Diffusers evidence records Apache-2.0.
- Immutable config SHA-256 is `4e5445f90b9edf1eeb75b6a793d53c66c2100df66c00bc778ac27a8a98aefb84`. The run used the official NF4 path because A100 compute capability 8.0 is below the official FP8 minimum of 8.9.
- 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 66,866,376 bytes. Their SHA-256 values are `708a08929835de9dc31c05b8123a21e19ef19b3a5a8ee31380597be952185f77`, `1a26969574247af242ae49f99dee4c9207738be06533142b895c8a90eec78af7`, `ebd08e1442efd0edacaac952c6272fc69ff43ee59e2225bc456707c77838e776`, and `c52a17b5ac5a19f685600d8387c8b80d0f6e732d85a0a54a44dd733d5a225771`.
- Four fixed-seed validation images are exact 1024x1024 RGB PNGs. The six-cell real/zero-shot/checkpoint review sheet SHA-256 is `a94c98966cb83123e87d281c17ae0674c27e732c35d39ead678566d4f440e714`.

## Privacy

Both preregistered screens pass. Against the corrected 32-field normal corpus, maximum DINO cosine is 0.845966 and minimum dHash distance is 21. Against the full corrected 128-field caption-neutral corpus, maximum DINO cosine is 0.845966 and minimum dHash distance is 21. Neither screen found an exact pixel match. DINO and dHash are privacy and 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 | Block max cosine | Full-resolution result |
| --- | ---: | ---: | ---: | ---: | --- |
| 32 | 2.999254 | 0.0 | 0.0 | 0.999986 | Pale repetitive micro-rosettes and oversized circular pseudo-glands, without recognisable gastric foveolar or gland architecture; reject |
| 64 | 3.077132 | 0.0 | 85.1 | 0.999972 | Enlarged repeating radial rosettes and isolated white clefts; more conspicuous non-gastric topology than step 32; reject |
| 96 | 3.523413 | 744.2 | 8,108.4 | 0.999875 | Dense repeated oval/radial units and extreme clear-region surrogate count; no orderly local normal gastric topology; reject |
| 128 | 4.215285 | 0.0 | 0.0 | 0.999594 | Fine but nearly uniform nested ovals with strong pink/blue-channel saturation and no credible cellular or gland diversity; reject |

All four checkpoints preserve or rearrange the zero-shot model's synthetic micro-rosette shortcut instead of converging toward the untouched M1 real normal field. Steps 32-64 add large radial structures; steps 96-128 collapse into dense, repetitive oval texture. The outputs lack credible nuclei, epithelial boundaries, lamina propria, and mixed-scale gastric glands. They are also visibly inferior to the already rejected Z-Image and Qwen controls in cellular detail. Sharpness increases modestly across checkpoints but cannot qualify structurally incorrect tissue.

Sydney-QA extracted raw and 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`. The near-one block repetition scores and unstable classical segmentation are descriptive engineering warnings, not clinical grades. 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 578 seconds. Conservative A100 instance uptime is 761 seconds, approximately US$0.78 at the project's US$3.673385/hour list-price assumption. The bounded `e2-standard-4` recovery worker ran for 386.55 seconds, approximately US$0.01 at US$0.134/hour. Total conservative compute is below US$0.80 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 model disk remains retained.

## Decision

No checkpoint advances. The failure is normal-gastric topology and category morphology, not integrity, privacy, rights-evidence capture, runtime, or GPU fit. The corrected Qwen, Z-Image, and FLUX.2 general-image LoRA routes have now each failed the same frozen normal-only gate. Another category-specific adapter on these backbones is not justified.

The next step is CPU-only: audit released pathology-native generators or pathology-specific super-resolution systems with an explicit submission-compatible model, code, and component licence chain. Any new GPU experiment requires a fresh normal-only preregistration against the same corrected split, untouched reference, privacy thresholds, Sydney-QA uncertainty, cost controls, and whole-checkpoint rule. Generic super-resolution remains ineligible because it can sharpen incorrect synthetic topology without restoring it.