# H01: geometry responds, tissue-quality gate does not pass

7 September 2026. Four-job run finished14:23:13UTC; recovered on14:53UTC wake. Single development source/seed, not clinical validation or generalization.

## Visual result

Human annotations improve visible organization over uniform initialization. Rotating the scaffold changes major white-channel and gland-boundary positions toward its supplied target. But all generated arms remain too smooth/painterly: oversized pale compartments, thick smeared nuclear-looking bands, and poorly resolved intervening tissue. **H01 fails its frozen tissue-quality advance gate**, despite a useful spatial-control signal. No seed/strength search or H03 orientation expansion of this recipe.

| Arm | Evidence | Verdict |
|---|---|---|
| Codec | Coarse contours and pale/white marks preserved | Supplied coarse cues survive reconstruction |
| Uniform | Washed-out fields, fragmented walls, poorly bounded pockets | Weak baseline |
| Human | Clearer profiles, traceable U-shaped right wall and enclosed pockets; smooth tiled compartments remain | Relative improvement, not resolved tissue |
| Rotated | Major white space/profile arrangement changes; boundaries merge and texture is inconsistent | Spatial response, incomplete tissue rendering |

Root inspected full-resolution named outputs and real reference; root review was not blinded. A separate Astra reviewer saw only anonymously named A/B/C images and morphology criteria. Its frozen ranking: **C > A >> B**, moderate confidence C/A and high confidence B weakest; none convincingly resolved tissue. Unblind:C=human,A=rotated,B=uniform. No quantitative crop score, significance or pathologist-review claim.

## Actual images

- Human-layout output
- Uniform control
- Rotated-layout control
- Sketch codec
- Real reference, not generated

## Verified execution

All4jobs/33denoiser calls, driverexit0. Exact full-model latent/PNG parity and all4 latent/PNG replay checksPASS. Recovered42portablefile hashesPASS; allcapturedtensorsfinite; identical epsilon/token tensors across generated arms,3different intended start latents. H01_A1_AUDIT.json owns independent local checks. ArchiveSHA88f28427288221f5bd8bed1cfc998de28606c8750f43eec29ab139c9dc719cf0.

First local checksum invocation used archiveparent instead of run-a1; corrected workingdirectory, then42/42passed without changing files. Source copying is NOT cleared. Reference-like arrangement persists even in uniform, with similar-position red/purple details across arms. No novelty/privacy claim; seven-reference multiscale copying screen remains required before promoting any later output. UnknownMPP stays unknown.

## Resources

OwnVM8097050405901662678 verifiedTERMINATED14:56:32.851UTC; boot14:16:38.320UTC,2394.531poweredseconds. Historical planningrate3.673385USD/hour gives2.44334USDcomputeestimate, excludingstorage/tax/agents; notinvoice. Actualworker~115seconds.30minwake left unnecessary postrunidle; future shortjobs need completion-aware retrieval or shorterwake. Own350GBdisk4073394244074830604 retained,storagecontinues. OriginalFLUXresourcesuntouched.

## Next diagnostic

Astra recommends H04: small adapter explicitly for human[lumen,epithelium,mucin], frozenPixCell/UNI, generation fromnoise. Target: learn spatial-mark-to-tissue translation rather than reinsert originalpixels. H02real-latent preservation could improve mainly by copying; deferred, not disproven. H03notadmitted becauseH01tissuegatefailed.

H04 CPUpreparation delegated to finiteSolhelper; **no H04GPU/training launched**. Root must review/freeze recipe/schema, zero-contribution latent/PNGparity, human/moved/disabled controls, fullruntime plan and copying evaluation beforelaunch. One-pairfit is technicalfeasibility only, notgeneralization from7correlatedreferences.