Experiment report library
Full public exports of the available research reports, including negative results and corrections. Operational paths have been omitted. Source report hashes are retained in the report index.
8–9 September programme
- A1-V2 result
- A1-TISSUE result
- A4-TOKENS: do layout-free tokens let the cartoon survive pass 2?
- A5-MATCH: if the tokens decide the layout, choose tokens whose real layout matches the cartoon
- B1-ORGANISMS: looking for real H. pylori on the HiESD gastritis slides
- C1-ENVELOPE: a classical morphometric realism instrument, calibrated on real gastric tissue
- C2-PROBES: pathology-encoder probes on the generated fields
- EVAL-BUILD: a rerunnable evaluation harness for the Phase-2 sweep
- EVAL-RUN: measuring the Phase-2 sweep
- F2-PAIRS: (label map, cartoon, real) pairs and a runnable structure-adapter package
- F2-TRAIN — the structure adapter, trained on an A100 and measured against the two-pass recipe
- F3 — adapter-guided repaint, and a coarse-label adapter
- GEN-CODE result, 8 September 2026
- SWEEP-RUN result, 8 September 2026
- SWEEP-V2: the v2 cartoon sweep, and what the generator did with it
Earlier experiments
- FLUX.2 Klein split-corrected normal challenger
- Image-path reopening audit
- M1 report — foundations and zero-shot baseline
- Pathology-native and pathology-SR route audit
- Q0: pinned Qwen-Image-2512 zero-shot baseline
- Q1: Qwen-Image-2512 LoRA infrastructure gate
- Q2 caption-neutral correction: tissue-scaffold gate
- Q2 no-overlay ablation: morphology gate
- Q2 tissue-prompt ablation gate
- Q2 v1: Qwen-Image-2512 morphology gate
- Q3 corrected exact-canvas gate report
- Q3 exact-canvas gate report
- Q3 detail denoising ablation
- Q3 corrected normal specialist
- Slide-split audit report
- Sydney-QA v1 implementation report
- Z-Image split-corrected normal challenger
- Z-Image exact-caption alignment
- Cartoon-started generation and lumen locking (7 September 2026)
- H01: geometry responds, tissue-quality gate does not pass
- H04: modest local improvement, no convincing layout independence
Selection and source evidence
- REAL-SOURCES — result
- SUBMIT-SELECT: how the ten were chosen, what they show, and what is wrong with them
- Copy screens on the ten final files
- INVENTORY — what has actually been downloaded, and whether it can serve the 70 real images
- SOURCES — public gastric H&E sources assessed against 4096x2048 @ 0.25 um/px