Organ-on-a-chip QC — reference replay and exploratory field scores
Research prototype. Benchmark numbers are on withheld half-session proxies, not identified physical chips. The field budget counts inputs consumed by the decision rule, not measured microscope or inference time.
Bundled Python-reference replay
These buttons replay probabilities computed by the Python tool on the bundled 1,024 × 768 images. This is not live inference and not exact reproduction of the full-resolution report.
Exploratory live uploads
Live uploads do not emit PASS/FAIL or chip confidence. Browser decoding and preprocessing have not been validated as operationally equivalent to the reference pipeline. Only field scores are displayed. Do not use this preview as a chip-quality or clinical decision.
Load 8-bit grayscale/RGB still PNG images (no JPEG/BMP, animation, profile, interlace, or transparency) to explore field probabilities.
Field probabilities
| Field | P(bad) | Used by reference rule |
|---|
Scope and limits
The Python reference attained 19/23 called half-session proxies (82.6%), with four wrong calls, three at confidence ≥ 0.9. Its minimum of eight fields was chosen on that test set. The untuned minimum scored 0.708 on 24 calls or 0.680 with every proxy called. The rule consumed 9.5 of 27.4 fields at the same observed forced accuracy 0.800; all available fields are still scored for the QC view. This is not a validated accuracy or time-saving guarantee.
Stored borderline images give a low-confidence PASS in the reference replay, while full-resolution Python input is inconclusive. Previous canvas resizing could change even the chip call. The new resampler remains fail-closed until its actual browser input path is verified; matching tensors alone is not sufficient.
Code, report and local Python reference · OOC Image Dataset (Zenodo CC-BY-4.0; descriptor CC-BY-SA). The 44 examples retain their attribution. Pillow resampling code is MIT-CMU; see license.