Submission: pangoleen — measured visible-evidence pipeline, 126.72 train, 0 catastrophic false approvals

[pangoleen]

Solution repo: https://github.com/pangoleen/mib-doc-challenge-solution

126.72 / 150 on the labeled training set (official scripts/evaluate.py, confidences 5-fold out-of-fold) · 128.46 / 150 on a held-out 200 · 127.19 / 150 re-measured inside the shipping Docker image · zero catastrophic false approvals in every run ever measured.

Runtime 2.76 s/PDF against the 6 s/PDF budget; the full 5,000-case validation set completes in 4.6 h against the 30,000 s cap. Image 0.56 GiB, bundled model artifacts 29.4 MiB, nothing downloaded at runtime.

Six deterministic stages. Adjudication is a hand-written rule cascade — nothing is learned about when to approve. Four small artifacts are fitted from the corpus (lexicons, a character-confusion matrix, per-source error rates, the calibrator); the calibrator only sets the emitted confidence, which never feeds adjudication and so cannot manufacture a false approval. The governing rule throughout: hidden text may block an approval, never grant one.

The solution repo's notes/ carries the full measurement record — every accepted and rejected experiment with the number that decided it, including an EV layer worth +4.42 train points at 23 false approvals that was measured and declined.

Attribution: multi-pass OCR pooling, the largest single gain here, was found by auditing #51 (arthurmichel00); the top-band crop variant came from #39 (zeroinfinity03). No code was taken from either — implementation, variant set, measurements and safety guard are ours, and four of the passes those pipelines rely on we measured and rejected.