Add submission: 0xUrsanomics

[0xUrsanomics]

Adds submissions/0xUrsanomics/ with predictions.jsonl (5,000 of 5,000 validation cases), MEMO.md, and SUBMISSION.md. No files outside that folder are touched.

Solution repository: https://github.com/0xUrsanomics/mib-doc-challenge-solution (public, MIT, Dockerfile at root).

Score on the labelled train split via scripts/evaluate.py: 101.42 / 150 (fields 40.37/50, classification 48.94/80, calibration 12.11/20, missing-case penalty 0/10). Predictions were produced by that image under --cpus 4 --memory 8g --network none --read-only, 0.091s per PDF against a 6s budget.

No model at runtime: PyMuPDF for text, tesseract for scan pages, hand-written rules for adjudication.

The trust layer is the part of the write-up I would point at first. Across the 188 train packets carrying a planted answer key, the key's field values agree with ground truth 93.1% of the time while its adjudication agrees 0 times out of 188, so the verdict is refused unconditionally and the field values are read only as a last resort, after adjudication has already returned. Details and known failure modes in MEMO.md.