MIB Doc Challenge solution (mkoker)

Offline document pipeline for the 8090 MIB intake challenge. Reads a directory of PDF case packets, writes predictions.jsonl. No network, no GPU, CPU only.

Core idea: the rendered page is the source of truth, not the PDF text layer. Every page gets rendered and OCR'd; text layer content is only trusted when it's confirmed visible in the pixels. That kills the hidden-text attacks by construction.

Layout

• Dockerfile, run.sh: the submission contract. Image takes <input_dir> <output_path>.
• src/extract.py: OCR-first extraction, visibility checks, visual detectors (illegible biometrics, green approval stamp).
• src/rules.py + RULES.yaml: rules engine. Every rule carries a citation to the field manual or the mined evidence with support counts.
• src/calibrate.py: confidence = observed tune-split accuracy of the fired rule path.
• tests/: synthetic attack PDFs (white on white, off-crop, hidden decoys, fake answer key) pinning the injection defense. pytest tests/.
• NOTES/: working notes, measurements, and the math behind design calls (abstention economics, calibration fits, adversarial audit, dropped detectors).
• split.json + scripts/make_split.py: the 801/199 tune/holdout split by sha256(case_id) used for all tuning and gating.

Run it

docker build -t mib-submission .
docker run --rm --network none \
  --mount type=bind,src=/path/to/pdfs,dst=/input,readonly \
  --mount type=bind,src=/path/to/out,dst=/output \
  mib-submission /input /output/predictions.jsonl

Local scores via the official runner and evaluator: DEV 111.53 / HOLDOUT 110.77 of 150, gap 0.76. Details in the technical memo in the challenge-repo submission folder.