MIB Doc Challenge - offline visible-evidence submission

Offline CPU submission for the 8090 MIB Document Challenge.

This solution scores 137.2283796025 / 150 on all 1,000 public training packets under the organizer's own evaluator:

Section	Score
Field extraction	45.1177777778 / 50
Adjudication	74.2900000000 / 80
Confidence calibration	17.8206018248 / 20
Total	137.2283796025 / 150

It emits 1,000 schema-valid rows with no omissions and no catastrophic false approvals. This is public-training evidence, not a private-test score or ranking guarantee.

Every push runs .github/workflows/public-contract.yml, which builds this repository from a clean clone, installs the pinned runtime with --require-hashes, runs the full test suite, builds the submission image, and asserts the fail-closed exit-64 contract. Reproducibility is checked, not claimed.

Approach

The stable core renders every page, recovers visible text with RapidOCR and targeted Tesseract fallbacks, links evidence across pages, resolves conflicting fields by document precedence, and applies conservative policy rules.

A separate final scoring layer then:

• repairs fee, name, visa, sponsor, arrival, and purpose fields;
• honors visible findings and damage cues;
• detects narrowly gated visible blue and red denial marks;
• promotes only tightly gated, cross-form clean packets;
• demotes unsafe approvals and hard-gates TRANSIT-7;
• blends confidence with frozen identity-free reliability features;
• applies a disclosed public-train-selected terminal ensemble that never creates approvals and whose transfer ablation did not show reliable positive mean lift.

Raster OCR and visible geometry are the primary evidence path. Sanitized native selectable text may nominate pages, regions, or field candidates. It is not authoritative for approval or denial: decision-changing selectable cues require independent raster or pixel corroboration. Some field recovery still uses sanitized selectable layout, so this is not a claim that arbitrary hidden text can never influence extraction.

There is no answer-key parser, opt-in switch, case-ID lookup, filename feature, network call, LLM, or VLM. The image runs with a read-only root filesystem and retains UID 0 solely so it can write the official runner's host-created output bind mount. It uses no network, GPU, API key, external service, or runtime model download.

Run

docker build -t mib-visible-submission .
mkdir -p /tmp/mib-output
docker run --rm --network none \
  --read-only --tmpfs /tmp:rw,nosuid,nodev,size=2g \
  --mount type=bind,src="/absolute/path/to/pdfs",dst=/input,readonly \
  --mount type=bind,src="/tmp/mib-output",dst=/output \
  mib-visible-submission /input /output/predictions.jsonl

The image accepts exactly two arguments: the input PDF directory and exact output JSONL path.

Use the organizer's exact runner and validator:

python3 scripts/run_docker_submission.py \
  --repo . \
  --input-dir /path/to/mib-doc-challenge/data/validation \
  --output /tmp/mib-output/predictions.jsonl \
  --manifest /path/to/mib-doc-challenge/data/validation_manifest.csv \
  --timeout-seconds 30000 \
  --require-complete

python3 scripts/validate_submission.py \
  --submission /tmp/mib-output/predictions.jsonl \
  --manifest /path/to/mib-doc-challenge/data/validation_manifest.csv \
  --require-complete

See MEMO.md for the score audit, transfer warning, and failure modes, and ATTRIBUTION.md for exact upstream commits.