Submission v3: arthurmichel00 — 129.05 official train / 130.50 holdout (supersedes #58)

[arthurmichel00]

MIB Doc Challenge Submission (v3 — supersedes #58)

Per the organizers' guidance that updated submissions are filed as a new PR: this replaces #58 (same author), which replaced #51. Delta vs #58: a constrained-candidate CTC fill channel for the four closed-menu fields — scoring every legal value against the recognizer's own frame posteriors instead of decoding its argmax — plus three text-level decode-repair mechanisms (fusion/ligature edit costs, a label-licensed truncation-prefix rule for clipped flag rows, and a cross-page sponsor digit vote). All four are flag-gated and were measured as one A/B arm on the full official train set: +0.26 train / +0.25 holdout, with every adjudication flip verified against gold. Two adjudications changed, both NEEDS_REVIEW → DENIED on gold-DENIED cases; no case moved toward APPROVED, and the catastrophic false-approval count is unchanged. Field-level audit: 30 changes right, 3 wrong, 7 neutral — the 3 wrong are itemized in the memo and LEVERS.md rather than left for a reviewer to find. Runtime moved from 5.64 to 5.81 s/PDF against the 6.0 budget; the added pass costs ~0.3 s on each of the ~20% of cases that trigger it, and that cost is disclosed alongside the points.

Links

• Solution repository (public, contains a Dockerfile): https://github.com/arthurmichel00/mib-doc-solution
• Superseded submissions: Submission: arthurmichel00 — measured, fail-closed document pipeline (128.69 official train / 130.25 holdout) #51 (128.69) and Submission v2: arthurmichel00 — 128.79 official train / 130.25 holdout (supersedes #51) #58 (128.79), same author.
• Submission form: re-submitted with this PR's link (earlier entries for Submission: arthurmichel00 — measured, fail-closed document pipeline (128.69 official train / 130.25 holdout) #51 and Submission v2: arthurmichel00 — 128.79 official train / 130.25 holdout (supersedes #51) #58 on file).

Summary

Official Docker-contract train score: 129.05 / 150 · fixed 200-case holdout: 130.50 (holdout ≥ train at all 9 measured milestones) · exactly 1 catastrophic false approval (the documented designed trap) · 0 fallback rows · 5.81 s/PDF vs 6.0 budget · 0.36 GiB image. predictions.jsonl generated by the frozen image over all 5,000 validation PDFs under the offline contract; passes validate_submission.py (5,000 valid, 0 missing). Approved-side consistency enforced structurally at emission and separately re-audited: 0 violations. No answer-key ingestion — the planted key is adversarial and never read as evidence. Built by AI agents in gated autonomous loops (30+ levers A/B-measured, 14 shipped; receipts in the repo's LEVERS.md and APPENDIX.md, including the full flip audit and timing ladder for this build).

Checklist

• I filled out the submission form linked above
• This PR only adds submissions/arthurmichel00/predictions.jsonl, MEMO.md, and SUBMISSION.md
• predictions.jsonl passes scripts/validate_submission.py against data/validation_manifest.csv
• My solution repository is public and includes a Dockerfile
• My Docker image runs offline (--network none) and accepts <input_pdf_dir> <output_predictions_path>
• My submitted runtime uses no LLMs, VLMs, cloud OCR, or network services
• Model artifacts fit the size limits in DOCKER_SUBMISSION.md
• No hardcoded validation answers and no manual per-case edits
• My memo describes my approach, failure modes, and what I would improve with another week

🤖 Generated with Claude Code

[arthurmichel00]

Superseded by #83 (same author) — v4 adds a generator-inversion last-resort reader for the closed menus, in-house likelihood fusion across views and pages, and a strip-scoped fine-tuned recognizer, measured +0.09 official train / +0.07 holdout with 0 adjudication flips and the catastrophic false-approval count unchanged at the single designed trap.