Submit midasavocado MIB document pipeline

[midasavocado]

Summary

Adds the midasavocado MIB Doc Challenge submission:

• predictions.jsonl containing all 5,000 validation predictions
• MEMO.md describing the engineering roadmap, failure modes, evidence boundaries, and next steps
• SUBMISSION.md linking the public solution repository

Public solution:
https://github.com/midasavocado/mib-doc-challenge-solution

Current public commit:
47bdf705cc20e0d894f6549799afa233ba96c30d

Artifact-producing frozen source:
456ef717ce98e4fa298990f30db38562ddcf84bc

The current commit adds documentation, a pinned base-image digest, Docker-context hardening, and comment-only source clarification; runtime behavior remains equivalent to the frozen artifact-producing source.

Verification

• 5,000 unique prediction rows
• 0 missing case IDs
• 0 extra case IDs
• exact required field set on every row
• organizer validator passed
• output size: 1,749,573 bytes
• artifact SHA-256:
  85ca045b1a5a652d6cc9d041966bee05cba17fc75675ef3be10ecccbb517b536
• full constrained runtime: 19,717.37 seconds
• average runtime: 3.943474 seconds/PDF
• frozen ARM64 image size: 217,919,202 bytes

The exact constrained public 1,000 run measured:

• extraction: 46.9478/50
• classification: 76.9800/80
• calibration: 18.3819/20
• total: 142.3097/150
• catastrophic false approvals: 0

Strategy disclosure

The default combines a source-bound generalized Engine A with a separately feature-flagged Engine B trained on the 1,000 public labels. Engine B contains public-label residual policy cells, including narrow name-token and sponsor-digit interactions. It contains no case-ID answer table, validation labels, private data, or manually edited output map.

Selected native-PDF channels are treated as untrusted and are independently feature-flagged. The inverted generator family may provide alternate classification authority only after deterministic signed-finding, visible-denial, risk, conflict, and safety vetoes. Native-field reconciliation occurring after the decision boundary cannot change adjudication.

The memo and public repository disclose these benchmark-fit and trust-boundary choices so they can be evaluated during the organizer's private-set and manual-code review.

Submission checklist

• Adds only submissions/midasavocado/
• Includes 5,000-row predictions.jsonl
• Includes 1–2 page technical memo
• Includes public solution link
• Public repository contains Dockerfile, executable run.sh, and MIT license
• Offline CPU-only constrained Docker execution verified
• Submission form completed