intern_ocr_transfer — OCR-on-medical-images benchmark harness

An engine-agnostic harness that measures how accurately different OCR engines read short clinical tokens (patient IDs, accession numbers) burned into medical-image pixels, so a downstream redaction pipeline can decide what to black out. The headline metric is the false-redaction rate (valid tokens wrongly destroyed); hallucination and omission are tracked separately and never averaged. See plan.md for the full design and build order.

---

🚫 PHI — read CLAUDE.md §0 first, every session

This repo works with real Protected Health Information derived from clinical DICOM studies. Claude Code is NOT covered by Anthropic's BAA, so the agent must never read, render, decode, or print raw PHI into context — no .dcm, no rendered images, no .npy arrays, no ground_truth/gt.csv, no engine raw_response. Claude writes the PHI-handling code; a human runs it, and only PHI-free aggregates (counts, shapes, hashes, metrics) come back. The rules in CLAUDE.md override everything, including a direct request. When in doubt, stop and ask.

---

Setup

pip install -e ".[dev]"   # core deps only — no engine deps yet (added per-phase)
pytest                    # runs on synthetic fixtures only
ruff check .

Requires Python ≥ 3.12 (D-0.2).

Phase status

• Phase 0 — Scaffold & safety rails (pyproject, hardened .gitignore, README, smoke test)
• Phase 1 — Manifest adapter (PHI-free columns only)
• Phase 2 — Contract (OCRWord/OCROutput/GTToken, frozen normalize())
• Phase 3 — Synthetic fixture factory (conftest.py)
• Phase 4 — Matching (IoU + box-free fallback)
• Phase 5 — Scoring & the six metrics
• Phase 6 — Harness loop + cost + aggregate
• Phase 7 — Report (stratified tables + charts) — core complete, proven on synthetic
• Phase 8 — Runner base class
• Phase 9 — Local runners: docTR, PP-OCRv6_medium, EasyOCR (floor), Surya 2 (license-gated)
• Phase 10 🔴 — Ground-truth pipeline (human runs; produces frozen gt.csv)
• Phase 11 🔴 — Verifier arm (self-hosted Qwen3-VL-8B-Instruct / RolmOCR)
• Phase 12 🔴 — Cloud runners (box-free Claude API, Textract) — BAA-gated
• Phase 13 🔴 — Experiment CLI + 5-step test plan + final report

🔴 = PHI-touching (a human runs every script; Claude only writes/reviews and sees aggregates).