EPIC: Train the P0 clinical PHI high-recall flagship

[maziyarpanahi]

Summary

section 6.1 P0 family OpenMed-ClinicalPrivacy-: the reference recall-first clinical PHI model (names, dates, ages>89, addresses, IDs, providers, facilities), benchmarked vs SHIELD (public) + i2b2/n2c2 (gated eval). Status: Privacy Filter families exist but there is NO single named, recall-gated clinical flagship and NO published clinical-PHI benchmark. This is the moat-defining model program; decompose before starting via recipe mode C (large teacher, recall-first), benchmarking SHIELD, certifying against the ramping G1a floor (>=0.99 v1.6 -> >=0.995 v2.0), G2, and G3.

Scope

• Decompose before starting: dataset assembly (public SHIELD + synthetic + DUA eval-only), recipe-mode-C teacher training, SHIELD benchmarking, and gate certification.
• Certify against G1a (ramping recall floor), G2 (name/address/date 98-99%), G3 (leakage 0) on held-out eval + leakage fixtures; publish the first clinical-PHI benchmark numbers against SHIELD.

Acceptance criteria

• Decomposition issues created before any training run.
• A named OpenMed-ClinicalPrivacy- checkpoint reports clinical-PHI numbers against the public SHIELD corpus.
• The checkpoint clears G1a (>=0.99 at v1.6), G2, and G3 with a signed gate report; manifest entry + model card generated.
• test suite green: .venv/bin/python -m pytest tests/ -q

Out of scope

• The benchmark/gate harness and SHIELD adapter (OM-031b, OM-020, consumed).

Files

• openmed/training/configs/large_teacher.yaml
• openmed/eval/datasets/public.py

---

Task: OM-051 · Milestone: v1.7 · Priority: P0 · Size: XL
Depends on: OM-038a, OM-038c, OM-031b, OM-020 · Blocks: —
Roadmap: section 6.1
Spec: PLANS/V2/EXECUTION/tasks/OM-051.md