Gate promotion of privately improved models

[maziyarpanahi]

Summary

A model trained across sites must not ship merely because the federated objective improved; privacy, leakage, critical recall, fairness, and artifact evidence are non-compensable.

Scope

• Combine privacy-account, direct-identifier recall, synthetic memorization, cross-site drift, lineage, license, and signed-evidence checks into a fail-closed decision.
• Publish separate privacy, utility, language, device, and assurance results so no aggregate score can hide a failed safety gate.

Acceptance criteria

• The implementation is deterministic and performs no mandatory network call.
• Logs, exceptions, reports, and fixtures contain no raw sensitive values.
• .venv/bin/python -m pytest tests/unit/eval/test_private_model_promotion.py -q

Out of scope

• A compliance certification or autonomous clinical decision guarantee.
• Bundling restricted datasets, credentials, or proprietary services.
• Unrelated changes to release publishing or existing public APIs.

Files

• openmed/eval/private_model_promotion.py
• tests/unit/eval/test_private_model_promotion.py
• docs/evaluation/private-model-promotion.md

---

Task: OM-180d · Milestone: v3.3 · Priority: P0 · Size: L
Depends on: — · Blocks: —
Roadmap: Audited V3 horizon expansion, 2026-08-13: OPENMED_V2_UNIFIED_ROADMAP.md section 6.4; OPENMED_BENCHMARK_PORTFOLIO_AND_TRUST.md section 8
Spec: PLANS/V2/EXECUTION/tasks/OM-180d.md