Privacy-policy coverage matrix¶

OpenMed can produce a deterministic crosswalk between bundled policy actions, the canonical label taxonomy, structured-column semantics, synthetic fixture identifiers, and focused tests. The artifact is a release-review aid; it is not a compliance certification or a clinical decision guarantee.

Generate the matrix¶

Generation uses bundled package resources and identifier-only coverage bindings. It does not load fixture payloads, inspect user data, or make a network call.

from pathlib import Path

from openmed.compliance import generate_policy_coverage

result = generate_policy_coverage(Path("audit-evidence/policy-coverage"))
print(result.manifest_path)
print(result.markdown_path)
print(result.matrix.coverage_percent)

The destination contains:

• policy-coverage.json: a machine-readable matrix with policy resource hashes, row identifiers, action names, counts, statuses, and binding hashes.
• policy-coverage.md: the same matrix rendered for review.

Repeated runs against the same local policy resources produce identical bytes. No timestamp, host path, fixture text, span, or source value is included.

Matrix semantics¶

Each row is one canonical label action in one bundled policy profile. The resource path points to the action key in the versioned policy JSON. Rows with an action other than keep are required rules. A required row is covered only when it has both a synthetic fixture identifier and a focused-test reference; missing required evidence raises UncoveredPolicyRuleError by default.

The matrix also records the structured semantic fields that map to each canonical label. The structured-field map is hashed independently so adding, removing, or relabeling a field changes the evidence fingerprint. The bundled fixture catalog uses category identifiers such as synthetic-policy-direct-identifiers; it deliberately contains no fixture payloads or sensitive values.

Review boundary¶

The matrix proves that policy definitions have declared local evidence links. It does not measure model recall, certify a jurisdictional control, or replace focused behavioral tests. Reviewers should inspect the referenced tests and run the narrow policy-coverage test before relying on the artifact.

The referenced focused test runs every matrix row through the public de-identification policy pipeline with synthetic category fixtures and checks the effective action or unchanged output for keep rules.