Language coverage diffs¶

openmed.core.language_coverage_diff compares declared coverage metadata offline; it does not run model evaluation or decide whether a locale is clinically fit.

Use diff_language_coverage(before, after) with versioned manifests:

from openmed.core.language_coverage_diff import diff_language_coverage

before = {"schema_version": "1.0.0", "packs": []}
after = {"schema_version": "1.0.0", "packs": [{
    "locale": "en-US", "scripts": ["Latin"],
    "identifier_classes": ["PERSON"], "surrogate_support": "approximated",
    "evidence_digest": None,
}]}
assert diff_language_coverage(before, after)["added"] == ["en-US"]

Locales canonicalize language case, script case and region case, including underscore aliases. Duplicate canonical locales reject. Capabilities are unique controlled tokens; surrogate states are missing, approximated, supported. An evidence digest is absent (None) or 64 lowercase hexadecimal characters. Manifests use schema version 1.0.0, with at most 1,000 packs and 256 entries in each capability list. Unknown metadata is not copied to output.

Added/removed locales are ordered. Shared locales may improve and regress simultaneously: gaining one script cannot conceal losing an identifier class. Replacing an evidence digest is a change, never inferred scientific improvement. Adding/removing evidence reports evidence availability only. Use render_language_coverage_diff for deterministic JSON. Outputs contain only locale codes and controlled change categories; no examples or model paths.