Unit-display normalization audit¶

OpenMed provides a deterministic, offline audit for localized unit display labels. It checks a synthetic or caller-owned locale table against the allowed canonical unit codes and an explicit alias table. The audit is a consistency check for exports and review tooling; it is not a compliance certification or a clinical decision guarantee.

Run an audit¶

The input table is keyed by locale and then by canonical unit code. Each value is the display label that the locale would render. Alias tables map a localized label back to its canonical code.

from openmed.clinical.units.display_audit import audit_unit_display_labels

report = audit_unit_display_labels(
    {
        "en": {"mg/dL": "mg/dL", "Cel": "Celsius"},
        "es": {"mg/dL": "mg por decilitro", "Cel": "℃"},
    },
    ("Cel", "mg/dL"),
    {
        "en": {"Celsius": "Cel"},
        "es": {
            "mg por decilitro": "mg/dL",
            "℃": "Cel",
        },
    },
)

report.summary
# {
#   "canonical_units": 2,
#   "conflict": 0,
#   "display_labels": 4,
#   "duplicate": 0,
#   "issues": 0,
#   "locales": 2,
#   "missing": 0,
# }

When alias_tables is omitted, the registered local clinical normalization lexicons provide the alias tables. Supplying the mapping explicitly is preferred for release audits because it makes the audit input self-contained. Locale identifiers are normalized to their primary language subtag, so fr-FR and fr cannot be silently audited as separate language packs.

Findings and privacy¶

The report marks each finding as missing, duplicate, or conflict:

• missing means a canonical code has no non-empty display label in a locale.
• duplicate means one normalized display surface is used for more than one canonical code in the same locale.
• conflict means an alias table is internally inconsistent, a label resolves to a different code, a label references an unknown code, or a non-canonical label is absent from the explicit alias table.

to_dict(), to_json(), and to_markdown() contain counts, normalized locale identifiers, issue reasons, and SHA-256 hashes of labels and codes. They never include submitted display labels or canonical code strings. The reproducibility hash covers that source-free payload, and repeated runs with the same inputs produce byte-identical JSON.

The implementation performs no network calls, emits no logs, reads no wall clock, and does not alter the registered lexicons. Keep source tables synthetic or apply the caller's data-access and PHI handling policy before invoking the audit.

Validation limits¶

Public issue/report constructors validate controlled reasons, locale syntax, SHA-256 fields, schema and the computed report hash. Nonempty catalogs require a locale; an empty locale table cannot claim successful coverage. Collections are bounded to 4096 entries each and text fields to 4096 characters. JSON indentation must be an integer from 0 through 8. Invalid inputs produce fixed errors without retaining source-bearing exception context. Hashes are consistency identifiers, not proof of authenticity or protection against guessing a small label vocabulary.