Clinical language readiness¶

The clinical readiness registry is independent of installed locale preprocessing packs. Being able to route a locale does not establish clinical capability.

openmed.clinical.languages.readiness provides immutable LanguageEvidence, ClinicalLanguageReadiness, and a private ClinicalLanguageReadinessRegistry. Evidence contains controlled kind/task/tier codes and SHA-256 references only. Do not put reviewer names, patient examples, model paths or free text into it.

For each declared task, supported requires lexicon, fixture, evaluation and native-review digests, an approved reviewer status, and device evidence for every declared model tier. Script and task inventories must be nonempty. Missing foundational evidence yields experimental. Foundational evidence with missing device evidence, no tier inventory, or declared limitations yields preview. Limitations always cap a complete entry at preview. Digests are references, not proof that a study was scientifically valid. A supported declaration makes no certification, diagnostic, or autonomous clinical decision claim.

from openmed.clinical.languages.readiness import (
    ClinicalLanguageReadiness, ClinicalLanguageReadinessRegistry,
)

row = ClinicalLanguageReadiness(
    language="ne", version="1.0.0", scripts=("Devanagari",), tasks=("entities",),
)
assert row.state == "experimental"
report = ClinicalLanguageReadinessRegistry([row]).report()
assert report["languages"][0]["state"] == "experimental"

Language underscore aliases normalize and duplicate canonical entries reject. Capabilities are sorted, evidence is bounded and task/tier bindings validate. Reports are deterministic and require no model, network, weights or real text. There is no process-global registration or implicit promotion of an installed pack. Native reviewers should examine referenced evidence and known limitations; this inventory does not replace that review.