Clinical transliteration provenance¶

openmed.clinical.languages.transliteration_graph keeps caller-supplied local transliteration alternatives without choosing one as clinically equivalent. It bundles no restricted terminology, model, remote provider, or automatic transliteration engine. Callers retain their spelling lookup locally.

Create TransliterationNode.from_text(term, language=..., script=...) to retain an exact SHA-256 term identity without storing spelling in graph output. Nodes bind language and script, so identical spellings in different scripts remain separate. Supply TransliterationRule objects with source/target nodes, a controlled rule-source code, evidence digest, confidence hint, and explicit reversibility. A rule marked reversible may be traversed backward; a lossy rule cannot. The declaration is caller evidence, not proof of linguistic correctness.

TransliterationGraph.candidates(source, language="en", script="Latin") returns all simple routes within bounded depth. Routes to the same target retain distinct provenance. Cycles terminate. Results order by descending product of declared confidence hints, then deterministic metadata. Products are ranking hints, not calibrated probabilities. Reviewers must resolve contradictory alternatives.

Use each path's to_dict() for fingerprints and controlled metadata only. No term spellings enter exports or errors. Plain hashes are not anonymization guarantees; do not publicly release graph artifacts derived from private patient identifiers. Graph size is capped at 10,000 rules; depth at 8, path count at 1,000 and traversal steps at 100,000. Exceeding a budget raises a controlled error rather than silently returning an incomplete candidate set. No network calls or normalization alter caller terminology. User-supplied vocabulary stays under its original license.