Synthetic prior-authorization packets¶

openmed.training.synthetic.prior_auth assembles deterministic synthetic notes, orders and forms sharing one opaque case identity and controlled diagnosis and treatment placeholders. It bundles no payer policy, insurer vocabulary, actual records or coverage decision service.

Declare AuthorizationCriterion(code, document_kinds, required=True) and synthetic AuthorizationEvidence(document_kind, code, present, distractor=False) items inside an AuthorizationCaseSpec. A criterion only uses evidence from allowed document kinds. Distractors remain labeled in document truth and never satisfy a criterion. Missing evidence remains missing. Conflicting true/false observations produce conflicting, never an arbitrary winner. Required missing/conflicting criteria produce abstain; explicit false evidence yields requirements_not_satisfied; complete positive evidence yields ready_for_review, not authorization approval.

generate_authorization_packet(spec, seed=...) produces byte-stable template text, exact source spans, traceable evidence references and gold statuses. Seeded opaque identities are stable for a declared recipe. validate_authorization_packet independently checks shared identity and placeholders, gold spans, references and review state. Inputs are bounded to 128 criteria and 1,024 evidence items; invalid or duplicate declarations reject with controlled errors. All fixture text is synthetically rendered from controlled codes; do not pass patient identifiers. No network calls, patient data, restricted terminology or autonomous clinical / payer decisions are involved. These templates test evidence-gap handling rather than medical plausibility or real-world prior-authorization success rates.