Synthetic FHIR records¶

openmed.training.synthetic.fhir_records.generate_fhir_records consumes existing ClinicalFact graph objects explicitly marked synthetic. It emits an offline R4 collection Bundle containing Patient, Encounter, Condition, MedicationStatement, Observation, Procedure and source-bound Provenance fixtures. No transaction requests, server calls, write-back, terminology download or patient identifiers are involved. Local placeholder coding uses urn:openmed:synthetic-facts.

Facts must contain only a controlled synthetic_code value, valid effective and observed dates, matching subject/encounter references, and existing causal parents. Missing encounters reject rather than being fabricated. Facts are sorted before assembly; patient/resource identities and Bundle full URLs are stable for an identical graph and seed. The existing R4 Bundle assembler handles reference rewriting. Every fact gets source provenance linking its opaque runtime identity. Encounter class AMB is a declared synthetic fixture default; clinical statuses are unknown where appropriate, not inferred from model confidence.

validate_synthetic_fhir independently checks referential closure, unique full URLs, required fields/types in OpenMed's bundled R4 subset, provenance shape, synthetic tags, time order and absence of write requests. Patient/Provenance receive fixture checks because the bundled base validator warns for unsupported resource types. Each encounter/clinical resource requires a distinct source-fact provenance binding; subject references must resolve to Patient and encounter references to Encounter with the same subject. Recorded provenance timestamps must be timezone-aware and cannot precede their target's declared effective time. For this fixture subset, date-only values represent UTC midnight when compared with timezone-aware timestamp instants. This convention does not infer actual clinical time zones or recover event times that were never declared. This is not full FHIR conformance, an implementation-guide certification or clinical validity. Reference-server compatibility remains the separate conformance workstream. The graph is capped at 1,000 facts and the validator at 3,000 entries. All data must be synthetic, with no restricted terminology or real-record import disguised as a fixture.