Synthetic cross-artifact truth alignment¶

validate_truth_alignment compares artifact-declared facts, times, labels and coordinates against shared ClinicalFact and EvidenceLocator records from the existing Journey contracts. SyntheticArtifactView groups those records with an existing ClinicalArtifact and a controlled note/FHIR/OMOP/image/waveform modality; it does not introduce another clinical truth graph.

All records must explicitly declare synthetic origin. Each view declares the same truth subset; use separate packages for intentionally partial views. The validator checks medication/value and unit changes, effective times, fact labels, missing/unexpected facts, evidence references, coordinate transforms and artifact bindings. Both the shared truth locator and the observed locator must bind to the view's artifact; copying a locator ID onto another artifact cannot qualify that artifact. It never fabricates missing modalities. Reports use hashed artifact/fact references and controlled mismatch categories, not submitted values.

Intentional corruptions require a named negative-control fixture, explicit expected mismatch categories and the negative_control partition. Such fixtures can pass their negative-control check but always remain valid_for_training=false. A normal partition rejects intentional corruption metadata. The validator checks declared consistency, not clinical plausibility, source-rights adequacy or privacy. No generator, network, model, or publication is invoked. Synthetic flags and opaque references are assertions requiring independent provenance review.

Shared fact IDs also retain their declared synthetic subject and encounter bindings. Changing either emits fact_identity without including identity values; an intentional identity corruption must use the named negative-control partition.