Synthetic longitudinal trajectories¶

openmed.training.synthetic.longitudinal creates deterministic multi-encounter synthetic fact graphs and observable notes using existing ClinicalFact, ClinicalArtifact and EvidenceLocator Journey contracts. It introduces no alternate runtime fact schema, clinical decision or private dataset dependency.

Declare bounded TrajectoryEvent items with controlled event and encounter codes, effective and observed relative days, and explicit parent event codes. Encounters must exist, temporal and observed order must permit each parent, and same-day causal cycles reject. Delayed results preserve their effective day separately from when the observation becomes available. No clinical progression is inferred; causality and fictional event codes come from the declared synthetic recipe.

generate_trajectory(events, seed=..., start_date=..., date_shift_days=...) produces a stable patient identity, runtime contract objects, exact note evidence spans, content hashes and synthetic provenance. Event evidence binds to the rendered value after its field label, including codes that also occur inside the label. Uniform date shifting retains patient / fact identities and parent relationships while moving dates, preserving intervals. Only 1,000 events and 128 parents per event are accepted. No model/network/file write is implicit. validate_trajectory independently checks the runtime graph, causal time, subject/provenance, observable hashes and evidence bindings.

The report() projection contains counts, digests/opaque references and controlled metadata, never document text. Fixture notes are fictional placeholders, not physiological progression, clinical utility, compliance or diagnostic evidence. Keep all synthetic text and recipes separate from protected evaluation data.