How could a generated release represent policy versions and their linked work records?

A release can separate policy_versions, clauses, work_items and obligations into linked tables. Policy versions can contain policy_id, version_id, supersedes_version_id, published_at and effective_at. Obligations can contain obligation_id, work_item_id, clause_id, status and evidence_record_id.

The distinction between policy_id and version_id preserves continuity without overwriting history. A clause link identifies the specific requirement behind an obligation, while the work-item link identifies where that requirement must be satisfied. Where applicability is undecided, the schema should represent that state explicitly instead of filling a version field with a guess.

JSONL can carry nested work-item snapshots for agent tasks. Parquet can support joins and cohort analysis across the same entities. A schema manifest should define timestamp conventions, enumerations, nullable fields and foreign-key expectations so format changes do not silently change meaning.

Generated-record provenance can include release_id and generation_batch_id without embedding source business records. A buyer can request validation results for orphaned links, invalid version chains and obligations lacking an authority reference. Those checks make the exported structure reviewable before it becomes part of a training pipeline.