How can generated records support executable propagation tasks?

A generated schema can separate `policy_versions`, `work_items`, `requirements`, `assignments` and `propagation_actions`. A work item can link to its governing policy through `policy_version_id`, while requirement records identify evidence that must exist before a particular transition.

Tool actions can expose business operations such as `assess_applicability`, `request_evidence`, `assign_review` and `complete_transition`. Each action should define required arguments, permission checks and state effects. A completion action can reject a request when the linked evidence requirement remains unsatisfied.

Use stable identifiers for joins and an explicit schema version for the package. JSONL can carry tool traces and messages; Parquet can carry larger related tables. Enum definitions and null semantics should explain whether an absent value means unknown, inapplicable or not yet collected.

A snapshot manifest can identify the initial state and the records included in a run. Evaluator hooks can inspect transition results and the final relationship graph. This supports tasks where success depends on changing connected business state, rather than placing a plausible explanation in a free-text answer while leaving the underlying work unchanged.