What file layout and fields suit a generated conditional approval release?

A release for conditional approvals could be delivered as linked JSONL files, one per record type, with Parquet copies for analytical use. Each file joins on stable identifiers.

- **decisions.jsonl:** decision_id, subject_ref, approver_role, decided_at, outcome, scope_text, policy_ref - **conditions.jsonl:** condition_id, decision_id, rule_text, owner_role, due_point, blocking, status - **evidence.jsonl:** evidence_id, condition_id, submitted_by_role, received_at, doc_type, extracted_fields, content_ref - **reviews.jsonl:** review_id, evidence_id, reviewer_role, finding, reason_code, reviewed_at - **changes.jsonl:** change_id, condition_id, change_type, authorised_by_role, effective_at - **steps.jsonl:** step_id, decision_id, step_name, depends_on_conditions, state

Links run one way: steps point at conditions, conditions at decisions, and evidence at conditions. A loader can rebuild the full case from a single decision_id.

Timestamps use ISO 8601 with an explicit offset, so the order of evidence against changes is unambiguous. Enumerations for status, change_type and reason_code are published in a schema file, so a validator can reject unknown values.

A buyer can also ask for a task manifest alongside the data, with one line per task: task_id, decision_id, start_state_ref, allowed_actions and expected_end_state. This keeps the data and the evaluation definition separate, so the same records can serve training and held-out testing.

The content is generated, so ask for a sample to inspect field completeness and referential integrity before committing.