What fields and links make up a generated approval record schema?

A release can ship as one file per table, in JSONL for readability or Parquet for volume. Each row carries a stable identifier, and links are explicit fields rather than text references.

- `approval_request`: `request_id`, `request_type`, `amount`, `currency`, `risk_flags`, `requester_role_id`, `related_request_ids`, `raised_at`. - `authority_rule`: `rule_id`, `request_type`, `amount_min`, `amount_max`, `risk_condition`, `required_role_id`, `aggregation_scope`, `valid_from`, `valid_to`. - `role_holder`: `role_id`, `holder_id`, `valid_from`, `valid_to`, `absence_status`. - `delegation`: `delegation_id`, `from_role_id`, `to_holder_id`, `scope_rule_ids`, `starts_at`, `ends_at`, `sub_delegation_allowed`. - `conflict_rule`: `rule_id`, `condition`, `effect`, such as exclude or require second approver. - `decision_event`: `event_id`, `request_id`, `approver_id`, `authority_basis` (a rule or delegation id), `decision`, `decided_at`. - `expected_outcome`: `request_id`, `valid_approver_ids`, `expected_decision`, `acceptable_escalations`.

The key link is `authority_basis` on each decision, because it lets a grader check whether the approver's authority was valid at `decided_at`.

Keep `expected_outcome` in a separate file so it can be withheld from training splits. Version the rule and delegation tables together, since a decision is only meaningful against the rules in force at its date.

All values are generated. Field names can be adapted to your own conventions as part of scoping.