What tables, keys and file formats would a generated intake release use?

A release can be delivered as linked tables, in JSONL for record-at-a-time loading or Parquet for analysis, with the same schema in both. Field names below are a starting point and can be adapted.

- `cases`: `case_id` (for example `APP-2187`), `received_at`, `channel`, `case_type`, `raw_text`, `status_initial`, `party_id`, `parent_case_id` - `parties`: `party_id`, `party_type`, `display_fields`, `duplicate_group_id` - `documents`: `doc_id`, `case_id`, `doc_type`, `text_rendition`, `declared_period`, `quality_flags` - `policies`: `policy_id`, `case_type`, `rule_text`, `required_items`, `effective_from`, `effective_to` - `routing_rules`: `rule_id`, `criteria`, `queue`, `priority`, `requires_approver_role` - `events`: `event_id`, `case_id`, `actor`, `action`, `payload`, `at` (empty at start, filled during runs) - `answer_keys`: `case_id`, `expected_status`, `expected_queue`, `expected_outstanding`, `rule_ids`, `expected_checks`

Foreign keys are plain identifiers, so a task is reconstructed by joining `cases` to its documents, party, parent case and the policy version in force at `received_at`.

Keep `answer_keys` in a separate file or store so it is never loaded into the agent's context by accident. Ask in a pilot for a schema file and a short loader example in your preferred language so you can validate keys and types before running anything.