What JSONL layouts could carry cases, documents, requirements and answer keys?

A release can ship one JSONL file per entity, with Parquet equivalents for larger volumes. The fields below are a starting layout to agree during scoping.

- `cases.jsonl`: `case_id` (for example CASE-0412), `product`, `attributes` (object), `opened_at`, `initial_status` - `requirements.jsonl`: `req_id`, `rule_ref`, `applies_when` (expression over case attributes), `accepted_types`, `max_age_days`, `satisfiable_from_records` - `documents.jsonl`: `doc_id`, `case_id`, `doc_type`, `party_id`, `issued_on`, `signed`, `legible`, `text`, `defect_tags` (held back from the agent view) - `held_records.jsonl`: `record_id`, `case_id`, `record_type`, `fields` - `events.jsonl`: `event_id`, `case_id`, `step`, `type` (customer_reply, deadline), `adds_documents` - `expected.jsonl`: `case_id`, `required_items`, `satisfied_items`, `missing_items`, `correct_action`, `end_status`, `clock_state`

Links run by identifier. Documents and events point to `case_id`, requirements are resolved through `applies_when`, and `expected.jsonl` is derived from the same requirement table, so the answer key and the policy cannot drift apart.

Keep the agent-visible view and the grader view in separate files or schemas. A manifest with a schema version and a seed per case lets you regenerate or diff releases.