What file layout and schema would a generated case set use?

A common delivery shape is JSONL for cases and records, with Parquet for larger tables if you want to query them in bulk. Each line is self-contained and carries IDs that link across files.

`cases.jsonl` might hold: `case_id`, `domain`, `goal`, `opened_at`, `source_record_id`, `target_record_id`, `variance` (amount, currency, direction), `policy_ids`, `allowed_tools`.

`records.jsonl` holds every record with `record_id`, `system`, `type`, `fields`, `created_at`, `posted_at`, `reversed_by` and `supersedes`.

`policies.jsonl` holds `policy_id`, `text`, and optional structured `rule` fields such as `tolerance`, `approver_role` and `method`.

`labels.jsonl`, kept separate from what the agent sees, holds `case_id`, `root_cause`, `acceptable_actions`, `forbidden_actions` and `expected_note_elements`.

For a simulated environment, add an `initial_state.json` per case and a `reset` call that restores it. Tool actions are described in a schema file with names, argument types and error codes, so the same definitions can be bound to your own agent framework.

Agree field names and enumerations in the pilot. Ask for a schema document and a validation script so you can check referential integrity, such as every `target_record_id` resolving to a record, before any run.