What file formats, schemas and tool actions can a set-up release use?

A set-up release can be packaged so engineers can load it with standard tooling.

**Record files.** One file per record type, for example `approvals`, `customers`, `accounts`, `services`, `tasks` and `audit_events`. JSONL suits nested records and streaming. Parquet suits columnar analysis and large volumes. Each row carries a stable ID and a `schema_version`.

**Schema files.** JSON Schema or an equivalent definition for each record type, including required fields, enumerations such as `account_status`, and foreign-key notes like `account.customer_id` referencing `customer.customer_id`.

**Task file.** One line per task:

- `task_id`, `instruction`, `start_state_ref` - `tools_allowed`, `policy_refs` - `fault_plan_ref` (optional) - `split` (train, validation or held-out)

**Tool manifest.** Names, argument schemas and return shapes for actions such as `search_customer`, `create_customer`, `open_account`, `activate_service`, `create_task` and `add_note`. This can be exported in the function-calling format your agent framework expects, after a mapping step agreed in the pilot.

**Answer key.** Held in a separate file: expected end-state assertions, required handovers, and the labels for any seeded faults. Keep it out of the agent's context.

**Manifest and checksums.** A release manifest listing files, record counts per file, generation seed and schema versions, so you can confirm you hold the complete release and regenerate it deterministically if scoped that way.