What file layout and keys would a schedule dataset use?

A schedule dataset works best as a set of related tables with stable keys, delivered as JSONL for readability or Parquet for volume. Agree the exact schema in scoping; the layout below is a typical starting point.

**Tables**

- `accounts`: `account_id`, `product_code`, `start_date`, `frequency`, `status` - `schedules`: `schedule_id`, `account_id`, `version`, `status`, `supersedes_id`, `change_reason`, `created_at` - `instalments`: `instalment_id`, `schedule_id`, `due_date`, `amount`, `state`, `settled_by_payment_id` - `payments`: `payment_id`, `account_id`, `received_at`, `amount`, `type` - `instructions`: `instruction_id`, `account_id`, `schedule_id`, `next_date`, `amount` - `cases`: `case_id`, `account_id`, `request_type`, `approver_role`, `outcome` - `policies`: `policy_id`, `product_code`, `rule_key`, `value`, `text`

**Task files**

- `tasks.jsonl`: `task_id`, `prompt`, `start_snapshot`, `allowed_tools`, `expected_outcome`, `checks`

**Conventions**

- Dates in ISO 8601 with an explicit timezone. - Money as integer minor units with a currency code. - Every foreign key resolves inside the release, and a validation script can confirm that. - Snapshots are named, so `start_snapshot` can be restored by the environment reset.

Ask for the validation script early. Broken keys are the most common reason a loaded dataset behaves differently from its description.