How would booking state be laid out in JSONL and Parquet files?

A common layout separates state from tasks.

**State as Parquet**

One file per table: `approvals.parquet`, `conditions.parquet`, `facilities.parquet`, `security_interests.parquet`, `schedules.parquet`, `ledger_entries.parquet` and `exceptions.parquet`. Columns use explicit types, with dates as ISO 8601 and amounts as integer cents plus a currency code to avoid floating-point drift.

**Tasks as JSONL**

One line per task:

- `task_id`, for example `TASK-0412` - `instruction`, the request given to the agent - `seed_snapshot`, the snapshot to load - `events`, a list of scheduled mid-episode changes by step number - `expected_state`, a reference to the target rows or a diff specification - `allowed_outcomes`, such as booked or returned with a reason code

**Policies as Markdown**

One file per policy, with an id referenced from tasks.

**Manifest**

`manifest.json` lists files, schema versions, row counts for that release, and the generation seed, so a loader can check it received everything.

JSONL and Parquet are common choices and other formats can be discussed. Confirm the layout against your loader in a pilot before committing to a full release.