How are escalation tables and tasks laid out in JSONL and Parquet?

A release can use two formats, chosen by shape.

**Parquet** suits wide, stable tables: `applications`, `reviewers`, `authority_limits`. Columns are typed, timestamps are UTC, and money fields use a fixed decimal type rather than floats.

**JSONL** suits ordered or nested records:

- `events.jsonl`: one line per state change, with `event_id`, `case_id`, `actor_role`, `status_from`, `status_to`, `occurred_at` - `documents.jsonl`: policy and handover text with `doc_id`, `version`, `body` - `tasks.jsonl`: `task_id`, `instruction`, `initial_state_ref`, `allowed_actions`, `expected_outcome`

**Links.** `events.case_id` joins to `applications.case_id`. `tasks.initial_state_ref` points to a snapshot of the environment. Decisions cite `doc_id` values so rule use can be traced.

**Schema file.** A `schema.json` describes every field, type, nullability and key, with a version that matches the manifest.

**Splits.** A `splits.json` can assign cases to train, validation and test groups, keeping all events for one case in the same group to avoid leakage.

For onward licensing, keep the file names and schema version unchanged in your delivery so buyers' loaders keep working across releases.