What formats and schema do escalation datasets and task packs use?

Format and schema are agreed per engagement, but a typical layout looks like this.

**JSONL** suits task definitions and event streams, one object per line:

- `tasks.jsonl`: `task_id`, `case_id`, `instruction`, `starting_state_ref`, `expected` (target role, reason codes, handover requirements) - `events.jsonl`: `event_id`, `case_id`, `actor_role`, `action`, `from_status`, `to_status`, `occurred_at`

**Parquet** suits relational tables you may query in bulk:

- `cases`, `parties`, `facilities`, `reviewers` - `authority_rows`: `role`, `product`, `limit_basis`, `limit_value`, `effective_from`, `effective_to` - `escalations`: `escalation_id`, `case_id`, `from_reviewer_id`, `to_role`, `reason_code`, `outcome`

**Documents** can be supplied as text or structured files, referenced by `document_id` from the case.

Identifiers follow a consistent style such as `CASE-0412` or `ESC-0319`, and foreign keys are explicit so joins are reproducible. A `manifest.json` lists files, schema versions, row counts for that delivery, the generation configuration and the foreign-key map.

The data is generated, so a validation script that checks referential integrity and date ordering is a sensible part of any sample. Ask for field names to be aligned to your own schema in a pilot rather than adapting your pipeline to ours.