What formats and handover details should a pilot agree?

Agree the delivery details in the pilot so your harness can load the output without rework.

**Files and formats.** Ask for JSONL for record files, and Parquet if you prefer columnar loading. Request a schema file listing each field, its type, allowed enumerations and which fields are foreign keys.

**Manifest.** A manifest should list every file, its record count, the release version and the schema version. Request a changelog, so a corrected sample can be compared with the first.

**Task definitions.** If you want a task pack, ask for a tasks file with task_id, decision_id, start_state_ref, allowed_actions, expected_end_state and accepted_reason_codes. Keep it separate from the records, so tasks can be regenerated without changing the underlying cases.

**Environment.** For a simulated environment, agree the tool list, such as get_decision, list_conditions, read_evidence, request_evidence, mark_condition_reviewed, escalate_to_approver and release_step. Agree how reset works: snapshot restore, fixed seeds for evidence arrival, and whether runs can execute in parallel.

**Evaluation hooks.** Ask how action logs are exposed and in what format, so your grader can read them without parsing free text.

**Feedback.** Agree how defects are reported, by case identifier and file, and how corrected versions are issued.

The content is generated, so build a validation step that checks referential integrity on every delivery.