What would a JSONL schema for mismatch cases and ground truth look like?

A pilot could deliver newline-delimited JSON, with Parquet as an alternative for larger tables. One file per object type keeps joins explicit.

**cases.jsonl**

- `case_id`, `opened_at`, `owner_role`, `status` - `record_a_ref` and `record_b_ref`, each with `system` and `record_id` - `stated_difference` with `amount`, `currency` and `unit` - `policy_ids`, a list

**records.jsonl**

- `system`, `record_id`, `entity_id`, `amount`, `occurred_at`, `updated_at` - `links`, a list of `{relation, target_id}` such as `settled_by` or `amends`

**policies.jsonl**

- `policy_id`, `rule_type`, `threshold`, `authoritative_system`, `approver_role`

**notes.jsonl**

- `case_id`, `author_role`, `written_at`, `text`

**ground_truth.jsonl** (withheld from the agent)

- `case_id`, `root_cause_code`, `correct_actions`, `acceptable_alternatives` - `required_evidence`, a list of record ids - `expected_final_state`, a diff against the seeded snapshot

**Rules for the schema**

- Identifiers are stable across files. - Timestamps carry an explicit time zone. - Nothing in `cases.jsonl` or `notes.jsonl` names the root cause.

Field names are a starting pattern. Agree the real schema with your harness during the pilot.