What file layout and schema would referral and decision records use?

A common delivery shape is one JSONL file per entity, with a manifest listing the files and schema versions for that release. A referral record might look like this:

`{"referral_id":"REF-0931","application_id":"APP-2187","from_role":"credit_officer","to_role":"senior_credit_manager","reason_code":"LIMIT_EXCEEDED","exposure_aud":380000,"limit_aud":250000,"created_at":"2026-03-04T10:12:00Z","status":"OPEN"}`

A decision record references `referral_id` and holds `outcome` (`APPROVED`, `DECLINED`, `APPROVED_CONDITIONAL`, `RETURNED`), `decided_by_role`, `conditions` as an array, and `decided_at`.

Integration points:

- **Parquet** for larger tables such as status history, with JSONL for text-heavy tables like handover notes - **Foreign keys** named consistently (`application_id`, `referral_id`) so joins need no mapping - **Tool actions** for an environment: `raise_referral`, `record_decision`, `return_to_reviewer`, `create_task` - **Enumerations** published in the schema, so graders can validate reason codes and outcomes - **Timestamps** in UTC ISO 8601, with status history as an append-only table

Ask for a JSON Schema file per entity so you can validate a load before training or running evaluations. If your pipeline expects different field names or a columnar-only format, state that when you scope so the delivery can be agreed in advance.