What file layout and fields would a generated sufficiency release use?

A sufficiency release can be delivered as linked JSONL files, or as Parquet if your pipeline prefers columnar reads. Keys are stable strings, so joins work without fuzzy matching.

- `cases.jsonl`: `case_id`, `case_type`, `jurisdiction`, `amount`, `opened_at`, `status_history[]`, `requirement_set_id` - `requirements.jsonl`: `requirement_set_id`, `requirement_id`, `description`, `validity_rule`, `allowed_substitutes[]`, `waiver_condition` - `documents.jsonl`: `document_id`, `case_id`, `doc_type`, `issuer`, `issued_on`, `expires_on`, `text`, `defects[]` - `links.jsonl`: `document_id`, `requirement_id`, `relation` (meets, partial, fails, irrelevant), `reason_code` - `messages.jsonl`: `message_id`, `case_id`, `direction`, `sent_at`, `requested_requirement_ids[]`, `in_reply_to` - `gold.jsonl`: `case_id`, `decision`, `unmet_requirement_ids[]`, `acceptable_actions[]`

Two design choices help evaluation. Keep `gold.jsonl` separate from the files the agent can read, so labels are never exposed by accident. And derive `relation` and `unmet_requirement_ids` from the policy fields, so changing a requirement can regenerate the labels consistently.

For Parquet, nested lists such as `status_history` can be stored as list columns. A schema file shipped with the release should state types, nullability and the enumerations for `decision` and `relation`, so a loader can validate before any run starts.