What table and file layout could a generated facility release use?

A release can be delivered as relational tables plus per-case trajectory files, so it loads into common training and evaluation tooling.

**Tables (Parquet or JSONL)** - `approvals`: `approval_id`, `limit_amount`, `currency`, `product_type`, `approver_role`, `conditions[]` - `offers`: `offer_id`, `approval_id`, `term_months`, `interest_only_months`, `signed_flag` - `facilities`: `facility_id`, `approval_id`, `status`, `status_reason`, `created_by_role` - `parties`: `party_id`, `legal_name`, `entity_type`, `alias_of` - `facility_parties`: `facility_id`, `party_id`, `role`, `effective_date` - `security`: `security_id`, `facility_id`, `asset_type`, `registration_ref`, `status` - `conditions`: `condition_id`, `facility_id`, `evidence_id`, `state` - `tickets`: `ticket_id`, `facility_id`, `from_role`, `to_role`, `summary`

**Case file (JSONL)** Each line holds `case_id`, `task_prompt`, `policy_version`, `seed`, `start_state_ref`, `expected_outcome` and `grader_id`.

**Trajectory file (JSONL)** Each line holds `case_id`, `step`, `role`, `action`, `arguments`, `observation` and `state_hash`.

**Conventions.** Identifiers follow one style per table, dates use ISO 8601, and amounts are stored as integers in minor units with a currency code. A schema file with types and nullability should ship with each release; scope the exact fields in a pilot.