What delivery formats and hand-off details should be agreed for a pilot?

Agree the delivery shape in the pilot brief so the sample loads into your tooling without rework.

**Formats.** JSONL suits case-level records and trajectories, which are nested and variable in length. Parquet suits large flat tables such as receipts, obligations and allocations, with typed columns. Policies fit well as JSON or YAML with stable rule identifiers. A buyer can ask for either or both.

**Layout.** One directory per release: `receipts`, `obligations`, `allocations`, `policies`, `events`, `cases`, plus `manifest.json` and a `schema` folder. Each file name carries a version.

**Keys.** `receipt_id`, `obligation_id`, `allocation_id`, `rule_id`, `case_id` and `account_id` act as foreign keys across files. Ask for a key-check script or a documented join order.

**Manifest.** It lists file names, record counts, schema version, generation seed, and which fields were produced by rule and which as free text. Ask for checksums so your pipeline can verify the transfer.

**Splits.** A `split` column on cases marks train, validation and test. Request that splits are made by scenario group, not by row, to limit leakage.

**Environment hand-off.** For simulated environments, ask for the tool schemas, the reset call, the grader interface and a seed list so runs can be replicated.

**Versioning.** Regenerated sets should carry a new version and a change note, so you can compare scores between rounds.