What delivery formats and first integration steps suit a pilot?

A pilot can be delivered in one or more forms, chosen during scoping.

**Files**

- JSONL for nested records such as decisions with conditions. - Parquet for larger tabular tables such as parties and documents. - A `manifest.json` listing files, schema versions, record counts for the pilot and a checksum per file.

**Task pack**

- `cases.jsonl` with `case_id`, `seed`, `starting_state_ref`, `expected_state_ref` and `tags`. - Policy files referenced by `policy_id`.

**Environment**

- Tool definitions for actions such as `search_party`, `create_account`, `set_status` and `create_task`, described in a JSON schema that fits common tool-calling formats. - `reset(case_id, seed)` and `snapshot()` endpoints, which a public MCP server or your own wrapper could expose.

**First step**

1. Load the manifest and verify checksums. 2. Check that every foreign key resolves. 3. Reset one case, run it by hand through the tools, and diff the result with the expected state. 4. Run your agent on that case and read the log.

Once one case works end to end, scale to the full sample. The data is generated and synthetic, so keep a review step for grounding and disclosure risk alongside the technical checks.