What delivery formats and environment setup should a pilot specify?

Put integration needs into the pilot brief so the first delivery runs in your stack without rework.

**Data files.** Choose JSONL for record-by-record loading, Parquet for analytics, or both. Ask for one file per record family, such as authority, delegations, requests, routing and gold, with documented primary and foreign keys.

**Versioning.** Request a release identifier in each file and a manifest listing files, schema versions and record counts, so you can detect drift between deliveries.

**Validation.** Ask for a schema definition, for instance JSON Schema, and a script or checklist to verify keys, date ranges and limit arithmetic. Run it in your own pipeline, not only theirs.

**Task pack layout.** Each task bundles a prompt, the record references it needs and a gold entry with condition tags.

**Environment option.** If you want a simulated environment, specify the tool interface (for example function-call signatures or an MCP-style tool list), whether write tools enforce authority or only log attempts, and the simulated clock behaviour.

**Reset and runs.** Agree on snapshot loading, seeded randomness and an append-only event log, so results are repeatable and can be compared across model versions.

**Output.** Define the results format your graders will emit, including per-condition fields.

The content is generated to your specification. Confirm formats and reset behaviour on the sample, along with grounding and disclosure risk.