How should a pilot delivery be packaged, versioned and connected to our harness?

Agree the packaging early, because it determines how quickly you can run the first evaluation.

**Files.** JSONL or Parquet per entity, plus `tasks.jsonl` and a `policies` file. Each delivery carries a `manifest.json` with schema version, file list, keys, generation seed and a content hash per file.

**Versioning.** Treat each delivery as immutable. A change, such as a corrected label, becomes a new version with a short change note, so past results remain interpretable. Pin your harness to a version and record it with every run.

**Tool schema.** If you want an environment, ask for a JSON schema for each action, for example `create_facility(approval_id, limit, term_months, repayment_frequency, balloon_amount)` and `hold_file(facility_ref, reason_code, route_to_role)`. Include error responses so the agent sees realistic failures.

**Reset.** A `reset(task_id)` entry point restores the start snapshot. Provide a state hash so you can verify it.

**Evaluation hooks.** Specify where the action log is written, how `grade(run_id)` is called, and which fields are volatile.

**Review loop.** Plan how label disputes and suspected disclosure concerns are raised and resolved. Confirm the details during scoping, not after delivery.