How do we connect pilot material to our agent harness?

A pilot is easier to adopt when delivery matches how your harness already loads work.

- **Delivery layout:** a folder per release with `cases.jsonl`, `parties.parquet`, `documents.parquet`, `policies.json`, `tasks.jsonl` and a `schema` file describing every field and key. - **Versioning:** each release carries a `release_id` and each task a `task_id`, so results can be tied to the exact material used. - **Splits:** a `split` field marks tasks as `dev` or `test`, and test tasks can be held out from training runs. - **Tool mapping:** a short table maps environment actions to your agent's tool definitions, for instance `set_status` to your case system's update call. - **Reset and export:** a command or API call loads a named snapshot and returns final state as JSONL.

Wire it up in three steps. Load the schema and run the key checks. Run a baseline agent against dev tasks and inspect the action logs. Then run the test split once per agent version and store reports with `release_id` and `agent_version`.

Licence terms govern where outputs can be stored and who can access them, so settle that before copying files into shared repositories. Keep pilot material apart from any real customer data in your own systems, so generated and real records are never mixed in one store.