How do I load a first approval sample into my agent harness?

A first sample can be wired in with a short sequence of steps, whichever agent framework you use.

1. **Load the tables.** Read the JSONL or Parquet files into a store your harness can query, such as SQLite or an in-memory dataframe set. Keep one table per record type and index on the identifier fields. 2. **Separate the keys.** Put the answer key file in a location the agent's tools cannot read. Only the grader should join to it. 3. **Pin the clock.** Store the task's `as_of` timestamp and make every date-sensitive lookup use it. Delegation and role-holder checks depend on this. 4. **Write the adapter.** Map your own tool names to the sample's actions, for example `lookup_authority`, `list_delegations`, `submit_for_approval` and `escalate`. Return rows in the shapes the schema defines. 5. **Snapshot and reset.** Copy the loaded tables before each run and restore them after. Without that, state from one task leaks into the next. 6. **Run a small batch.** Start with ten tasks, log every call, and read the traces by hand before looking at scores. 7. **Check the grader.** Confirm that a deliberately wrong run, such as approving without authority, fails for the reason you expect.

If field names or tool shapes clash with your conventions, raise this during scoping so the release can be adapted. Keep the sample under version control, because later releases will need to be compared with it.