What technical steps make up a useful first end-to-end trial?

A useful first technical step is to run one small slice end to end before asking for more.

1. **Agree the interface.** Settle the tool names, argument types and error codes. If your agent framework uses its own tool format, ask for definitions in that format or in a neutral JSON schema you can convert. 2. **Receive the slice.** Expect `cases.jsonl`, `records.jsonl`, `policies.jsonl` and a separate `labels.jsonl`, with Parquet for any large tables, and a schema document. 3. **Validate.** Run a script that checks every foreign key, that each variance matches the recomputed difference, and that no label text appears in agent-visible fields. 4. **Load the environment.** Bind the tools to the simulated state. Call `reset` twice for the same case and compare state hashes. 5. **Run a baseline.** Use your current agent, store traces in the agreed JSONL shape, and score by component: cause, action, evidence, routing and note. 6. **Review failures.** Sort failed runs into data problems, tool problems and agent problems before changing anything. 7. **Feed back.** Send the failure list to MirrorFirm so the next slice targets the gaps.

Keep identifiers stable between slices, such as `CASE-0412`, so results can be compared over time. Scope, size and delivery terms are settled in the pilot agreement.