What should I request in a sample or pilot of approval material?

Put these requests in writing before the pilot starts, so the sample can be judged against agreed criteria and not impressions.

- A schema file per entity, with field types, keys and allowed values, plus a manifest with hashes and the generation seed. - A starting snapshot containing organisation, rules, delegations and open requests, loadable with one command in your harness. - A task set with prompts, accepted end states and labelled traps, split into a development group and a held-out group. - Validity labels and rule references for every decision, stored apart from what the agent sees. - A short description of how the policy text and structured rules were produced, and where gaps or assumptions remain. - A written list of what is out of scope, including domains, authority patterns and any process steps not modelled. - A statement of permitted uses under the licence, covering evaluation and training separately. - A contact route for questions found during inspection, and an agreed way to report defects.

Add your own acceptance tests to the list, such as recomputing authority from the tables or checking that resets produce identical states.

Keep the first request small. A sample you can fully inspect teaches more than a large one you can only skim, and it makes the decision about wider scope easier.