What integration contract should be agreed before a handover pilot starts?

Agree a small integration contract before loading the generated pilot. It should identify `release_id`, `schema_version`, `tool_contract_version`, `initial_snapshot_id` and `scoring_version`. Store these identifiers with every run so a result can be traced to the exact combination of material and evaluator behaviour.

Specify the agent-facing interface: files, retrieval tools or stateful business actions. For each tool, define input validation, error shape and whether the action changes state. Also define which clock the environment uses, since waiting conditions should not depend accidentally on the evaluator's machine time.

Reset behaviour needs an explicit boundary. State whether reset clears queued events, generated messages, cached retrieval results and action identifiers, as well as the main records. A partial reset can make a later agent inherit consequences from an earlier run.

Finally, agree how changes are introduced. A schema migration, revised policy or scoring correction should carry a new identifiable version and an explanation of its effect. Ask for a load-and-reset smoke run as the first integration step, followed by inspection of the resulting state and run manifest.