How should an environment distinguish process completion from evidence completeness?

Model business state and evidence state separately. A utility service order can have `service_status: completed` while its evidence assessment remains `support_status: incomplete`. Collapsing these into one status makes it impossible to test whether an agent notices the distinction.

A generated environment could expose service records with `service_order_id`, `completion_event_id` and `completed_at`. A separate assessment table could contain `requirement_id`, `service_order_id`, `support_status` and `missing_support_reason`. Links connect the assessment to the operation without making either field a substitute for the other.

Provide distinct tool actions such as `read_service_order`, `list_supporting_records` and `record_evidence_finding`. The final action should update the assessment, not reopen or complete the underlying service order. This keeps the agent's evidence work within its assigned responsibility.

An evaluation hook can compare the two state spaces after each action. A useful failure signal is an agent changing operational status merely to make the evidence package look consistent. Reset should restore both state spaces so repeated runs begin with the same discrepancy.