How should environment state separate business activity from evidence assessment?

Represent the operational process and its review as linked but separate state objects. A utility service case can have fields such as `case_id`, `service_status` and `completed_at`. Its evidence assessment can hold `assessment_id`, `case_id`, `assertion_id`, `coverage_status` and `review_status`.

This separation matters when an agent discovers a documentation gap. Updating `coverage_status` to `insufficient` should not reopen the service case or change its completion timestamp. Equally, an operational correction should not silently convert an already submitted assessment into an accepted one.

Expose actions according to those boundaries. `attach_evidence` changes assessment links. `submit_assessment` moves the review into a submitted state. An operational correction, if included in scope, should use a separate action with its own authority checks and recorded effects.

A release can supply JSONL fixtures for both object types and a schema describing permitted transitions. An environment reset should restore the operational baseline and the initial review state together. Evaluation hooks can then compare the business history before and after assembly, detecting agents that obtain an apparently clean result by modifying facts they were only asked to assess.