How can an environment test whether actions actually produce evidence?

A simulated environment could maintain separate operational and evidence event streams. An operational event records what an actor did; an evidence event records what became available to a reviewer. Keeping these streams separate prevents the presence of a document from automatically proving that the associated action occurred.

An operational event could contain `event_id`, `case_id`, `action_type`, `actor_role` and `occurred_at`. An evidence event could contain `evidence_id`, `source_event_id`, `captured_at`, `storage_status` and `visible_to_role`. The link establishes the intended relationship without removing the need to inspect capture failures.

Tool actions such as `perform_check`, `save_attestation` and `assemble_pack` could change different parts of state. A check may complete while its attestation fails to save. An agent assembling the pack must detect that the reviewable trace is incomplete.

A reset could restore the initial case and both event streams. Evaluation hooks could inspect whether the final pack contains evidence for each claimed action and whether the agent reports capture problems.

Licensed resellers should describe this distinction in integration notes so customers understand what the environment measures beyond document retrieval.