Why should an evaluation distinguish policy validity from agent knowledge?

A generated policy record can carry `policy_version_id`, `effective_from`, `effective_to` and `published_at`. A delivery event can separately record `recipient_role`, `available_at` and `acknowledged_at`. These fields represent different facts: when a rule applies, when it becomes accessible and when someone confirms receipt.

A utilities service request can link its decision event through `decided_at` and `policy_version_used`. The evaluator can then inspect whether the agent used the rule applicable to that event and whether the replacement rule was visible through its permitted tools at the time.

This distinction matters when a notice arrives after the effective time. A benchmark should not silently treat inaccessible information as something the agent ignored. It can instead test the response after discovery, such as identifying decisions that need authorised review.

JSONL event records can preserve both business timestamps and observation timestamps. An evaluation hook can reconstruct the policy view available at each action, rather than scoring against the final document collection. That makes the task a test of reasoning under changing information and prevents later knowledge from distorting the assessment of earlier behaviour.