Why should a policy task expose events as well as the latest record state?

A current-state record can show the policy now attached to a case while hiding the sequence that produced its decisions. For connected agents, that sequence matters: a decision made before publication may be valid even though the same action would be invalid after notification.

A generated event stream can include event_id, entity_id, event_type, occurred_at, received_at and policy_version_id. Keeping occurred_at separate from received_at allows the environment to represent delayed delivery without pretending the business event happened later.

Events such as policy_published, case_assessed and change_notice_received can link to the same case. The agent can inspect these through a read_events action alongside a read_case action. The tools should state whether the returned stream is complete or limited to a particular observation window.

This structure supports a specific training distinction: what changed in the business versus what the agent could know at the time. Evaluation can inspect the evidence available before each action, rather than judging every historical action against the final snapshot. A replay can then preserve that information boundary for a fresh agent run.