Which runtime hooks make policy propagation failures diagnosable?

Instrument each tool call with an action trace containing `episode_id`, `action_id`, `actor_role`, `target_id`, `policy_version_used`, `pre_state_hash`, `post_state_hash` and `result_code`. Keep evaluator-only context separate from tool responses so instrumentation does not reveal hidden answers.

Run invariant checks after state-changing actions. A utility environment might require every confirmed connection to reference a completed assessment whenever the governing policy demands one. If an agent confirms the connection and later repairs the assessment link, a final-state check alone could miss the period of invalid operation.

Useful hooks include `on_action_completed`, `on_policy_activated` and `on_episode_finished`. They can emit structured violations with the affected entity, governing clause and first failing action. An episode-level report can distinguish a prevented invalid attempt from a violation that actually entered state.

MirrorFirm could generate invariant definitions and expected trace conditions alongside the environment specification. Builders should establish which checks their runtime enforces and which only score behaviour. That separation matters: a rejected action is evidence about agent judgement, while an accepted invalid transition may also expose a simulator enforcement gap.