Which evaluation hooks make an interactive complaint pack reproducible for downstream buyers?

An interactive pack can expose `reset_environment`, `get_observation`, `submit_action` and `get_evaluation` operations. A reset request should identify the scenario version and initial-state identifier, with a seed where the environment includes controlled variation. The reset response can return a state fingerprint for later comparison.

Each action trace can record `run_id`, `step_id`, `tool_name`, sanitised arguments, action result and resulting state fingerprint. Evaluation hooks can inspect whether an action satisfied a prerequisite, attempted an unauthorised transition or left a required obligation incomplete. Keep evaluator-only state out of ordinary tool responses.

MirrorFirm could generate environments with these interfaces or provide an equivalent replay specification for buyer-hosted execution. A reseller should preserve the mapping between environment version and scoring version when distributing updates.

Reproducibility also requires identifying external effects. If a tool simulates sending an acknowledgement, the trace should record the simulated delivery result rather than depend on a live messaging service. Downstream customers can then replay a failed run, inspect the exact transition that caused the failure and distinguish agent behaviour from changes in the test environment.