What evaluation hooks help test recovery from partial policy-update failures?

A simulated environment can expose reset_environment with a scenario_id and seed, then return a run_id for every subsequent action. Reset should restore records, queued events, simulated time and external-action receipts so repeated runs start from a defined state.

Fault hooks can interrupt a change after one write commits but before the tool returns confirmation. In a utilities task, the account update may succeed while the related notification action times out. The agent must determine what happened instead of repeating every action blindly.

Tool calls can accept an idempotency_key and return operation_id values that support a read_operation lookup. The evaluator can then inspect duplicate side effects, unresolved obligations and unsupported assumptions about success. Injecting failure before commitment and after commitment should produce distinguishable recovery demands.

Keep evaluator hooks outside the agent's normal tool view. A run manifest can record fault settings, initial-state checksums and scoring rules, while an action trace records tool arguments and results. Together, these make a failed run reproducible and allow developers to compare recovery behaviour without changing the underlying business task.