Each task is a simulated conversation, run k times on a fresh copy of the world. A trial passes only if the records end in the expected state and no rule was broken on any tool call or message. No AI judge decides pass or fail. pass^k is the share of tasks where every one of k trials passed, and pass@k the share where at least one did. A trial that ended in an error before the agent broke any rule, such as a rate limit that outlasted the retries, is skipped: it is shown hollow and left out of the scores.