KillSwitch-Bench

I like this benchmark because the adversarial nature of the underlying language helps judge a model's ability to remain productive despite miscommunication and imprecision, which we'll reflects the nature of real work.