Researchers propose a new approach to medical vision-language models that focuses on calibrated triage rather than autonomy. They evaluate nine confidence estimators and find that standard metrics mislead in comparing their performance. The study suggests that a good estimator can help a competent model defer safely where it is competent, but none of the estimators can manufacture reliability where the base model lacks it. --- Why it matters: This research matters to engineers working on medical AI because it highlights the importance of calibrated triage and confidence estimation in ensuring safe and reliable decision-making. The study's findings have implications for the development of medical vision-language models and their deployment in clinical settings. Source: https://arxiv.org/abs/2606.15910