Despite extensive research on neural network calibration, existing methods typically apply global transformations that treat all predictions uniformly, overlooking the heterogeneous reliability of individual predictions. Furthermore, the relationship between improved calibration and effective uncertainty-aware decision-making remains largely unexplored. This paper presents a post-hoc calibration framework that leverages prediction reliability assessment to jointly enhance calibration quality and...