We propose a computationally efficient localized PAC-Bayes prior directly integrated into training, yielding tight generalization certificates, individual prediction guarantees, robustness, and calibration for deep networks.