Calibration-Aware Uncertainty Cascades for Heterogeneous Model Routing

CAUC provides a post-hoc framework to decouple deployment policies from specific model pools by independently calibrating confidence scores across heterogeneous models. This allows for a unified reliability scale to decide whether to accept an early prediction, invoke a larger model, or combine outputs based on validation data.