Resilient Metacognition: Adaptive Calibration for Trustworthy Synthetic Agents

Artificial agents deployed in dynamic environments must not only estimate their own reliability but also adapt when those estimates drift under stress. While previous work established metacognitive reliability as a foundation for epistemically trustworthy systems, resilience requires more than static calibration. In this paper, we investigate adaptive calibration, the online mechanisms that allow agents to dynamically adjust their confidence to remain trustworthy under perturbation, distributional shift, and adversarial pressure. We design a series of controlled experiments comparing agents, static temperature-scaled calibration, and adaptive calibration with dynamic temperature updates. Using metrics such as Expected Calibration Error (ECE), Brier score, AUROC, and risk coverage, we compare three conditions - no calibration, static calibration, and adaptive calibration and show that uncalibrated agents consistently display overconfidence, static calibration improves reliability but fails under shifting conditions, and adaptive calibration offers resilience by reducing miscalibration at the cost of raw predictive accuracy. Stress tests under out of distribution and adversarial inputs confirm that adaptation is critical for sustaining epistemic trust. Our findings suggest that calibration resilience is a core ingredient for building synthetic agents that remain epistemically robust in real-world deployment. By demonstrating that adaptive calibration preserves trustworthiness even in unstable regimes, this study extends the metacognitive framework toward practical, scalable design principles for resilient artificial minds.