Parent-Guided Adaptive Reliability (PGAR): A Behavioural Meta-Learning Framework for Stable and Trustworthy AI

PGAR adds a "parent" supervisory layer that monitors AI system stability in real-time, using three specific signals to detect problems (incident detection), catch overconfident predictions, and remember how to recover from past failures—creating a continuous reliability score between 0 and 1 that automatically adjusts how quickly the system learns.

The framework reduces learning rate during unstable moments and restores it as conditions improve, which the researchers prove mathematically (via Lyapunov stability analysis) will keep the system's adaptation bounded and controlled, addressing a key concern about AI systems becoming unpredictable under stress.

Empirical testing demonstrates measurable improvements in calibration and reduced variance in loss, meaning the AI system makes more honest predictions about its own confidence levels and behaves more consistently—critical for applications where users need to trust the system's self-assessment.

PGAR is designed as a lightweight "plug-in" that works with existing AI systems without requiring complete redesign, making it practically adoptable for safety-relevant applications where interpretable reliability traces (visible records of why the system is or isn't confident) matter for accountability.

The approach prioritizes interpretability and stability over raw performance, suggesting a responsible AI design philosophy that values trustworthiness and transparency over maximizing accuracy alone.