Interpretation Correctability in Adaptive AI: Evidence, Verification, and Calibration of Personalized Beliefs

This research summary presents six controlled experiments on memory integrity, interpretation, verification, and calibration in adaptive AI. The work examines a distinction that becomes increasingly important as AI systems learn about people over time: preserving what a person said does not necessarily mean preserving a justified understanding of that person. Across the experiments, failures emerged during memory revision, interpretation, verification, contextual application, and behavioral use. The final experiment tested decomposed verification across entailment, temporal validity, and contextual scope. Results differed substantially across models: Qwen detected all corrupted interpretations but rejected all valid interpretations, while NVIDIA Nemotron 3 Ultra detected all corruptions and preserved seven of eight valid interpretations in the same experimental setup. These findings motivate Interpretation Correctability as a distinct evaluation target for adaptive AI: whether a system can maintain useful personalized inferences while keeping them grounded, current, contextually bounded, calibrated, and correctable. Reproducibility materials, notebooks, protocols, limitations, and experiment records are available at:https://github.com/GesaSchneider1/interpretation-correctability