Tri-Info uses information theory to detect failures in Vision-Language-Action models by analyzing action diversity, temporal consistency, and state coupling. It achieves 83% accuracy on real-world tasks across six models and three environments, outperforming prior methods and maintaining performance without retraining.