Machine learning models are typically evaluated based on predictive accuracy, yet real-world systems ultimately rely on decisions derived from those predictions. This work introduces Decision Stability Score (DSS), a post-hoc, model-agnostic metric that quantifies the robustness of decisions under input perturbations. We show both theoretically and empirically that prediction accuracy and decision stability are fundamentally misaligned. DSS is governed by prediction margins relative to decision boundaries rather than error magnitude, explaining why models with lower prediction error can produce less stable decisions. Experiments on Formula 1 telemetry data and the UCI Heart Disease dataset demonstrate consistent misalignment across models and domains. These findings highlight the need to evaluate machine learning systems not only based on predictive accuracy but also on the reliability of downstream decisions.
Priyanshu Nayak (Sun,) studied this question.