ACT/HOLD rule

The ACT/HOLD rule is a decision-making framework designed for safety-critical assistive systems, particularly in contexts like Activities of Daily Living (ADL) where incorrect interventions can pose risks. At its core, it operates by evaluating the reliability of a model's prediction for a user's next action. Instead of relying on raw model confidence, which often misrepresents true correctness, the rule leverages *calibrated probabilities*. These probabilities, achieved through post-hoc calibration, ensure that the predicted confidence accurately reflects the empirical likelihood of the prediction being correct. The mechanism is straightforward: if the calibrated reliability surpasses a predefined threshold, the system "ACTs" and provides assistance; otherwise, it "HOLDs," refraining from intervention. This approach is crucial for transforming a confidence threshold into a quantifiable safety parameter, enabling verifiable and safer behavior in assistive control loops. It is primarily used in robotics, human-robot interaction, and pervasive computing for applications requiring reliable and safe automated assistance.