CUBICS: A New Method for Situation-Aware ML Safety Estimation

A new arXiv paper (2608.16564) introduces CUBICS, a method for situation-aware performance estimation of safety-relevant machine learning components (MLCs). The paper addresses the challenge of ensuring ML safety in safety-related applications by proposing the use of field data from shadow mode or safety envelopes to build proven-in-use arguments. Traditional Bayesian approaches model failures as a simple Bernoulli process with a single global failure probability, which is inadequate for MLCs whose performance depends on context. CUBICS aims to improve statistical evidence by considering coverage of relevant situations, including edge cases, and avoiding the infeasibility of building a single integrated statistical model for the entire system. The paper is categorized as a new announcement on arXiv, with the abstract outlining the motivation and proposed approach. The method is designed to provide more accurate and context-aware performance estimates, which is crucial for the safe deployment of ML in critical applications.