Quantifying Confidence in Assurance 2.0 Arguments

Researchers from SRI International and City, University of London have introduced a novel method for quantifying confidence in Assurance 2.0 arguments, a framework critical for certifying the safety of complex systems like AI. The paper, 'Quantifying Confidence in Assurance 2.0 Arguments' by Robin Bloomfield and John Rushby, presents a probabilistic approach that is simple, systematic, and mathematically sound. It specifically addresses how much confidence a safety decision requires versus how much the supporting argument actually provides, a central question in high-stakes domains.

The method works by exploiting the structured decomposition of claims within a safety case. It offers different analytical approaches depending on the degrees of independence and diversity among sub-claims, and how those sub-claims eliminate concerns that undermine confidence in their parent claims. Crucially, it uses only elementary probabilistic constructions, such as Fréchet bounds, making it accessible to engineers without requiring deep expertise in probabilistic analysis. The authors demonstrate that their approach is resilient to counterexamples that have undermined previous confidence quantification methods.

This tool is positioned as a supplement to the primary Assurance 2.0 evaluation criteria of logical indefeasibility and dialectical examination. Its practical value lies in enabling engineers to evaluate cost-versus-confidence tradeoffs for different risk levels and to assess the overall balance of confidence across an entire structured argument. This provides a more rigorous, quantifiable foundation for making critical decisions about the deployment of AI and autonomous systems where safety is paramount.