Current AI regulation, while establishing a clear demand for safety in high-risk systems, critically lacks a quantitative definition of "acceptable risk" and a technical methodology for verification. This gap, highlighted by the impending enforcement of the EU AI Act, leaves developers facing conformity assessments without the necessary tools to produce concrete safety evidence, particularly for opaque AI models. This paper introduces a foundational solution by proposing a two-stage framework that transforms AI risk regulation into a rigorous engineering practice, drawing parallels with the established aviation certification paradigm. For the first time, a clear path is laid out for demonstrating AI safety before deployment, addressing a critical void in the burgeoning field of AI governance.
From Normative Mandate to Quantifiable Guarantees
The proposed framework begins with Stage One, where a competent authority formally establishes two key parameters: an acceptable failure probability, denoted as $δ$, and an operational input domain, $varepsilon$. This act is not merely administrative; it carries direct civil liability implications, shifting the locus of risk definition to a normative, legally accountable body. This crucial step moves beyond abstract principles to define concrete, measurable safety targets for AI systems.
RoMA and gRoMA: The Verification Instruments
Stage Two of the framework introduces the RoMA and gRoMA statistical verification tools. These instruments provide a definitive, auditable upper bound on a system's true failure rate. Crucially, they operate as black-box methods, requiring no access to the internal workings of the AI model and thus scaling to arbitrary architectures, including complex statistical inference engines that resist white-box scrutiny. This innovation provides the missing technical instrument for AI safety verification, enabling developers to generate quantitative safety evidence that satisfies existing regulatory obligations and integrates seamlessly with current legal frameworks.
