Runtime Responsibility Boundaries: A Missing Control Layer in Regulated Human-AI Systems

Current AI safety approaches focus on model capability, accuracy, and compliance artifacts. However, many real-world failures occur at runtime when systems continue to interact after responsibility becomes unclear, shifted, or unanchored. This paper introduces Runtime Responsibility Boundaries (RRB) as a missing architectural control layer. RRB addresses five interacting dimensions that determine whether AI output remains permissible: time as state, human state uncertainty, context risk escalation, system role binding, and responsibility anchoring. The framework provides structural enforcement of halt conditions, handoff protocols, and deterministic non-continuation for high-risk AI systems in healthcare, governance, and regulated domains. RRB has been validated in anonymized institutional deployments and reviewed by MIT, Fraunhofer Institute, EU Commission, and University of Oxford.