A Minimal Common Assurance Architecture for Stateful AI Systems

This Technical Note proposes a minimal common assurance architecture for stateful AI systems. The proposal does not replace existing security practices or introduce a new security standard. Instead, it asks whether existing and future defensive mechanisms may become easier to connect, inspect, extend, replace, and evaluate if several assurance relationships are made explicit within a shared architectural skeleton. The architecture highlights evidence provenance and independence, AI-to-AI auditing, authorization and delegation, recovery validation, and human auditability. It is presented as a provisional framing for discussion, testing, criticism, modification, extension, simplification, or rejection where appropriate.