EFA/AI-SAFE2 7-Layer Protocol Stack (E7)

This paper proposes an OSI-equivalent reference architecture for AI systems aligned to the Ethical Functionality without Agency (EFA) and AI-SAFE2 paradigms. Whereas the OSI model decomposes network communication into standardized layers with explicit interfaces and fault isolation, contemporary AI deployments lack an equally rigorous, safety-centered decomposition that preserves human authority and enables auditable assurance. We introduce the EFA/ AI-SAFE2 7-Layer Stack (E7), a layered model proposing to make it “technically impossible” for a system to violate its ethical constraints at runtime, rather than merely documenting a policy against such violations. The layers: (1) Physical and Supply Chain, (2) Compute and Platform, (3) Knowledge and Evidence, (4) Model Behavior, (5) System Orchestration, (6) Policy, Ethics, and Controls, and (7) Mission and Authority. The stack is designed to (a) prevent “agency leakage” by making decision rights and accountability explicit at the top of the system, (b) translate ethical and legal obligations into testable constraints, and (c) support layered debugging and governance through defined artifacts such as decision-rights matrices, constraint catalogs, provenance controls, and operational telemetry. We argue that E7 enables consistent assurance across model, data, workflow, and infrastructure changes, mitigates threshold-drift risks endemic to voluntary safety regimes, and provides a defensible basis for audit, certification, and organizational accountability without attributing moral agency to AI systems.