The Calculus of Sovereign Delegation An Architectural Proposition for Separating Machine Intelligence from Human Authority

Abstract As artificial intelligence transitions from generating text to executing actions through tools, APIs, workflows, and autonomous agents, the nature of human-machine delegation changes. A prompt is not a contract. It is a semantic instruction issued to a probabilistic system. Yet organizations are increasingly granting high-stakes access to AI systems based on the assumption that model alignment, human oversight, and post-hoc audit logs are sufficient to preserve authority, accountability, and safety. This paper introduces The Calculus of Sovereign Delegation as an architectural proposition for governing the gap between human intent and machine execution. It builds on two prior EFA arguments: first, that only humans believe, confer authority, and bear moral responsibility; and second, that AI systems can produce synthetic personhood artifacts that lead humans to assign credibility, authority, or responsibility as if the AI possessed real agency, belief, or moral standing. The paper argues that safe AI delegation requires the structural separation of machine intelligence from human and institutional authority. Technical agency is not moral agency. Model alignment is risk reduction, not governance. Prompting is instruction, not enforceable authority. Human-in-the-loop oversight is insufficient unless the human remains identified, viable, empowered, and evidentially attached to the decision. This paper further presents the Cheung–Rupp Limit of Dimensional Governance: deterministic governance cannot be reliably guaranteed over the full high-dimensional probabilistic state of a generative model without materially narrowing or distorting its utility. Therefore, enforceable governance should attach to the discrete output, decision, tool call, or action at the point where it becomes inspectable, admissible, and capable of binding real-world consequence. The Ethical Functionality without Agency (EFA) Canon is presented as one architectural expression of this calculus. Sovereign Substrate Admissibility defines what may enter the governed environment. The Runtime Enforcement Layer determines what may bind. The Standardized Evidence Framework records decision-native evidence. The Human Ethical Agent of Record preserves non-delegable moral and legal agency. The Ring of Fire and Governance Heartbeat monitor whether delegated action remains structurally governable over time. Keywords AI governance; sovereign delegation; structural governability; runtime enforcement; human accountability; agentic AI; EFA Canon; REL; SEF; HEAR; synthetic personhood; AI control plane; decision governance.