The Institutional Layer of Artificial Intelligence: Authority, Durable Commitments, and Accountable Action in Large-Model Systems

Large models can generate proposals, interpret evidence, and coordinate work, but these capabilities do not by themselves establish the authority to act, preserve commitments across deployments, or provide procedures for contesting decisions. This paper develops an institutional architecture for large-model agents. An institution is represented by authenticated roles, versioned rules, persistent obligations, resource accounts, an execution gate, and a review process. We distinguish uncertainty about the world from uncertainty about authority, and derive the minimum decision loss when institutionally different situations are observationally indistinguishable. We characterize the memory required for exact enforcement through continuation-equivalent histories, establish policy-independent safety under complete mediation and atomic state updates, and separate action permissions from the capacity to fulfill accepted commitments. A decision-theoretic analysis identifies when automatic execution, verification, or refusal is optimal, while a common-cause model quantifies the limits of replicated review. An agentic investment workflow illustrates the architecture through authorization, concurrent orders, revocation, settlement obligations, and rule changes. Finite-state enumeration and analytical calculations validate the illustrations; no large-model benchmark results are claimed. The contribution is a unified account of institutional sufficiency, grounded in electronic institutions, access control, automata, and decision theory. Institutional guarantees remain conditional on the correctness of their rules, enforcement mechanisms, and external interfaces; computational compliance alone does not establish social legitimacy.