Governance-Oriented Agent Architecture (GOAA): A File-SystemLevel Entropy Governance Paradigm Centered on Natural-Person Decision Authority, with a Full-Tier Deployment Pathway from Individuals to Enterprises

Abstract

Large language model (LLM)-driven agents face systemic disorder as they move from prototype to production: rules decay through interaction, behavior drifts through operation, and systems bloat through evolution. This paper proposes the Governance-Oriented Agent Architecture (GOAA), an entropy governance paradigm centered on the final decision-making authority of natural persons, with file-system-level control authority as its physical substrate. GOAA is built on four axioms—Decision Authority, File Ontology, Inherent Ambiguity, and Entropy Growth—that delineate the governance boundary; it replaces the logical control plane with physical files as the governance substrate, and establishes regular human adjudication and a file-based four-layer memory as the governance loop of cognitive entropy. The paper further provides a dual-source entropy framework (technical entropy and cognitive entropy) and a full-tier deployment pathway from individuals to enterprises. Empirical results from approximately 8.16 billion tokens of real-world operation and three case studies demonstrate that rules consolidated via human adjudication remain stable across sessions, and that the file-based governance substrate is economically viable for long-term deployment. This paper serves as the systemic overview of Cognitive Lag (Yin, 2026; Zenodo 10.5281/zenodo.22031207) and provides a systematic complement to the governance gap in long-cycle agent deployment. Keywords: agent governance, entropy, file-based governance, regular human adjudication, cognitive entropy.