Who Governs AI While It Is Thinking?

Abstract

Artificial intelligence governance often focuses on whether a system should be deployed and whether its final output is acceptable. This essay examines the changing computational episode between those points—the period in which evidence, models, resources, authority and potential consequences may change while computation remains in progress. It proposes Persistent Reference Infrastructure: a runtime-governance architecture in which a substantively useful Reference state is established before exceptional risk is known and retained, revalidated or updated across material transitions. The Reference state provides a continuing comparative basis for evaluating later computation rather than requiring the evolving process to serve as its own sole producer, reviewer and validator. The architecture separates three runtime decisions: whether the next controlled operation may proceed, whether an existing candidate remains eligible for use, and which eligible candidate or other governed disposition should determine the outcome. Higher-fidelity processing remains conditional and may be suppressed when the Reference result is sufficient, while consequential operations remain subject to pre-commitment enforcement. The essay also develops the Economic Wall of Accuracy, a dynamic resource-allocation boundary at which the next available resource may reduce consequential risk more effectively through verification, authorization, escalation, restraint or enforcement than through further computation. The framework does not claim that comparison guarantees correctness or that every AI deployment requires the same controls. It presents a testable governance architecture for consequential AI systems and outlines failure modes, certification considerations and empirical comparisons against simpler alternatives.

Keywords