Who Governs AI While It Is Thinking? Comparative Runtime Governance and the Economic Wall of Accuracy

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Artificial-intelligence governance increasingly addresses model evaluation, routing, output filtering, human oversight, and post-deployment monitoring. Agentic systems, however, can change materially during execution as evidence, model choice, resource use, authority, and proposed actions evolve. This paper proposes comparative runtime governance for consequential AI deployments. Its central architectural claim is a Reference-state continuity requirement: before exceptional risk is known, the system establishes a substantive Reference (baseline) candidate or state, preserves sufficient lineage to relate that state to later material changes, and uses that comparative basis when reassessing candidate eligibility before consequential commitment. Higher-fidelity processing remains conditional rather than automatically authoritative. A persistent episode state records decision-relevant candidate changes, resources, evidence provenance, authority, prior interventions, and external effects. A runtime governor separates three decisions that are often collapsed:1) whether another controlled operation may proceed;2) whether a candidate remains eligible; and3) which eligible candidate or disposition should determine the outcome.A protected execution gate makes those decisions effective before consequential action.The paper also develops the Economic Wall of Accuracy as a prospective allocation rule: once mandatory authority and integrity constraints are satisfied, the next available resource should be directed toward whichever permissible intervention—further computation, differentiated verification, additional evidence, human review, restraint, or enforcement—is expected to reduce consequential risk most effectively. The proposal is positioned against runtime assurance, ongoing authorization, adaptive computation, model routing, and recent agentic-AI governance work, and is presented as a testable conceptual architecture rather than a demonstrated performance improvement.