Capability-Offset Deployment: A Recursive Frontier-Guardian Architecture for AI Control, Economic Feasibility, and Public Governance

Abstract

This paper proposes Capability-Offset Deployment (COD), a recursive socio-technical framework for controlling increasingly capable artificial intelligence systems. The central principle is that no frontier AI capability with potentially systemic consequences should be granted greater operational freedom than society possesses demonstrated capability to supervise and control. COD distinguishes the technological frontier from the deployable frontier and proposes that newly available capability should first increase supervisory capacity over lower deployable tiers before becoming eligible for broader deployment. The framework does not assume that a later model generation is automatically safer or more trustworthy. Instead, it defines a verified supervisory capability offset across safety-critical domains, requiring adversarial evidence, independent evaluation, enforcement mechanisms, and explicit limits on uncertainty, capability leakage, false positives, utility loss, and adaptive evasion. The paper situates COD in relation to AI Control, weak-to-strong supervision, bootstrapped monitoring, capability and control buffers, defense-in-depth safety architectures, open-weight risk management, confidential computing, and frontier-safety governance frameworks. It also analyzes economic feasibility, monitoring costs, latency, opportunity costs, open-access implications, international coordination, institutional legitimacy, and regulatory design. The proposal is presented as a conceptual research and policy preprint rather than as empirical proof. It identifies conditions under which COD should be modified, restricted, or rejected, and argues that the core question should remain open between public regulation, private self-regulation, mixed governance models, and international verification regimes.