Oversight-Centered Metrology and Control for Agentic Systems: Costly Interrupt Channels, Claim Margins, and Deployment-Relevant Evaluation

This paper develops an oversight-centered metrology and control theory for agentic systems deployed in real socio-technical workflows. Rather than treating a human as a privileged oracle, the framework models humans, automated verifiers, static analyzers, delayed outcome labels, physical sentinels, operational signals, and external auditors as costly interrupt channels with explicit error, delay, availability, and congestion properties. The paper argues that the performance and safety of agentic systems are not properties of a base model alone. They depend jointly on the deployment context, tool interface, monitoring regime, review policy, and the channels available to interrupt, correct, veto, rollback, or quarantine behavior. To make deployment claims scientifically legible, the manuscript introduces an operational oversight-evaluation specification, workflow-level estimands, and a claim-justification margin that combines estimation error, transport mismatch, and audit-distortion budgets. The theory covers static and dynamic complementarity, delay before irreversibility, routing error, review congestion, selective review, overseer-state effects, alarm fatigue, strategic silence, suppressed-risk debt, benchmark-to-workflow transport, audit gaming, active probing, dependence-aware redundancy, sequential monitoring, and ambiguity-budgeted safe control. Its central metrological claim is that evaluation reports for agentic systems should not report benchmark scores alone. They should report explicit context signatures, channel families, costs, delay budgets, transport assumptions, audit diagnostics, and fail-closed claim status. The goal is a deployment-relevant science of oversight for AI agents and other agentic systems.