Timing, Redirectability, and Runtime AI Oversight: The Sampling-Rate Hypothesis

This paper presents a theory-first framework for runtime AI oversight centered on pre-commitment monitoring, proxy faithfulness, and intervention feasibility. Its core claim is narrow: monitoring can improve intervention success only when a system can be observed, interpreted, and redirected before a hazardous trajectory reaches commitment. The framework organizes runtime oversight around three requirements: usable signal, sufficient remaining time, and retained intervention authority. It introduces Safety Slack, S_t, as a design margin comparing usable oversight capacity with effective hazard burden, and develops a phase-sensitive account of escalation through contact, attention, recognition, impulse, and commitment. The manuscript also distinguishes latent theoretical targets, operational proxy estimates, runtime control estimates, and decision-oriented adequacy margins. Optional formal supports from Optimal Stopping, Structural Causal Models, Information Theory, Control Barrier Functions, Semi-Markov timing, and adversarial monitoring are included as theoretical scaffolds, but the framework remains an empirical research scaffold rather than a safety guarantee. The intended contribution is a falsifiable structure for testing whether pre-commitment runtime oversight improves intervention success over output-only or post-commitment monitoring under realistic limits of proxy quality, latency, redirectability, adversarial pressure, and monitoring overhead.