Timing, Redirectability, and Runtime AI Oversight: The Sampling-Rate Hypothesis (Version 2)

This paper proposes the Sampling-Rate Hypothesis, a conceptual framework for runtime AI oversight centered on the temporal relation between supervisory monitoring and the internal evolution of safety-relevant hazards. The central claim is that runtime oversight can contribute to safety only when observation, interpretation, and intervention occur quickly enough relative to the characteristic pace of hazard formation. To formalize this intuition, the paper introduces a stylized effective-capacity framework incorporating monitoring cadence, proxy faithfulness, discernment efficiency, end-to-end latency, phase-dependent redirectability, latent vulnerability, background leakage, and adversarial pressure. A central argument of the framework is that runtime safety depends not only on whether hazardous trajectories are detectable, but also on whether they remain redirectable before external commitment. The paper further argues that hazard formation may be computationally overlapping rather than cleanly sequential in some architectures, that monitoring may partially reshape hazard dynamics through self-interference and observer effects, and that mechanistic interpretability may provide one useful proxy source for operational estimates of hazard-transition intensity. Stylized toy simulations are discussed as qualitative consistency checks rather than as complete empirical validation. Version 2 updates the manuscript structure, clarifies the interpretive status of the formalism, refines the discussion of adversarial complications and robustness, revises the toy-simulation framing, and improves figure placement and presentation.