Determine the deployment-time damage rate

Determine the damage probability p_d for an accelerated call, namely the rate at which an accelerated call damages its executed action chunk, because this quantity is not known before deployment and is required to convert the coverage analysis into an absolute safety verdict.

Background

The paper models expected executed damage over a reference period using p_d, the probability that an accelerated call damages its action chunk. The ratio between different reference periods can be evaluated without knowing p_d, because the parameter cancels, but an absolute safety determination requires its value.

The authors explicitly state that p_d cannot be determined from the pre-deployment measurements considered in the paper. They therefore use bounds and monitor-off behavior to estimate how the safety tolerance might be affected, rather than establishing a deployment-independent absolute guarantee.

References

We do not know $p_d$ --- \S{sec:map} is the finding that nothing before deployment does --- but it multiplies both terms, so it cancels from the ratio between periods:

— The Unexpired Plan: A Free Monitor for Accelerated Diffusion Policies  (2610.05747 - Zhao et al., 5 Oct 2026) in Section 5.2, “Coverage buys safety where you cannot price it” (paragraph beginning “Turning the ratio into a verdict needs p_d”)