Additional inference compute can increase the number of correctly resolved security‑assurance tasks, yet repeated success, unique coverage, accepted evidence, and operational protection are distinct quantities. We introduce a resource‑constrained framework that separates these notions.
For repeated conditionally independent attempts with latent success probability $\Theta$, coverage is defined as $$ C_n = 1 - \mathbb{E}\big[(1-\Theta)^n\big] $$ and its limiting value equals $1 - P(\Theta = 0)$.
Positive pairwise outcome correlation alone does not impose a ceiling below one. We construct two models that share the same mean success rate and pairwise correlation but exhibit different asymptotic coverage.
This result is distinguished from the effective sample size used to estimate a mean, showing that finite‑budget observations generally cannot identify an asymptotic support ceiling.
We then relate coverage to fallible evidence checking, proper scoring of factual grounding, complete resource accounting, service capacity, and a response model that incorporates mitigation delay.
A conceptual defensive architecture separates evidence analysis, adjudication, and operational authority, reducing single‑point failure risk.
The evaluation protocol specifies held‑out tasks, paired comparisons, negative cases, and uncertainty reporting to assess system performance.
The contribution is a coherent theoretical synthesis together with counterexamples that expose invalid extrapolations, rather than an empirical scaling law. All numerical illustrations are analytic; no hardware benchmark or attacker‑defender equilibrium is claimed.
Review