Risk-Gated Mechanization: Severity, Verifiability, and Decisions Invariant to the Executor's Error Law

When should a step of agentic work be replaced by a deterministic mechanism? Earlier papers in this series answered with escape probabilities and costs, and treated every escaped fault as equally bad. We add severity and formalize a two-way principle, which reads informally as: mechanize a high-severity step whose correctness can be verified, and do not mechanize one whose correctness cannot. Read literally, the principle overreaches; we give its exact conditions. Severity sets a risk budget; it does not decide alone. An option is admissible when its worst-case loss over the models compatible with the evidence fits the budget. Mechanization is mandatory when closure is admissible and every other available option is not, and forbidden when a compatible closed model exceeds the budget; when no option is admissible, the step needs more evidence, an exogenous gate, or a redesign. Verification is a sufficient way to make closure admissible, and an evidence-compatible closed model whose expected loss exceeds the budget, such as an indistinguishable unsafe twin, forbids it. The bounds that decide are an executor's conditional escape floor and ceiling and a mechanism's certified ceiling. They hold uniformly over every error law, dependence structure, adaptive routing, and mixture of executors that satisfies them; a stochastic-dominance theorem extends this invariance to every nondecreasing loss. When evidence-indistinguishable worlds prefer opposite actions, every evidence-based rule has worst-case regret that grows linearly with severity. A martingale budget gives a scheduler an anytime guarantee under adaptive routing. Persistence enters through convex order: within conditionally independent processes, an uncertain mechanism fault mass is worse than fresh errors of the same mean for every convex loss, by the convexity of Bernstein polynomials, while the probability of any escape moves the other way. Production mixing averages floors; conjunctive review intersects blind sets, and one reviewer dominates another for every producer exactly when its blind set is contained in the other's. With linear loss and constant severity, the results reduce to the earlier mechanization criterion. The decision-theoretic and probabilistic tools are classical; the contribution is their application to the mechanization decision. These are mathematical models, not measurements.