Audited Selective Verification for Risk-Controlled N-1 Thermal Contingency Screening under Deployment Shift

What the math gives to ML

The paper provides a concrete risk-controlled selective-verification mechanism: use a cheap score to skip expensive neural-network checks, then audit a random subset of skipped cases with the true verifier and accept only thresholds whose exact binomial upper confidence bound is below a prescribed violation budget. Its key transferable asset is that the safety guarantee depends on random auditing and calibrated confidence bounds, not on surrogate accuracy, so it remains valid under deployment shift when the audit is sampled independently of scores. This can turn expensive inference-time verification, robust prediction, tool calls, or human review into an adaptive-computation layer with a measurable risk-versus-compute frontier.