Diverse Scenarios, Selective Evaluation: Certified Spectral-Risk Decisions under Costly Robustness Tests

Which robustness tests must we actually run before a risk-sensitive choice is provably fixed? We study this evaluation bottleneck downstream of diversity-driven scenario discovery. Each candidate has scenario losses known initially only through certified intervals; an expensive exact test reveals one loss, and decisions are scored by empirical Expected Shortfall (ES) or coherent spectral risk. We derive exact revelation frontiers that make ES certification polynomial, keep full verification polynomial for fixed spectral-band count, and remain polynomial for arbitrary-band spectra when evaluation intervals form common-point/clique blocks; deletion distance $\chi$ from that geometry gives $O(3^\chi S^4\operatorname{poly}(L))$ fixed-parameter complexity. We also characterize the exact risk-active scenario support and obtain an instance-relative adaptive acquisition guarantee. Experimentally, a benign screen certifies the exact top three with $99/4480$ exact evaluations; all 11 fixed-rank stress controls recover the reference top three with zero audit violations; and a rigorous nonlinear campaign validates all 2880 initial enclosures and $27/27$ decisions, with $3.19$--$3.34\times$ speedups in small/medium regimes. Overall, RiskHedge turns large robustness repertoires into selectively evaluated, exactly certified risk-sensitive decisions.