Validation-Selected Post-Hoc Calibration of Joint Crash Risk in Multi-Asset Scenario Generators

Multi-asset scenario generators can match marginal and correlation summaries yet misrepresent joint downside events. We propose a validation-selected, base-generator-agnostic corrective transport for finite scenario samples. In training-reference probability-integral-transform (PIT)/logit coordinates, a low-capacity rectified flow learns from tail-biased source--target pairs, while a chronological internal suffix selects strength from candidates including the identity. Across rolling out-of-sample (OOS) tests on two Fama--French industry panels and one sector exchange-traded-fund (ETF) panel, a deliberately misspecified stress proxy's reference-threshold lower-tail root-mean-square error (RMSE) falls by 9.3\%-12.2\% and joint-crash gap by 17.4\%-17.6\%; all six improvements remain pointwise supported across tested circular block lengths. The selector leaves the proxy unchanged in 35.5\%-38.7\% of windows. A self-ranked audit is consistent with a marginal/threshold contribution but does not support broad rank-dependence improvement. No monthly mean-conditional value-at-risk (CVaR) interval excludes zero. Three unadjusted daily Gaussian intervals do, but one simply rescales turnover; these tests do not establish a general decision gain. The results support selective repair of severe reference-threshold misspecification, but not general dependence or decision improvement.