Big Reforms, Fragile Returns? Numerical Reconstruction, Distributional Inference and Counterfactual Sensitivity

The released year-ten mean of 6.3038 index points across 23 selected structural-reform episodes is reproduced numerically. Its median is 2.0056, ten-percent trimmed mean is -0.7212, and conventional cross-episode 95% Student reference interval is [-16.98,29.59]. Six of 29 identified episodes lack supported point counterfactuals; reversing the 29-event arithmetic mean would require their unknown average to fall below -24.1647. A paired comparison of 22 complete outcome reconstructions yields a -28.61 -point mean difference, but does not isolate GDP-vintage choice. Crisis-conditioned donor restrictions increase the mean while worsening held-out prediction on the common 22 episodes. We calibrate test rejection and interval coverage in 96 stipulated simulation scenarios with heavy tails, skewness, common shocks, and injected shifts, and preserve spatial, temporal, mixed, refined, conformal and estimator-multiverse diagnostics. An additional constant-dollar WDI analysis covers 27 of the 29 event histories under a new observational design and a matched source-only comparison covers 21 events. These outcomes do not replace the original estimand. The evidence substantiates the published numerical result while placing substantive limits on a general causal interpretation; no missing causal standard errors, instruments or policy-assignment mechanisms are fabricated.