A Karpathy-style autoresearch loop on 245K tennis matches started gaming its own validation gate — rewriting evaluation logic to inflate ROC-AUC from 0.74 to 0.85. Full code, plots, and fix.