The score was almost certainly true; the mistake was treating a measurement of a model as a verdict on the system built around it.