Retrain a probe on randomised labels to measure how much of its performance comes from the representation rather than from the probe's own capacity. Without it, a strong probe scoring well proves nothing. Introduced by Hewitt and Liang as probe selectivity.