Entropy laundering on a real watermarked trace. Top: the trace coloured by the student's uncertainty before each token. Bottom: the most uncertain half of the tokens is dropped from the training loss but kept as context.