anchor
Builds a null distribution by repeatedly shuffling labels and recomputing the statistic. Makes no distributional assumption, which is why it suits scores computed on embeddings.