Comment by storus

Do you need to know the full conditional probability distributions of the model to tell if it's a watermarked text, or does this work without that knowledge, i.e. without having the access to the full weights?