PHLE method hits AUC of 0.84 in decoding hidden model intent

A LessWrong post describes persona-header likelihood evaluation (PHLE), which uses base pre-trained models with an appropriate header to decode subliminal signals in model outputs. Using 30 examples, PHLE differentiated number sequences from a model prompted with "You love owls" vs "You love eagles" at an AUC of 0.84, while prompted classifiers stayed near chance at an AUC of 0.59.