A method for analyzing how concepts are represented in a model's internal hidden states by examining their geometric properties.