A preprint reports near-perfect speech-deepfake source classification on a closed 17-generator benchmark, with feature-importance rankings that broadly agree with SH