Pre-registration: when does dynamics-aware tokenization pay off? (SFVT / TSB-AD-U)

Abstract

Pre-registration of the predictions, frozen cutoffs, and analysis plan for testing whether the k-means token transition rate predicts when successor-field tokenization (SFVT) outperforms position-based tokenization on TSB-AD-U anomaly detection. Written before any anomaly-detection result was computed; no anomaly label was read in producing the accompanying atlas. Includes the atlas over 867 TSB-AD-U series and named datasets where we predict the method will show no advantage.