Performance failures in sustained attention tasks are typically treated as a single phenomenon. In this study, trial-level behavioral data from a large Sustained Attention to Response Task (SART) dataset are used to examine the temporal structure preceding different types of failures. A short-horizon drift signal is computed from reaction time (RT) sequences and aligned to failure events. Two distinct patterns emerge. Lapses (slow failures) are preceded by progressive slowing of responses, while fast commission errors (impulsive failures) are preceded by progressive speeding. These trajectories exhibit opposite slopes over the ten trials leading into failure events and remain stable after within-subject normalization, indicating that the effect is not driven by RT scale differences. Commission errors further separate into fast and slow subtypes, with slow errors showing intermediate dynamics. This demonstrates that error categories in sustained attention tasks are not homogeneous and should not be modeled as a single failure process. The results support a dynamical interpretation in which failures arise from distinct temporal regimes rather than a single noise-driven mechanism. Lapses reflect gradual disengagement, while impulsive errors reflect loss of inhibitory control. These findings provide a framework for analyzing performance breakdowns based on temporal structure rather than aggregated error rates.
Thomas S. Mitchell (Fri,) studied this question.