Fig. 5: nTc marks the moment of internal decision commitment.

a, Example trial. The inferred time of commitment is far more precise when neural activity is used (nTc, orange line) than when it is inferred solely from sensory stimulus timing and choice behaviour (dashed grey line). b, Distribution of estimated nTc values relative to stimulus onset. Among the 34.7% of trials in which commitment times could be detected, nTc varied widely relative to the onset of auditory click trains. The decline in nTc frequency over time reflects randomized stimulus durations (0.2–1.0 s). c, Distribution of nTc values relative to movement onset to report the decision of the animal (exiting centre fixation port). As in b, nTc timing also varies widely across trials. The leftmost bin includes trials in which the nTc occurred more than 1 s before movement. d, Supporting the interpretation of nTc as a decision commitment and, despite the highly variable timing of nTc, sensory evidence presented before nTc affects the decision of the animal but evidence presented after nTc does not (weight of clicks on choice inferred using logistic regression). Trials for which the estimated time of commitment occurred at least 0.2 s before stimulus offset and 0.2 s or more after stimulus onset were included for this analysis (9,397 of 55,057 trials across 115 sessions with 12 rats). The green line is the prediction from the MMDDM model fit to the data. e, Behavioural accuracy was lower in trials in which nTc could not be identified. Predictions were made by fitting MMDDM to the data, simulating trials from the fitted models and applying the same nTc detection procedure as that used for real data. Dashed reference lines at abscissa = 0 and ordinate = 0.5. f, nTc was more likely to be identified in trials with stronger evidence. For each evidence strength bin, the fraction of trials with an identified nTc was divided by the overall trial fraction across all bins, which was lower in the data than in the model predictions. Black circles and green lines indicate the mean across sessions. Black error bars and green shading indicate the 95% bootstrap confidence of the mean. Dashed reference lines at ordinate = 1.0.