The pupil regression reported in the main text (Figure 5b) examined the relationship between precision-weighted prediction errors (ɛ3; PE’s) and pupil size across all trials in the experiment. A strength of this analysis is that it represents the pupil response when each participant was actually surprised, and does not impose knowledge of the task structure. Nonetheless, to examine the relationship between precision-weighted PE and pupil size in the volatile and stable periods of the task we conducted the same regression analysis (see main text and online methods) but separately for the 72 ‘stable’ trials and ‘72’ volatile trials (see Figure 1) towards the end of the experiment. (left) In the stable period there is no relationship between precision-weighted prediction errors and pupil size in either group or no differences between the groups. (right) The relationship between precision-weighted PE’s and pupil size in the ASD participants (blue) is apparent 1000ms after the outcome appears in the volatile period of the task. Blue solid line shows where the ASD participants differ from zero and black dotted line shows where the ASD participants differed from the NT participants. Shaded region represents standard error of the mean. Consistent with the analysis of learning rates in the volatile and stable task periods (Figure 3c), this suggests that the ASD participants tend to show aberrant noradrenergic surprise about volatility, in response to volatility (e.g. over-updating learning about volatility and over-engaging noradrenergic responses to surprise about volatility, in the face of environmental volatility). However, we caution against the low trial numbers included in this analysis (72, vs a maximum of 456 in the analysis reported in the main text) and the fact that one control participant did not have enough good trials in the volatile period to be included in this analysis, so participant numbers are also reduced (ASD=11, NT=13).