A) Stimulus layout showing the upper-left and lower-right arrays of gratings. Participants indicated which array had higher contrast by pressing the up or down arrow key with their left or right index finger, with the response mapping counterbalanced across blocks. B) Grand average waveforms in C1 electrodes for upper targets (blue) and lower targets (red) in the main task blocks (darkest shade), the easy blocks (intermediate shade) and the blocks in which only the target array appeared (lightest shade). The topographies show the C1 (80-90 ms) for upper and lower-field targets in the easy condition. C) Timeline of two example trials, one with the 400 ms deadline and a late response (top) and the other with the 600 ms deadline and a timely response (bottom). The feedback only told participants whether or not they were on time. Accuracy feedback was given summarily at the end of each block.

Choice-predictive signals used as covariates to control for in C1-choice regression models.

A) Butterfly plot showing the subtraction of lower-chosen trials from upper-chosen trials, including topographies indicating the identified signals. In cases where the topography included both positive and negative poles, electrodes were chosen from both poles as separate signals. B) Time-frequency plots of the same subtraction applied after converting to the time-frequency domain, showing topographies of the six choice-predictive signals identified.

Accuracy and C1-choice relationship as a function of response time.

The sliding RT windows encompassed a range of 20 percentile points each, centred at each unit percentile from 10% to 90%. (A-B) Response time distributions for all trials (A) and for each deadline condition (B). (C-D) Accuracy (percent correct) across the same RT windows. Black circles at the bottom of panel C show RT windows where accuracy was significantly above chance level. Red and blue circles at the bottom of panel D show the same for the 400 ms and 600 ms deadline respectively, while grey circles show windows where there was a significant accuracy difference between them. (E-F) C1 coefficients in logistic models of choice, for the same series of RT windows, multiplied by -1 for the figure so that positive values would correspond to choice-congruent C1 amplitude. Circles at the bottom indicate significant windows in the same way as in panels C-D. See Figure 3 Figure supplement 1 for a regression performed with C1 as the dependent variable, comparing effects of current and previous choices. Error shading in panels C-F indicate standard error of the mean.

Models predicting C1 amplitude as a function of choices on the current trial (A) and the previous trial (B).

In both cases, coefficients were multiplied by -1 so that positive values indicated when C1 amplitude was congruent with the choice. C1 amplitude was predicted by both current and previous choices within a similar restricted RT window.

Percentage of trials in which the upper target was chosen as a function of choices on the previous trial.

Response locked CPP in the 600-msec deadline condition for blocks of hard trials with 55% accuracy (dark green), easy trials with ceiling performance (mid-green) and trials in which only one of the two stimulus arrays appeared (light green), shown for RTs faster (A) and slower (B) than the median.

The mid-green dotted lines show the time points corresponding to mean target onset latency (relative to response), C1 latency, and the emergence of evidence-dependent CPP buildup, demonstrating a delay of approximately 70 ms from the C1 to evidence-dependent buildup. The dark green dotted line shows mean target onset latency in hard blocks.

A) Butterfly plot of the target-evoked VEP in hard blocks showing topographies of the main components observed. B-C) As above but for the subtraction of lower-target trials from upper-target trials (B) and the subtraction of chose-lower trials from chose-upper trials (C). Note that the scale of the y-axis in these panels is smaller than in panel A by a factor of 10. D) Percentage of choices in hard blocks accurately classified by linear discriminant analyses (LDA) applied to choice outcomes. Each classifier was trained on a 10-msec window and then applied to that same window, separately for each participant. To account for overfitting, we shuffled choices and refitted the LDA to demonstrate chance performance (dotted black line). The black dots at the bottom indicate time windows where the real classifier performed significantly better than chance.