Fig. 1: Task.

A Example task trial. Participants chose between a safe stimulus (OS) or a gamble which probabilistically led to one of two outcome stimuli (O1 or O2). Information required for value computation was provided in discrete stages to require computation to occur at a specified time-point. Participants were first informed of the point values for all the three outcomes. Participants were then presented with one of four possible probability stimuli (P1, P2, P3 or P4) on which they had been pretrained, indicating four different probability combinations. They then decided whether to accept or reject the gamble. Rejecting led to collection of the safe outcome OS along with its trial-specific associated points. Accepting led to collecting either O1 or O2 along with the trial-specific associated points. All outcome and choice stimuli were represented by decodable visual stimuli. Note that in the example trial, the gamble was accepted. Stimuli in the real experiment were photographs. B Outcome probabilities. The chances of collecting O1 versus O2 upon accepting the gamble depended on which probability stimulus was presented. Probability of reaching O1 was .2, .4, .6, and .8 for P1, P2, P3 and P4 respectively, and p(O2) = 1 - p(O1). These probabilities were extensively pretrained. C Outcome rewards. On each trial either O1 or O2 was designated to be the ‘trigger’ outcome, whose value was selected from three levels (45, 65, or 75 during gain blocks or −45 −65 or −75 on loss blocks). The non-trigger outcome was always 0. OS was selected from 4 levels (20, 32, 44, 56 during gain blocks or −20, −32, −44, −56 during loss blocks). To discourage habitual responding on repeated choices, a variable amount of common noise (between 0 and 20) was added to all outcomes. Finally, a random value (between −6 and 6) was added to each outcome separately. A, C House and scissor images were obtained from svgrepo.com where they are published under MIT licenses. They were respectively created by Adam Whitcroft and scarlab.