a, Bayesian model comparison using BIC scores revealed that the Maintain model explained the data better than a Decay model in which a free parameter captured how much the animals would ‘forget’ the unavailable option as well as two risk-sensitive models where the probabilities could be distorted by the animals. BICs are summed across 20 sessions. b, Distortion parameter for the simple distortion model across animals/sessions as well as the mean. The blue line represents the mean distortion parameter across sessions (n = 20). c, To compare the Decay and Maintain model, we simulated choice data using the best fitted parameter of each model and compared choice rates pre and post reversals (~110th trials) (left panel) as well as switch behaviors after a win or a loss (right panel); n = 20 sessions. The real data are represented in black.