04model playground
this is the fitted model running live in your browser on the same task. pick a fitted agent, or move the parameters yourself, then step through a game: the buttons show the model's sampling policy, and the panels show its bayesian posterior, its distorted memory, and the final-choice policy it would use if the game ended now.
occlusions follow the text version of the task (each button visible with probability 0.5), whose exact optimum is the paper's 67.8%. parameters are the paper's cross-validated fit averages. the dashed lines in the bottom panel are the model's final-choice policy \(\sigma(\kappa_f \mathbf{h}_t)\) if the game ended at round \(t\); with a stubborn memory they keep favoring early evidence even when the posterior has moved on.