Fig. 1: Task design and behavior.

a Schematic sequence of events in example trial. Visual evidence samples (polar angles from vertical reference line, of small discs at 5° of eccentricity) are drawn from Gaussian distributions. After the sequence, participants report their inference about the underlying source (distribution mean) by moving a joystick. After six samples, participants either report a categorical judgment about the source, or receive a categorical cue (see main text). After cue presentation, participants make a pre-assigned (left or right, blocked) button press, to match Choice and Cue conditions in terms of motor output. b Range of generative means (μ) and standard deviation (σ). Red font indicates the near-zero generative mean trials used for consistency effect analyses (−4° to 4°) (c) Accuracy of the intermediate choice and reliability of the cue for the near-zero trials indicated in (b). d Proportion of “right” choices as a function of generative mean (Choice; first 6 samples). For comparison, the fraction of “right” Cues are plotted against the first 6 sample mean for Cue. Lines: fits of a sigmoid function on group data. Dots are group averages. e−i Quantification of estimation biases. e Distributions of estimates as a function of generative mean for Cue condition. Estimates were first z-scored and binned before pooling probability densities across participants. f Same as (e), for Choice. g Distributions of estimates collapsed across generative means for Cue (gray) and Choice (black). P-values, Hartigan’s dip test of unimodality. h Mean estimates for trials with ‘left’ and ‘right’ choice or cue categories on near-zero trials. Data points, participants; p-values, two-sided permutation tests. i Difference of mean estimates between ‘right’ and ‘left’ categories (as in (h)), as function of the strength of category evidence (binned). Lines, group average; error bars, SEM; statistics: interaction condition x evidence strength, two-factorial ANOVA, ηp² = 0.05; 95% CI: [−0.096, 0.195]). Inset, ideal observer prediction (see main text). N = 30 participants for all panels except (c, e, i): N = 27 (excluding three participants with different spacing of generative means not suited for the corresponding analysis; Methods).