Extended Data Fig. 7: Summaries of candidate models and data from planning experiments.

Each row corresponds to a measurement of attention to obstacles from a planning experiment: Awareness judgements from the initial memory experiment, the up-front planning experiment, and the critical mazes experiment; recall accuracy and confidence from the critical mazes experiment; and the binary hovering measure and hovering duration measure (in log milliseconds) from the two process-tracing experiments. Each column corresponds to candidate processes that could predict attention to obstacles: fixed parameter value-guided construal modification obstacle probability (VGC, our model), trajectory-based heuristic search hit score (Traj HS), graph-based heuristic search hit score (Graph HS), distance to bottleneck states (Bottleneck), successor-representation overlap (SR Overlap), expected distance to optimal paths (Opt Dist), distance to the goal location (Goal Dist), distance to the start location (Start Dist), distance to the invariant black walls (Wall Dist), and distance to the centre of the maze (Centre Dist). Note that for distance-based predictors, the x-axis is flipped. For each predictor, we quartile-binned the predictions across obstacles, and for each bin we plot (bright red lines) the mean and standard deviation of the predictor and mean by-obstacle response (overlapping bins were collapsed into a single bin). Black circles correspond to the mean response and prediction for each obstacle in each maze. Dashed dark red lines are simple linear regressions on the black circles, with R² values shown in the lower right of each plot. Across the nine measures, value-guided construal tracks attention to obstacles, while other candidate processes are less consistently associated with obstacle attention (data are based on n = 84215 observations taken from 825 independent participants).