Great idea! Have you considered storing triplets <range, correct number, selected number> for each try and making image plots of these (x/y coordinates are correct/selected numbers, color of each pixel represents frequency) for multiple users for each range? I think the image might reveal interesting properties of human eyeballing, like near-perfect accuracy around 50%, but with less obvious correlations.