The decision to stop "counting" explains numerical underproduction in primates

Sequential numerical production requires generating a target number of actions, but the mechanisms governing the termination of counting sequences are unclear. We trained two rhesus macaques to produce instructed numerosities by generating the corresponding number of actions in response to sign numerals or dot arrays. Although action sequences were controlled for temporal parameters, both monkeys performed accurately across both stimulus formats. However, their performance accuracy showed a systematic asymmetry: underproduction occurred more often than overproduction, contrary to predictions from perceptual numerosity judgments governed by Weber-Fechner scaling. This asymmetry was mirrored in reaction times, with longer latencies for underproduced sequences and shorter latencies for overproduced ones. A stochastic accumulation-to-bound model in which a decision variable evolves across counting steps and triggers termination when a threshold is reached can account for both accuracy and reaction-time patterns. These results provide evidence for a sequential stopping process that contributes to numerical production, relating counting behavior to general principles of decision formation.