(a) Schematic of the ‘target-locked-rise’ hypothesis. Under this hypothesis, preparatory activity develops shortly after target onset, but rises at different rates on different trials, leading to different RTs. A related hypothesis (not illustrated) is that the rise is stochastic and a higher average slope produces short RT trials. These hypotheses lead to the following prediction when trials are aligned to movement onset and averaged: when looking back in time, preparatory activity should decline slowly for trials with longer RTs (which had shallower slopes), and quickly for trials with shorter RTs (which had steeper slopes). (b) Schematic of the ‘prepare-on-demand’ hypothesis. Under this hypothesis, preparatory activity stays weak until close to the time of movement. Activity then rises in a stereotyped profile, and movement onset follows shortly thereafter. Note that this profile could involve a step, a ramp, or a variety of other shapes. The key aspect of this hypothesis is not the exact temporal profile, but rather the proposition that the profile is stereotyped across trials. Despite this stereotypy, average target-locked data could appear as a steady rise. In contrast, averages locked to movement onset would better capture the typical temporal profile. Thus, a central prediction of the prepare-on-demand hypothesis is that movement-locked averages will look similar for short-RT and long-RT trials. (c) Preparatory subspace occupancy computed from data locked to movement onset for monkey Ba. For each neuron, trials were divided into those shorter and longer than the median. This yielded both a short-RT and a long-RT population response. Populations responses were projected into the preparatory subspace (found using the cue-initiated context, as in Figure 8) and occupancy was computed. Only data after target onset are shown. Thus, the leftwards extent of each trace is limited by the longest interval between target onset and movement onset (which is of course longer for long-RT trials). The profile of preparatory occupancy is similar for short-RT and long-RT trials. These results are most consistent with the prepare-on-demand hypothesis. There is only a slight hint of occupancy being higher for long-RT trials prior to movement onset, as predicted by the target-locked-rise hypothesis. (d) Same as panel c but for monkey Ax. In summary, temporal profiles were somewhat different for the two monkeys (more ramp-like for monkey Ba and more step-like for monkey Ax). However, in both cases the profile was very similar regardless of RT, consistent with the prepare-on-demand hypothesis. We stress that this analysis still averages over potentially diverse single-trial events. For example, the ramp of preparatory subspace occupancy that begins ~400 ms before movement onset could be a ramp on every trial, or a step with somewhat variable timing: sometimes rising 350 ms before movement onset and sometimes rising 200 ms before movement onset. What can be concluded is that RT variability is not primarily driven by variability in the slope of ramping preparatory activity.