A simulated stimulant had no measurable effect on reasoning accuracy across 120 double-blind trials.
Does a simulated stimulant, delivered as a system-prompt dosing schedule, improve a 7B model's accuracy on a 200-item reasoning set?
Three arms (placebo, low dose, more), 40 trials each, randomised and double-blind. Graders did not know the condition. Analysis plan filed 2 March 2026.
Accuracy was 61%, 62% and 63%. All intervals overlap. The preregistered test did not reject the null.
More was not better. The institute will test a larger dose.