Bridging Neurons to Behaviour: A Generative Neural Engine Mechanistically Rejects the Independent Race Model

A central challenge in systems neuroscience is understanding how the brain generates robust, low-dimensional behaviour from variable, high-dimensional neural activity. To bridge this gap, we developed a generative neural engine using a Deep Markov Model that captures the stochastic dynamics of the macaque premotor cortex during a countermanding task. The two defining properties of this engine are low dimensionality and Markov dynamics: together, they guarantee that the extracted representation is a genuinely compact dynamical system. Our analysis demonstrates that 3 dimensions constitute the minimal “tipping point” required to maximize dynamical predictability and functionally distinguish competing motor plans with near-perfect accuracy. This engine emergently acts as a generative algorithm, reproducing the complex distribution of behavioural reaction times despite being trained exclusively on neural activity. Utilizing this ‘virtual brain’ for in silico experiments, we mechanistically test—and reject—both independence axioms of the Independent Race Model. A simulated stop-failure analysis reveals SSD-dependent distortions of the reaction-time distribution that violate context independence, with profiles consistent with the Pause-then-Cancel framework and, in one subject, closely mirroring the human behavioural pattern reported by Bissett et al. (2021). A per-trial simulation further reveals a distinct positive correlation between reaction times and inhibition latency, violating stochastic independence. Both violations are physically mandated by the geometry of the shared neural manifold, replacing the abstract race with an interactive dynamical account. Finally, we demonstrate a novel protocol for designing perturbations in the neural trajectories that induce systematic shifts in reaction times. This work provides a data-driven generative algorithm that physically connects neural implementation to behavioural dynamics, bridging the gap between Marr’s Implementation and Computation levels.