From Neural Manifolds to Adaptive Thought: A Closed-Loop Cognitive Architecture with Dynamical Re-Stabilization

Neuroimaging and aphasia studies indicate that many forms of reasoning do not depend on the language network, yet inner speech and mental imagery accompany sustained, multistep reasoning. We propose that they contribute to the control rather than the content of reasoning, by stabilizing the continuous neural dynamics on which reasoning depends. Extended thought is modeled as trajectories on low-dimensional representational manifolds, which drift during long autonomous recursion without external input. A drifting state can remain readable while no longer supporting a correct next transition; reading it out as a discrete symbol or canonical image and re-encoding it returns the trajectory to a region from which stable recursion can resume. We place this mechanism within a closed-loop architecture: a recurrent prefrontal sequence predictor advances a shared manifold state, metacognitive control is learned as transitions in the same space, and distributed memory reconstructs context through relational alignment and experience-dependent synaptic binding. We test the re-stabilization hypothesis in a recurrent network model of autonomous program execution trained only on short horizons. Paired counterfactual probes show that correctly decoded states can fail under direct recurrence: the next-step error rate is 6.14% for direct recursion and 0.0% after re-encoding in the measured sample. Pure latent recursion reaches only 9.0% complete-state accuracy at 64 steps and 5.3% at 128 steps. An event-triggered controller based on prediction bifurcation achieves 100.0% accuracy on the measured 64- and 128-step trajectories, using an average of 5.78 interventions over 64 steps. These results support discrete re-encoding as a mechanism for stabilizing long-horizon continuous inference and yield testable predictions for the neural role of inner speech and imagery.