Precision-Controlled Active Inference Accounts for Sensory Reweighting in Quiet Standing

Human quiet standing depends on the context-dependent reweighting of vestibular, proprioceptive, and visual information. Posturography has empirically characterized this phenomenon, but a minimal generative-control account of how changes in sensory reliability propagate from state estimation to postural action remains incomplete. Here, we test a minimal continuous-time active inference model of quiet standing. In this model, sensory reweighting is implemented as channel-specific precision control over prediction errors. A one-link inverted pendulum receives vestibular, proprioceptive, and visual observations, estimates posture using a generalized-coordinate variational free-energy objective, and selects ankle torque by minimizing the same objective under an upright generalized sensory goal. Context changes alter the relative precision of sensory channels, without changing the plant, action optimizer, or goal dynamics. Across controlled sensory perturbations, reducing the precision of an unreliable visual or proprioceptive channel reduced perturbation-driven postural shifts by approximately 82%. An automatic-differentiation-based state-update gradient contribution closely matched the reduction, identifying the mechanistic locus of reweighting in the perceptual update. A linear reliability-to-precision law, lambda(c) = 1 + 7c, monotonically controlled sensory contribution, and a fixed-precision ablation showed that global precision reduction did not produce reweighting: the behavioral bias remained comparable to the equal-precision condition unless precision was changed selectively across channels. These results support the claim that postural sensory reweighting can be understood as relative, context-selective precision control in a continuous active inference loop.