Affect as a Formal Description of Intelligence Dynamics

We propose Kansei Geometric Dynamics: treating affect/intentionality (kansei) as an endogenous dynamical driver that can steer semantic inference without rewriting the underlying semantic substrate. We instantiate this idea as the Unified Generative Field (UGF), a fiber-bundle extension where an affective gauge charge couples to a connection on a cognitive-state manifold. Curvature introduces a Lorentz-like transverse term in an overdamped flow (“Semantic Hall Drift”), enabling controllable deviation from passive gradient descent and offering a mechanism for avoiding statistically strong but pragmatically misaligned semantic traps. On the Müller–Brown non-convex landscape, moderate charge increases target-reaching probability from 17.9% to 48.6% under matched descent speed. On a GloVe-based semantic graph, kansei-modulated transport redirects the polysemy of stump from a political neighborhood toward embodied affordances with lower path cost. We discuss normalization (LayerNorm/RMSNorm), a dialectic view of hallucination, and an interpretation of in-context learning as exogenous charge injection.