Agency emerges from asymmetric access to motor information in a minimal predictive framework

The ability to distinguish self-generated actions from externally caused events, commonly referred to as the sense of agency, is a fundamental property of biological agents. Despite extensive work within predictive coding and related frameworks, the minimal computational conditions under which agency becomes well-defined remain unclear. Here we propose that agency emerges from competition between predictive models operating under asymmetric access to motor information. We introduce a deliberately minimal predictive framework consisting of two systems that receive identical sensory input but differ solely in access to motor commands. One system implements a motor-conditioned forward model, whereas the other implements an environment-based predictive model without motor access. Agency is quantified by an agency index defined as the difference between prediction errors, allowing attribution to be interpreted as implicit model selection within a statistical inference framework. We show that stable agency attribution emerges exclusively under asymmetric information access. Symmetric predictive architectures, whether motor-aware or motor-blind, yield identical prediction errors and render agency attribution computationally undefined. By varying sensorimotor delay, we identify a relative critical delay at which agency collapses when consistency is violated.