Is Free-Energy Minimisation a Generative Principle or a Privileged Representation of Self-Maintaining Dynamics?

Abstract

The free energy principle (FEP) is used both as a representation of organised dynamics and as a candidate account of the mechanisms that produce them. We separate these claims in artificial systems whose controllers were specified by proportional–integral–derivative control, reward learning, or viability selection rather than variational free-energy minimisation. Trajectory-only observers fit explicit Gaussian free-energy flows and matched alternatives, then faced four tests: drift representability, restriction binding, state sufficiency, and stochastic consistency. For any fixed sensory feature map, a dissipative FEP flow is exactly an affine flow whenever the fitted state matrix has a negative-definite symmetric part. This restriction was non-binding for the evolved controller; for PID, however, a nearly singular state direction made the recognition map strongly dependent on the regularisation floor despite a small overall projection. Increasing feature flexibility did not provide a general predictive advantage: a radial-basis observer slightly improved PID prediction but substantially worsened intervention extrapolation for hard and smoothed Q policies. Adding omitted sensory history recovered low error under delay (trajectory normalised mean-squared error at most 0.046 for PID and evolved controllers). A scale-invariant covariance-shape diagnostic distinguished matched from deliberately mismatched internally stochastic Langevin controls, but the deterministic target suite supplied no internal diffusion from which to infer a stochastic mechanism. Smooth non-maintaining controls also admitted accurate reconstructions. Empirical content therefore lies not in post-hoc free-energy notation alone, but in independently tested dynamical restrictions, state variables, stochastic predictions, and prospective generalisation.