The Description Dressed as a Law: A Structural Audit of the Free Energy Principle and the Case for Perceptual Control

Abstract

The Free Energy Principle is the dominant unifying proposal in the contemporary brain and behavioural sciences. From a single premise — that any system persisting as a distinguishable thing can be described as minimising variational free energy — its proponents recover perception, action, attention, learning, and the boundary of the organism itself. This paper does not dispute the framework's mathematical seriousness, and states its case at the strength its authors currently give it before making any argument against it. The argument is narrower and, if it holds, more damaging. The framework declares a division between an unfalsifiable principle and a falsifiable periphery of process theories, and this paper argues that the division is not kept under load. Three faults are set out, each tested against a different kind of evidence. The first is a category error: the collapse of what a system wants into what it expects, and the inability of precision weighting to distinguish a goal being set from a goal merely being predicted, except after the fact. The second is empirical: the base imperative implies that an environment without variation is optimal, while the clinical and electrophysiological record of environments in which homeostatic needs are met and variation is withdrawn demonstrates the reverse. The third is structural: the framework's formal content has been substantially revised under criticism three times, its advertised empirical reach was extended over the same period, and nothing has been published relating the second to the first. The paper then argues that the phenomena the framework reaches for first — stabilisation against a varying environment, the persistence of goals under disconfirmation, and the rejection of disturbances that were never modelled — are accounted for more economically by a closed negative-feedback architecture specified before the framework existed, together with an operational procedure for identifying what a system controls that the framework has no equivalent of. What that architecture does not yet have is set out in full. Two features of the method are worth flagging. Where the framework has an answer, the answer is conceded before the objection is made, and in two cases the concession is total. And the paper states the conditions under which each of its arguments would be shown wrong — two of which have already been met, with the corresponding claims withdrawn in the body of the text rather than reinterpreted. Version 2.0 supersedes version 1.0 (doi:10.5281/zenodo.21309527), which presented four faults rather than three and contained a misstatement of the framework's mathematics that is corrected and recorded in Section 7.