Grounding the Free Energy Principle: A Prior the Dark Room Cannot Satisfy

The Free Energy Principle says that anything which persists acts to minimize the surprise of itssensory exchanges against a model of preferred states. Its hardest unsolved problem is wherethose preferences come from. The standard answer, that an agent prefers the states it typicallyoccupies, is clean and produces three well-known troubles: the dark-room objection, that asurprise-minimizer should seek the most predictable place there is and stay; the circularity worry,that preferences read off behavior predict whatever the agent does; and the boundary problem,that the formalism does not say which Markov blanket to draw. This paper proposes a differentprior. Instead of preferring its typical states, the agent prefers high-P̂O states, where P̂O is itsown embedded estimate of Potential Order, the per-resource capacity of an arrangement to evolvetoward greater resistance to decay and to spread that capacity. We adopt a Boltzmann priorover P̂O and make one structural claim the headline rests on: its steepness is set by the agent’sintegration capacity, the depth and maturity of the dispositional kit it discriminates states with,while the agent’s metabolic budget enters separately, as the capacity to act on the prior ratherthan as its sharpness. That separation is the load-bearing move. Folding budget into steepnesswrongly predicts that an exhausted agent, its prior flattened, becomes indifferent and contentto stay in the dark room; keeping budget in the action layer predicts what animals actuallydo, preferring restorative states as sharply as ever while reaching only the nearest. The darkroom then falls out as a corollary rather than a separate repair: a low-P̂O place is dispreferreddirectly, with no curiosity term bolted on. We separate the scoring criterion from the agent’sestimate of it, which dissolves the circularity; show on a minimal two-state world that the P̂Oprior and the typical-state prior yield different posteriors after a single inference step; distinguishthe proposal from the prior-preference vector of standard active inference; and run the accountthrough the active-inference machinery, where the familiar split between seeking and avoidingsurprise falls out of one structure. We stake the framework on a dissociation test, that steepnesstracks integration while budget gates reach, and add a first-exposure test that separates the twopriors maximally where the typical-state prior has nothing to say.