Language of Stress and
Predictive Processing
Where We Converge and Where We Diverge
Pace, J. C. (2026). The Language of Stress and Predictive Processing: Where we Converge and Diverge. FigShare. DOI: 10.6084/m9.figshare.31286308
Overview
Predictive processing and the free energy principle are the most developed frameworks in contemporary cognitive science, and any account proposing that value rather than information is the brain's primitive owes them a careful answer. This paper is that answer. It is written to concede as much as the evidence requires before claiming anything.
A substantial portion is given to what the two frameworks share and to phenomena where no distinct claim is made. Attention capture by a faint but relevant signal, expert selective attention, engagement with fiction, novelty seeking, empathic response to people we are close to: predictive processing handles each of these, and the paper says so rather than offering them as support. It also withdraws claims made in an earlier version, including an argument that the cocktail party effect exposed a gap in precision weighting. It did not, and the withdrawal is recorded.
What remains is a disagreement about what a valenced state is. On the predictive account, value enters as priors over expected states and prediction error is minimized; on the account offered here, a valenced state is the mattering itself rather than a representation of it. The paper traces that single difference through six positions where the two frameworks touch: what the system treats as primitive, what it defends, how deviations are weighted, whether aversiveness persists once error resolves, whether a route's pull is laid down by history or computed from the current model, and the identity claim itself. These are argued as one ontological difference seen from six angles rather than six separate findings, which the paper notes is the more credible version of the claim.
The disagreement becomes checkable at one point. Precision, in predictive processing, is estimated reliability. Rigidity, here, is how much depends on an expectation holding. They come apart in a specific case: a belief held weakly but staked heavily should update readily on a precision account and should resist on this one. That case is aimed at precision-weighted belief updating. It does not by itself separate the framework from active inference’s prior preferences, which are expectations held against evidence and made true by action; what would separate them there is whether a preference updates with what the organism actually occupies, and the paper says that comparison still needs its own treatment. Each of the framework's six predictions is stated with its falsification condition, and where the other side has a strong reply, it is given in full before any response. The central prediction can be run now in a setting where the stake is assigned from outside rather than measured; its self-relevant form needs beliefs ordered by stake, not an instrument. Only one prediction, in its strong form, waits on an instrument for comparing rigidity across people, and that instrument does not yet exist.
Version 2.1 corrects where the paper places the regress argument against active inference. It had pointed to what sets the precision of a preferred-state prior. Preferences are part of how each policy is scored, so they sit on the evaluation side. The regress meets active inference one step later, at the precision over policies. It sets how decisively the best-scored policy governs without changing any policy’s evaluation, and where habits enter as a prior over policies it can change which policy governs. The paper now says so, and names the empirical condition under which the account fails for human deliberation: a dopaminergic gain that proves separable from valuation. The identity prediction is also restated so that it is tested by felt relief when a demand is resolved, not by report, which narrows a falsifier.
Version 2.2 follows The Prioritization Regress (PhilArchive). It says what the central prediction does not separate, which narrows what the paper claims for it, and the version moves for that reason. It also says that continuous-time active inference, taken alone, settles reflexively; that a brain implementing policy selection as competition with no separable gain would be a candidate implementation of the framework rather than a rival; how the framework’s one structure compares with outcomes sharing a unit within one generative model; and what active inference’s own account of valence holds (Hesp et al., 2021).