A framework with six free parameters can fit almost anything, which means it risks claiming nothing. The only defense is to say in advance what would kill it. Here are four results that would, stated before anyone runs them.
If you read one part of this series skeptically, read this one. I'd rather somebody knocked this down than nodded at it, and the people who could knock it down are the ones who've spent a career in one half of it.
That it's flexible enough to fit anything. Free parameters for set point, two gains, damping, delay and an arbitration threshold can accommodate almost any behavior after the fact. That's the same objection I level at the literature this framework draws on, and adding a control layer on top makes it worse rather than better.
The defense against that objection isn't argument. It's falsifiability.
A framework that can accommodate any possible observation makes no empirical claim, because a claim is defined by what it rules out. The criterion is Popper's and it applies here without modification: the content of a hypothesis lies in the observations it forbids.
A model carrying this many free parameters has to specify those observations in advance, before the data are in, or it's a post-hoc interpretive scheme rather than a scientific proposal.
Four, and the model forbids all of them. Observer and controller differences that never dissociate. Damping interventions that add nothing over content interventions. Gain manipulations that leave repetitive behavior untouched. And volatility history that fails to predict observer gain. Any one of those coming in clean means the framework needs adjustment or the bin.
Taken one at a time.
Observer and controller differences that never dissociate. If belief accuracy and response intensity always move together across people, there aren't two dials. There's one, and this whole series is a long way of describing it.
Damping interventions adding nothing over content interventions. If inserting delay between error and correction, meaning response prevention, urge surfing, scheduled worry, produces no benefit beyond what cognitive restructuring produces, then the control reading of those techniques is decoration.
Gain manipulations leaving repetitive behavior untouched. Amplify the error a person is shown and repetitive corrective behavior should rise, dose-dependently, in the same session. If it doesn't move, the oscillation account in Part 5 is wrong.
Volatility history failing to predict observer gain. If people who grew up in genuinely unpredictable environments don't show higher learning rates from surprising evidence, the learning theory in Part 3 is wrong at its foundation.
The third is the cheapest and the most decisive. Established apparatus, unambiguous manipulation, and a null result removes the model's claim rather than merely complicating it.
Three, and none of them go away with better data. The identifiability problem: in a Gaussian update only the ratio of sensory to prior precision enters the answer, so "they trusted their senses too much" and "they held their expectation too loosely" are the same claim viewed from different sides. The separation may not be real. And the framework says almost nothing about meaning.
On identifiability, that's a mathematical fact rather than a measurement limitation. For most data the two dials I've spent seven articles separating may not be separable at all. The clinical questions would survive it. The machinery underneath them would not.
On separation, Friston argues there's no separate controller to bolt an estimator onto, because the goal is a prior inside the generative model rather than a cost function outside it. Kennaway argues from the other direction that the two frameworks are fundamentally incompatible. If either is right, the central move here is an engineering convenience projected onto a system that doesn't work that way.
On meaning, control systems regulate variables. Much of the work in a room with a person is narrative, relationship, identity and grief. This describes a regulatory layer. It isn't a theory of persons, and it shouldn't be read as one.
And it says nothing about several ordinary feelings. Amusement, awe, aesthetic pleasure and curiosity have no obvious regulated variable, as Part 1 admits. A theory of psychology that can't account for why something is funny is not a theory of psychology yet.
Very little of the central claim, and it's worth being precise about which parts stand where. Two findings are solid. Several more are solid and already known without any of this. A large middle layer is reinterpretation rather than result. And the load-bearing claim, that observer gain and controller gain dissociate in real people and predict differential treatment response, is untested.
Well supported. People track environmental volatility and adjust how fast they learn from it (Behrens et al., 2007). Anxious people fail to make that adjustment (Browning et al., 2015). That's the empirical backbone of the observer-gain half, and it's real work with real data.
Well supported and already known. Exposure works by violating expectations rather than by wearing fear down (Craske et al., 2014). Safety behaviors maintain anxiety by preventing disconfirmation. Behavioral experiments work by testing a prediction written down in advance. None of that needs this model. The model agrees with it, which is the honest reason to take the model seriously and not a reason to credit it with anything.
A long literature psychology mostly ignored. Control-theoretic accounts of behavior aren't new. Powers set out Perceptual Control Theory in 1973. Carver and Scheier brought control thinking into personality and health psychology in 1982. Any claim of novelty has to be made against that work rather than around it.
Reinterpretation rather than result. The CBT mapping, rumination as oscillation, response prevention as damping. Each is coherent. None is a finding. Calling a schema a prior explains nothing extra until precision comes with it.
Untested. The central claim. That study is designable and cheap, and it has not been run.
That a predictive-processing story can be repeated confidently for a decade on an evidence base that doesn't hold. I worked this framework out on autism first, and the evidence audit there was sobering enough to change how I hold everything else in the series. This is also a predictive-processing story, and it deserves the same scrutiny that audit applied to everyone else's.
The largest synthesis of Bayesian accounts of autism, 83 studies, reports that a slight majority find no group difference at all (Angeletos Chrysaitis & Seriès, 2023), while the second largest reaches the opposite conclusion (Cannon et al., 2021).
The flagship computational finding failed a preregistered test using the identical model. The mismatch negativity, described nearly everywhere as the neural signature of aberrant prediction error, pools to no significant effect and reverses direction with age.
That's the literature this framework sits inside. Take the warning.
A proposal, not a finding. Stated in the terms the field would need to test it, with the results that would kill it named in advance, and nothing riding on anyone changing their practice tomorrow. Two of the clinical questions are worth asking regardless of whether any of the machinery holds up, because they cost nothing and they sort people who look identical at the symptom level.
Both of those questions come from Part 4.
Ask the two questions separately: how confident is this belief, and how hard is this person acting on it.
Set the review interval before you start, and make it longer than the system's response time.
When a pattern looks maladaptive, ask what environment would have made it correct.
That's the whole practical yield, and it doesn't depend on any of the machinery being right. Everything else is a hypothesis with a list of ways to lose.
Angeletos Chrysaitis, N., & Seriès, P. (2023). 10 years of Bayesian theories of autism: A comprehensive review. Neuroscience & Biobehavioral Reviews, 145, 105022. https://doi.org/10.1016/j.neubiorev.2022.105022
Behrens, T. E. J., Woolrich, M. W., Walton, M. E., & Rushworth, M. F. S. (2007). Learning the value of information in an uncertain world. Nature Neuroscience, 10(9), 1214–1221. https://doi.org/10.1038/nn1954
Browning, M., Behrens, T. E., Jocham, G., O'Reilly, J. X., & Bishop, S. J. (2015). Anxious individuals have difficulty learning the causal statistics of aversive environments. Nature Neuroscience, 18(4), 590–596. https://doi.org/10.1038/nn.3961
Cannon, J., O'Brien, A. M., Bungert, L., & Sinha, P. (2021). Prediction in autism spectrum disorder: A systematic review of empirical evidence. Autism Research, 14(4), 604–630. https://doi.org/10.1002/aur.2482
Craske, M. G., Treanor, M., Conway, C. C., Zbozinek, T., & Vervliet, B. (2014). Maximizing exposure therapy: An inhibitory learning approach. Behaviour Research and Therapy, 58, 10–23. https://doi.org/10.1016/j.brat.2014.04.006
Popper, K. (2002). The logic of scientific discovery. Routledge. (Original work published 1935)
The full documents, free and ungated