What You Predict Is What You Defend
Two frameworks, each with a hole exactly where the other has something solid. Join them and the estimate stops being a reading and becomes the target. That single move gives you two dials that come apart cleanly, and they predict different things about who gets better with which kind of help.
This is the only genuinely new claim in the series. Everything in Part 1 and Part 2 already existed, in two literatures that mostly don't cite each other.
What happens when you connect the two models?
The estimate becomes the set point. The estimator combines what you expected with what you sensed and produces an estimate of the current state. That estimate is also your updated prediction of what should be happening, and that prediction is what the controller compares against. There's no separate goal sitting off to one side. The thing you predict is the thing you defend.
Look at what that does to the loop. Sensory input arrives and goes to two places, not one. It goes to the controller, which compares it against the set point and decides how hard to act. And it goes to the estimator, which weighs it against the prior and decides whether to move the estimate at all.
The control literature never had a principled account of where set points come from. The prediction literature never had an account of why the organism cares. Bolted together, each one answers the other's question.
What are the two gains?
Observer gain, K, is how much new evidence is allowed to move your estimate relative to what you already expected. Predictive processing calls it precision. Controller gain, L, is how hard the system acts once an error exists, and it has nothing to do with what you believe. A person can hold a wildly inaccurate estimate and respond calmly, or hold an accurate one and respond violently, and you can't tell which from the surface behavior.
The separation is not my invention. It's the separation principle in control theory, worked out by Kalman in 1960 and formalized by Joseph and Tou the following year: for a linear system with Gaussian noise, the optimal estimator and the optimal controller can be designed independently (Kalman, 1960; Joseph & Tou, 1961).
What's new here is the claim that because the estimate becomes the set point, K determines where the target sits, not just how confident you are about it.
That's the whole argument, and it's worth sitting with. In standard control, the observer gain governs your uncertainty and the set point is handed to you from outside. Here the observer output is the set point, so a change in K moves the target itself. Two people with identical controller gain, facing identical circumstances, will be defending different things if their observer gains differ.
Which is what makes the two dials clinically different rather than merely separable on paper.
Why doesn't knowing something's irrational change how it feels?
Because insight lands at one level and the feeling is generated at another. Changing a high-level belief doesn't automatically change a low-level estimate. "I know it's irrational and I still feel it" isn't a failure of insight or a lack of effort. It's what it feels like when the part of you that talks has changed its mind and the part of you that runs the estimate is still using the old weighting.
If an emotion is the error signal, and the estimate feeding that comparison is an inference rather than a measurement, two things follow. Both happen constantly, and anyone who does this work has watched them happen.
Emotions respond to reframing at all because inferences move when the evidence or the expectation moves. A thermostat doesn't feel differently about 62 degrees once you explain the situation to it. A person does, and that difference is the whole reason talk therapy can work.
And reframing has a ceiling, for the reason in the capsule above.
A parameter learned over fifteen years from consistent evidence won't be revised by one good afternoon of argument. Which is why insight is a weak lever and repeated disconfirming experience is a strong one, and why the behavioral experiment beats the discussion about the behavioral experiment.
Where do these settings come from?
Experience, and mostly the statistics of one particular environment. Volatility teaches observer gain: in an unpredictable world a surprise is signal, so you learn to weight new evidence heavily. Consequence teaches controller gain: where small problems reliably became large ones, high gain is the correct setting. Overshoot teaches damping, because a system punished for waiting learns not to wait.
None of these parameters are fixed at birth. There may be an initial set point, but each one gets estimated from experience after that, assuming the system is working properly.
Someone who grew up in a chaotic house learned to weight incoming evidence heavily, because in that house it was heavily weighted evidence. That is not a distortion. It's an accurate parameter estimate from a real sample.
People demonstrably track this. Behrens and colleagues showed that human learners adjust their learning rate to the volatility of the environment, speeding up when the world becomes unstable (Behrens et al., 2007). Browning and colleagues then showed that anxious individuals fail to make that adjustment (Browning et al., 2015). Those two findings are the most solid empirical ground this framework stands on.
Volatility and consequence set the parameters together, not separately, and laying them on two axes is more useful than either alone.
| | Low consequence | High consequence | |---|---|---| | Stable environment | Dampened reaction, little prediction adjustment | Reliable prediction, strong reaction | | Unstable environment | Fast prediction updating, reaction either way | Rapid prediction updating, strong reaction |
An aside. Attachment theory fans may want to consider this model in the development of stable, anxious and avoidant styles.
The two problem quadrants aren't the ones people expect. Stable and high-consequence produces someone reliable and intense, which looks like competence right up until the environment changes. Unstable and high-consequence produces someone who updates fast and reacts hard, which is exhausting to run and entirely correct for the world that built it.
What does this say about dysregulation?
That most of it is a correctly learned parameter running in an environment that no longer matches it. Incongruent rather than broken. The setting was right where it was learned, the environment changed, and the setting didn't. That reframe changes the clinical conversation from correction to recognition, and it's usually answerable in a session: what environment would have made this the right setting?
It shows up two ways, and they need different repairs.
An ineffective K means the estimate itself is off, so the target sits in the wrong place and everything downstream is aimed at it. An ineffective L means the estimate is fine and the correction is disproportionate, or adjusts in the wrong direction entirely.
Same presentation. Different repair.
Aaron Beck, who built cognitive therapy in the 1960s and gave us the idea that what you believe about a situation drives how you feel about it, would call the learned part a schema: a stored belief about yourself or the world, formed early and applied automatically ever after (Beck & Haigh, 2014).
He'd be right. What the model adds is where a schema comes from mechanically. It's the stored prior, built from the statistics of an environment the person actually lived in, still being emitted as a prediction in an environment that no longer matches it.
One more thing, and it's the engineering half of my background talking. Expectation runs in front of evidence, in a brain and in a boardroom. We decide and then manufacture the logic. The only discipline I know that corrects for it is the one you'd use on an airframe: estimate from experience, test, evaluate the result without flinching, and let the data outrank the argument you fell in love with.
That standard applies to this framework too. Part 7 is where I hold it to it.
Next, though, is what any of this changes in a room with a client.
References
Beck, A. T., & Haigh, E. A. P. (2014). Advances in cognitive theory and therapy: The generic cognitive model. Annual Review of Clinical Psychology, 10, 1–24. https://doi.org/10.1146/annurev-clinpsy-032813-153734
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
Joseph, P. D., & Tou, J. T. (1961). On linear control theory. Transactions of the American Institute of Electrical Engineers, Part II: Applications and Industry, 80(4), 193–196. https://doi.org/10.1109/TAI.1961.6371743
Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82(1), 35–45. https://doi.org/10.1115/1.3662552
The full documents, free and ungated