11 OF 22 PART THREE — WHEN IT GETS STUCK

Prediction Updates

Priors change — but not through insight. Understanding how the nervous system actually revises its models explains why so much that looks like progress doesn't hold, and what conditions make genuine change possible.

6 minute read

Every human being carries a vast collection of predictions about the world — about what is safe, what is dangerous, who can be trusted, what their own feelings mean, what tends to happen when they try things. These predictions are not primarily conscious beliefs. They are the system's operating model: built from experience, continuously updated, running mostly beneath awareness, generating behavior, emotion, and interpretation in real time.

Understanding how these predictions actually update — the specific conditions required for a prior to revise — is one of the most practically important things this framework addresses. Because almost everything that goes wrong in attempts to change patterns comes from a misunderstanding of this mechanism.

How priors are built

Priors form through repeated exposure to patterns. When the system encounters the same combination of conditions producing the same outcome enough times, it builds a prediction: "when X, then Y." This happens automatically, across all layers of the compression stack — from the procedural level (this sensation predicts danger) to the conceptual level (conflict predicts rejection) to the narrative level (this is just who I am).

Early priors have privileged status. They were formed when the learning system was most plastic, most dependent on the environment for calibration, and most in need of reliable predictions for survival. This is why attachment priors are so durable — they were formed under conditions of maximum urgency and maximum learning sensitivity. The nervous system was trying to solve, as fast as possible, the most important problem it would ever face: how to secure proximity to the people its survival depended on.

Priors don't get replaced by understanding. They get revised by experience that genuinely contradicts them — registered at the layer where they live, with enough precision and repetition that the system updates its model rather than filing the exception.

The precision problem

One of the most important variables in prior updating is precision — how confident the system is in its current prediction. High-precision priors are resistant to update. They are resistant because the system has a lot invested in them — they represent hard-won learning about how things reliably work. When incoming experience contradicts a high-precision prior, the system has two choices: update the prior, or discount the experience as an anomaly.

Most of the time, with high-precision priors, the system discounts the experience. This is the mechanism behind "one step forward, two steps back" — the familiar pattern of meaningful therapeutic progress followed by a return to baseline. The new experience registered, but the prior's precision was high enough that it was filed as an exception rather than as grounds for revision. Nothing actually updated.

This is also why insight alone so rarely produces lasting change. Insight is a conceptual event. It operates at Layer 7–9 of the compression stack. The prior it is trying to address may live at Layer 3 — embodied, pre-verbal, inaccessible to language. The insight is real and may even be accurate. It just isn't happening at the level where the prior lives.


What actually produces updates

Prior updating requires prediction error at the right level of the stack — a mismatch between what the system predicted and what actually happened, registered at the layer where the prior is held. Not read about. Not understood. Experienced.

What updates priors
New experience that genuinely contradicts the prediction, registered in the body, repeated enough times that the system can't file it as an anomaly. Completing — not just managing — the feared outcome. Disconfirmation at the layer where the prior lives.
What doesn't update priors
Understanding why the prior formed. Knowing it's "irrational." Being told it's inaccurate. Experiencing exceptions while the prior's precision is high enough to discount them. Relief strategies that reduce the signal without contradicting the prediction.

This is what makes therapeutic exposure work when it works — not because the person is "facing their fears," but because they encounter the thing they have been predicting as catastrophic, and it doesn't produce the predicted outcome, and that prediction error is registered in the body, and with enough repetition, the prior revises. The model updates. The next time the cue appears, the prediction is different — because the system has actually learned something, not just understood something.

Safety as a precondition for updating

Here is a crucial and often missed requirement: prior updating is resource-intensive. It requires the system to hold prediction error open — to sit with the mismatch between what was expected and what is happening — rather than immediately routing to a strategy that resolves the pressure. That requires capacity. It requires enough tank that the discomfort of the mismatch doesn't immediately trigger avoidance or shutdown.

This is why safety is not just a nice-to-have for therapeutic work. It is a literal mechanistic requirement. A system that is in the strategy stream — managing live pressure, burners on, tank depleted — cannot hold prediction error open long enough to update. It routes to relief. The relief works. The prior is untouched.

The Conditions for Genuine Update

Sufficient tank: Enough capacity to hold prediction error without immediately routing to relief. The system has to be able to stay with the mismatch.

Contact at the right layer: The new experience has to reach the layer where the prior lives. Conceptual experiences don't update procedural priors. Embodied experiences — in the right relational context — can.

Enough repetition: The system needs multiple disconfirmations before a high-precision prior revises. Single exceptions get filed as anomalies. Patterns get incorporated.

Low enough arousal: When arousal is too high, the learning system narrows — it encodes the specific details of the specific event rather than extracting a generalizable update. Moderate arousal, in safety, is the optimal learning state.

Implications for clinical work

Many people come to therapy having tried very hard to change — having understood their patterns clearly, having committed to different behavior, having worked with genuine effort and intention. And the patterns persist. This is not character. It is mechanism. The priors that are running their experience were not formed through understanding and cannot be revised by it.

Genuine prior revision requires the conditions listed above. And creating those conditions — building the tank, reducing unnecessary load, creating safety, establishing the relational context in which new experience can actually register — is not preliminary to the real work. It is the real work. Or rather, it is the ground without which the real work has nowhere to land.

← Previous
10 · Deconstruction
Next →
12 · Relief vs. Completion