🧠 Cognition Theory v1

What is this?

Cognition Theory v1 is a 16-chapter framework organized into four blocks: foundations, core mechanisms, advanced discussion, and applications and boundaries. It asks a simple question — how are our judgments actually formed? — and answers it with a single model of the cognitive system.

Its one-sentence claim: cognition is not "seeing the world", it is "constructing a model that can be used to predict the world."

The theory deliberately avoids handing out a list of "correct beliefs". Its goal is a skeleton of a theory that can be tested against reality and revised repeatedly — including revision of its own highest-level assumptions. Because it is written in substrate-neutral terms (information in, model built, prediction out, error-fed update), the same vocabulary can be used to describe human cognition and machine learning systems; the text itself focuses on the cognitive subject.

Core ideas: one chain, eight symbols

The entire theory unfolds from this chain: R → I → G → FG → MG → PG → EG → UG

Reading warning: the arrows do not all mean "earlier in time". R → I describes reality generating a receivable signal, while G → FG describes assumptions acting on filtering; the two happen in parallel. G is drawn between I and FG only to show that "information" and "basic assumptions" are two inputs meeting at the filter.

The eight axioms

1. Reality is not cognition. The model is not the world — otherwise filtering, error, and update would be meaningless.
2. Cognition must filter. A finite system cannot process unlimited information. Filtering is a necessary condition for cognition, not a mistake.
3. Basic assumptions determine part of the filtering rule. Different assumptions produce different information worlds — but assumptions influence filtering rather than fully determining it.
4. Classification compresses reality. An infinite set of objects is mapped onto a finite set of categories. Classification is a downstream result of filtering, and an independent compressive function.
5. A model consists of objects and relations. Objects without relations form a list, and a list cannot predict. This is also why a new discovery is often a new relation between two old objects, not a new object.
6. Models are tested through prediction. Prediction is the only serious channel through which a model touches reality. Without it, a model can only describe and cannot be falsified.
7. Prediction error is a signal, not an error. Error tells the system to inspect something. It does not by itself prove that the model is wrong.
8. The system must allow its own basic assumptions to be corrected. If G cannot be revised, every error below it is locked in at the top level.

A terminology distinction is emphasized throughout: the axioms of this theory are rules for how we explain cognition; G is the set of high-level assumptions actually held by the cognitive subject being explained. The two are not the same thing.

Two levels of update

Second-level change is rare and expensive, but when it happens it usually triggers a large rearrangement of everything built on top of the old assumptions.

How we get misled: seven types of cognitive error

Errors are classified by the link in the chain where they occur, so that "I was wrong" can be located precisely instead of treated as one vague failure.

A critical warning: prediction error is not the error itself. A near-zero prediction error does not prove the model is valid — a model can predict the wrong thing consistently, or be shielded from the cases that would break it.

How to judge reliably: cognitive ability, not knowledge volume

The theory defines cognitive ability as the combination of five independent abilities: modelling, prediction, error identification, model revision, and basic-assumption revision.

• It is explicitly not a single score, a rating system, or a way to rank two people.
• A complex model is not the same as high cognitive ability. Someone with an elaborate model who cannot detect errors, revise the model, or revise assumptions may end up more trapped than someone with a simple model who can correct it.
• Reliability comes from keeping the loop open: predict, compare with reality, locate the error by layer, update — and remain willing to update the layer that rarely moves.

How the 16 chapters are organized

Suggested reading paths: general readers §1 → §2 → §3 → §14 → §16; advanced readers §1 → §3 → §5 → §7 → §9 → §11; researchers read all sixteen in order.

Key takeaways

• You do not see the world; you run a model of it. Confusing model with reality is the root of systematic bias.
• Filtering is not a defect. The real question is not "why do people filter?" but "by what rule do they filter?"
• Prediction is the test. A belief that cannot generate a checkable prediction is only a description.
• Error is information about where to look, not proof that the model failed.
• The deepest errors are the cheapest to keep and the most expensive to fix. Basic assumptions change rarely, but change everything downstream.
• Judge reliability by correction capacity, not by the sophistication of someone's model.

Who it is for

• Readers who want a structured way to examine their own judgments and disagreements.
• People working with AI who need a shared vocabulary for models, context, prediction, and error.
• Anyone who has to distinguish "my model is wrong" from "my deepest premise is wrong".