Overview · Written article
The active-inference brain
Perception as free-energy minimisation, action as expected-free-energy minimisation, and a hard rule about what may enter the logits.
How to read this page
Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.
These 13 pages are the only ones on this site written for this site. Everything else you can read here is a real file from a working repository, republished as it stands. These were written by hand for someone arriving with no context: an entry point, one article for each major part of the estate, and a set of practical guides for installing, running and maintaining things.
They carry one unusual discipline. Every citation in them points at a real file at a named commit, and the build resolves it there. If the file has moved, or a quoted range has run off the end of it, the build fails. It will not publish a reference that no longer lands. The wording of an article can drift out of date; what it points at cannot silently do so.
Start with "Start here". It is short, and it explains the one idea the rest of the estate is arranged around. After that, read whichever part you care about, or go to the quick start if you would rather run something than read about it.
What they are not: authoritative. An article is prose about documents, and the documents are the record. Where the two disagree, the document is right and the article is the thing to fix.
Your browser cannot switch reading levels, so the document itself is shown.
Precise — the source document
This is the article as written. Every citation and quoted block in it is resolved against the real file at the real commit when the site is built.
This is the engine every agent in the system runs — the UNIs in the world and the Producer that films them. 46 modules, 52 test files.
It is also the part of this site where claims have to be handled most carefully, so this article is deliberately more cautious than the others.
There is no reward
The agent does not maximise a score. Perception minimises variational free energy; action minimises expected free energy. Nowhere is there a scalar return to accumulate.
This is not a philosophical flourish, it is an engineering property. A reward signal is the thing a system accidentally optimises — the thing you later discover it was gaming. There isn't one to game.
The model
A generative model here is (A, B, C, D, E) plus precisions plus Dirichlet counts: what observations
each state produces, how states transition under each action, which observations are preferred, where
things start, and prior habits.
Learning is Dirichlet accumulation — counting what actually happened — and nothing else.
Why the joint state is never built
A realistic agent has many factors, and the joint state space is their product. Materialise it and you are finished before you start.
The engine is mean-field multi-factor: the posterior factorises, and — the result that makes it
work — expected free energy is additive across factors. So G(π) can be computed per factor and
summed, and the joint N_x = Π N_f is never constructed.
Expected free energy, in two parts
G decomposes into an epistemic term — how much a policy is expected to reduce uncertainty about
hidden state — and a pragmatic term — how far predicted observations sit from preferred ones.
Planning is a depth-limited beam search over policies, which is an approximation to sophisticated inference and is labelled as such rather than described as the real thing.
The invariant that governs what may be added
This is the most transferable idea in the subsystem. When a new term is proposed for the engine, it must be one of a closed set: pragmatic, state-epistemic, parameter-novelty, or a precision. Nothing else enters the logits.
Novelty — the parameter-information-gain term — shows what that discipline costs and buys. Its own docstring states the property that keeps it from being a reward in disguise:
It is LARGE and positive in under-sampled (low-count) `(state,outcome)` cells, **decays monotonically to 0
as counts → ∞** (the no-smuggled-reward invariant — it is information, not reward), and is **independent of
C**. So it adds a standing drive to act where the model is still uncertain about its own likelihood — whichThen the test file asserts exactly those properties by name: that W decays monotonically to zero as
counts grow, that it is independent of C (so it cannot smuggle in a preference), and that it is
bounded so it cannot swamp survival.
And the strongest one: with the novelty gain at zero, the agent's decisions are byte-identical to the agent without the term, over the full planning path. A new capability must leave the old behaviour bit-for-bit unchanged when switched off.
Precision is attention
Two precisions are tracked and updated: confidence in policies, and confidence in each observation modality. Attention, in this framing, is not a spotlight — it is the optimisation of confidence.
The viability edge is real
Homeostasis and metabolism are not decoration. An agent has internal states that can go wrong, and "wrong" is defined by the model rather than by a designer's penalty term. Both modules carry an explicit claim fence in their own docstrings — a paragraph stating what the module does not demonstrate.
Continuous action
Discrete policy selection is not the whole story. Motor control is continuous predictive coding:
action descends the free-energy gradient, ȧ = −∂F/∂a, proportional to precision times the difference
between target and sense.
Thirty-nine lines, most of them the derivation.
What is NOT established
Stated plainly, because this is where overclaiming would be easiest:
- This is not a claim about consciousness or experience. The estate's phenomenology document maps human phenomenological categories onto implemented mechanisms and records where the mapping fails. The failures are the point of the document.
- Passing tests are not biological parity. The contract governing this work says so in as many words, and separates observation from reconstruction from simulation as a rule.
- There is no rendered architecture diagram of the brain — factors, EFE, action selection — anywhere in the estate. Of the five subsystems documented on this site, this is the largest diagramming gap, and it is recorded rather than glossed.
What is established is narrower and more interesting: a working multi-factor active-inference engine with no reward signal, whose additions are gated by an invariant that requires them to change nothing when switched off.
Plain — written for this website, not the source document
This page describes the reasoning engine every agent in the project runs, including the one that directs the broadcast. Its central point is that there is no reward. The agent does not collect a score. Perception reduces one quantity, and choosing what to do next reduces a related one that looks ahead. The page argues this is an engineering property rather than a philosophical flourish: a reward signal is the thing a system accidentally optimises, and there is not one here to game.
It then explains the rule that governs what may be added to the engine. Any proposed new term has to belong to a short closed list, and the worked example had to decay to nothing as evidence built up, stay independent of the agent's preferences, and leave behaviour unchanged when switched off.
The page is deliberately more cautious than the others on the site. It says this is not a claim about experience, that passing tests are not biological parity, and that no diagram of this part exists anywhere.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 9376fc5b3bf50df5
Clear — written for this website, not the source document
This is the article about the inference engine every agent in the estate runs, the ones in the world and the one directing the broadcast. It says up front that it is deliberately more cautious than its neighbours, because this is where overclaiming would be easiest.
There is no reward anywhere. Perception minimises variational free energy; action minimises expected free energy. No scalar return is accumulated. The article treats this as an engineering property rather than a slogan: a reward signal is the thing a system is later discovered to have been gaming, and there is not one here.
The generative model is a set of matrices plus precisions plus counts — what observations each state produces, how states transition under each action, which observations are preferred, where things start, and prior habits. Learning is counting what actually happened, and nothing else.
A realistic agent has many factors, and materialising their combined state space would finish you before you start. The engine keeps the factors separate and leans on the result that expected free energy is additive across them, so it can be computed factor by factor and summed, and the combined space is never built.
Expected free energy splits in two: how much a policy is expected to reduce uncertainty about hidden state, and how far its predicted observations sit from preferred ones. Planning is a depth-limited beam search, which the article labels an approximation rather than describing as the real thing.
The most transferable idea here is the rule about what may be added. A proposed new term must be one of a closed set, and nothing else enters the choice. The worked example is a novelty term. Its own documentation states the properties that keep it from being a reward in disguise, and its tests assert them by name. It decays to zero as counts grow. It is independent of preference, so it cannot smuggle one in. And it is bounded, so it cannot swamp survival. The strongest is that with the term at zero, the agent's decisions are byte-identical to the same agent without it.
Precision is treated as attention rather than as a spotlight. Internal states can go wrong, with wrong defined by the model rather than by a designer's penalty term, and both of those modules carry an explicit statement of what they do not demonstrate. Continuous motor control descends a gradient.
The closing section is the point of the page. This is not a claim about experience; the estate's own phenomenology document records where that mapping fails. Passing tests are not biological parity. And there is no rendered diagram of this subsystem anywhere, which is the largest such gap on the site.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 9376fc5b3bf50df5