nucleation

CLAIM-337: Active inference already runs an internal loop: mental action selects among higher-level cognitive states and sets the precision of lower-level observations, structure learning reshapes the generative model without action, and a judgement is behaviour, so a judgement–observation loop may be action and perception one Markov blanket inward

4 min read

Properties9
StatusProposed
Version2
Rolecounter
Warrantinterpretive
Tagsprobabilistic-modeling, cognition, neuroscience
Linedistributed-agency, pragmatic-transport
GroundsLIT-1191 — Towards a computational phenomenology of mental action: modelling meta-awareness and attentional control with deep parametric active inference, LIT-1195 — Attention or salience?, THEORY-662 — The feeling of fluency is an inference about the precision of one's own processing, triggered by a change in precision against expectation; because its cause is internal and not given to introspection, a felt discrepancy is attributed to the most salient external source, as familiarity, liking or a sense of error, THEORY-564 — A bounded generative model can learn its own structure with no machinery beyond inference and learning: spare hidden-state slots with flat likelihoods get filled when an observation fits none of the learned causes, and Bayesian model reduction on the learned prior over states can return unneeded slots, though in a toy concept-learning task it recovers the true number of causes in only 45–80% of runs with four to seven causes, and never with two or three, unless the likelihood is given, THEORY-577 — Our judgments that consciousness poses a hard problem are behaviour, so explaining them is an easy problem that can be solved without mentioning consciousness; and a theory that names a mechanism as the basis of consciousness owes an account of how that mechanism produces those judgments, which IIT and biological theories struggle to give, since a functional duplicate with zero integrated information, or in silicon, would make the same reports
Objects toCLAIM-341 — The loop that keeps a lucid agent's model in touch with what it models runs between judgement and observation: each judgement reshapes the observer whose later observations it then judges, and action on the world is not part of that loop, so active inference's action–perception cycle misplaces where the agent is active
Cited as “complements” byCLAIM-541 — Wu-wei is better read as control placed and timed where a system's own dynamics do most of the work than as the absence of control: Ott–Grebogi–Yorke control of chaos is a formal instance, in which small perturbations timed by the system's own wandering hold it on an unstable orbit, and the control, though small, continues for as long as the orbit is held; so reported effortlessness has at least three distinct sources, no conflict to override, conflict settled cheaply by learned accommodation, and conflict met as it arises by small well-timed feedback, which may feel alike and differ in mechanism

Graph View

The claim

The objection, at full strength. Active inference does not confine action to the body.

  • Mental action. Sandved-Smith et al. (LIT-1191) “cast mental action as policy selection over higher-level cognitive states and add a further hierarchical level to model meta-awareness states that modulate the expected confidence (precision) in the mapping between observations and hidden cognitive states” (read in NOTE-958). At the attentional level, mental action changes the agent’s true attentional state, and so what it then observes of itself: action and perception one level inward. At the perceptual level it does not change content. The stimuli are “predetermined, with a deviant stimulus shown every 20th time step”, and mental action changes only their precision. Parr and Friston (LIT-1195, read in NOTE-956) separate attention, the precision given to data, from salience, the epistemic value of an action that samples them. They never call attention an action. Covert precision changes what is observed only by biasing “planning in favour of an overt saccade”.
  • Judgement fixing what is observed. Brouillet and Friston (THEORY-662): the feeling of fluency is an inference about the precision of one’s own processing, and because its cause is internal, a felt discrepancy is attributed to “the most salient external source”, as familiarity, liking or a sense of error. No body is in that loop (NOTE-785).
  • The observer reshaped without action. Smith et al. (THEORY-564): a bounded generative model “can learn its own structure with no machinery beyond inference and learning”. Spare hidden-state slots fill when an observation fits none of the learned causes. During that learning, reporting policies are disabled (NOTE-485). Active inference thus has CLAIM-341’s diachronic loop, in which judgement reshapes the observer, inside one generative model whose likelihood is learned.
  • Judgement is behaviour. Chalmers (THEORY-577): “Our judgments that consciousness poses a hard problem are behaviour.” If judgements are behaviour, keeping judgement while cutting action relabels the action rather than removing it.

In a deep model, a higher level’s outputs are a lower level’s priors and its inputs are the lower level’s posteriors, so the lower level is, formally, the higher level’s environment. The judgement is mental action at the upper level, and the observation is what the lower level returns. “Cutting action out” draws the blanket around the upper level and calls what crosses it judgement.

What it does not say

  • It does not save the diagram. Every item above runs with no action on the world. If the loop is redescribed as mental action, precision and structure learning, the redescription concedes CLAIM-341’s point that action on the world is not part of it, and the disagreement is about whether the diagram should say so.
  • Where the blanket goes is a modelling choice, made before the blanket is found (THEORY-073; NOTE-447: an external intervention has “exactly the same formal properties” as an internal cause). So “one blanket inward” is always available, and the objection costs little to make and says little.
  • The readings narrow the objection. In Sandved-Smith et al., mental action changes perceptual precision, not perceptual content, which is the line CLAIM-341 draws. Nothing in that model is learned, so the structure-learning strand rests on THEORY-564 alone. Parr and Friston give attention no power to change content short of an overt saccade, which is action on the world.

Note of 2026-10-29: what holds it back

Its grounds THEORY-662 and THEORY-564 are Proposed: no fitted test of Brouillet and Friston’s fluency inference, and Smith et al.’s structure learning is one simulation. It also grounds on THEORY-577, Proposed, but uses only that reading’s first clause, that problem reports are behaviour.

Note of 2026-10-31: the owner’s “sounds close to FEP”

At U8 of the exchange “Analyze nucleation concepts” the owner said the assistant’s “Inquiry is the reflexive transformation of the apparatus of individuation” “sounds close to FEP”. A8 introduced Smith et al.’s structure-learning paper as prior art. The record holds it (LIT-615, read in NOTE-485), and it is one of this claim’s grounds. A8’s caution is the one this claim’s “What it does not say” already makes: free-energy minimization passes “too quickly from an admissible inferential description to the claim that the described system physically performs inference”. The owner’s other comparison at U8, that the record’s unit of explanation “feels very ‘wu-wei’”, is not filed. The record’s nearest entries are CLAIM-358 (local accommodation lowers the amortized cost of arbitration), CLAIM-162 (flow reported as effortless) and LIT-910 (Effortless Attention, Unavailable).

Note of 2026-11-02: wu-wei, coded

The owner’s “very wu-wei” came back in the exchange “Assess SDT Coverage” (2026-11-02), this time of Ott–Grebogi–Yorke control of chaos (U17). This note’s predecessor recorded it as appearing once and not filed. On its second appearance it is filed as CLAIM-541, Proposed, beside CLAIM-162 and CLAIM-358.

Backlinks