CLAIM-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

The claim

The owner, at U1: “Friston’s ‘Free Energy’ approach is generally correct, but the cybernetic model is wrong. Rather than cycling between Action <--> Perception, we can cut ‘Action’ out of the loop completely and constrain attention to the internal feedback loop, which decomposes nicely into Judgement (qua measurement that generates interpretations) <--> Observation, which connects us back to our model and also makes our complaint with Friston’s formulation precise.”

The record’s model already has this loop. CLAIM-297 writes the operator model as a circuit, “frame → retrieval → interpretation → endorsement → future retrieval policy”. It is two coupled processes over a meaning state x, operator weights w and queries q:

• meaning: x_{t+1} = Σᵢ w_{i,t} P_i(q_t)ᵀ x_t;
• influence: w_{t+1} = G(w_t, x_t, o_t, E_t);
• queries: q_{t+1} = Q(x_{t+1}, w_{t+1}, g).

Read with the owner’s terms, the query is the judgement’s measurement context, the operators applied under it are the judgement, and x is the interpretation it yields. The interpretation, through endorsement and the next query, fixes what will be retrieved and so what the next judgement observes. Nothing in the circuit acts on the world. The world enters only as evidence o_t in the update of influence. CLAIM-275 holds the half that concerns order: an operator can intervene on the meaning canvas before it has the interpretation its intervention makes available.

“Generates” is read diachronically. A judgement need not produce the observation it judges. It changes the observer, which produces the later observations. The record holds this for perception: perceptual learning can install frames that change which objects a perceiver individuates from the same input, so “cognition can shape the perceptual system over time while leaving open whether any single percept is cognitively penetrated” (THEORY-593). And for cognition: percepts are used only inside a “cognitive envelope”, where working memory holds “recoded, task-dependent abstractions rather than the percepts themselves” (THEORY-454). The observation the loop runs on is already a product of earlier judgement.

The complaint against active inference is then precise in Pearl’s terms (THEORY-419). Active inference treats perception as conditioning on observations that arrive at the blanket’s sensory states, and places all of the agent’s influence on what it observes in action, which samples the data (LIT-1181, read in NOTE-960). The paper states the contrast itself: “the model (agent) is fixed and the aim is to minimize surprise by re-sampling the data”, and agents “can only act by changing hidden states, not the sensory mapping per se”. The claim holds that the agent’s main influence on what it observes is an intervention on itself, the observer, made by judging. Action on the world is one further influence. It is not part of the loop.

Judgement is not action

The objection that a judgement is itself behaviour (THEORY-577), and so action relabelled (CLAIM-337), is answered by where the judgement acts. Action, in the active-inference sense, changes external states, which then change sensory states. A judgement changes the judge’s own state, the canvas and the retrieval policy (TERM-026, CLAIM-275). Both are operations. What separates them is their target, and the claim is that the loop that keeps a model in touch closes through the second.

QBism is the sharpest contrast (LIT-1182, read in NOTE-955). It agrees that a measurement creates its outcome rather than revealing it: “A measurement does not, as the term unfortunately suggests, reveal a pre-existing state of affairs. It is an action on the world by an agent that results in the creation of an outcome — a new experience for that agent.” But QBism keeps judgement and measurement apart. A judgement is the agent’s state assignment (“a quantum state assignment is also a personal judgment”). A measurement is “any action an agent takes to elicit a set of possible experiences”, and the agent’s “own direct internal awareness of her own private experience” is the one thing it does not model. Its loop is judgement, action, experience, update: the action–perception shape this claim rejects, with the self left out of the measurement. This claim makes the judgement itself the measurement, and the judge the thing it changes.

Evidence that judging changes the judge

The record had none from outside its own exchanges until it registered three works. Two have been read.

• Feldman and Lynch (LIT-1196, read in NOTE-957). Measurement creates state, not only a response: “If the researcher asks a question about attitude (or belief, intention, etc.), and no such cognition already exists, the measure creates one”. And the diachronic form is explicit: “Once the answer has been computed, it can be retrieved to answer subsequent questions”. This is the route TERM-039 and CLAIM-127 leave open, an arrow from the context into the interpreter’s state, argued in classical terms. Three limits:
  • the paper has no data of its own and calls its account speculative;
  • it values the created state mainly for its effect on behaviour;
  • it gives judgement no special place: “all real-world experiences with an object have the potential to shape subsequent responses to it”.
• Slovic (LIT-1187, read in NOTE-949): preferences are “constructed—not merely revealed—in the elicitation process”. Every result compares elicitation procedures. That shows the measurement context entering the response, which is the route A129’s graph already has. It never follows a constructed preference into a later judgement, so it supports the synchronic reading this claim sets aside, not the diachronic one it holds. Construction is claimed only for “important, complex, and perhaps unfamiliar” options.
• Morwitz, Johnson and Schmittlein (LIT-1189, Unavailable): merely asking purchase intent changes later purchasing. Known from the abstract only.

All three are classical in mechanism (retrieval, accessibility, construction) and do not bear on contextuality. Feldman and Lynch is the one that supports the claim as held.

What keeps the loop honest

CLAIM-339 objects that, without action, the loop confirms itself. CLAIM-342 answers from the model: what anchors the loop is evidence whose arrival or scoring does not pass through the interpretation it tests (CLAIM-330, CLAIM-141). Action is one source of such evidence, not a necessary one.

The synchronic form, not claimed

The stronger reading of U1, in which each judgement produces its own outcome, is not part of this claim:

• It is open. Whether eliciting a judgement changes the interpreter, and not only the response, is the route TERM-039 and CLAIM-127 leave open, to be settled by sequential designs.
• The record demoted it once before. CLAIM-109 said a frame “participates in determining which meaning becomes actual”, and it was demoted to CLAIM-009 when its argument (ARG-002) was undercut by CLAIM-041.
• Contextuality may not help. If the synchronic form is to differ from classical state change, it needs contextuality. The weak sense is the starting point of Contextuality-by-Default, not a finding (THEORY-559). The sheaf sense cannot arise for one system judged context after context (THEORY-010). Spekkens’s sense can (QUESTION-067). The record holds no behavioural evidence of contextuality proper (CLAIM-325, THEORY-623).

What it does not say

• It does not say action is idle. Every lucid locked-in patient built their model while acting (CASE-107), and the claim is about the loop that keeps a built model in touch, not the one that built it.
• It does not reject free-energy minimization. The owner’s complaint is with the cybernetic diagram, not with the bound. Three readings already make the complaint that active-inference models assume perception and action rather than deriving them (NOTE-450, LIT-598, LIT-615).
• It does not say judgements are quantum measurements (ARG-002, CLAIM-041).
• It does not deny that active inference can redescribe the loop, as mental action on precision together with structure learning, and no action on the world (CLAIM-337). That redescription concedes the point about action and keeps the disagreement about the diagram.
• It does not file the owner’s U2 statement that the two lines are one model. It follows that statement in choosing CLAIM-297 as the model’s statement of the loop.

Rivals

• Observation and control as one loop through the environment. NOTE-121 defines an observer by feedback closure, “the observer’s outputs must alter the environment, which then alters later inputs”. The claim denies that condition.
• Meaning from action. Barrett: “meaning does not trigger action, but results from it” (NOTE-434). CLAIM-275 refuses to fix either order. This claim holds that the order through the world is not the one that keeps the model in touch.
• Sense-making from regulation. Di Paolo: a norm the system itself has needs adaptivity, which detects tendencies toward the viability boundary “and acts so as to reverse them” (THEORY-628).
• Agency defined by action. Agents are “the systems that would adapt their policy were their actions to affect the world differently” (THEORY-527). On that definition a judge that cannot act is not an agent. The claim is about a lucid judge, not about agency.
• Perceptual control. Powers: “behavior is the control of input, not output”, and above the lowest level a system’s output “is not a muscle force, but a reference level” for the level below (LIT-1180, read in NOTE-963). In that paper every loop closes through the environment, down to the spinal systems that control muscle effort. But the theory’s first statement has a loop with no action: an “imagination connection” by which a reference signal returns as the system’s own perception, so that the system can “plan an action ‘mentally’ without actually performing it” (LIT-1188). That loop is self-satisfying, since the “match” is “automatically ensured”. It is the synchronic form this claim does not claim. Powers adds the anchor problem in the same passage: the hierarchy “could not perform very reliably” if its actions were “completely ‘short-circuited’” by it, and Part II lists switching a system “entirely to the imagination-connection” as one of the ego’s defences (LIT-1197). Unlike the 1973 paper, 1960 does reshape the observer. The reorganizing system’s first act is to build new perceptual input functions, “tentative definitions of new variables”, driven by error in bodily intrinsic signals, “measures of a set of physiological states” (NOTE-954). And Part II tests a change to the observer by imagining: a golfer drops a new idea because “a moment spent imagining doing so would reveal” the conflicts it makes (NOTE-959). So perceptual control theory has both of the claim’s ingredients. The imagination loop changes nothing in the perceiver. Reorganization changes the perceiver, but is driven by bodily error, not by judgement. The claim’s loop, judgement reshaping the observer, falls between them. The reorganizing system’s inputs may include “mean error signal in the hierarchy”, a scoring signal the learned hierarchy does not set, which is QUESTION-121’s candidates 2 and 4.
• The epistemic cut. Pattee (LIT-1190, read in NOTE-952) does not put measurement first. Measurement and control are converses: “measurement transforms physical states to symbols in memory, while memory-stored controls transform symbols to physical states”, and “control of construction came first”. So he belongs with the rivals that keep action, here control, in the loop. Two of his points still bear. Semantic closure, in which the reader of the symbols is built from them, has the shape of this claim’s diachronic loop. And his condition for objectivity, that measurement must not influence the laws it measures, is close to CLAIM-342’s anchor.

Note of 2026-10-18: two appearances in the exchange of 2026-10-17, one beyond what Slovic shows

A21 §14, a recursive judgment model. A21 wrote

j_t = J(x_t, c_t, n_t), c_{t+1} = F(c_t, j_t, a_t, w_t), n_{t+1} = G(n_t, j_t, feedback_t),

and: “The judgment is both an output of the current evaluative organization and a possible input to its future organization.” That is this claim’s loop, a second appearance, in the normative setting, and A21 cited the counter (CLAIM-337) and not this claim. A21 drew the counter’s warning correctly: “recursive cognition and internally generated policy changes are not automatically evidence of distinctive volitional authorship.” Two differences from this claim. A21’s context update takes the action a_t, so its loop runs through action, which this claim denies is needed. And it adds an update of the standards n, which this claim does not have; that part is the will-organization line’s revisable governance, and A21 §14’s “our multiscale theory cannot always assume a fixed normative target” is QUESTION-112’s moving decision family.

A37, a dynamic reading of Slovic. A37 linked NOTE-949 to this claim and concluded that “an endorsement mechanism may not be aggregating fixed preferences. It may be participating in the production of the preferences it aggregates”, with the cycle context → local preference construction → endorsement → future context. The link is the record’s, and the cycle is this claim’s diachronic form. But this claim’s own limit on Slovic stands: every result “compares elicitation procedures”, and Slovic “never follows a constructed preference into a later judgement, so it supports the synchronic reading this claim sets aside, not the diachronic one it holds.” The step from constructed preference to future context is A37’s, not Slovic’s. Feldman and Lynch (LIT-1196) remain the one reading that supports it.

The status is unchanged.

Note of 2026-10-21: the loop re-derived as governance of the observer

A16 §8 of the exchange of 2026-10-20, “Governance of the observer”, re-derived this loop (“The governor may partly determine the evidence on which future governance decisions depend”) without citing it, and without CLAIM-330 and CLAIM-340, which say when it is self-confirming. Its “first-order” and “second-order epistemic adequacy” (preserving the answers to a fixed query family, against preserving how the family is selected) appears once. The exchange took the loop on to institutions in CLAIM-404.

Note of 2026-10-29: filed with its concessions

Two counters object to this claim, and its text already says what it grants each.

• CLAIM-339, that a loop cut off from action loses one route for evidence it did not select. “What keeps the loop honest” grants it: action is one source of such evidence, not a necessary one. CLAIM-342, Active, answers the rest.
• CLAIM-337, that the loop is active inference’s action and perception one Markov blanket inward. “Judgement is not action” answers its premise that a judgement is behaviour by where the judgement acts. “What it does not say” grants that active inference can redescribe the loop as mental action, precision and structure learning with no action on the world. The claim keeps the disagreement about the diagram, which CLAIM-337 itself says is all the redescription leaves.