nucleation

CLAIM-447: Model-based wanting is a pursuit process's ranking of the futures it predicts under candidate interventions, P(S | do(a)) against P(S | do(a₀)), while cue-triggered wanting is learned salience that need not compute those futures; a process's causal model, its wanting and its practical efficacy are separable, so learning to want differently, learning that a want will not pay, and learning not to act on it are three different changes

4 min read

Properties11
StatusProposed
Version2
Rolethesis
Warrantinterpretive
Tagsmotivation, causality, emotion-and-affect
Lineorganizational-affect, will-organization
Rests onCLAIM-437 — 'Wanting' and 'liking' are saliences relative to different, overlapping processes, pursuit and consummation, each ranking states by their bearing on its own concerns; valence is then relative to a perspective, every perspective in the negotiation has its own, the system's valence is that of the perspective whose ordering is the negotiation's potential, and ambivalence is strong saliences of opposite sign with no potential absorbing them
GroundsTHEORY-062 — Model-based and model-free predictions both operate in instrumental and in Pavlovian learning, and which one controls behaviour follows their relative reliability, not the preference of a supervising system, THEORY-895 — Incentive sensitization: repeated drug exposure can make the mesolimbic 'wanting' system lastingly hyper-reactive to drug cues and contexts, so that cue-triggered incentive salience grows, persists for years and even increases during abstinence, while 'liking' need not rise and may fall; sensitized 'wanting' can oppose the person's cognitive desire, and incentive salience can run without any reportable change in feeling
ComplementsCLAIM-283 — The precedence a motive claims over action and the authority the rest of the organization grants it are different quantities: a motive can claim control without receiving it, and be granted control it does not claim, 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
Corroborated byTHEORY-893 — Reward has separable components: 'liking', its hedonic impact, measurable without report by orofacial taste reactivity in humans, apes, monkeys and rodents, and 'wanting', its incentive salience; brain stimulation and sensitization can raise 'wanting' without raising 'liking', so motivation can grow while pleasure stays flat or falls, and 'liking' reactions can occur without any subjective feeling of pleasure
Cited as “complements” byCLAIM-519 — Competent behaviour does not establish an internal world model: over a fixed interface and a finite horizon, an agent with no model, whose actions are looked up from the history so far, reproduces the interaction distribution of an agent that has one exactly; a policy can identify predictive structure without any particular internal component realizing it, so compression and generalization support comparing candidate architectures, and establishing an internal mechanism needs evidence that discriminates candidate implementations, such as targeted interventions

Graph View

The claim

The owner, 2026-10-23: “really this is a do() operation and the future state that’s being anticipated is what might or might not happen given a candidate intervention.”

CLAIM-437 makes ‘wanting’ salience relative to a pursuit process but does not say what that process evaluates. This claim says. A pursuit process i in state s holds a model of what an intervention would bring about, P_i(S′ | s, do(A = a)), and ranks it against an alternative, often doing nothing:

a ≻_i a₀  ⟺  P_i(S′ | s, do(a)) ≻_i P_i(S′ | s, do(a₀)).

The ranking is of predicted distributions over futures, so the record’s ordinal commitment holds: no real-valued utility is assumed. Where one is available, the familiar case is a positive difference of expected utilities under the two interventions. do() matters because the process evaluates what the intervention would cause, not what it would expect to see if the action merely happened (THEORY-062’s model-based prediction). The intervention need not act on the world. It can act on attention, a retrieval policy or the observer itself (CLAIM-341).

Two levels. Model-based wanting computes those futures. Cue-triggered wanting is learned salience on a state–action pair that need not contain a prediction at all, which is what incentive sensitization shows: pursuit that persists against the person’s own prediction of the consequences (THEORY-895, THEORY-893). That the cached level is amortized prospective evaluation is a hypothesis about how it was acquired. It is not a claim about what it now computes.

Three components, not one. Prediction, appraisal of the predicted future, and motivational salience can each be present without the others. One can predict that exercise would help and value the help with no urge to exercise; a cue can move one strongly with no favourable prediction.

Wanting is not will. A process can want an intervention and have no authority to make it. Which proposed intervention becomes effective is the will-organization’s business (CLAIM-283). So there are three separable questions: what future a process anticipates (its causal model), what intervention is salient to it (its wanting), and what intervention becomes effective (its place in the organization). Learning to want differently, learning that what one wants will not pay, and learning not to act on it change different ones.

Wanting need not presuppose conflict. A single process can favour an anticipated future that no other process contests. Conflict makes precedence observable. It is not what makes a want.

The owner’s sentence, “I want something because I predict I will occupy a more desirable state after obtaining it”, becomes: I want something, in the cognitive sense, insofar as I evaluate the futures attainable through obtaining it as preferable to the relevant alternatives. That keeps desirability apart from probability: one can want recovery with no confidence that any treatment achieves it.

Dated part of 2026-10-24: governance changes what can be wanted

Later in the same exchange (U20), with CASE-124’s legibility state in hand: what enters P_i is only what the process can acquire and represent. Governance therefore reaches wanting by two paths. The practical path is that authority changes which interventions are feasible or affordable. The epistemic path is that legibility changes which futures can be represented credibly enough to be evaluated at all. A process can keep valuing an outcome, safe engagement for an anxious agent say, while becoming unable to generate a credible, actionable forecast of reaching it (CLAIM-446). A desire can lose practical efficacy without losing its evaluative content, and a change of legibility can make an old desire actionable without changing the concern behind it.

What it does not say

  • That all wanting is model-based, or conscious, or endorsed.
  • That the cached level descends from the model-based one.
  • How a pursuit process is individuated apart from the dissociations it is invoked to explain. That is still TERM-072’s open problem.

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

It rests on CLAIM-437, Proposed, and grounds its two levels in THEORY-062’s arbitration of model-based and model-free control and in incentive sensitization (THEORY-895), both Proposed. No reading the record holds dissociates cue-triggered from model-based wanting by a manipulation that leaves desirability and cue associations fixed, which its first two defeaters turn on.

Backlinks