The question
Two instances from one exchange.
Hedonic. A pursuit process keeps selecting a reward and reaching it, while the consummatory evaluation of the reward declines. In an adaptive organization that evaluation can inform pursuit and, where apt, change the authority of the processes responsible. In a captured one the pursuit process keeps its authority despite evidence that the outcome no longer rewards. CLAIM-437’s wanting and liking then differ not only as two evaluations but in their capacity to govern. Incentive sensitization already explains persistent wanting by a change in the incentive mechanism (THEORY-895). The question is whether divergence can also be maintained by asymmetries of evidential access and governance, which would be the record’s addition and is not established by the sensitization literature.
Epistemic. A process can hold a belief (TERM-079) that keeps predicting and inferring accurately while the organization denies it enough authority to change collective judgements or actions: a specialist right and unheeded.
Why the record needs it
It is the point where the will line, the affect line and the legibility line meet. CLAIM-403 separates access, recognition, credibility, uptake and participation in revising standards. The cases here sit past recognition and short of uptake, and an account that cannot tell them from failures of recognition will read every unheeded evaluation as an unregistered one.
What would answer it
A model or study with the evaluation measured independently of choice (hedonic reactivity or experienced-utility reports; independently scored forecasts), the organization’s registration of it measured separately (its self-model, its record, its reports), and an intervention on the authority of the evaluating process with the evaluation and its registration held fixed. The hedonic version: an outcome that becomes progressively less pleasant, or a dangerous activity that becomes safely enjoyable, with wanting, liking and governance responsiveness recorded separately, compared across an ordinary adaptive learner, a learner with persistent pursuit salience, and one whose pursuit authority also shapes the evidentiary regime.
Related
- The hedonic quantities as the exchange separated them (CLAIM-437’s note of 2026-10-24): hedonic impact, anticipated pleasure, incentive wanting, hedonic authority.
- Relief after avoidance is reduced distress, not positive pleasure and not evidence of danger (CLAIM-446); safe engagement’s pleasure is often counterfactual and unsampled.
Note of 2026-10-25: four rivals for failed uptake
A31 of the exchange of 2026-10-24, continued, proposed this question’s probe as its Experiment 6, without citing it, with four mechanisms and the intervention that tells each apart:
- incorrect hedonic prediction, separated by reliable outcome evidence;
- cue-triggered incentive salience, separated by reducing cue-driven drive;
- habitual control, separated by changed contingencies and retraining;
- governance-mediated suppression, separated by restoring the evaluation’s access to decision.
A31: “A sophisticated habit or incentive-learning model may explain much of the result without any extra governance state.” It also adds an avowal equivalent to “I no longer enjoy this”, made available to the other processes, to test whether an accurate report is enough to revise pursuit. That is CLAIM-403’s step from recognition to uptake, inside one agent. Not run. The question is unchanged.
Note of 2026-10-26: a synthetic separation of denied uptake from nonregistration
The lab’s v2.6 (dmarx/PerspectiveLab@74621fb experiments/affective_dissociation/) runs the note of 2026-10-25’s four rivals, with a fifth class in which the negative evaluation is never registered. Liking is an exogenous outcome value, −1 in every class after devaluation, read by an assay that does not use action. Anticipated pleasure and registration have readouts of their own, registration’s right nine times in ten. Pursuit is a Bernoulli choice whose logit adds anticipated pleasure, cue drive, habit and another pursuit motive, and subtracts 1.2 × negative liking × registration × authority. The parameters are set by hand so that all five classes pursue with the same one-step probability after devaluation, 0.668188, the logistic of 0.7.
- In the denied-uptake class the negative evaluation is registered and its authority is 0. In the unregistered class it is not registered and its authority is 1.
- Restoring authority to 1, with anticipated pleasure, liking and registration unchanged, lowers pursuit in the denied-uptake class to 0.378 and leaves the unregistered class at 0.668. Restoring registration does the reverse. Each of the other three classes falls to 0.154 under its own intervention and under no other.
- Restoring authority moves no other class only because every other class already has authority 1. Halving authority, one of the held-out targets, raises pursuit from 0.668 to 0.786 in the three classes whose evaluation is registered with full authority, and lowers the denied-uptake class’s to 0.525. Only the unregistered class is untouched. An intervention on authority moves pursuit wherever a registered evaluation enters action, not only where authority is blocked.
- With the probes and readouts, a Bayesian classifier over the five classes names 195 of 200 held-out worlds correctly. A hybrid of cue drive and denied uptake, left out of the catalog, is assigned to a catalog class in all 40 of its test worlds, 30 of them to denied uptake, with a mean top posterior of 0.868. The adequacy check alarms in 9 of the 40. A mixture is the case this question most needs told apart, and the construction does not tell it.
- In a longitudinal run of 60 seeds, with no resets and liking stepping from +1 to −1, the denied-uptake, unregistered and cue classes follow the same pursuit course, to within 10⁻⁹, for all 36 steps after devaluation. Passive history does not tell denied uptake from nonregistration here. Only the intervention does.
Within a specified synthetic class, then, the construction meets the first test under “What would answer it” for the hedonic instance. The evaluation is measured apart from choice, its registration apart from both, and an intervention on authority holds them fixed. But the classes, and what registration and authority mean, are written into the generator, and the lab’s test of the separation checks what the generator says. Liking drops in one step, not progressively. No learner’s pursuit authority shapes the evidence it gets. The epistemic instance is not touched. The lab bridges this question to CLAIM-437, whose processes the model does not represent (that claim’s note of 2026-10-26), and its same-kernel tie is CLAIM-427’s point (its note of 2026-10-26). No Active claim or reading answers the question (ADR-049), and nothing is split from it. The question is unchanged.
Note of 2026-10-26 (later): predictors fitted to the same experiences
The lab’s v2.7 (dmarx/PerspectiveLab@c54617e experiments/trained_affect/) replaces v2.6’s hand-set classes with generator worlds and fits predictors to them. Each episode runs 72 uninterrupted steps, and liking drops from +1 to −1 at step 23. All eight worlds share one logistic law of pursuit: learned anticipated value, cue drive, habit, and a term for negative liking times registration times authority. They differ in its coefficients. The uptake coefficient is −1.9 in the governance world and −0.20 to −0.30 in the value, habit and sensitization worlds. Of the held-out worlds, one mixes cue drive with governance, two add a threshold or a lag to the uptake term, and one makes the assays noisier. An outcome is observed only when it is pursued.
Seven predictors are fitted by penalized logistic regression to the same 8,640 choices from four source worlds, with penalties chosen on 2,880 validation choices from the same worlds. Each is scored by its one-step probability MSE against the generator’s law, on 80 held-out episodes per world. Registration and authority are not measured. The experimenter sets them at each step, and every predictor is given them as inputs. The predictors are fitted offline to recorded choices; they are not agents that learn by acting.
- The note of 2026-10-25 quoted A31: a sophisticated habit or incentive-learning model “may explain much of the result without any extra governance state”. Here a cue-and-habit predictor with 4 coefficients does best in the habit world (0.01152) and in the sensitization world (0.00896). The predictor that adds the uptake term, a fifth coefficient, scores 0.02581 and 0.01466.
- In the governance world the uptake predictor (0.01586) loses to a 7-coefficient predictor with main effects of liking, authority and registration and no interaction (0.01378), and to a 15-coefficient one with cross terms (0.01205). It wins only on the held-out cue–governance hybrid, 0.00423 against 0.00511 for the 15-coefficient predictor (paired seed-cluster difference −0.00088, 95% interval [−0.00118, −0.00060]).
- In the governance world over steps 30 to 50, the wanting proxy averages +0.35, the liking assay −1.008 and the probability of pursuit 0.393.
- Restoring registration at full authority, with the generator’s latent states held fixed at steps 32 and 50, changes pursuit by −0.388 in the governance world, −0.039 in the habit world and 0 in the lagged-authority world. These contrasts are computed from the generator’s equations and were never given to the predictors.
So the classes are again told apart by an intervention the experimenter makes, and a predictor fitted to the same experiences does not need the uptake term to predict it. Registration is not measured apart from the experimenter’s setting, which is the second requirement under “What would answer it”. Liking still drops in one step. No world’s authority shapes its evidence: choice gates what is sampled, in every world alike. The epistemic instance is not touched. No Active claim or reading answers the question (ADR-049), and nothing is split from it. The question is unchanged.