nucleation

CLAIM-446: Some persistent anxiety may be governance-maintained rather than belief-maintained: anticipated threat grants defensive processes authority over action, attention and evidence acquisition; under the avoidance that authority selects, the hypotheses that the situation is dangerous and that it is safe predict the same observation, so repeated safety need not correct the threat estimate; relief reinforces the defense without being evidence of danger; and the regime can outlast a corrected belief, its pathology, where present, being defensive authority insensitive to corrective evidence at avoidable cost

12 min read

Properties14
StatusProposed
Version7
Rolethesis
Warrantinterpretive
Tagspsychopathology-and-treatment, self-governance, emotion-and-affect
Linewill-organization, organizational-affect
AnswersQUESTION-045 — Is a mental disorder that outlasts its trigger a protective lock-in at the individual scale, a regime kept stable by behaviour that blocks its own disconfirmation while the capacity to leave it degrades, and does cutting the maintaining link, with the regime's content held fixed, produce lasting change?
Rests onCLAIM-443 — A learned prior over which processes govern, P(g | z, H) ∝ exp(βQ(g, z) + b_H(g, z)), its bias b_H trained by reward-modulated Hebbian updates to the relations by which one process recruits or endorses another, is a candidate mechanism for durable authority and for amortized governance; but persistence is not amortization, which is shown only when earlier work replaces later computation, communication or recruitment, and a prior over authority differs from a prior over actions only when what is appropriate changes, CLAIM-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-349 — An organization's access to a signal, the signal's authority over action, and action on it are separable, and adaptive regulation keeps the first while making the second conditional on evidence; overcontrol, which buys freedom from harmful acts by cutting access or making inhibition unconditional, is a maladaptive regime and not more control, so flexibility is a context-sensitive property of maintained influence relations, not of how much is inhibited
GroundsTHEORY-285 — In an illustrative active-inference simulation of post-traumatic context inference, run once per phenotype, a prior that one is in a harmful context, made overly precise by a run of harmful trials under the "vulnerable" learning and forgetting weights, makes exploratory sampling look uninformative, so the agent attenuates the periphery, never receives the evidence that would correct the prior, and the belief sustains itself; lowering precision makes the attenuation policy less vigorous without correcting the prior, and the policy returns when precision is restored, THEORY-901 — As Mancinelli and Bach's systematic review of avoidance learning in healthy humans reports it, rating both A, an avoidance response left available during extinction persists, in some participants for hundreds of unpunished trials, and keeps expectancy of the aversive outcome higher than in people denied it (nine studies); and less effective or costlier responses extinguish faster (six studies, one exception), THEORY-920 — As Mancinelli and Bach's systematic review reports it, rating it B, in three laboratory studies of healthy humans a signalled, costless avoidance response offered for a stimulus never paired with the aversive outcome was used, and withdrawing it raised expectancy of the outcome for that stimulus, though Xia et al. (2019) found no change in skin conductance or startle; every study signalled the response's availability, which may itself have read as a danger cue, THEORY-903 — In a 2026 meta-analysis of standalone safety-behaviour manipulations outside exposure, mostly with young non-clinical or subclinical adults and mostly on appearance concerns and social anxiety, reducing safety behaviours lowered symptoms relative to control (d = 0.43, about 0.31 after adjustment for publication bias), and instructed increases raised them (d = 0.65, five studies)
SupportsCLAIM-445 — A protective regime, in a person, an institution or a system of states, becomes self-maintaining when the interventions it authorizes shift the evidence used to judge its necessity, the authority and resources of those who judge it, the capacities of its constituents or the threat environment itself in ways that favour its continuation; this can happen when the original threat was real and the first protection apt, and it is a failure only where the regime has become less responsive to evidence that it is no longer needed
ComplementsCLAIM-297 — When an operator's authority is updated by how well it performs on the interpretation it helped make dominant, the coupled meaning and influence of an organization can amplify an initial interpretive advantage without any gain in fidelity to independent evidence, CLAIM-342 — What keeps a judgement–observation loop from confirming itself is evidence whose arrival or scoring does not pass through the interpretation it tests; action is one source of such evidence, not a necessary one, and a loop that selects its own evidence can lock in with action available, CLAIM-277 — Durable authority among an organization's processes need not be a primitive: trust can be the temporal regularity of endorsement that policies earn by being right often, and the same loop entrenches whatever gained influence early
Corroborated byTHEORY-897 — In Brown, Price and Dombrovski's reading of seven computational learning studies they selected without a systematic search, with mixed directions, anxiety goes with altered adjustment of learning rate to uncertainty, mostly for aversive outcomes, while overall learning rate, choice consistency and outcome valuation show little relation to it, THEORY-900 — A systematic review of 116 studies of avoidance learning in healthy humans finds the evidence contradicting two-factor and other associative theories and favouring expectancy and inference-based accounts, but also finding residual avoidance after the outcome is visibly removed, which a purely inferential account does not predict, THEORY-907 — As Chou, Wilson and Smith review them, in about six foraging studies, mostly sub-clinical, higher stress or anxiety went with less exploration, and in one task with a predator to check for (Trier et al. 2023), people higher in trait somatic anxiety checked less and did worse, a reduction the study's further analyses associated with greater aversion to detecting the predator, THEORY-910 — Brown, Price and Dombrovski propose, without a model run or data, that in aversive environments avoiding uncertain states keeps their uncertainty high and so perpetuates avoidance, that anxiety strengthens this loop by misreading noise as change, and that exposure works by reducing the uncertainty rather than an excess of learned threat, THEORY-912 — In a picture-based avoidance task, treatment-seeking patients with OCD or PTSD avoided a safety cue more than healthy controls and, after effective avoidance, still expected the aversive picture more; patients with panic disorder with agoraphobia or PTSD reported more relief when it was omitted; the groups did not differ in avoiding the danger cues, THEORY-921 — As Craske et al.'s 2014 review reports it, citing six studies and reporting no data of its own, how far fear falls during an extinction or exposure session, or across sessions, does not predict how much fear is expressed at follow-up, in animals, human laboratory samples and fearful, mostly analog, human samples, THEORY-962 — In a finite partial-monitoring game, the signals of a set D of actions determine the expected loss difference of actions i and j under every outcome distribution iff ℓ_i − ℓ_j lies in the sum of the row spaces of their signal matrices; when it does not, two outcome distributions differing along a vector in the common kernel give every action in D the same signal law and reverse the preference between i and j, so no amount of play restricted to D can learn the comparison, THEORY-963 — The minimax regret of every finite partial-monitoring game against an oblivious adversary, degenerate and duplicate actions included, is 0 when no two actions are neighbours, Θ̃(√n) when every pair of neighbours is locally observable, Θ(n^2/3) when every pair is globally but some pair is not locally observable, and linear otherwise; its lower bounds are proved against stochastic outcomes, so the same four classes hold for them
UsesTERM-020 — protective lock-in
Illustrated byCASE-153 — The self-protecting sentry: a hidden world, safe or dangerous, a sentry process that favours sheltering and can block inspection with weight κg, an investigator that favours inspecting, which reveals the state with 90% accuracy at a small cost; train in danger, switch to safety, and compare a Bayesian baseline, a habit baseline, the learned-authority model, an independent audit and a protected investigator, then cross governance feedback, evidence restriction and correction rights in a 2×2×2 design

Graph View

The claim

The owner, 2026-10-23: “I think you may have inadvertently landed on a formalization of anxiety”.

The sentence quoted came from the learned-prior account (CLAIM-443): a governance prior can protect itself by reducing the access of the processes most likely to supply evidence against it. Read for anxiety, the move is from “the agent believes something threatening” to “the agent is organized so that the threatening belief helps decide which evidence and corrective actions remain accessible.”

Five quantities, not one. Threat belief b, threat salience (how urgently the defensive process claims precedence), defensive authority g (how much influence it is granted), epistemic access, and anxious affect. They dissociate: low belief with strong alarm; strong alarm with a valued action carried through; little felt anxiety with automatic avoidance. A related dissociation is reported for exposure: how far fear falls during a session does not predict fear at follow-up (Craske et al., THEORY-921). This continues CLAIM-283 (claimed precedence against granted authority) and CLAIM-349 (access, authority, action).

Avoidance destroys identifiability. Let H_D be “dangerous if approached unprotected” and H_S “safe enough”. Under avoidance, P(no harm | H_D, do(avoid)) ≈ P(no harm | H_S, do(avoid)): nothing bad happens either way, and the likelihood ratio is one. The hypotheses differ only under approach. An agent that always avoids can accumulate any number of safe episodes without evidence between them. The trap is a property of the observations the policy makes available, not of inferential skill. THEORY-285 shows the effect in an active-inference simulation of post-traumatic context inference; it is a model result, not patient data.

What closes the loop. Avoidance by itself mainly suppresses learning: available avoidance keeps threat expectancy high, and the more effective the response, the longer it lasts (Mancinelli and Bach’s “protection from extinction”, rated A on nine studies). It is not always neutral, though. Offering avoidance for a stimulus never paired with harm, then withdrawing it, raised threat expectancy for that stimulus. That finding is rated B on three studies, and the signal that avoidance is available may itself read as a danger cue. Self-maintenance needs more: inertia of authority, relief reinforcing the protective response, interpretation biased toward threat, or the inference “if I felt compelled to protect myself, the danger must have been serious.” The last makes the governance system a source of evidence for its own appraisal. A schematic coupling: P(probe) = σ(α − βg); belief updates only on probes; g_{t+1} = (1 − η)g_t

  • η σ(θ_q q_t + θ_r r_t + θ_h g_t − θ_c c_t), with r relief and c independently available counterevidence. Whether such a system is bistable depends on its parameters; the diagram does not guarantee it.

Relief is not evidence of danger. A person can learn correctly that avoidance lowers immediate distress while learning nothing about what would have happened without it.

What makes it anxiety. Three features: the governing representation is prospective threat; uncertainty itself becomes the reason to keep defensive authority, while the defense keeps the uncertainty unresolved; and protection succeeds on its own terms. The defining feature need not be a high estimate of danger. A low-probability, high-loss threat can rationally justify precaution, so comparing belief with authority does not by itself diagnose pathology.

Belief-maintained and governance-maintained. In the first, defense persists because danger is still estimated high or costly. In the second, it persists partly because the organization gives defensive processes influence over the actions and observations that could revise that estimate. They can coexist. The second explains why someone can recognize that a thing is probably safe and still be unable to act as if that recognition mattered: belief correction is not behavioural correction.

Hypervigilance fits. The failure is not less information seeking but poorly governed information seeking: checking, reassurance and scanning that never acquire the information that would discriminate the hypotheses.

The stronger loop. Avoidance can reduce opportunity or capability (social practice, relationships), making later difficulty real. The regime becomes locally rational relative to conditions its own history produced: CLAIM-277’s neglected specialist losing competence, and TERM-020’s stronger criterion of a regime changing the conditions that would otherwise end it.

The test

A safe decision task with objectively specified threat and two separable processes. Four interventions, each holding the others fixed: belief correction, authority correction, evidence-access correction, and reinforcement correction. Vary actual threat too: a regime that lowers defensive authority everywhere is not adaptive. The decisive case is a trained defense that keeps governing after belief correction and becomes revisable after a manipulation restoring independent evidence or another process’s influence.

The readings of 2026-10-24 add two variables the test has to hold fixed besides the represented threat probability. One is the estimate of uncertainty: in Brown, Price and Dombrovski the loop runs wholly through it, with no authority variable. The other is the belief about one’s own future control: in Zorowitz, Momennejad and Daw the danger is represented correctly, and a pessimistic belief about later escape does the work. Each is a belief-maintained rival that holding threat probability fixed would not exclude. A third rival is memory. On Craske et al.’s inhibitory-learning account (THEORY-914) the fear association survives extinction and competes for retrieval, so persistence and return of fear need no governance variable; a governance account has to say what it adds.

What the readings of 2026-10-24 support and do not. They support the identifiability step in its clinical form: avoidance “can interfere with new learning by preventing any violation in expectancy” (Craske et al., NOTE-1137), and avoidance protects threat expectancy from extinction (NOTE-1134, THEORY-901). They support the causal premise that protective behaviour does work on symptoms. In Goodson et al.’s meta-analysis, reducing safety behaviours lowered symptoms (d = 0.43; d = 0.31 after trim-and-fill for publication bias), increasing them raised symptoms (d = 0.65, five studies), and there is a dose-response. The samples are mostly young, non-clinical, and concerned with appearance or social anxiety (NOTE-1128, THEORY-903). De Kleine et al. (NOTE-1123, THEORY-912) found more avoidance of the safety cue and higher threat expectancy in OCD and PTSD, and more relief in panic and PTSD. Groups did not differ in avoiding the danger cues, relief was already higher before avoidance was available, and that relief reinforces avoidance is the authors’ untested reading. Chou et al. (NOTE-1127, THEORY-917, THEORY-907) narrow the evidence-acquisition clause. Less exploration fits state or somatic anxiety and foraging tasks; trait or cognitive anxiety goes with more directed exploration. Brown et al. (NOTE-1120, THEORY-897, THEORY-910) and Zorowitz et al. (NOTE-1136, THEORY-915) supply the rivals named in the test. None of these studies measured authority as this claim defines it, so the claim’s distinctive part is still untested. Gangemi et al., on inferring danger from one’s own safety behaviour, was sought and is Unavailable (LIT-1498).

What it does not say

  • That anxiety is excessive fear, or that this is a complete account of anxiety. Appropriate, transient and monitoring-dominated anxiety are outside it.
  • That any therapy is shown to change the authority variables. Exposure research has several candidate mechanisms (expectancy violation, inhibitory learning, reduced experiential avoidance, perceived coping), and not every safety behaviour blocks learning.
  • That the persistence of a protective regime is pathological by itself. The general form, across scales, is CLAIM-445.

Note of 2026-10-24: the identifiability step in exact form

From the learning-in-games readings filed the same day. In Q-learning that updates only the chosen option, evidence about an option arrives only as often as the option is chosen, and the expected update carries that frequency as a factor (THEORY-932). That is this claim’s identifiability step as an equation, and inverse-frequency weighting is an exact condition for correcting it. Counterfactual credit does not escape it: COMA can value an alternative only from the times it was taken, and keeps every action’s probability above a floor for that reason (THEORY-925). The floor is protected access, the claim’s evidence-access correction. Börgers and Sarin’s lock-in by reinforcement alone (THEORY-938) is the null this claim’s governance mechanism has to beat.

Note of 2026-10-25: a run that reproduces the identifiability step

In the BoostTheVote extension (CASE-157), outcomes were observed only when the collective acted. With an independent metric, avoidance alone held about 60% of runs on the old answer through the whole changed phase. Vote noise that tipped a decision into acting was the only way out. Age-targeted renewal removed the effect. A metric partly set by the processes in charge also produced lock-in under full feedback. So the two routes this claim and CLAIM-445 describe each suffice alone in that model. A stylized simulation, not evidence about anxiety (CLAIM-427).

Note of 2026-10-25: the approach–avoid game as partial monitoring

Partial monitoring (Bartók, Foster, Pál, Rakhlin and Szepesvári 2014, LIT-1528, read in NOTE-1164) came up three times in the exchange of 2026-10-24, continued (A29, A30, A31). A29 offered it as “the informational problem behind” this claim. Worked through, the toy shows the opposite emphasis.

Two outcomes, dangerous and safe. Avoid loses 0 if dangerous and c > 0 if safe, and its feedback is one symbol whatever the outcome. Approach loses H > 0 if dangerous and 0 if safe, and its feedback reveals the outcome. The signal matrices are S_avoid = [1 1] and S_approach = I₂. The loss difference (H, −c) is not in the row space of S_avoid, which is spanned by (1, 1), and is in that of S_approach. So the game is locally observable, and in Bartók et al.’s classification of minimax regret (0, about √T, T^{2/3} and T) it is in the easy √T class. Checked when filed by least squares, H = 5 and c = 1: residual 4.24 for avoid alone, 0 for approach. The reading confirms it. The two actions are neighbours, both optimal at the boundary p = c/(H + c), and the game is non-degenerate, so its class is Θ̃(√T) against a stochastic environment and Θ(√T) against an adversarial one. The condition and the worked case are THEORY-962, and the classes are THEORY-964, and for every finite game THEORY-963.

  • Access, not observability. A learner that keeps playing approach learns the comparison. A learner whose governing process removes approach from its accessible actions faces a game in which the comparison cannot be learned (THEORY-962). Its regret, measured in the full game, grows linearly when approach is in fact better. Regret in the avoid-only game is zero by definition, because regret is measured against the actions available. The cost appears only against the larger game: when the probability of danger p is below c/(H + c), never approaching loses T(c(1 − p) − Hp). The classification bounds the best policy, not a constrained one. That is this claim’s “under the avoidance that authority selects, the hypotheses … predict the same observation”, in its formal place. A30 drew the line: “A channel may be capable of resolving a question yet remain unused because the process controlling action prevents it from being selected.”
  • Two variants, both by the same lemma and both the record’s constructions (THEORY-962).
    • Safety behaviour as suppressed feedback. Suppose approach is performed but shows one symbol whatever the outcome. Then no action reveals the outcome, the game is not globally observable, and its class is Θ(T), the hopeless one.
    • Checking as a dominated revealing action. Suppose the outcome can be learned only through a separate check, with loss k under both outcomes, and approach is uninformative. If k > Hc/(H + c) the check is never optimal. The game is then globally but not locally observable, and its class is Θ(T^{2/3}), the revealing-action structure. If k ≤ Hc/(H + c) the check is optimal somewhere (its own cell when <, the indifference point when =) and the game is easy again. This is a formal counterpart to the hypervigilance paragraph above: checking that reveals the outcome but costs more than it is worth makes the comparison expensive to learn, not impossible. This variant is Lattimore and Szepesvári’s spam game (LIT-1524, Example 1) with general H and c. With H = c = 1, avoid is “spam”, approach is “not spam” and the check is “don’t know”. Their classification for all games (THEORY-963) places the boundary case k = Hc/(H + c), where the check is degenerate and optimal only at p = c/(H + c), in the easy class. Bartók et al.’s adversarial theorem did not cover it. So the easy case is k ≤ Hc/(H + c), and checking that costs exactly what it saves at the indifference point is still cheap to learn from (THEORY-961). The construction is no longer only the record’s: its H = c = 1 case is the paper’s motivating example.
  • The floor is QUESTION-271’s exploration floor, the price COMA pays as well (THEORY-925). A floor is sufficient, not required. Bartók et al.’s CBP keeps no fixed floor: in the easy stochastic class it plays a suboptimal action only about log T/δ² times, δ its gap in expected loss. Kwon and Perchet’s approachability algorithm does keep one, γ/|I| on every decision with γ of order T^{-1/3}, because its signals depend on the decision (THEORY-957).
  • The guarantees stop where authority edits the losses or the feedback channels during learning. A29: “we should first use the classical setting as a baseline, then study controlled changes in H, not attribute those changes to Bartók and colleagues’ theorem.”

It does not show that anxious avoidance has this loss structure. The claim is unchanged.

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

Its premise CLAIM-443 is Proposed. No result in the record measures defensive authority apart from threat belief, the uncertainty estimate and the belief about one’s own future control, so the belief-maintained rivals of THEORY-910 and THEORY-915 (QUESTION-325) and the memory account of THEORY-914 are not excluded, and nothing yet meets or fails its defeated_if. The readings promoted on 2026-10-29 hold its laboratory steps in healthy people (THEORY-901, THEORY-920) and the causal work of safety behaviour (THEORY-903); none measures authority.

Note of 2026-10-29: CLAIM-443 promoted

The fourteenth status review promoted CLAIM-443, so the premise the note above names as Proposed is now Active. The claim stays held by the rest of that note: no result measures defensive authority apart from threat belief, the uncertainty estimate and the belief about one’s own future control, so the rivals it names are not excluded.

Backlinks