nucleation

CLAIM-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

6 min read

Properties11
StatusActive
Version3
Rolethesis
Warrantinterpretive
Tagsself-governance, learning-and-conditioning, motivation
Linewill-organization, distributed-agency
Rests onCLAIM-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
SupportsCLAIM-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
ComplementsCLAIM-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, CLAIM-344 — Resolve is a maintained organization of precedence, commitment and revision that keeps a temporally extended policy effective across changes in locally dominant evaluations while keeping the capacity to revise it; mere slowness or persistence is not resolve, CLAIM-364 — Processes conditioned never to object coordinate without override at the same behaviour and supervisory cost as processes that negotiate, so the two part only under intervention, when a constituent dissents, or in their history, CLAIM-444 — A replicator dynamic on the distribution of authority among an organization's processes is the special case of the meaning–authority–legibility model in which the organization's interpretation is the authority-weighted mixture of its processes' and evidence reaches scoring without a maintained regime of its own; the replicator tools (Bayesian reading, potential and harmonic games, stochastic stability) then apply, and the record's separate interpretation and legibility states earn their place exactly where interpretation departs from that mixture or legibility keeps a memory of past authority, CLAIM-358 — Learned local accommodation can reduce the amortized cost of recurrent arbitration: where a recurring conflict among an agent's processes admits a stable accommodation, settling it among the processes can cost less over repeated encounters than escalating each instance to a higher-level process that mediates or overrides, though local negotiation and the relations it needs are not free
Corroborated byTHEORY-913 — In small fixed-effect meta-analytic pools without correction for publication bias, responses specified in if-then plans ran more immediately (d = .77, k = 7), held up under a concurrent cognitive load (d = .85, k = 7) and ran without conscious intent (d = .72, k = 3) compared with responses under a goal intention alone, and the planned cues were detected, attended to and remembered better (d = .80, k = 6); the authors read this as control of the response handed in advance to the cue, THEORY-904 — Gollwitzer and Sheeran (2006) conclude, narratively from a handful of studies that measured or manipulated the goal intention, that implementation intentions affect performance only when the goal intention they serve is strong and activated, so the planned response is not elicited mechanically by its cue; in 2024, across eight tests, plan effects were larger with strong, activated or self-concordant goal intentions than with weak, unactivated or self-discordant ones (d = .79 for the difference), and in 31 physical-activity tests split at the median of intention, planning helped with strong intentions (d = .31) and had no effect with weak ones (d ≈ 0), THEORY-918 — Across 642 experimental tests to September 2019, forming implementation intentions improved outcomes, with uncorrected effects on behaviour (d = .27), cognitive performance (.47), affect regulation (.66) and social judgement (.52), but publication bias was substantial: the pooled d = .36 falls to a model-averaged d = .15 under robust Bayesian meta-analysis, which still found extreme evidence of an effect (BF10 = 319), with large heterogeneity, THEORY-908 — Between tests, uncorrected for publication bias, implementation intentions in the contingent if-then format had larger effects than plans that schedule when, where and what without linking cue to response (d = .43 against .29); tests in which the plan was rehearsed had larger effects (.45 against .33, with no gain beyond one rehearsal), tests in which it was recorded had smaller ones (.31 against .44), and adding how or for how long to a time and place went with about half the effect (.24 against .46), THEORY-922 — Among 56 of 88 patients in one DBT intensive outpatient programme, kept because they finished a four-month cycle and gave enough diary data, skills-homework completion in months 1–3 correlated with larger falls in suicide and substance-use urges (r ≈ .35) and phone-coaching frequency with a larger fall in self-harm urges (r = .31); neither correlated with change in how often the behaviours occurred, and only homework separated patients with and without a target behaviour in month 4, THEORY-923 — In one simulated T-maze task with hand-set parameters, run once per condition, an active inference agent's habit, learned by counting its own state transitions and policy choices, emerged from goal-directed behaviour, and once acquired it persisted, and reinforced itself, after the reward was devalued, while an agent devalued before acquisition switched at once, THEORY-909 — As reviewed by Gerstner et al. (2018) from five papers, mostly in slices and one laboratory per region, pairing pre- and postsynaptic activity changed the synapse when a neuromodulator or plateau potential followed within about 1 s in striatum, within 5 s (noradrenaline, LTP after pre-before-post pairing) or 2.5 s (serotonin, LTD after post-before-pre pairing) in visual and prefrontal cortex, and within about 2 s in CA1, and not at longer delays; the reviewers read this as a decaying synaptic eligibility trace and as experimental support for three-factor learning rules on behavioural time scales
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, CASE-155 — Authority as a replicator population: for a distribution π of control over processes with mean qualities q, the expected Maynard-Cross update is π_a q_a / v, Bayes' rule when q is each process's likelihood; the Maynard dynamic is the Taylor dynamic slowed by 1/v, so the two share their orbits; mean quality climbs at the rate Var_π(q); and when a process's score rises with its own share, ℓ′ = ℓ + h + c(2π − 1) has an unstable threshold π* = (1 − h/c)/2, above which a process the evidence disfavours captures the organization

Graph View

The claim

The owner, 2026-10-23: “what if we just treat this as a kind of momentum term? or like a hebbian prior?”

The two suggestions are different mechanisms on different timescales.

Momentum, read as switching cost. g_t = argmax_g [Q(g, z_t) − κ·1(g ≠ g_{t−1})]. A governing allocation resists small transient changes of salience. This is inertia or hysteresis rather than momentum in the optimization sense.

A Hebbian prior. Let g be an allocation of authority among the processes and z the context. Governance is selected as

P(g | z, H) ∝ exp(β Q(g, z) + b_H(g, z)),

with Q the current evaluation of the allocation and b_H a learned, history-dependent bias toward allocations used before. The bias is trained on the relations by which a process j recruits or endorses a process i:

ΔE_ji = η x_j y_i (R_t − R̂_t) − λ E_ji,

co-activation gated by an outcome signal, with forgetting. This is a schematic three-factor rule, not a claim about any neural implementation. In a familiar context the prior already puts a process in a position to govern, so the case for it need not be rebuilt. CLAIM-277’s durable authority from repeated endorsement becomes this prior; CLAIM-358’s note of 2026-10-23 gives the accounting for when it pays.

Worked example (DBT). An avoidance process A, a relationship process B and a pause-and-assess procedure R. At first B has influence only if R is recruited deliberately and at cost. After practice, A’s own activation recruits R. Neither A nor B changes what it values. What changes is the conditional probability that one process’s activation recruits another with a regulatory role (CLAIM-349’s note of 2026-10-23).

Two limits

Persistence is not amortization. A strong prior can change which allocation wins while every process still does all the same costly evaluation and negotiation. Amortization is shown only when earlier work replaces later computation, communication or supervisory recruitment. In CLAIM-358’s terms, reuse pays under a declared cost measure and comparable outcomes when N(c_d − c_π) > F + C_maintenance + C_error, with C_maintenance and C_error totalled over the N encounters, and the inequality has to be measured, not inferred from stable behaviour.

A prior over authority is not a prior over actions. “When distressed, pause” is an action policy. “When distressed, make this other process’s evaluation available before acting” is a relation of influence. They look the same in familiar situations and come apart when the appropriate action changes: a recruited process with the information and capacity can choose a different response, and a cached action cannot.

Naive Hebbian learning entrenches. Whatever is recruited often gets more chances to succeed and is reinforced whether or not it deserved authority, and a routine that controls which alternatives are considered reinforces its own use. The prior encodes a history of reinforcement under a sampling regime, not competence. This is CLAIM-277’s warning and CLAIM-364’s, and it is where the same exchange found a candidate account of anxiety (CLAIM-446).

The test

Train two systems to the same action. In one, reinforce the cue-to-action mapping; in the other, reinforce recruitment of a process that chooses among actions. Then change which action is appropriate. Separately, compare recurrent cost before and after training: responses that become more persistent with no computation or supervision saved are not evidence for amortization. Learned local accommodation and a compiled supervisor should also fail differently when lower-level interactions and higher-level deliberation are disrupted selectively.

What the readings of 2026-10-24 add. Friston et al. (NOTE-1138, THEORY-923) is a published analogue of the selection rule, π = σ(Ê − F − γG), but over action policies, not over who governs. Its learned prior grows by counting the policies pursued, with no reward gate. That is closer to the entrenchment warned of below than to the rule above. Its devaluation simulation runs this claim’s test on the cached arm: once formed, the habit persists. Every policy is still scored on every trial, so the paper shows persistence, not amortization. Gerstner et al. (NOTE-1130, THEORY-909) narrow “three-factor rule”. Only the hour-scale tag of synaptic consolidation is well established. Eligibility traces on the scale of seconds have recent support, mostly from brain slices, and the rule above lacks the decaying trace that learning from delayed outcomes needs. With reward minus expected reward as the third factor, frequency alone is not reinforced, which answers part of the entrenchment point below; what is never recruited is still never credited. Implementation intentions (NOTE-1124, NOTE-1131; THEORY-913, THEORY-904, THEORY-918, THEORY-908) are the nearest studied case. An if–then plan “delegat[es] the control of goal-directed behavior to specified situational cues”, which is a prior over actions, not over who governs. It works only while the goal is active, a third arrangement this claim had not named. The effect is reliable but small: d = .65 in the 2006 meta-analysis, uncorrected for publication bias, and about .15 after the 2024 review’s robust Bayesian correction. The 2006 tests of efficiency and conserved capability measure later effort saved, not only persistence, though for a cached action. What happens when the planned response proves wrong is the authors’ open question (“We do not know yet what happens”), and it is this claim’s test. Edwards et al. (NOTE-1121, THEORY-922) associate DBT homework with falling urges, not with less frequent behaviour, and phone coaching did not significantly predict outcome. They measure neither skill use nor its cost. Linehan’s manual and Amestoy et al.’s mediation study were sought and are Unavailable.

What it does not say

  • That this is how DBT works. The DBT example is a functional interpretation, and the record holds no reading of Linehan.
  • Which learning law is right: Hebbian updates, Bayesian policy priors and learned arbitration can produce the same behaviour from different dynamics.
  • That a learned governance arrangement that is cheap and stable is appropriate (QUESTION-087).

Note of 2026-10-29: promoted

The fourteenth status review promoted the claim, at the status review’s ruling (the owner delegated these rulings), after moving CLAIM-358 to complements and the LITs out of grounds. Its test, a change of the appropriate action after training a cached and a recruited arm, is still unrun; Friston et al.’s devaluation run is the cached arm alone.

Backlinks