The competence that vanished overnight

Most accounts of skilled performance begin too late. They begin when a person has already perceived the relevant scene, selected the relevant facts, kept them available, understood the social stakes and decided that effort is worth spending. From that point, reasoning can look like a self-contained operation. The hidden assumption is that the other faculties supply neutral inputs and then stand aside.

Leela’s case breaks that assumption. A regulated organism does not first build a complete, objective model and then ask emotion, memory or culture what to do with it. Physiological needs affect arousal and control. Perception is organised around possible action. Affect changes salience and cost. Memory reconstructs a usable past from partial traces. Other people become evidence, threat, support or authority. Language packages distinctions that a group can share. Culture stores methods no individual rediscovered.

Three kinds of constraint

The stack becomes useful when “everything interacts” is replaced with three testable forms of dependence.

Regulation is an operating-envelope constraint. Research on allostasis and interoception identifies large-scale systems linking bodily regulation with cortical and subcortical processing, while work on stress and sleep shows that altered state can impair attention, working memory and flexible control.134 The empirical findings do not prove one complete theory of intelligence. They do establish that higher-order performance is not independent of bodily condition.

Perception, affect and memory are routing constraints. A scene offers many potential actions and interpretations; affective value biases competition; remembered structure supplies expectations; and the selected action changes the next sample. Cisek’s affordance-competition account is one strong model of perception and action being prepared in parallel rather than in a strict perceive-then-decide sequence.6 Pessoa’s network account makes the complementary point that emotion is interlocked with perception, motivation, cognition and action rather than housed in a separate “emotional module”.7

Social cognition, language and culture are coordination constraints. They allow one person to inherit a distinction, ask for missing evidence, divide labour, challenge an interpretation and use procedures accumulated across generations. Shared intentionality and cumulative-culture research shows that human performance can depend on packages of teaching, imitation, prosocial coordination and communication rather than on isolated individual discovery.1011

This reframes the initial puzzle. Leela did not become globally less intelligent. The task route became narrower. A familiar diagnosis that could have been reached through several evidence paths became dependent on the colleague’s remark. The intervention target is therefore not “improve reasoning” in the abstract. It is to restore the envelope, change the evidence route or add an independent coordination path.

flowchart TD
  A["The competence that vanished<br/>overnight"] --> B["The competence that vanished<br/>overnight"]
  B --> C["A stack made of loops"]
  C --> D["Interventions that expose the<br/>stack"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#5B5CF6,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

A stack made of loops

The word “stack” suggests a vertical order: physiology at the bottom, culture at the top. That picture is useful only if it is immediately corrected. Human intelligence is full of return paths. Learned concepts alter perception. Social safety changes regulation. Tools alter action possibilities. Repeated action changes memory. Cultural practice changes what counts as a relevant distinction. The stack is vertical in dependency but recurrent in operation.

Regulation sets the range, not the answer

A regulated body maintains temperature, energy, arousal, hydration and other variables within ranges compatible with continued action. Cognitive work draws on that organisation but cannot be reduced to it. The key claim is narrower: regulation changes the range of control policies that can be sustained. Under acute stress, flexible prefrontal operations can give way to more habitual or immediately salient responses; sleep loss can alter vigilance, learning, emotional reactivity and social behaviour.345

This does not mean “the body decides everything”. A skilled person can compensate, a team can redistribute work, and a familiar routine can remain stable. It means that competence has a state-dependent operating envelope. A test taken in quiet, safety and daylight samples one part of that envelope. A night shift with interruption, uncertainty and social pressure samples another.

Perception and action define each other’s options

Imagine a cup on a table. The visual properties are the same whether your hand is free, gloved, injured or holding a tray. Yet the immediately available actions are different. If perception serves action, those differences should matter before a detached description is complete. A cup that can be grasped, avoided, handed over or used as a weight is not represented only as a bundle of colour and shape.

The practical implication is simple. A person may fail to “see” the relevant option because the interface, permission, physical layout or social role makes the option unavailable. Adding more information will not necessarily help. The route may need a new action affordance: a visible override, a permitted pause, a second display, a way to ask for help.

Affect is a priority system, not decorative colour

Affect is often described as noise added to otherwise rational thought. That framing assumes the system already knows what deserves attention. In a finite organism, it does not. Potential threat, loss, reward, novelty and social consequence compete for limited processing. Affective weighting helps allocate that processing. The same mechanism can protect or distort. A sudden hiss should dominate attention near a gas line; an angry email should not dominate a complex risk calculation merely because it is vivid.

Network evidence supports treating emotion and cognition as deeply interlocked rather than as separate streams that meet only at the end.7 The contested question is how best to characterise the computation. The stack thesis does not require one theory of emotion. It requires only the more modest observation that affect changes selection, effort, memory and action, so a reasoning failure may be a weighting failure.

Memory is a reconstruction service

Memory does not deliver a complete archive into a reasoning chamber. It reinstates fragments, relations and learned expectations according to the current cue and goal. Event-cognition research treats perception, action, language and memory as interfaces around structured events rather than sealed processes.9 Cultural-memory research adds that what people rehearse, narrate and preserve is shaped by social practices.15

That gives memory two roles in the stack. It supplies task structure, and it inherits the biases of earlier selection. If Leela originally encoded a fault pattern under a simplified training scenario, retrieval may confidently return the simplified pattern. More memory is not automatically better. The useful question is whether the current cue retrieves the distinction needed now.

Language coordinates without owning thought

Language is powerful enough to create a recurring mistake: fluent expression is treated as the visible surface of thought, so the two are assumed to be the same faculty. Evidence from neuroimaging and neurological cases supports a substantial separation between high-level language processing and many forms of reasoning. A recent synthesis argues that language is primarily adapted for communication rather than being the medium of all thought.13 That claim remains debated, but it supplies an important boundary for this paper: language is one layer in the stack, not its executive.

Its main contribution here is coordination. Language lets people point to an absent cause, define a threshold, transmit a procedure, assign authority and challenge a claim. Conversation can also align later attention and interpretation. In one study, groups that built consensus showed greater subsequent neural alignment than groups that did not, including on novel material from the same narratives.14 Alignment is not truth. It is evidence that communication can retune later processing.

Culture is stored search, with inherited blind spots

No individual derives rail signalling, algebra, surgical asepsis or constitutional procedure from first principles. Culture preserves solved subproblems in tools, norms, diagrams, roles and institutions. Comparative work on cumulative culture found that children’s progress on a staged puzzle was associated with a package including teaching, imitation and prosociality that was not observed in the same form in chimpanzees or capuchins.11 The broader cultural-niche argument is that humans often succeed by acquiring locally adaptive practices from others, even when the individual cannot independently justify every step.18

The cost is inherited error. A practice can preserve a useful distinction or freeze a mistaken one. Cultural differences in attention and categorisation show that habitual environments can shape what is foregrounded, although the size, universality and interpretation of such differences require care.12 Culture therefore increases reach while changing the prior search space.

Interventions that expose the stack

A dependency map earns its place only when it changes an experiment. Merely listing faculties produces an attractive taxonomy but no causal leverage. The stack predicts a specific pattern: perturbing a required lower or lateral dependency should alter several apparently higher capacities, while preserving tasks that bypass that dependency. A successful rescue should restore the blocked route without retraining every local competence.

Worked example 1: the four-light panel

This example separates competence from access. A correct verbal explanation after the trial does not prove that the relevant representation controlled the action. Conversely, a wrong button press does not prove that the rule was absent. The intervention must target the route between them.

Worked example 2: the smallest useful formalism

An additive model scores a task by averaging readiness across required layers. Let r_i be readiness of layer i between 0 and 1, and w_i be the task’s dependence on that layer. The additive score is easy to understand:

Some tasks allow that. A speaker can compensate for imperfect vocabulary with gesture; a team can compensate for one person’s memory with a checklist. Other tasks contain a narrow chain. To represent that distinction, use a soft bottleneck:

Consider two profiles with the same additive mean of 0.75. The balanced profile gives every required layer 0.75. The uneven profile gives one required layer 0.25 and the other seven about 0.821. With τ = 0.12, the uneven profile’s soft bottleneck is about 0.493. For a chain-sensitive task with κ = 0.75, its predicted effective score is 0.557, while the balanced profile remains 0.75.

The model’s value is not numerical realism. It forces two questions that an average hides: which layers are genuinely required, and how substitutable are they? A negative result is useful. If the uneven profile performs as well as the balanced one, the task has a bypass, the supposed dependency was misidentified, or the bottleneck model is wrong.

Worked example 3: the configured rail-control task

Return to Leela. The configured system includes more than her brain. It includes the signalling display, shift log, procedure, radio protocol, colleague, authority boundary and simulated track state. The task is to isolate a section safely under uncertainty. A faculty atlas begins by marking what each layer contributes and what observable failure would indicate a blocked path.

A practical protocol uses four matched trials. Trial one is baseline. Trial two adds mild sleep restriction or a validated fatigue proxy without changing task content. Trial three restores external counting and forces an independent evidence request before action. Trial four keeps the support but removes the colleague’s opinion. Measure not only final accuracy, but evidence sampled, time to first hypothesis, hypothesis revisions, communication turns and action latency.

The stack predicts a pattern. Fatigue should have a larger effect when the task requires interference control and social disagreement. Externalising the sequence should rescue memory load but not necessarily social conformity. Hiding the colleague’s opinion should reduce one bias but may slow useful coordination. Different rescues should produce different fingerprints. A generic “cognitive ability” explanation predicts less selectively unless it specifies the same dependencies.

Worked example 4: the counterexample that prevents a rigid stack

A concert pianist can play an overlearned scale while carrying on a simple conversation. The same pianist may fail to improvise a novel modulation under identical distraction. The familiar routine relies on a highly practised sensorimotor and procedural path. The novel task requires broader working memory, affective control and perhaps explicit symbolic reasoning. A strict vertical model would predict both should fail together. A task-specific route model predicts the dissociation.

This is the strongest boundary on the thesis. Faculties are partly dissociable, routes can be redundant, and expertise can reorganise them. Large-scale brain networks can become more segregated or integrated as tasks become automated, consistent with changing dependency structure rather than one permanent hierarchy.8 Double dissociations also do not automatically prove clean modules; distributed systems can produce them after selective damage.17

Failure case: when coordination amplifies error

The upper layers of the stack do more than add capability. They can create a self-reinforcing failure. A cultural prior directs attention; selected evidence is named in shared language; conversation aligns interpretations; the aligned story shapes later memory; and repeated retrieval strengthens the prior. Every individual may be locally coherent while the group becomes globally wrong.

stateDiagram-v2
  [*] --> Claim
  state "Three kinds of constraint" as Claim
  state "Thought experiment 1: the<br/>rested replacement" as Test
  state "Regulation sets the range, not<br/>the answer" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Build the faculty dependency atlas

A useful atlas is not a personality profile. It is a map of one configured task. Begin with the outcome that matters, then work backwards through the paths that make that outcome reachable. For Leela, the outcome is not “good reasoning”. It is safe isolation of the correct rail section, using authorised evidence, within the available time.

The map should record five things for each claimed dependency: the contribution it makes, an observable trace, a perturbation that should selectively weaken it, a rescue that should selectively restore it, and the inference that remains prohibited. A layer belongs in the atlas only when changing it could change a prediction.

Executable lab: compare an average with a bottleneck

The live atlas below implements the minimal model from Part III. Choose a task, move the eight readiness sliders and compare the weighted average with the constraint-aware score. The intervention ranking asks a deliberately small question: which single 0.10 increase would most improve the current task model?

A positive result permits a targeted test. It does not prove the named layer caused the real-world failure. A negative result can mean redundancy, an incorrect task map, an insensitive intervention or a false bottleneck thesis. The smallest useful implementation is one task, three to five observable dependencies, one reversible perturbation and one matched rescue.

function evaluate(readiness, task) {
  const entries = task.weights.map(([name, weight]) => ({
    readiness: readiness[name], weight
  }));
  const totalWeight = entries.reduce((s, x) => s + x.weight, 0);
  const additive = entries.reduce(
    (s, x) => s + x.weight * x.readiness, 0
  ) / totalWeight;

  const weightedExponentials = entries.reduce(
    (s, x) => s + (x.weight / totalWeight) *
      Math.exp(-x.readiness / task.tau), 0
  );
  const bottleneck = -task.tau * Math.log(weightedExponentials);
  const score = (1 - task.kappa) * additive +
    task.kappa * bottleneck;

  return { additive, bottleneck, score };
}

The evaluator encodes two contestable assumptions: task dependencies can be assigned relative weights, and chain sensitivity can be represented by a soft minimum. It cannot establish a latent faculty structure, diagnose impairment, measure general intelligence or decide whether any system is conscious.

type Layer = "regulation" | "perception" | "action" | "affect" |
  "memory" | "social" | "language" | "culture";

type Dependency = {
  layer: Layer;
  readiness: number;       // invariant: 0 <= readiness <= 1
  weight: number;          // invariant: weight >= 0
  observable: string;
  perturbation: string;
  rescue: string;
  prohibitedInference: string;
};

type TaskAtlas = {
  outcome: string;
  tau: number;             // invariant: tau > 0
  kappa: number;           // invariant: 0 <= kappa <= 1
  dependencies: Dependency[];
};

Calibrate routes, not people

The sliders create an immediate temptation to ask where their numbers come from. The wrong answer is to assign them from impression and then treat the result as measurement. Readiness should be anchored to a task and an observation. A perception value might summarise detection under a declared contrast and noise level. A memory value might summarise correct event-order retrieval after a fixed delay. A social value might summarise whether an independent judgement survives a confident peer cue. The number is shorthand for an experimental condition, not a hidden quantity discovered inside the person.

Calibration should begin within the same person or team. Between-person comparisons introduce differences in experience, strategy, trust and tool use before the dependency map is stable. Repeat matched task variants, preserve the outcome definition and change one causal property at a time. Once the route is understood, broader samples can test whether the pattern travels. The first target is a reproducible intervention fingerprint, not a universal norm.

A five-trial dependency protocol

Use a short sequence that can disconfirm the proposed bottleneck. The sequence is designed for simulation, training or low-risk process improvement. It is not a licence to impose fatigue, distress or impairment.

1. Baseline. Run the configured task with ordinary supports. Record the final outcome, evidence sampled, hypothesis revisions, communication turns, action sequence and recovery from error.
2. Targeted perturbation. Alter one safe task property linked to the claimed dependency. Add visual clutter for perception, a harmless delay for event-order memory, or a scripted confident cue for social weighting.
3. Selective rescue. Keep the perturbation but restore the proposed path. Increase contrast, externalise the timeline, or require a private judgement before conversation.
4. Matched alternative. Add support of similar effort that should not repair the target path. Extra rule explanation is a useful control when the claim concerns cue selection rather than missing knowledge.
5. Bypass. Reconfigure the task so the target dependency is no longer required. An automated event counter may bypass memory load; an independent sensor may bypass contested testimony.

A strong pattern has four parts. The perturbation changes the predicted trajectory. The selective rescue restores it more than the matched alternative. The bypass reduces sensitivity to the perturbation. Unrelated tasks remain comparatively stable. If only final accuracy changes, interpretation remains weak because speed-accuracy trade-offs, motivation or guessing can mimic the result. Trajectory measures show where the route diverged.

Consider the social-cue hypothesis in Leela’s task. At baseline, she samples the timing trace before asking the colleague. Under a confident but wrong cue, her first hypothesis appears earlier, she samples fewer independent sources and revises less often. A private initial judgement restores source diversity, while extra procedural text does not. When the colleague’s comment is hidden, the difference disappears. This fingerprint supports a coordination bottleneck more strongly than a generic fall in accuracy would.

Now imagine the opposite result. The confident cue changes nothing, but externalising event order restores performance across social and solo conditions. The atlas should move weight from social cognition to memory or perception. If both interventions rescue performance, the task may contain parallel routes. If neither does, the named layers may be proxies for regulation, motivation or an unmeasured interface problem. Failed selectivity is not a nuisance result; it is the mechanism correcting the map.

From a score to an intervention claim

The live model ranks a hypothetical 0.10 improvement, but a real decision requires three additional tests. Is the intervention observable through a path independent of the outcome? Is it reversible or containable? Could it shift another layer in the opposite direction? A checklist may reduce memory demand while increasing visual clutter. A mandatory challenge may reduce conformity while increasing social threat. Rest may improve regulation but change staffing and time pressure. The atlas must represent these interactions rather than counting every support as a free gain.

For that reason, the intervention claim should be written in causal form: “Under task configuration C, changing X while holding Y stable should alter trajectory measure M, and rescue R should restore M.” State the rival explanation beside it. The claim is strengthened when the rescue fingerprint repeats, weakened when an equally plausible alternative produces the same pattern, and abandoned when the predicted dissociation fails. This is the point at which a colourful faculty map becomes an experimental instrument.