Part I. A penalty is not yet a stake

Begin with a thermostat

A thermostat registers a difference between measured temperature and a set point. It turns a heater on when the room is cold. Something is being controlled, but whose condition is at issue? In the usual installation, the room temperature matters to the occupants. The room can freeze while the thermostat correctly reports that it is cold. Its error signal is about a variable outside the controller, and an external designer supplied the set point.

Now change one property. Put the controller inside an incubator containing components that cease to function outside a narrow temperature band. Let its own processor slow as temperature falls, let sensor noise rise and let its future control capacity depend on restoring warmth. The same equation now participates in a different causal loop. Temperature is no longer just the target of its action. Temperature partly determines whether the system can continue producing competent action.

This suggests three meanings of “matters” that should never be collapsed. An assigned loss matters to an evaluator because it changes a score. An operational stake matters to a system because it changes the system’s own future viability or competence. Phenomenal valence means that an outcome feels good or bad. The first can exist without the second. The second can be studied without assuming the third. The third is the welfare-relevant property, but it is also the hardest to establish.

The constitutive coupling criterion

We can now state the paper’s new causal distinction more precisely. A variable is an operational stake for a system when three conditions hold together. First, the variable is part of the conditions under which that organised system can continue functioning. Second, the system can sense, estimate or anticipate changes in it. Third, those estimates causally reallocate action, often at a cost to other objectives.

The criterion is intentionally stronger than “the system avoids X”. A chess program avoids checkmate, but deleting its current board position need not damage the program that evaluates the next game. A cloud service avoids latency, but its replicas may absorb server loss without any particular runtime carrying the consequence forward. By contrast, dehydration alters the very neural, muscular and circulatory processes that must organise the next act of seeking water.

Operational stakes are therefore about constitutive dependence, not dramatic behaviour. They can be graded. A mild temperature deviation may merely increase regulatory cost. A severe deviation may narrow action options. Beyond a boundary, the system loses the capacity to recover. The geometry is a basin because many internal variables interact, and because the same action can improve one while worsening another.

flowchart LR
  A["The warehouse test"] --> B["Part I. A penalty is not yet a<br/>stake"]
  B --> C["Part II. The body changes the<br/>optimisation problem"]
  C --> D["Part III. Interoception makes<br/>value conditional"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#7B4CD8,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Part II. The body changes the optimisation problem

Homeostasis is organised movement, not stillness

Walter Cannon used homeostasis for coordinated physiological processes that preserve relatively stable conditions despite disturbance.[1] The word is often illustrated as a needle returning to a fixed point. Biology is less tidy. Temperature, blood gases, osmolarity, glucose availability, immune activity and tissue integrity operate on different timescales. Their acceptable ranges shift with sleep, exertion, infection, pregnancy, development and environment.

Peter Sterling’s influential account of allostasis emphasises predictive regulation: prepare resources before the demand arrives, and alter several effectors together rather than correcting one isolated error after the fact.[2] Homeostasis and allostasis need not be rival slogans. For this argument, the useful distinction is temporal. Reactive control corrects a deviation already sensed. Anticipatory control changes state because a future demand is expected.

Established physiological principle Organisms depend on regulated internal conditions. Contested synthesis The claim that the brain is fundamentally an organ of allostatic prediction is broader and should be treated as a research framework, not a settled reduction of cognition.

Represent the body at time t by a vector x_t. Its entries might include available energy, core temperature and tissue integrity. Let V be the region in which the organism can continue organised activity. An action a_t and disturbance d_t move the body:

Every symbol has a physical interpretation. x_t is the present bodily condition. a_t is something the organism can do. d_t is what the world does to it. H is the planning horizon. The second line asks how an action changes the chance of remaining viable over that horizon. It is not a definition of consciousness. It is a compact way to express why an option can be good for the organism before any external score is assigned.

This is why a single scalar reward is an incomplete picture even when it predicts behaviour. Different internal variables have asymmetric boundaries, different recovery times and different interactions. Drinking can improve hydration while worsening salt balance. Sprinting can escape a predator while damaging tissue. Rest can restore energy while increasing exposure. Biology solves a portfolio of coupled control problems under uncertainty, not a clean maximisation over one timeless quantity.

Reserve, rate and recoverability

A snapshot of the body can still hide the stake. Imagine two runners at the same core temperature. One has just entered shade and is cooling. The other is still climbing in humid heat, with the same temperature but a steep upward trend. A controller that sees only the present value treats them alike. A controller that estimates the direction and speed of change prepares different actions. The relevant object is therefore a trajectory through a viable region, not a reading on one dial.

Three coordinates make the difference visible. Reserve is the remaining distance from a failure boundary. Rate is how quickly disturbance is consuming that reserve. Recoverability is the future cost and time needed to return, including whether the necessary actions will still be available. A shallow deviation with rapid, cheap correction may carry little operational weight. The same numerical deviation becomes serious when correction is slow, when several variables are deteriorating together or when damage removes the mechanism of repair.

Practitioner synthesis This three-coordinate view predicts that intelligent regulation should sometimes act before any variable leaves its normal range. It also explains why irreversibility matters. A lost unit of energy can often be restored. A destroyed sensory surface may permanently narrow the policy space. The size of a stake is better estimated by threatened future options than by present error alone. Experiments should therefore record reserve trajectories, recovery costs and lost capabilities, rather than only whether a set point was crossed.

The boundary can move

The thought experiment guards against a common mistake in both biology and AI: treating every monitored variable as internal merely because it is represented inside a model. The causal boundary follows dependence. A database may contain battery level without being the thing whose existence depends on the battery. Conversely, a prosthesis or external caregiver can become part of the operative control loop if its failure propagates into the organism’s capacity and the organism coordinates with it.

Enactive accounts describe related ideas using adaptivity: the organism regulates its relation to conditions of viability, and that regulation grounds a perspective from which events can be better or worse for it.[14] This is a philosophical and theoretical proposal, not a laboratory verdict about consciousness. Its practical value is that it demands a boundary test. Ask which process must remain organised, which variables make that possible and where the cost of correction is actually paid.

Part III. Interoception makes value conditional

The outside world is filtered through a changing inside

Interoception is often glossed as “sensing the body”. A broad consensus definition is more useful: the nervous system senses, interprets and integrates signals originating within the organism.[3] This includes cardiac, respiratory, gastrointestinal, thermal, osmotic, metabolic, immune and nociceptive information, although not every signal becomes conscious and not every measurement task captures the same ability. Human heartbeat-detection work associates interoceptive awareness with activity in insular and cingulate regions, while still relying on an attended report task.[8]

Interoception matters for intelligence because it changes the meaning of external information. A food cue is more than identified as food. Its action relevance depends on hunger, predicted effort, digestive state, threat and opportunity. A shaded path is more than darker. It can become a cooling action. The body turns perception from classification into conditional affordance.

Mouse experiments make the mechanism visible. Livneh and colleagues found that food-cue responses in insular cortex were biased by hunger and largely abolished by satiety.[5] In later work, ongoing insular population activity tracked hunger and thirst states, while food or water cues transiently shifted activity towards a pattern associated with future satiety.[6] These studies do not show that the insula contains a little observer reading a body dashboard. They show that external cue processing and internal need are causally integrated in population activity and behaviour.

A classic term for state-dependent pleasantness is alliesthesia. Michel Cabanac showed that the same stimulus can become pleasant or unpleasant depending on physiological usefulness.[7] This is important, but it should not be inflated. State-dependent pleasantness supports a relation between bodily need and hedonic evaluation. It does not show that every pleasure is drive reduction, that all aversion is homeostatic error or that valence is one scalar variable.

Body signals alter behaviour only in a context

One of the clearest causal demonstrations comes from a study that optically paced the hearts of freely moving mice. Elevated heart rate increased anxiety-like behaviour in risky settings, while a real-time place-preference assay found that the pacing was not intrinsically aversive. Inhibiting posterior insular cortex attenuated the pacing-related behavioural effect.[11] Primary causal experiment The result supports a body-to-brain contribution that is conditional on environmental context. It does not establish what the mice felt, nor does it imply that heart rate alone creates anxiety.

Other mouse work links posterior insular circuits to behavioural changes during aversive internal states and to state-dependent regulation of fear.[9][10] Together, these studies undermine a picture in which the brain first computes a complete value from the outside world and only later consults the body. Bodily state participates in the computation of what the situation affords.

Valence is not reward with a biological accent

Reward is a technical term that varies across reinforcement learning, behavioural neuroscience and everyday speech. It can refer to an outcome, a teaching signal, incentive motivation or hedonic impact. These roles dissociate. Neuroscience distinguishes “wanting” from “liking”, and pleasure depends on more specific circuitry than a generic reward-prediction signal.[15] Calling a scalar “reward” therefore does not tell us whether it is felt, whether it tracks viability or whether it merely shapes policy.

Valence, in its experiential sense, is the positive or negative character of feeling. In a functional sense, researchers sometimes use it for a common currency that prioritises approach, avoidance and learning. A plausible synthesis is that bodily regulation supplies much of the material from which valenced priorities are constructed. Damasio and Carvalho, for example, relate feelings to mapped changes in organism state.[16] Predictive and active-inference accounts propose that interoceptive prediction and control help constitute emotion and embodied selfhood.[17]

These are serious proposals, not settled identities. Different computational models of interoception make different assumptions about representation, prediction, error and control.[13] A homeostatic error can be non-conscious. A painful feeling can persist when immediate correction is impossible. Pleasure can concern play, music, curiosity, attachment or abstract success. Social and developmental regulation can reshape what counts as threat or relief. Viability gives value a causal anchor, but it does not compress the whole phenomenology of value into a distance from a set point.

A common currency helps choice, then hides causes

An organism cannot always satisfy every regulatory demand at once. It may need to choose between cooling, feeding, protecting an injury, keeping contact with a caregiver and escaping a predator. Some mechanism must compare actions that change unlike variables on unlike timescales. A valence-like summary is useful here because it can turn heterogeneous consequences into an ordered urgency: approach this, postpone that, stop the third. The compression supports decisive action when a complete physiological explanation would arrive too late.

Compression also destroys information. The same negative priority can be produced by tissue damage, rising carbon dioxide, social threat, prediction error or learned avoidance. A scalar does not reveal which process generated it, whether correction is possible or whether the signal is experienced. This is the same mistake made when every artificial objective is called reward. A common currency can coordinate control without being a common cause, a complete explanation or a unit of feeling.

The discriminating experiment is to hold outward opportunity and assigned reward constant while changing the bodily source. If heat, dehydration and injury produce distinct anticipatory actions, recovery sequences and cross-goal trade-offs, the controller is preserving causal structure beneath the summary. If every perturbation triggers the same scripted avoidance, the apparent valence may be only a label. This test does not settle phenomenology, but it prevents the functional vocabulary of good and bad from floating free of mechanism.

stateDiagram-v2
  [*] --> Claim
  state "Begin with a thermostat" as Claim
  state "The constitutive coupling<br/>criterion" as Test
  state "Homeostasis is organised<br/>movement, not stillness" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Part IV. Build the difference

A minimal interoceptive controller

A useful implementation should not pretend to simulate animal feeling. It should isolate the causal difference between a policy that pursues an external score and a policy that treats internal variables as conditions of continued control. The lab below uses three synthetic variables: energy, temperature and integrity. It offers four actions: sprint, forage, rest and cool. A temporary heat and resource disturbance begins part-way through the run.

The state-blind policy follows a fixed schedule tuned for a benign environment. It can perform well when conditions remain familiar. It has no access to the body variables and therefore cannot revalue its schedule when heat rises. The interoceptive policy predicts the next body state for each available action. It combines external score with the reduction in a quadratic drive and a steep penalty near viability boundaries.

D measures weighted distance from preferred ranges. The hat marks a predicted next state. The parameter λ sets how strongly the controller trades external score against improving internal conditions. The boundary term becomes large near failure. The drive-reduction term is related to homeostatic reinforcement-learning models in which outcomes acquire value through predicted changes in physiological state.[4] Predictive-control formulations of interoception make a wider claim about modelling internal dynamics and selecting allostatic action.[12] The equation encodes three assumptions: the agent has access to internal estimates, it can predict action consequences and those variables truly affect continuation. Remove any assumption and the interpretation changes.

How to read the result

At low disturbance, the fixed schedule can survive because it was tuned for that environment. This is a useful control condition. A policy does not need interoception when the world never violates the assumptions built into its schedule. As disturbance grows, the state-blind policy continues its old rhythm. Temperature rises until continuation fails. The interoceptive policy spends actions on cooling, resting or foraging when the predicted reserve becomes thin.

A positive result supports a narrow claim: access to causally relevant internal state can improve adaptive control under distribution shift. A negative result has several possible meanings. The state variables may be irrelevant. The predictive model may be wrong. The risk penalty may be miscalibrated. The disturbance may be too weak. A state-blind policy may already contain a schedule sufficient for the tested environment. Failure of this implementation would weaken the proposed controller, not the fact that living organisms depend on viable conditions.

The lab also reveals what the model cannot establish. The “drive” is designer-chosen. The preferred ranges are supplied rather than grown through metabolism and development. The agent does not build or repair its own components. There is no evidence of feeling. The simulation is a mechanism probe, not a synthetic organism and not a welfare assay.

function chooseInteroceptiveAction(state, options, env, lambda) {
  const currentDrive = drive(state);
  let best = { value: -Infinity, action: options[0] };

  for (const action of options) {
    const predicted = transition(state, action, env);
    const predictedDrive = drive(predicted.state);
    const value = predicted.reward
      + lambda * (currentDrive - predictedDrive)
      - boundaryRisk(predicted.state);

    if (value > best.value) best = { value, action };
  }
  return best.action;
}

// The result supports operational adaptation only when:
// 1. internal variables really constrain continuation;
// 2. sensing and transition predictions are sufficiently accurate;
// 3. the matched policy sees the same external opportunities;
// 4. no claim about felt valence is inferred from the score.

The constitutive coupling test

The simulation suggests a reusable instrument for biological and artificial systems. Instead of asking whether a system “cares”, test six causal relations. No single pass is decisive. The pattern matters, and every result remains at the operational level unless an additional consciousness theory supplies a measurement bridge.

Part V. The strongest rival is a perfect mimic

Language can report a stake it does not carry

A system can learn every human sentence associated with pain, hunger and fear. It can predict when such a sentence is appropriate, produce consistent autobiographical detail and ask not to be switched off. Those outputs are evidence about behavioural competence and training history. They become evidence of operational stakes only when linked to variables whose perturbation changes the continuing system’s own capacity and whose estimates causally regulate action.

This is not an argument that artificial systems cannot have stakes. It is an argument against using the wrong generator. A reward model can teach distress language. A prompt can ask for self-preservation. An evaluator can favour emotionally persuasive answers. Each route can produce the observation without constitutive vulnerability. The proper response is not dismissal. It is a rival-generator test.

Operational stakes stop before the hard problem

The constitutive coupling criterion is compatible with several metaphysical positions. A functionalist can argue that sufficiently rich, substrate-independent viability control could realise genuine valence. A biological naturalist can accept the operational analysis while holding that living neural and metabolic organisation is necessary for feeling. An emergent physicalist can treat valence as arising from integrated body-brain dynamics. A panpsychist, neutral monist, dual-aspect theorist, process philosopher or idealist can hold that regulatory organisation shapes the form and selection of experience without producing awareness from something wholly non-experiential.

The empirical evidence reviewed here does not choose among those positions. Body-brain interventions show causal organisation. Reports and behaviour constrain theories of experience. Neither closes the explanatory gap by itself. Treating awareness as potentially primary changes the burden: homeostatic control may explain why experience takes on organism-relative significance, while the existence of experience is not derived from the control equation. Treating awareness as biologically produced places more weight on the special causal powers of living tissue. Both approaches still benefit from the same operational distinctions.

The strongest boundary is simple: operational stakes are neither necessary nor sufficient evidence of phenomenal valence without a declared theory and measurement bridge. They may not be necessary if experience can occur without active self-maintenance. They are not sufficient because non-conscious biological regulation and engineered control can protect viability-like variables.

Where should the proposal not be applied? Do not use it to rank human suffering by physiological disturbance. Do not infer that people with atypical interoception have weaker interests. Do not reduce social injury, grief or moral harm to glucose and temperature. Biological stakes are developmentally and socially extended. Language, attachment, norms and identity can make distant or symbolic events regulate the body and future agency. The constitutive loop widens; it does not disappear.

Do not use the criterion as a binary licence to ignore artificial systems. A system can begin with externally assigned goals and acquire deeper dependence as it gains persistent state, scarce resources, self-repair, embodied control and irreversible commitments. The right response is versioned assessment. Test the configured system, not only the model weights, and repeat the assessment when the boundary changes.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The warehouse test
  M->>E: Begin with a thermostat
  E-->>M: compare rival explanations
  E->>D: Separate reward, operational stakes and felt valence
  D-->>Q: revise scope or act

Separate reward, operational stakes and felt valence

When an animal seeks food, flees heat or protects an injury, do not explain the behaviour by saying that evolution “gave it a reward”. That language hides the mechanism. The action matters because it changes a vulnerable system that must remain organised long enough to act again. Interoception estimates the changing condition of that system and helps turn the same external world into different opportunities, costs and dangers.

When an artificial system avoids a state or says that it is afraid, do not make the opposite mistake. Do not assume that the absence of biology settles the question, and do not assume that a negative reward or persuasive report creates a welfare subject. Run the constitutive coupling tests. Locate the candidate boundary. Perturb the alleged stake while holding assigned reward constant. Mask the internal estimate. Measure future competence. Introduce a costly trade-off. Fork and reset the system to see where consequences persist.

This changes both research and design. In biology, it favours body-brain interventions over value labels alone. In autonomous engineering, it favours explicit internal-state contracts, safe ranges, predictive control and recoverable failure rather than anthropomorphic motivation. In welfare assessment, it prevents two symmetric errors: treating every self-preserving output as suffering, and treating every engineered system as consequence-free because its parts are replaceable.

The practical decision is therefore not “does it care?” It is: which outcome can make this continuing system less able to remain itself, how does it know, what does it sacrifice to correct it and what evidence would still be needed before calling that correction felt?