The same box asks different questions

A cognitive test is usually described by what the experimenter built: two cups, a delay, a mirror, a string, a screen. The animal encounters something else. It encounters colours it can or cannot separate, odours that leak or vanish, movements it can or cannot perform, rewards with unequal value, and timing that fits or conflicts with its mode of action.

Jakob von Uexküll’s operational insight was that an animal’s environment is not simply the inventory around it. It is the closed relation between what can become a meaningful perceptual cue and what the organism can do in response. He called the perceptual side the Merkwelt and the effect side the Wirkwelt.¹ This paper uses that relation without adopting his wider metaphysics or his rejection of Darwinian explanation. It also does not claim access to what another animal’s world feels like.

The experimental consequence is sharper than the familiar advice to make tasks “ecologically valid”. A naturalistic task can still be a bad comparison. A captive-born animal may master an artificial symbol, and a wild-looking apparatus may contain uncontrolled scent or effort cues. What matters is whether the animal receives the evidence needed for the target operation, can express the resulting choice, and cannot succeed through an easier route.

The hidden assumption inside “same test”

When researchers insist on the same cups, colours, distances and response rule, they usually protect one valuable property: procedural standardisation. Yet they can silently assume that the interface is neutral. Under that assumption, species differences in the score are credited to memory, inhibition, numerical judgement or self-recognition, while sensory and motor differences are treated as random noise.

That assumption fails when interface costs are systematic. A colour pair may be easily separable for one species and nearly metameric for another. A five-second delay may be short relative to one animal’s sampling cycle and long relative to another’s. Reaching through a slot may be a familiar action for a primate and an awkward, fear-inducing one for a different body plan. A food reward can produce eager engagement in one individual and satiety or caution in another.

Reviews of non-human cognitive testing have therefore emphasised that reliability and validity cannot be read directly from task labels.² Comparative psychometrics adds a further warning: tasks intended to measure the same construct do not always correlate as a single ability should.³ A poor result can be a genuine limit, an interface failure, a state effect, or a mismatch between the task and the proposed construct.

One score can carry four different claims

A score begins as a behavioural fact: under a particular configuration, this animal selected one option on a stated fraction of trials. The next step is a construct claim: the selection depended on numerical discrimination, delayed retention, inhibition, causal inference or another proposed capacity. A stronger step is a mechanism claim: the animal solved the problem through a specific representation or computation. The strongest step concerns experience: what, if anything, the animal perceived or understood from its own point of view.

These steps do not rise automatically with accuracy. A dog can follow a scent without inferring exclusion. A fish can perform a mirror-contingent sequence without the experiment deciding whether the relevant construct is self-recognition, body checking or a learned relation. An animal can also possess the target capacity while failing to express it because the interface blocks the cue, the action or the consequence.

The interface audit governs the bridge from behaviour to construct. It asks whether the target route was available and whether a rival route could produce the same score. It does not identify the internal algorithm merely by removing obvious confounds, and it does not infer phenomenology from successful performance. Those later claims need their own dissociations, interventions and theory.

This separation explains why reliability is necessary but insufficient. Repeating a biased task can yield a stable species difference. A task can also produce individually reliable scores while several nominally similar tasks fail to converge on one construct. Comparative psychometrics treats that pattern as evidence about the measurement model, not as an inconvenience to be averaged away.³ The first question after a score is therefore not “Which species is better?” It is “Which level of claim did this configuration actually earn?”

A recent honeybee study addressed a related debate by recalculating stimulus properties across spatial frequencies relevant to bee vision. The analysis weakened a proposed spatial-frequency account of the bees’ choices and supported a biology-aligned approach to stimulus design.⁸ The important lesson is not that every bee performs human-like arithmetic. It is that a confound has to be defined in the animal’s perceptual variables, not ours.

flowchart TD
  A["The elephant at the wrong screen"] --> B["The same box asks different<br/>questions"]
  B --> C["Preserve the causal demand,<br/>not the furniture"]
  C --> D["Biology translates the question"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#B96D00,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Preserve the causal demand, not the furniture

Strip a task down until only the relation that makes the correct answer correct remains. That relation is the causal demand. Cups, screens, odours, symbols and gestures are implementations. They may change across species, provided the intervention structure does not.

Consider a delayed quantity choice. The causal demand might be: observe which of two sets contains more items, retain that relation across a delay, then choose the corresponding location after non-numerical cues have been controlled. The demand is not “look at black dots”, “wait with hands still”, or “tap the left panel”. Those are interface decisions.

A translation is valid when the same counterfactual makes the answer change. Swap the quantities and the correct side must swap. Hold quantity constant while changing total area and the answer must not change. Increase the retention interval and only the memory burden should change, not odour concentration or effort. Remove the target relation and performance should fall towards the level predicted by remaining cues.

Five invariants for a translated test

A practical translation can be tested against five invariants. They are more useful than asking whether two apparatuses look alike.

Information invariant. Each subject receives evidence sufficient to distinguish the task states, at a comparable signal-to-noise level established by control trials. A visual quantity task and an olfactory quantity task can satisfy this invariant even though the physical cues differ.

Counterfactual invariant. The same conceptual intervention changes the correct answer. Swap larger and smaller quantities, remove the delay, reverse the reward contingency, or break the mirror contingency. The translated task must respond to those interventions in the corresponding way.

Action invariant. The response requires a choice, not an unrelated feat. A species may indicate selection by approach, gaze, trunk reach, peck, body orientation or a trained symbolic response. What must remain is the commitment structure: selecting one option should exclude or clearly distinguish the other.

Consequence invariant. Rewards, losses and opportunities preserve their decision role. “Wait for the better option” is not preserved when one species faces a valued live prey item and another faces two nearly equal pellets. Preference has to be measured, not presumed.

Uncertainty invariant. The subject should face the same ambiguity relevant to the hypothesis. If one interface provides a salient local cue that perfectly identifies the correct cup, it has subsidised the decision. If another obscures the target cue below detection threshold, it has taxed it.

Translation must be constrained before the target result

Changing an interface after a species performs poorly creates an obvious danger: the researcher can keep adapting the task until the preferred result appears. A valid translation is therefore not a rescue operation. It is a preregistered family of implementations whose membership is fixed by the invariants above and by independent calibration data.

Begin with controls that do not require the target operation. Estimate the smallest colour contrast, odour difference or spatial separation the subject can discriminate in the apparatus. Establish that each response option can be expressed with similar reliability. Measure reward preference, satiation and latency. Check that symbols or locations have stable learned meanings. These measurements set the operating range before the main trials begin. They do not guarantee equality between species, but they reveal when one interface lies outside a subject’s usable range.

Then create at least two implementations of the same causal demand. A quantity relation might be expressed through visible items and through sequential sounds, or through two odour sources whose intensity and reward mass are independently controlled. A delay demand might be tested through waiting in place and through returning after an occlusion. The critical test is not whether the two scores are numerically identical. It is whether the same interventions produce the same qualitative fingerprint: reversing the target relation reverses choice, removing the delay removes the delay effect, and orthogonal variation of non-target cues does not control the answer.

Successful translation is a pattern of invariance across interventions, not a claim that two sensory worlds are equivalent. If only one implementation yields the effect, that divergence is information. It may expose a hidden tax, a shortcut, or a genuine boundary on the construct. The researcher should model these rivals rather than select the friendlier task.

Computational validity makes a related move by asking whether tasks share relevant information-processing structure rather than superficial form.⁴ The proposal here adds a practical break condition: if an interface translation changes which counterfactual determines the correct response, it has left the comparison class. A scent cue can replace a visual cue while preserving quantity selection. It cannot replace quantity itself with reward magnitude and still count as the same demand.

The boundary matters because a cognitive test is a coupled system: organism, apparatus, experimenter and scoring rule. Moving a response from limb movement to gaze does not automatically make the test better. It can introduce new calibration errors or observer degrees of freedom. The correct question is whether the new coupling removes irrelevant demand without adding a route that predicts the answer.

Biology translates the question

The umwelt is not a licence to explain every result after the fact. It becomes scientific when it generates controls, predicts which redesign will matter, and specifies what should remain unchanged.

Worked example 2: configured biological systems

Quantity through a trunk, delay through a chamber

Plotnik and colleagues presented Asian elephants with food quantities that could be discriminated through smell alone.⁷ This does not prove that elephants possess the same numerical representation as humans. It does something more disciplined: it demonstrates that a quantity comparison can be presented through a sensory channel central to the animal’s behaviour. It also creates new controls. Researchers can vary ratio, total mass, odour concentration, container position and sampling time to ask which property drives choice.

In another body plan, Schnell and colleagues adapted a delay-maintenance task for common cuttlefish. The animals learned visual symbols indicating that a chamber would open immediately, after a delay, or not at all. They chose between prey types whose preferences had been established, and some maintained delays from roughly 50 to 130 seconds. Longer delay performance covaried with reversal-learning performance in that study.⁹

The defensible conclusion is behavioural: under this configured task, cuttlefish could forgo an immediately available, less preferred option while waiting for a preferred option. Calling it “the marshmallow test” is memorable but can hide translation work. The symbols, prey, chamber mechanics and stopping rule were not decorative accommodations. They constituted the route by which delay maintenance became testable for a cephalopod.

Ecological tuning is not a free pass

Cleaner wrasse offer a more difficult case. In their natural “market”, visitor clients leave if not served, while resident clients wait. Laboratory versions replace clients with coloured plates carrying food. A computational analysis by Prat, Bshary and Lotem distinguished versions that could be solved through simpler chaining from a more natural sequence structure requiring configural chunking in their model. The useful tuning parameter depended on ecological frequencies, and under some conditions even an optimal parameter could not produce the expected preference.¹⁰

This is not evidence that every cleaner fish literally implements the model. It demonstrates a design principle: the distribution of trials can be part of the interface. A task can preserve visible options while changing how often combinations occur, thereby rewarding a different learning rule. “Same choice” is not “same problem” when the temporal statistics have changed.

Worked example 4: failure and edge case

The mirror shows the boundary of the method

The mark test illustrates both the necessity and the insufficiency of ecological alignment. Cleaner wrasse feed on ectoparasites that can appear as small dark marks. In a follow-up study, mirror-experienced fish scraped their throats when given a brown mark visible through the mirror, whereas blue and green marks did not elicit the same response. The study added controls for mirror experience, mark colour, tactile irritation, viewing another marked fish and mirror relocation.¹¹

Those controls increase confidence that the configured fish performed the behavioural sequence specified by the mark protocol. They do not compel one interpretation of self-awareness. The authors themselves frame the result as pressure either to expand the distribution of mirror self-recognition or to revise what the mark test establishes. A species-appropriate interface can validate performance without settling the ontology of the capacity.

stateDiagram-v2
  [*] --> Claim
  state "The hidden assumption inside<br/>“same test”" as Claim
  state "One score can carry four<br/>different claims" as Test
  state "The ultraviolet number board" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Build an interface-tax audit

The audit is not a score for animal intelligence. It is a decision instrument for the experiment. It asks whether a positive result is bypassable and whether a negative result had a fair route to become visible.

Step 1: write the demand as an intervention

Do not begin with “the animal must press the triangle”. Begin with the smallest causal statement. For a quantity task: changing only numerosity should change the correct choice; changing area, density, odour mass or side should not. For a delay task: increasing delay should alter the retention or waiting burden while reward preference and option availability remain fixed. For a mirror task: mark-directed action should depend jointly on body location, mirror contingency and mark relevance.

Step 2: draw the target route and every plausible bypass

List the evidence the animal must detect, the transformation it must perform, the decision it must express, and the event the apparatus records. Then add any variable that could predict the registered answer without the target operation: scent, local enhancement, experimenter gaze, colour luminance, motor side bias, reward noise, repeated-trial statistics or an observer’s discretionary scoring.

Step 3: estimate taxes with independent controls

Each interface dimension needs a control whose success does not require the target capacity. A perceptual control asks whether the subject discriminates the relevant cues. An action control asks whether it can reliably express each option. A motivation control establishes reward preference and engagement. A timing control measures sampling and response latency. A learning-history control checks whether the symbols and apparatus have become stable rather than merely familiar to one site or individual.

A failed target trial is interpretable only when the controls show that the route was open. Passing all controls does not prove the capacity. It removes a class of explanations for failure.

Step 4: keep taxes and shortcuts separate

The audit records five suppressive taxes: perceptual access, action reliability, motivational engagement, temporal fit and learning-history fit. It records shortcut exposure separately. Combining them into one “ecological validity” score would be dangerous because a high tax and a strong shortcut can cancel in the observed accuracy while leaving the result uninterpretable.

Step 5: predeclare the inference gate

The same observed score can demand opposite decisions depending on the control record. A near-chance result after the subject fails a basic discrimination control is not a weak negative; it is a failed measurement attempt. The same result after low-tax controls, adequate engagement and a verified response route can lower confidence in the capacity, but only for the tested range and configuration. A high score with an intact shortcut remains positive behaviour while leaving the target construct underdetermined.

Write those outcomes before the target data are visible. A useful gate has at least five exits: bounded positive, positive but underdetermined, informative negative, uninformative negative, and redesign. Each exit should specify the allowed sentence in the paper. For example, “subjects selected the larger olfactory quantity under controlled odour presentation” is narrower and more defensible than “elephants understand number”. “Performance did not exceed chance at delays from 20 to 60 seconds after all interface controls passed” is stronger than “the species lacks memory”.

The null result also needs a destination. If the target effect disappears when a shortcut is broken, the study has identified the route supporting the original behaviour. If it disappears after an interface change that was supposed to preserve all five invariants, either the translation failed or the construct is implementation-sensitive. If it survives both interface replacement and shortcut removal, confidence in the bounded construct claim increases even though the mechanism may remain open.

Preregistration protects translation from becoming outcome-dependent accommodation. Meta-research in comparative cognition has documented how claim strength and statistical practice can outrun what individual designs establish.¹² An inference gate makes the restraint operational. It forces the team to decide in advance which evidence changes the animal claim, which evidence changes the task claim, and which result changes neither because the measurement route was never secured.

Core algorithm

The browser lab runs the following compact model. Copy it into any modern JavaScript console or Node.js runtime. The seeded simulation makes repeated runs reproducible for the same configuration.

function simulateAudit(config) {
  const clamp = (x, lo = 0, hi = 1) => Math.max(lo, Math.min(hi, x));
  const channel = [
    config.perception,
    config.action,
    config.engagement,
    config.timing
  ].map(value => clamp(value)).reduce((a, b) => a * b, 1);

  const targetP = 0.5 + (clamp(config.competence, 0.5, 1) - 0.5) * channel;
  const shortcutShare = clamp(config.shortcutShare);
  const observedP =
    (1 - shortcutShare) * targetP +
    shortcutShare * clamp(config.shortcutAccuracy, 0.5, 1);

  let seed = config.seed ?? 125;
  const random = () => {
    seed = (1664525 * seed + 1013904223) % 4294967296;
    return seed / 4294967296;
  };

  let correct = 0;
  for (let i = 0; i < config.trials; i += 1) {
    if (random() < observedP) correct += 1;
  }

  const gate = shortcutShare > 0.15
    ? "positive result underdetermined"
    : channel < 0.55
      ? "negative result may be interface-limited"
      : "bounded interpretation possible after real controls";

  return { channel, interfaceTax: 1 - channel, observedP, correct, gate };
}

What a positive simulation permits: it shows that the configured assumptions can produce an interpretable score when channel quality is high and shortcuts are low. What a negative simulation means: low observed accuracy can arise without lowering latent competence. What it cannot establish: the real value of any parameter, the presence of a representation, subjective experience, or a species ranking.

type EvidenceRef = {
  protocolId: string;
  measure: string;
  result: number | string;
  uncertainty: string;
};

type InterfaceAudit = {
  causalDemand: {
    targetOperation: string;
    interventionInvariants: string[];
    prohibitedShortcuts: string[];
  };
  speciesInterface: {
    perceptualChannel: string;
    choiceExpression: string;
    consequence: string;
    timing: string;
    trainingHistory: string;
  };
  controls: {
    perception: EvidenceRef;
    action: EvidenceRef;
    motivation: EvidenceRef;
    timing: EvidenceRef;
    learningHistory: EvidenceRef;
  };
  shortcutBreaks: EvidenceRef[];
  inferenceGate:
    | "bounded-positive"
    | "positive-underdetermined"
    | "negative-informative"
    | "negative-uninformative"
    | "redesign";
};

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The elephant at the wrong screen
  M->>E: The hidden assumption inside “same test”
  E-->>M: compare rival explanations
  E->>D: Translate the interface before ranking the minds
  D-->>Q: revise scope or act

Translate the interface before ranking the minds

The practical decision is no longer “Should every species receive the same task?” It is “Which relations must remain fixed, which surfaces must change, and what evidence shows that the translation did not add or remove a route to success?”

For a new study, the smallest useful implementation is modest. Choose one causal demand. Build one perceptual control, one action control, one motivation control and one timing control. Name the strongest shortcut and break it experimentally. Predeclare how each pattern changes the claim. This is more valuable than adding another species to an apparatus whose interface has never been audited.

For an existing literature, the audit changes how negative and positive findings are weighed. A negative result with failed sensory controls should lower confidence in the task, not automatically in the animal. A positive result that disappears when a local cue is removed should lower confidence in the proposed mechanism, not in the reality of the behaviour. A result that survives translated interfaces and counterfactual interventions deserves more weight than one repeated only through identical equipment.

Large collaborative projects can then standardise the right things. ManyPrimates showed that multi-site teams can preregister protocols and assemble broad samples, first with 176 individuals from 12 species at 11 sites and later with a delayed-response dataset of 421 non-human primates across 41 species.⁵⁶ The next step is not to abandon standardisation. It is to standardise the causal demand, control battery, data record and inference gate while allowing justified interface translations.

There is also a risk of over-accommodation. Researchers can redesign a task until a preferred species passes, then call the final interface “natural”. The antidote is prospective prediction. Before collecting target data, state which control is expected to fail, which redesign should repair it, which target effect should remain, and which result would count against the interface explanation. The audit is useful only when it can be wrong.

The shift is small in wording and large in consequence. We stop asking animals to enter a human apparatus and then treating their difficulty as a property of their minds. We build a translation aperture, measure what it costs, and preserve the causal question through it. Only then does comparison begin.