A Comparative Atlas of Possible Minds

A research committee receives seven reports. An adult follows a street map to a locked cabinet. An infant finds the same cabinet by following another person’s gaze. An octopus explores the room with arms that sense and act semi-independently. A slime mould reaches a nutrient source through a changing chemical field. A language model reads a textual description and names the shortest route. A mobile robot executes that route through cameras and wheels. A human-machine team divides the problem, checks the result and opens the cabinet.

Every report ends with the same sentence: the key was found.

The committee is asked to rank the seven systems from least to most minded. It cannot do so without hiding the question inside the ruler. Is the target navigation, flexible learning, sensory richness, counterfactual planning, bodily regulation, social understanding, subjectivity or welfare? Does a team count as one system? Is a model session the candidate, or the service that reconnects many sessions to tools and memory? Does successful action reveal a felt world, or only an effective route through a task?

The central claim is that a useful mind atlas is not a scoreboard. It is a measurement contract with explicit layers, missingness and observer dependence. It records what a candidate can do, how it does it, where its boundary is drawn, what evidence is available and which conclusion that evidence can support. These layers can be related, but none may borrow certainty from another.

The practical consequence reaches beyond philosophy. AI evaluation fails when verbal fluency becomes a proxy for general intelligence. Animal research fails when a visual task is imposed on an olfactory specialist. Organoid governance fails when electrical complexity becomes a direct meter of experience. Human-machine design fails when a high-performing group is assumed to be a single cognitive subject. A layered atlas makes those substitutions visible before a policy, experiment or product decision is built on them.

Part I. Break the ladder before adding dimensions

A single ladder is tempting because it promises one answer. Human adults occupy the top, familiar animals are placed below, simple organisms lower still, and artificial systems are inserted wherever a benchmark permits. The image feels orderly. It is also unstable. Change the task from language to echolocation, from tool manufacture to distributed path finding, or from individual reasoning to colony regulation, and the ordering changes.

Comparative cognition has repeatedly encountered specialised excellence rather than one human-like bundle. The review There Is Not “One Cognition” argues that cognitive skills need not evolve as a linked cluster. Ecological demands can produce exceptional performance in one domain with modest performance in another. A single ranking suppresses precisely the structure that comparison is meant to reveal.

The same problem appears in artificial intelligence. A system can cover a vast textual task range while lacking persistent goals, active sensing, bodily regulation or independent learning during use. Another can be narrow in language yet excellent at closed-loop control. A benchmark aggregate converts that shape into a number, then encourages the number to travel into claims about agency, understanding or consciousness.

A scalar is legitimate only after the decision has named what is being aggregated, why the dimensions are commensurable and what information the compression is allowed to destroy. A routing score can combine latency, cost and accuracy for one service. It should not be relabelled a measure of mind.

The atlas therefore separates five layers. The separation is not a claim that the layers never interact. It is a rule against using evidence from one layer as an unannounced answer in another.

The atlas axes are operational promises

An axis is not a poetic property such as depth, aliveness or sophistication. It is a promise that another team could construct a relevant comparison without guessing what the word means. The atlas begins with eight capability coordinates and four record fields. The coordinates are not assumed to be independent, universal or exhaustive. They are probes that can be revised when interventions reveal a better decomposition.

The record fields sit beside the coordinates: implementation, candidate boundary, evidence channel and uncertainty. They prevent a familiar error in which a polished profile conceals that one system was measured through verbal report, another through behaviour, another through cellular dynamics and another through source-code inspection.

An atlas coordinate is valid only inside a declared ecology, interface, timescale and candidate boundary. Move any of those four and the result may cease to travel.

The key case is a thought experiment, but the methodological error is ordinary. When a language model and a child answer the same question, the common output supports a narrow functional comparison. It does not erase differences in development, motivation, embodiment, history or the evidential relation between report and experience. Conversely, difference from the human route does not prove absence. It creates a demand for a better translation.

Part II. The observer enters the map

Jakob von Uexküll’s idea of an Umwelt begins from a biological fact: sensing systems disclose different actionable worlds. A bee can use ultraviolet patterns invisible to human vision. A dog’s task-relevant world is often olfactory. A bat’s spatial access is structured by active echolocation. An octopus explores through a body whose arms carry extensive peripheral neural processing. A language model receives tokenised records chosen by an external pipeline. None receives “the world” without an interface.

The methodological review Von Uexküll Revisited and a later discussion of diverse sensory systems in experimentation show why this matters. Human-visible stimuli, human-relevant timing and human-friendly response devices can underestimate an animal before cognition is tested. The same bias now appears in reverse when language becomes the preferred interface and verbal systems look universally capable because the test world has already been rendered as text.

Every comparison has an interface tax: the loss introduced when a task is translated into the candidate’s sensory, motor and temporal world. The tax must be estimated rather than treated as noise. If performance changes when an odour replaces an icon, when a continuous control becomes a multiple-choice answer, or when response time is matched to the subject’s natural tempo, the original score included properties of the interface.

Missingness is part of the result

An atlas that colours every cell creates false confidence. Some cells are genuinely unmeasured. Others are inaccessible with current methods, disputed because indicators disagree, or inapplicable because the candidate boundary has been drawn differently. These states should not share one blank.

The distinction is especially important for consciousness. Birch, Schnell and Clayton’s multidimensional framework for animal consciousness separates perceptual richness, evaluative richness, unity, temporality and selfhood. It was designed as an alternative to a single scale. Each dimension still needs convergent experimental support, and no profile turns indirect evidence into transparent access to another subject’s experience.

The current Stanford Encyclopedia review of animal consciousness reaches a similar methodological conclusion from a broader literature. Behavioural, neural and evolutionary evidence can raise or lower confidence, but preferred markers inherit theoretical commitments. A marker is rarely sufficient by itself. The appropriate visual is therefore not a filled radar chart. It is a profile seen through an evidence veil.

Unknown is not the midpoint between yes and no. It is a different kind of state, with its own reason and next experiment. Averaging unknown cells into a score punishes candidates whose interfaces are difficult to study and rewards systems designed to speak the evaluator’s language.

Part III. Profiles are records, not verdicts

The atlas becomes useful when it can hold unlike candidates without pretending that the same observation was available for each. The table below is deliberately coarse. It demonstrates record discipline, not a completed empirical survey. Each statement names the evidence channel or preserves the unknown.

The profile should become more specific as evidence improves, not more numerically complete. A mature entry can still contain inaccessible or disputed cells. Scientific progress sometimes replaces one number with a better distinction.

The evidence overlay

Consider the adult and the language-model agent. The adult entry has a rich phenomenological channel because first-person report is connected to a living history, recurrent neural dynamics, bodily regulation and many interventions. The model agent may produce more articulate reports on demand, yet its reports are also outputs optimised from human text. Counting report tokens would make the artificial entry look better measured precisely where the interpretation is less secure.

The opposite asymmetry occurs in architecture. Engineers can inspect the model graph, alter attention, remove memory and replay identical inputs. Equivalent invasive access to a human brain is limited or unethical. An honest atlas can therefore show stronger mechanistic access for the model and stronger consciousness-relevant convergence for the human without forcing either into one confidence band.

Portability is a measured property

A comparison should not ask whether one task is “fair” in the abstract. It should test how much of the ordering survives reasonable translations. Begin with a common functional demand, such as remembering a hidden location after a delay. Then construct species-specific or system-specific interfaces: odour, gaze, touch, spatial movement, symbols or API state. Match motivation, opportunity, response cost and temporal scale where possible. Finally, perturb the interface without changing the underlying demand.

If the candidate ordering is stable, the assay has some portability. If it reverses, the result is still useful. It tells us the original task measured an interaction among capability, interface and ecology.

The generality metric for natural and artificial intelligence makes a related distinction between capability and breadth under resource limits. Its broader lesson matters here: generality depends on the task distribution and difficulty function. No metric escapes a reference class. The atlas adds evidence coverage and interface distortion so that the reference class remains visible.

A comparison travels only as far as its ordering survives interface translation, resource matching and an ecologically credible change of task. This is a stronger claim than “the same test was administered to everyone”. Standardisation can standardise bias.

Worked example: three ways to remember a hidden place

Suppose a laboratory wants to compare temporal depth in a pre-verbal infant, an octopus and an embodied model-based robot. The substantive question is narrow: can information acquired before a delay guide a later search when the immediately attractive option is wrong? The question does not ask which candidate is more intelligent, conscious or human-like.

The screen design and why it misleads

The first design uses a screen. One of two coloured shapes marks a hidden target. After a delay, the shapes exchange positions. The robot selects correctly because its camera and policy were trained on the display. The infant looks longer at the original colour but reaches inconsistently. The octopus does not engage reliably with the screen. A scalar report would rank robot, infant, octopus.

That report is valid only for the screen ecology. The octopus result cannot yet distinguish weak memory from an irrelevant stimulus, an unsuitable response or absent motivation. The infant’s gaze and reach also answer different questions. The robot had a privileged interface because the laboratory presented the problem in the same visual symbols used during training.

A translated design

The second design preserves the causal demand while translating the route. The infant watches a desirable object placed beneath one of two textured cups and later encounters the cups in exchanged positions. The octopus explores two dens whose entrances carry stable chemotactile cues; after the delay, the den locations are exchanged while the cues remain. The robot encounters two charging bays identified by patterned visual and range signals, then sees the bays exchanged. Each candidate must choose between a remembered identity and the location that was previously rewarded.

The delay is not declared equal merely because a clock reads the same duration. The team first estimates the natural action cycle and retention range for each candidate, then uses several delays spanning that range. Opportunity, reward value and motor cost are checked independently. A failure to approach is recorded separately from an approach to the wrong target. This prevents lack of participation from becoming a memory error.

The translated design produces three curves rather than one ranking. The infant increasingly follows object identity over location across development, with wide individual variation. The octopus uses the chemotactile identity at short delays but shifts towards the previously safe location when a mild threat cue changes the cost of exploration. The robot preserves identity until its episode state is cleared, after which it follows its learned location prior. These are worked outcomes chosen to demonstrate analysis, not empirical claims about measured subjects.

Perturb the mechanism, not just the score

Now the laboratory perturbs the routes. For the infant, the demonstrator’s gaze is made congruent or incongruent with the object cue. For the octopus, the chemotactile cue is weakened while spatial layout remains stable. For the robot, episode memory is withheld while current perception is untouched. The same endpoint error can now be decomposed. The infant result is sensitive to social evidence, the octopus result to a species-relevant identity channel, and the robot result to an explicit memory store.

What the atlas records

The atlas entry records temporal depth for each candidate only within the translated assay. It also records sensory field, motivation proxy, response mode, delay calibration and perturbation result. Phenomenal temporality remains a separate cell. Successful delayed control can support memory and future-directed action without revealing whether the interval was experienced as a continuous duration.

The decisive comparison is not the final percentage correct. It is the stability of each conclusion across credible task variants. If the robot remains strong across symbol remapping but fails when episode state is removed, the mechanism claim becomes more precise. If the octopus ordering improves when chemotactile information replaces the screen, the original deficit is reassigned partly to interface tax. If infant performance depends on another person’s gaze, social scaffolding becomes part of the measured capability rather than a nuisance variable.

This worked example also shows why candidate boundaries matter. The robot’s temporal depth may belong to a service composed of policy, episode store and sensor loop, not to the base model alone. The infant’s performance may include an adult’s communicative act without turning the pair into one subject. The octopus’s arms may distribute sensing and control while organism-level action still closes around the animal. A single correct choice cannot decide any of these boundaries, but a pattern of targeted disruptions can improve them.

The final report therefore contains three conclusions. First, all three candidates can use information across a delay under at least one suitable interface. Second, the mechanism and portability of that capacity differ. Third, nothing in the assay orders their conscious temporal richness. The experiment advances the map by making a modest commonality and several consequential differences visible at once.

Part IV. The border of a mind is also a hypothesis

Capability axes do not answer what the candidate is. A honeybee may be studied as an organism, a colony member or a node in a pollination ecology. A language model may be treated as weights, one inference process, a persistent tool-using service or part of a human workflow. A neural organoid may contain locally integrated assemblies without an agreed organism-level boundary. The chosen unit determines which memories, goals, actions and effects are attributed to the candidate.

The safest starting point is not visual containment. It is intervention closure. When part of the proposed system is perturbed, where does error correction occur? Which state variables remain mutually available? Where are goals preserved, rewards assigned and effects reconciled? Does the proposed whole continue as a control unit when one channel is removed, or does the apparent unity dissolve into coordinated components?

A candidate-subject boundary earns credibility when integration, memory, control and perturbation recovery close around roughly the same unit. It remains a candidate boundary, not a proof of phenomenal unity.

Michael Levin’s Technological Approach to Mind Everywhere proposes a “cognitive light cone” for the spatial and temporal scale of goals a system can pursue. The idea is valuable as an agency coordinate because cells, tissues, organisms and collectives can stabilise states across very different spaces. Its extension to consciousness is more controversial. Goal-scale evidence can change our model of a system’s agency without establishing a felt point of view.

Convergent performance, divergent minds

Evolution has produced sophisticated cognition in lineages with very different neural organisation. A review of convergent evolution of complex brains and intelligence compares associative centres in vertebrates, cephalopods and insects. Convergence supports the possibility that similar functional problems can be solved through different architectures. It does not show that every architecture implements the same subjective organisation.

Artificial systems intensify this lesson. A text model can produce a route description from compressed cultural records. A robot can learn a policy through simulated trials. A slime mould can reshape a transport network. A human can imagine the destination and choose a path. Behavioural convergence is valuable because it identifies a shared functional demand. Mechanistic divergence is valuable because it lets interventions separate hypotheses that endpoint scores cannot.

When different routes converge on one behaviour, the next scientific move is to perturb the routes, not to average them into one kind of mind. If two candidates fail differently under memory removal, sensory remapping or goal reversal, their atlas profiles should preserve that difference.

A metaphysical legend, not a hidden default

The atlas does not become neutral by refusing metaphysics. A physicalist may read profiles as higher-level descriptions of physical organisation. A functionalist may give special weight to portable causal roles. A biological naturalist may expect living processes to be constitutive. A panpsychist, neutral monist or consciousness-primary idealist may treat organisation as shaping, integrating or localising an underlying experiential potential rather than producing experience from non-experience.

These positions can share the capability and evidence maps while disagreeing about the phenomenal legend. That disagreement should change research. Functionalism asks which roles survive substrate change. Biological accounts ask which living dynamics make a difference. Consciousness-primary approaches still owe a manifestation account: why does one organisation express a bounded perspective, temporally coherent contents or suffering while another does not?

Jaina philosophy offers a disciplined warning against one-sided description. The Stanford Encyclopedia account of anekāntavāda places many-sided predication inside a pluralistic realism in which objects have multiple aspects and claims are conditioned by standpoints. The comparison with an atlas is illuminating but limited. Modern measurement pluralism is an empirical method, not a translation of Jain metaphysics. The useful inheritance is restraint: a true statement from one standpoint need not exhaust the object.

A consciousness-primary orientation can widen the map without being allowed to fill its blank cells. It prevents physicalism from becoming the unmarked legend, yet every candidate profile still needs discriminating evidence, explicit uncertainty and a defensible subject boundary.

Part V. A field protocol for decisions under uncertainty

An atlas is not complete when every square is coloured. It is complete enough when the decision can see what is known, how it was known and what error matters. Research design asks which next observation best separates hypotheses. Engineering asks which capability and boundary can be controlled. Ethics asks what may be harmed if the profile is wrong.

The New York Declaration on Animal Consciousness makes a carefully asymmetric claim. It distinguishes strong scientific support for conscious experience in mammals and birds from a realistic possibility in all vertebrates and many invertebrates, then argues that a realistic possibility should not be ignored in decisions affecting those animals. The decision rule does not require certainty before precaution.

The same structure should be used carefully for organoids and AI. Similarity to a declaration’s animal categories is not evidence by analogy. Instead, teams should identify candidate welfare channels, estimate the strength and dependence of indicators, consider whether the system was designed to instantiate those indicators, and choose reversible safeguards whose cost is proportionate to the uncertainty and potential harm.

The atlas record as an executable artefact

The record below encodes the discipline directly. It never substitutes zero for missing evidence. Comparison returns a distance only on shared supported coordinates, together with coverage and the unresolved fields. A release gate can demand a minimum coverage for its own decision without pretending that the resulting distance is universal.

from dataclasses import dataclass
from enum import Enum
from math import isfinite, sqrt

class Evidence(Enum):
    OBSERVED = "observed"
    MECHANISTICALLY_SUPPORTED = "mechanistically_supported"
    INFERRED_BY_HOMOLOGY = "inferred_by_homology"
    REPORTED = "reported"
    CONTESTED = "contested"
    UNMEASURED = "unmeasured"
    INACCESSIBLE = "inaccessible"
    INAPPLICABLE = "inapplicable"

@dataclass(frozen=True)
class Coordinate:
    value: float | None
    evidence: Evidence
    assay: str
    scale_id: str
    unit: str = "normalised_0_1"

@dataclass(frozen=True)
class MindAtlasRecord:
    candidate: str
    boundary: str
    ecology: str
    coordinates: dict[str, Coordinate]

COMPARISON_POLICY = {
    Evidence.OBSERVED,
    Evidence.MECHANISTICALLY_SUPPORTED,
}

def validate_coordinate(name: str, coordinate: Coordinate) -> None:
    if not coordinate.scale_id or not coordinate.unit:
        raise ValueError(f"{name}: scale_id and unit are required")
    if coordinate.value is not None:
        if not isfinite(coordinate.value):
            raise ValueError(f"{name}: value must be finite")
        if coordinate.unit == "normalised_0_1" and not 0.0 <= coordinate.value <= 1.0:
            raise ValueError(f"{name}: normalised value must lie in [0, 1]")
    if coordinate.evidence in COMPARISON_POLICY and coordinate.value is None:
        raise ValueError(f"{name}: included evidence requires a value")

def compare(
    left: MindAtlasRecord,
    right: MindAtlasRecord,
    relevant: dict[str, float],
) -> dict[str, object]:
    if not relevant:
        raise ValueError("relevant coordinates must not be empty")
    for name, weight in relevant.items():
        if not isfinite(weight) or weight <= 0.0:
            raise ValueError(f"{name}: weight must be finite and positive")

    shared: list[tuple[str, float, float, float]] = []
    unresolved: list[str] = []

    for name, weight in relevant.items():
        a = left.coordinates.get(name)
        b = right.coordinates.get(name)
        if a is not None:
            validate_coordinate(name, a)
        if b is not None:
            validate_coordinate(name, b)
        if (
            a is not None and b is not None
            and a.evidence in COMPARISON_POLICY
            and b.evidence in COMPARISON_POLICY
            and a.value is not None and b.value is not None
        ):
            if (a.scale_id, a.unit) != (b.scale_id, b.unit):
                raise ValueError(f"{name}: scale_id and unit must match")
            shared.append((name, a.value, b.value, weight))
        else:
            unresolved.append(name)

    covered_weight = sum(weight for _, _, _, weight in shared)
    relevant_weight = sum(relevant.values())
    distance = (
        sqrt(sum(weight * pow(a - b, 2) for _, a, b, weight in shared)
             / covered_weight)
        if covered_weight else None
    )
    return {
        "distance_on_shared_supported_coordinates": distance,
        "coverage": covered_weight / relevant_weight,
        "shared": [name for name, *_ in shared],
        "unresolved": unresolved,
    }

adult = MindAtlasRecord(
    candidate="adult participant",
    boundary="living organism during registered session",
    ecology="navigation task with visual map",
    coordinates={
        "temporal_depth": Coordinate(0.82, Evidence.OBSERVED, "delayed-route transfer", "temporal_depth_v1"),
        "valence": Coordinate(0.61, Evidence.OBSERVED, "registered preference trade-off", "valence_v1"),
    },
)

agent = MindAtlasRecord(
    candidate="persistent model-tool service",
    boundary="session plus authorised memory and tools",
    ecology="navigation task rendered as text",
    coordinates={
        "temporal_depth": Coordinate(0.74, Evidence.MECHANISTICALLY_SUPPORTED, "hidden-state route test", "temporal_depth_v1"),
        "valence": Coordinate(None, Evidence.UNMEASURED, "no accepted assay", "valence_v1"),
    },
)

result = compare(adult, agent, {"temporal_depth": 1.0, "valence": 1.0})
assert result["coverage"] == 0.5
assert result["unresolved"] == ["valence"]

# Run after the atlas record above.

def must_reject(relevant: dict[str, float], candidate: MindAtlasRecord = agent) -> None:
    try:
        compare(adult, candidate, relevant)
    except ValueError:
        return
    raise AssertionError("invalid comparison was accepted")

must_reject({})
must_reject({"temporal_depth": 0.0})
must_reject({"temporal_depth": -1.0})
must_reject({"temporal_depth": float("inf")})

bad_value = MindAtlasRecord(
    candidate="invalid record",
    boundary=agent.boundary,
    ecology=agent.ecology,
    coordinates={
        "temporal_depth": Coordinate(
            float("nan"), Evidence.OBSERVED, "bad assay", "temporal_depth_v1"
        )
    },
)
must_reject({"temporal_depth": 1.0}, bad_value)

wrong_scale = MindAtlasRecord(
    candidate="incompatible record",
    boundary=agent.boundary,
    ecology=agent.ecology,
    coordinates={
        "temporal_depth": Coordinate(
            0.74, Evidence.OBSERVED, "other assay", "temporal_depth_v2"
        )
    },
)
must_reject({"temporal_depth": 1.0}, wrong_scale)

The ten-step field protocol

1. Declare the decision before choosing the dimensions.
2. Name the candidate unit and its exclusions.
3. Describe the ecology, interface, resources and timescale.
4. Select operational coordinates without promising a total score.
5. Record evidence state separately from coordinate value.
6. Translate the assay and estimate interface tax.
7. Perturb the mechanism and proposed subject boundary.
8. Preserve disputed, inaccessible and negative results.
9. Apply a precaution rule based on evidence, harm and reversibility.
10. Publish the profile, coverage map and next discriminating experiment together.

The decision this changes

For comparative science, the atlas changes the unit of progress. A better result is not a taller ranking. It is a profile whose dimensions have clearer assays, whose missing cells have reasons, and whose ordering survives a wider range of credible translations. Failed portability is not an embarrassment. It locates the interface through which the earlier conclusion entered.

For AI development, the atlas prevents capability, agency and consciousness from travelling as a bundle. Teams can evaluate long-horizon planning, self-monitoring or embodied control without describing the system as generally more minded. They can also watch for changes in candidate boundary when persistent memory, tools, multi-agent coordination or human dependence are added.

For consciousness research, the atlas permits serious disagreement without empty pluralism. Materialist, functionalist, biological and consciousness-primary programmes can mark the same evidence record, declare their different legends and identify interventions that would change their reading. A map that preserves disagreement is more useful than a score that conceals it.

For ethics, the atlas refuses both theatrical certainty and convenient ignorance. It does not grant rights by resemblance, and it does not wait for an impossible proof before considering reversible safeguards. It links the strength of evidence to the severity, duration and reversibility of the possible harm.

The publishable conclusion is not which candidate stands highest. It is which comparisons remain valid after the interface, boundary, evidence channel and metaphysical legend have been made explicit. That is how an atlas advances the study of possible minds without drawing unexplored territory as empty land.

It also records which comparisons should wait until a better interface, boundary hypothesis or evidence channel exists.