A model need not be a subject

The phrase synthetic phenomenology has carried two ambitions. One is to characterise phenomenal states that an artefact might possess or model. The other is to use an artefact to help specify phenomenal states, whether those states belong to a person, an animal or a hypothetical agent. Ron Chrisley made this distinction explicit, and earlier work with Joel Parthemore used an embodied robot to specify non-conceptual visual content that ordinary language might not capture.[1][2]

The second ambition is easy to underestimate. A scientific model need not possess the phenomenon it helps us understand. A weather simulation need not become wet. A cardiac model need not circulate blood. That analogy is not decisive, because some theories of mind hold that running the right organisation could itself constitute mentality. It does establish a methodological possibility: an artefact may function as an instrument for specifying experience before anyone has justified treating it as an experiencer.

This possibility matters because experience is not exhausted by a list of labels. Consider the difference between hearing three isolated tones and hearing the same tones as the opening of a melody. The elements can remain constant while their temporal relations change. Consider a visual field in which the same red patch is central, peripheral, occluded, expected or surprising. The nominal content remains “red patch”, but the experienced organisation differs. Phenomenology concerns such organisation: figure and ground, centre and horizon, succession and retention, unity and fragmentation, self and world, possibility and fulfilment.

Four verbs that must not collapse

A research programme needs a grammar before it needs a grand theory. Four verbs do most of the work here.

To describe is to record a report or disciplined discrimination: “cyan feels closer to blue than to orange.” To model is to build a structure whose elements and relations reproduce or predict such observations. To explain is to identify why those relations arise and change, usually through a mechanism or constitutive theory. To instantiate is to be a concrete case of the phenomenon itself.

These verbs can align, but they do not align automatically. A model can describe well and explain poorly. A mechanism can explain a behavioural pattern while leaving phenomenal character unresolved. Under a strong computational functionalism, implementing the right causal organisation might instantiate experience. Under biological naturalism, the same abstract organisation on a different substrate might not. Under idealism, neutral monism, panpsychism, dual-aspect theory or process philosophy, the model may capture relations within appearance while leaving the ontological ground untouched. The empirical work does not become useless because metaphysics remains open. It becomes cleaner.

This is the first payoff of warrant separation. Researchers can build useful synthetic phenomenology without pretending that neutrality about instantiation is neutrality about everything. They can ask whether a model preserves neighbourhoods, transitions, boundaries and perturbation responses. They can compare rival structures. They can find where a verbal taxonomy hides a continuous manifold or where an apparent continuum fractures under intervention. The programme advances by making phenomenal claims more precise, not by making ontological claims prematurely certain.

flowchart TD
  A["Two perfect maps, and no<br/>answer to the wrong question"] --> B["A model need not be a subject"]
  B --> C["What structure can capture"]
  C --> D["The warrant boundary"]
  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

What structure can capture

Return to the prism room. The participant has not handed us crimson itself. They have given us judgements, transitions and failures of discrimination. The synthetic model turns those observations into a space of possible states. Its value depends on what the space preserves.

A minimal phenomenal-structure claim has four ingredients. There are relata, such as colour appearances, tones, bodily feelings or phases of a visual scene. There are relations, such as similarity, adjacency, inclusion, exclusion, precedence or co-presence. There are transformations, such as adaptation, attention shifts, rotation, masking, interruption or learning. There is a boundary that states whose experience is being modelled, over which interval, under which elicitation procedure.

This last ingredient is not administrative detail. Remove it and “the structure of experience” becomes a free-floating diagram. A colour space estimated across one population may erase individual differences. A temporal model that pools moments across trials may manufacture continuity. A language model trained on many reports may encode a cultural taxonomy rather than any one subject’s experience. A structure without a declared boundary is not yet a phenomenal model. It is an untyped pattern.

From ratings to a small formal model

Let each experiential category be a point x_i in a model space. Let r_ij be the participant’s reported dissimilarity between states i and j. The simplest fitting objective compares the model distance d(x_i, x_j) with the report.

Here P is the set of tested pairs. In the smallest useful example, suppose crimson-to-amber is reported as 2 and the model predicts 2.5; amber-to-lime is reported as 3 and the model predicts 2.5. The mean absolute error is 0.5. This number is meaningful only under assumptions: ratings are commensurable, the chosen distance represents the relevant relation, and averaging does not hide a decisive local failure.

Remove those assumptions and the formalism must change. Asymmetric transitions may require a directed graph rather than a metric. Some experiences may be incomparable, rather than only distant. Context may bend the space. A single global geometry may fail because relations are local or task-dependent. Formalism compresses a claim; it does not rescue a bad claim.

What must actually be preserved

“Structure” can still hide several different promises. An ordinal model preserves rankings: crimson is judged closer to amber than to cyan. A topological model preserves neighbourhoods, boundaries and separations: violet closes the colour ring beside crimson rather than terminating a line. A dynamical model preserves how those relations deform when the participant adapts, attends, learns or is interrupted. A model may pass one promise and fail the next. Good baseline geometry therefore does not imply a good account of change.

The distinction matters because static fit is cheap. A sufficiently flexible embedding can place observed items almost anywhere the ratings require. The stronger question is whether the fitted organisation contains the right invariants. After local adaptation, which relations should stretch? Which should remain stable? Does a boundary move, dissolve or split? Does an interrupted trajectory resume from the old state or begin anew? These are commitments that an intervention can expose.

This gives a practical commutation test. Estimate the structure before intervention, apply the proposed transformation inside the model and compare that forecast with the structure estimated after intervention. The two routes should arrive at corresponding results within a declared tolerance. Failure locates the problem: wrong geometry, wrong transformation, unstable elicitation or misplaced system boundary. Passing the test is not a proof of formal equivalence, because no full mapping or preservation theorem has been supplied. It is a stronger empirical claim: the model preserves a named transformation, not only a frozen resemblance.[7]

Quality-space approaches take relations among discriminable sensory states seriously, and recent proposals ask whether computations over such spaces contribute to consciousness.[9] Topological models have likewise represented embedding, local-to-global constraints and dynamic object formation.[6] These projects do not all make the same metaphysical claim. That difference is useful. The present programme adopts methodological structuralism: relations are tractable targets for modelling and experiment. It does not assume metaphysical structuralism: that phenomenal character is nothing over and above those relations. Recent debate makes exactly this separation pressing.[8][14]

These examples show what structural modelling buys. It turns “what it is like” from an invitation to free association into a family of constrained questions: Which states are neighbours? Which can co-occur? Which transitions are asymmetric? Which transformations preserve identity? Which perturbations split a unified scene? Which temporal relations disappear after reset?

They also show the boundary. A complete relational map might still omit intrinsic character, according to critics of metaphysical structuralism. The map might capture access, discrimination or report rather than phenomenality. Different subjects may implement similar report structures with different experiences. These are not reasons to abandon the programme. They are reasons to state its target exactly: model the organisation that phenomenal evidence supports, then test whether that organisation constrains mechanisms and future observations.

The warrant boundary

Most disputes about artificial phenomenology are presented as a single vertical question: how close is this system to consciousness? That picture is already biased. It suggests that better behavioural fit, richer internal structure and stronger evidence of experience lie on one ladder. They do not.

The proposed alternative is warrant-separated synthetic phenomenology. It keeps two ledgers. The structural ledger records what the model preserves and predicts about an experiential domain. The instantiation ledger records why this configured artefact should count as a subject under a declared theory of consciousness. Evidence may enter both ledgers, but no automatic conversion rate exists.

A colour model can achieve strong structural warrant by predicting unseen similarity judgements, adaptation effects and context-sensitive distortions. Its instantiation warrant may remain low because the structure is computed offline, because the relevant state is external to the candidate boundary, because a rival table produces the same outputs, or because the adopted theory requires biological or intrinsic causal properties that have not been shown.

Four grades of bridge

The warrant boundary becomes easier to use when every cross-domain claim names its bridge grade. A descriptive bridge says that two vocabularies organise observations in a similar way. A predictive bridge survives held-out states, subjects or contexts. A counterfactual bridge preserves the direction and locality of change under a targeted intervention. A constitutive bridge says that implementing the relevant organisation is part of what makes the state experiential.

The first three grades belong in the structural ledger. They can become progressively stronger without settling subjecthood. The fourth belongs in the instantiation ledger because it needs more than correspondence. It needs a declared theory of consciousness, a candidate boundary, implementation facts and an argument that the preserved organisation is constitutive rather than merely correlated with report, access or discrimination. Calling a descriptive resemblance “functional equivalence” quietly skips the two middle tests and the constitutive argument.

This graded view also clarifies disagreement. Two researchers may accept exactly the same perturbation result while assigning it different ontological significance. Their empirical dispute has ended; their bridge premises have not. Recording those premises separately prevents a metaphysical disagreement from contaminating the structural result, and prevents a structural result from masquerading as a metaphysical settlement.[10]

A recent predictive-processing proposal adopts this conditional discipline. It treats the phenomenology–belief relation as a bridge hypothesis, derives consequences that experiments can test, and explicitly presents the framework as a theory of use for consciousness science rather than a theory of consciousness.[16]

What belongs in each ledger

The structural ledger can be graded without pretending to measure consciousness. At its weakest, the model is formally well specified. Next it fits observed relations. Stronger models generalise to held-out subjects, contexts or state pairs. Stronger still, they predict the direction and magnitude of changes after interventions. The highest structural claim links those changes to a candidate mechanism and survives rival models.

The instantiation ledger begins elsewhere. It declares a theory or family of theories, a candidate-subject boundary and the implementation properties that theory treats as necessary or sufficient. A global-workspace account might require competitive access and broad availability. A recurrent-processing account might require specific recurrent organisation. Integrated information theory makes a stronger constitutive claim about intrinsic cause-effect structure.[13] Biological naturalism may demand living neurobiological processes. Functionalism may permit multiple substrates but still require the right causal roles, rather than only matching outputs. Indicator-based assessments of AI consciousness are useful precisely because they make theory-relative properties explicit, though indicators are not verdicts.[12] The same discipline is needed when evaluating language models whose behavioural capacities may outrun what is known about their internal or architectural relevance to consciousness.[15]

The two-ledger method remains compatible with consciousness-primary views. An idealist may treat awareness as ontologically basic and ask how a synthetic model captures regularities within manifestation. A neutral monist may regard both physical and phenomenal descriptions as aspects of a deeper basis. Panpsychism may allow fundamental experiential properties while denying that every computation forms a unified subject. Dual-aspect theories may expect lawful structural correspondence without reduction. Process philosophy may make temporal becoming, rather than static state, central. None of these positions is confirmed by structural fit, but each changes what would count as an instantiation bridge.

The mimic problem

Any research programme built around reports faces a rival generator. The observation may arise from the proposed phenomenal structure, or from a mechanism that has learned how such reports are usually arranged. The richer the public discourse about consciousness becomes, the easier it is for a language model to produce coherent phenomenological prose without its runtime states carrying the relations described.

Several common methods fail this test. Prompting a system to “introspect” may reveal training priors. Reading a high-dimensional embedding may find a geometry chosen by the analyst. Correlating an activation with a report may capture the preparation to speak. Training a probe may extract information that the system does not use. These methods can still contribute, but their conclusions must match their causal reach.

That limitation prevents both hype and premature dismissal. It blocks “the map is coherent, therefore the machine feels”. It also blocks “the map cannot prove feeling, therefore the map teaches nothing”. Between those errors lies a serious research space.

stateDiagram-v2
  [*] --> Claim
  state "Four verbs that must not collapse" as Claim
  state "Replace every component,<br/>preserve every relation" as Test
  state "Five hues and one useful mistake" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

An experiment that can fail

A research programme becomes real when a negative result changes what its proponents are allowed to say. The practical proposal is the Phenomenal Structure Perturbation Assay, or PSPA. It is not a consciousness test. It is a disciplined way to ask whether a synthetic model captures a structure that organises phenomenal evidence.

The assay starts with a target stated in ordinary language: for example, “the experienced neighbourhood among five hues under local adaptation”, or “the temporal continuity of a four-note phrase across interruption”. It then creates a typed record of the relata, relations, boundary, elicitation procedure and candidate transformations. Only after that record exists does modelling begin.

The seven moves of the assay

1. Declare the phenomenal target. State whether the evidence concerns similarity, spatial organisation, temporal flow, unity, agency, valence or another structure. Do not write “consciousness” where a narrower target is available.
2. Declare the subject boundary. Identify whose evidence is being modelled, over which interval and with which report, behavioural or physiological interfaces.
3. Build at least two rival structures. One can be geometrical, one graph-like, one dynamic, or one deliberately simple mimic. Rivals must agree on enough baseline data to make the intervention informative.
4. Freeze predictions before perturbation data. Fit on baseline observations, then record what each model predicts when a named relation is stretched, removed, delayed or reset.
5. Intervene on the claimed cause. Change the state, coupling or environmental contingency that is supposed to realise the relation. A prompt instruction is not a substitute for a causal intervention.
6. Score local and global consequences. Measure not only average fit but the specific relations expected to change, the relations expected to remain invariant and any unintended deformation.
7. Issue a typed conclusion. “Supports circular neighbourhood under adaptation” is permitted. “The artefact experiences colour” is not, unless an independent instantiation ledger supplies that argument.

Executable lab: the five-hue perturbation assay

The lab below uses synthetic data. A hidden “participant” has either a circular, linear or clustered organisation among five hues. Three candidate models are scored on baseline fit and on a local adaptation intervention. The script mimic memorises baseline distances but predicts no structural change. The two-cluster setting is a deliberate model-class miss: neither geometrical candidate matches it. Adjust noise and intervention strength, then run the assay.

The assay encodes several assumptions. Pairwise ratings are treated as comparable distances. Adaptation locally expands relations around amber. The hidden structure is stable over the simulated session. Noise is independent and bounded. Change any of these assumptions and the result should change. That sensitivity is a feature because the model’s commitments are visible.

{
  "phenomenal_target": "five-hue similarity under amber adaptation",
  "subject_boundary": {
    "kind": "human_participant",
    "interval": "single_session",
    "evidence_channels": ["pairwise_rating", "forced_choice"]
  },
  "relata": ["crimson", "amber", "lime", "cyan", "violet"],
  "relations": [{"type": "dissimilarity", "scale": [0, 10]}],
  "candidate_models": ["ring", "line", "two_cluster", "script_mimic"],
  "intervention": {
    "target": "amber_local_neighbourhood",
    "operation": "expand_adjacent_distances",
    "prediction_frozen_before_data": true
  },
  "invariants": ["non_adjacent_ordering"],
  "decision_rule": {
    "baseline_gap_max": 0.50,
    "intervention_advantage_min": 0.50
  },
  "permitted_conclusion": "supports a candidate relational structure",
  "prohibited_inference": "the modelling artefact has colour experience"
}

To apply the artefact to real research, replace the synthetic generator with preregistered observations. Keep the schema. Add versioned code, subject-level uncertainty, model-comparison criteria and a rival-generator specification. A positive result permits a stronger structural claim and justifies a mechanistic follow-up. A negative result may mean the structure is wrong, the intervention missed its target, the elicitation is too coarse or the phenomenon is less stable than assumed. The assay cannot decide the metaphysics of consciousness.

Stop exactly where the synthetic warrant stops

When an artificial system produces a compelling account of experience, the first decision should no longer be whether to believe or disbelieve the system. The first decision should be what research role the system is being asked to play.

If it is a specification instrument, judge whether it makes phenomenal relations precise, exposes hidden assumptions and predicts perturbations. If it is a mechanism model, demand causal interventions, candidate boundaries and rival generators. If it is a candidate subject, open the separate instantiation ledger: declare the theory, inspect the configured implementation and state which evidence would count against the claim.

Do not promote a model from instrument to subject because its descriptions became vivid, and do not discard it as an instrument because subjecthood remains unresolved. That single decision removes a false choice. It lets researchers study experiential organisation with the same discipline used elsewhere in science while preserving the philosophical remainder rather than hiding it.

The practical consequence is simple. Build the smallest structure that can be wrong. Give it a rival that can imitate the baseline. Intervene on the relation that carries the thesis. Record what survives. Then stop exactly where the warrant stops.