The coordinate science keeps dropping

Place three bells around a seated listener. One is in front, one behind and one above. A microphone array can recover frequency, amplitude and source location in room coordinates. The listener does not receive those coordinates as a neutral list. A sound arrives from here, behind the shoulder, above the head, approaching or receding. The body supplies an origin. Head movement changes the relation. Attention makes one stream figure and the others background.

Now rotate the listener while leaving every bell fixed. The room description is unchanged. The listener-centred organisation changes. A theory that predicts only the room has preserved the object but lost the transformation. This is the smallest useful intuition for subjectivity: subjectivity adds an index to the description, identifying the system for which, from where and through which capacities a state is present.

Perspective includes more than spatial viewpoint. It includes sensory format, bodily orientation, temporal scale, current concern, attentional selection and the history through which a signal becomes discriminable. None of these alone is consciousness. Together they specify a transformation from world states into a candidate field of appearance.

The hidden assumption in many objective accounts is that information is exhausted by what a state is about. Yet two systems can carry information about the same wall while organising distance, direction, salience and action differently. “Both represent the wall” is therefore too coarse. It behaves like saying two maps are equivalent because both contain London, while ignoring projection, scale, orientation and route cost.

This yields the paper’s central proposal, the Perspective Preservation Principle: a candidate explanation of experience should predict how the relational organisation of appearance changes when perspective-bearing variables are intervened upon, while relevant world and task variables are held fixed.

The principle is deliberately weaker than a theory of consciousness. It does not say which physical, biological or fundamental property makes experience exist. It says what an explanation must not erase. A model may be physicalist, functionalist, biological, emergentist, idealist, neutral-monist, panpsychist, dual-aspect or process-based. If it speaks about a specific experience, it still owes a mapping from changed conditions to changed appearance.

Three descriptions must therefore be kept apart. A world description states what is present in coordinates that do not privilege the observer. A presentation description states how differences are organised for a system with a particular body, sensory route and history. A record description states what that system later says, selects or does. They can agree without being interchangeable. The microphone can locate a bell correctly, the listener can hear it behind the left shoulder, and the response sheet can record “rear-left”. Each description answers a different question.

This separation prevents two symmetrical mistakes. The first treats presentation as decorative because world and behaviour are already modelled. The second treats the participant’s words as the presentation itself. In the rotated-listener case, the decisive datum is not the phrase “behind me”. It is the lawful reorganisation produced when bodily orientation changes while the bells do not. The words are one way to sample that reorganisation. A pointing response, head turn or pairwise similarity judgement may sample it differently. Subjectivity becomes scientifically tractable when the invariant is a transformation pattern, not a privileged sentence.

Objectivity need not mean deleting the origin. In navigation, a camera-centred direction becomes useful because the system knows how it transforms when the camera turns. Removing the camera model would not produce a purer account; it would make the measurements uninterpretable. Experience research faces a similar choice. When viewpoint is merely a nuisance, researchers can average across it. When the target is how a scene is present, that averaging removes the dependent variable. The proper objective move is to specify the origin, vary it and search for relations that remain stable across observers or response modes.

This also explains why first-person character adds data without becoming an infallible authority. A participant may misname a direction, forget a transition or accept a suggestion. Across repeated rotations, however, their pointing, discrimination and description may still exhibit a coherent transformation. The public result is not private access borrowed by the scientist. It is an intersubjectively testable constraint on any model claiming to explain the episode. Objectivity is earned through controlled translation between perspectives, not by pretending that perspective was never present.

flowchart TD
  A["The coordinate science keeps<br/>dropping"] --> B["A report is a second event"]
  B --> C["Make perspective an intervention"]
  C --> D["What survives the strongest<br/>objections"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#D95545,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

A report is a second event

Ask someone why they chose the left face rather than the right. Their answer may sound direct: warmer eyes, a more trustworthy expression, a feeling of familiarity. Yet the causal process that produced the choice may include influences the person did not detect. Nisbett and Wilson assembled evidence that people often lack direct access to higher-order processes and instead generate plausible explanations from available cues and cultural theories.⁴

That result is sometimes overextended. Failure to report the cause of a choice does not show that a person cannot report whether one image looked brighter, whether attention narrowed, or whether a pressure began in the chest and spread. The target matters. “What appeared?” and “What produced the appearance?” are different questions with different warrants.

Jonathan Schooler sharpened a related distinction between experience and meta-consciousness, the explicit re-representation of what is being experienced. He described temporal dissociations, such as suddenly noticing that one’s mind had been wandering, and translation dissociations, where verbal reflection changes or misdescribes a subtle state.⁵ The report is therefore not a passive readout. It is a new cognitive act.

Empirical work supports calibration rather than blanket trust or blanket dismissal. Fleming and colleagues measured how well confidence discriminated correct from incorrect perceptual decisions and found substantial individual differences associated with anterior prefrontal structure, though the association was correlational.⁶ Maniscalco and Lau formalised metacognitive sensitivity so it could be separated from first-order task performance and response bias.⁷ A confident participant can be poorly calibrated; an accurate participant can use confidence cautiously. “Introspection” is not one capacity.

Micro-phenomenological methods try to improve access by returning participants to a singular episode, asking non-leading “how” questions and separating description from explanation. Petitmengin and colleagues reported that guided description greatly increased detection of covertly swapped choices in a choice-blindness task.⁸ A later self-inquiry replication using a standardised form did not reproduce that detection advantage, although it elicited richer process descriptions.⁹ The disagreement is informative: the instrument, training, interviewer and order of operations are part of the measurement.

Neurophenomenology takes the further step of seeking reciprocal constraints between first-person structure and third-person dynamics. Varela proposed this as a research programme rather than a quick metaphysical solution.² Lutz and colleagues used trial-by-trial experiential categories to guide analysis of neural synchrony during a visual task, showing how first-person distinctions can alter what is sought in physiological data.³ Later practical work has stressed a trade-off: thin reports scale and align with repeated measurements, while thick interviews preserve detail but are slower and harder to standardise.¹⁰

Make perspective an intervention

The debate changes when “subjectivity” stops naming an inaccessible remainder and starts naming a family of controlled transformations. The simplest design holds the scene and task constant, changes one perspective-bearing variable, and asks what pattern should change if the candidate explanation is right.

The formal version needs only enough notation to prevent an equivocation. Let the candidate experiential organisation at time t be:

The crucial separation is between Φ, the proposed route of presentation, and Γ, the report channel. If a delayed probe changes the answer, the change may belong to Γ rather than Φ. If rotating the body systematically reorders near and far while report timing is fixed, the change is evidence about Φ. Without this separation, every report error is misread as an error of experience, and every fluent report is misread as transparent access.

The Perspective Preservation Audit compares relational structures rather than assuming a natural unit of experience. Participants can sort episodes by similarity, make forced comparisons, mark transition order or use a trained scale. These responses form a dissimilarity matrix. A candidate model generates its own matrix across the same conditions. Rank agreement asks whether the model gets the ordering right. Normalised stress asks how much geometric mismatch remains after allowing one overall scale factor.

The smallest numerical example uses three pairwise reports. The perspective model’s residuals are 0.05 in each pair. The stress calculation is:

Executable lab: Perspective Preservation Audit

The browser lab below contains synthetic reports for four conditions: baseline, shifted viewpoint, sensory remapping and narrowed attention. Select which variables the candidate model is allowed to use. The lab computes Spearman rank agreement and normalised stress. Increase report-channel distortion to see how a measurement path can weaken an otherwise good bridge.

A positive result permits a narrower claim than “the model explains consciousness”. It permits the claim that the selected variables preserve a specified relational pattern across specified interventions. That is enough to guide further experiments and reject coarser accounts. A negative result can mean the candidate model is wrong, the subject boundary is wrong, the report instrument is unstable, or the chosen grain is mismatched. The design must contain controls that separate these possibilities.

The same logic can be run outside the browser with the reference implementation below. It uses only Python’s standard library and contains explicit invariants: the observed and predicted vectors must have equal length, distances must be non-negative and a zero-variance prediction cannot earn rank agreement.

from __future__ import annotations

from dataclasses import dataclass
from itertools import combinations
from math import sqrt
from typing import Iterable, Sequence


@dataclass(frozen=True)
class Condition:
    name: str
    viewpoint: float
    sensory_code: float
    attention: float


def average_ranks(values: Sequence[float]) -> list[float]:
    indexed = sorted(enumerate(values), key=lambda item: item[1])
    ranks = [0.0] * len(values)
    cursor = 0
    while cursor < len(indexed):
        end = cursor
        while end + 1 < len(indexed) and indexed[end + 1][1] == indexed[cursor][1]:
            end += 1
        rank = (cursor + end + 2) / 2.0
        for position in range(cursor, end + 1):
            ranks[indexed[position][0]] = rank
        cursor = end + 1
    return ranks


def pearson(left: Sequence[float], right: Sequence[float]) -> float:
    left_mean = sum(left) / len(left)
    right_mean = sum(right) / len(right)
    left_c = [value - left_mean for value in left]
    right_c = [value - right_mean for value in right]
    denominator = sqrt(sum(v * v for v in left_c) * sum(v * v for v in right_c))
    return 0.0 if denominator == 0 else sum(a * b for a, b in zip(left_c, right_c)) / denominator


def spearman(left: Sequence[float], right: Sequence[float]) -> float:
    return pearson(average_ranks(left), average_ranks(right))


def pairwise_distances(
    conditions: Sequence[Condition],
    include: tuple[bool, bool, bool],
    weights: tuple[float, float, float] = (0.80, 0.90, 0.65),
) -> list[float]:
    distances: list[float] = []
    for first, second in combinations(conditions, 2):
        differences = (
            first.viewpoint - second.viewpoint,
            first.sensory_code - second.sensory_code,
            first.attention - second.attention,
        )
        squared = sum(
            (difference * weight) ** 2
            for difference, weight, active in zip(differences, weights, include)
            if active
        )
        distances.append(sqrt(squared))
    return distances


def normalised_stress(observed: Sequence[float], predicted: Sequence[float]) -> float:
    if len(observed) != len(predicted) or not observed:
        raise ValueError("Observed and predicted distances must have equal, non-zero length")
    if any(value < 0 for value in (*observed, *predicted)):
        raise ValueError("Dissimilarities must be non-negative")
    energy = sum(value * value for value in predicted)
    if energy == 0:
        return 1.0
    scale = sum(o * p for o, p in zip(observed, predicted)) / energy
    residual = sum((o - scale * p) ** 2 for o, p in zip(observed, predicted))
    return sqrt(residual / sum(o * o for o in observed))


def audit(observed: Sequence[float], predicted: Sequence[float]) -> tuple[float, float, bool]:
    rank_agreement = spearman(observed, predicted)
    stress = normalised_stress(observed, predicted)
    passed = rank_agreement >= 0.80 and stress <= 0.25
    return rank_agreement, stress, passed


conditions = [
    Condition("baseline", 0.00, 0.00, 0.00),
    Condition("view-shift", 0.35, 0.00, 0.05),
    Condition("sensor-remap", 0.05, 1.00, 0.15),
    Condition("attention-narrow", 0.00, 0.15, 0.75),
]
observed = [0.30, 0.92, 0.52, 0.88, 0.56, 0.81]
predicted = pairwise_distances(conditions, include=(True, True, True))
print(audit(observed, predicted))
# Expected, approximately: (0.94, 0.04, True)

Code caption: the artefact tests whether a candidate set of perspective variables preserves synthetic first-person relations under intervention. It cannot establish that the candidate system is conscious or that the variables constitute experience.

stateDiagram-v2
  [*] --> Claim
  state "The sonar transplant" as Claim
  state "Perspective-indexed transformation" as Test
  state "The description-limit exercise" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

What survives the strongest objections

The first objection says that subjective occurrence is private, while scientific data must be public. There is a useful distinction here. The occurrence is available from the participant’s perspective; the recorded response, choice and intervention history are publicly inspectable. Science does not gain direct possession of the occurrence. It constructs and tests a measurement bridge. Public method does not require pretending that every target property is itself public in the same way as the record.

A stronger functionalist or illusionist objection says that once every disposition, report and control process is explained, no further fact remains. Dennett’s heterophenomenology treats reported worlds as data without granting their descriptions automatic authority.¹² The audit is compatible with that stance. It makes the functional target harder, because the theory must reproduce intervention-sensitive perspective structure, rather than only generate plausible consciousness talk. Whether anything ontologically additional remains is not decided by the fit.

The opposite objection says that relational structure leaves out intrinsic quality. A perfect geometry of colour similarities may still fail to convey red to someone without vision. This is precisely the boundary Nagel identified. The Perspective Preservation Principle does not turn structural description into acquaintance. It proposes an objective phenomenology in a modest sense: concepts that preserve some organisation of experience across subjects who may not share the experience itself. A residual may remain even after every relation is modelled.

Physicalism in its strongest form need not deny first-person character. It claims that the perspective-indexed organisation is wholly constituted by physical organisation, even when the useful description is relational rather than microscopic. Functionalism adds that the relevant causal roles may survive a change of substrate. Biological naturalism denies that role preservation is automatically enough, because the causal powers of living tissue may be constitutive. Emergentism predicts that the relevant organisation appears only when interactions reach a particular grain or regime. The audit cannot settle these views from output matching. It forces each view to name what a replacement must preserve and where a break should appear.

Consciousness-primary and dual-aspect positions reverse or soften the usual explanatory direction. Idealism can treat the objective account as an abstraction within experience rather than its ontological foundation. Neutral monism and dual-aspect theory can treat subjective and objective structures as two organisations of a more basic reality. Panpsychism grants an experiential aspect broadly but still owes an account of combination and the individuation of this perspective. Process philosophy makes events and relations primary rather than enduring substances. None gains a free prediction. Each must still explain why rotating a body, remapping a sense or delaying a probe changes this pattern rather than another. The audit is ontologically non-decisive but explanatorily demanding.

Report-based studies also risk measuring the neural and cognitive machinery required to answer. No-report paradigms attempt to separate activity related to experience from activity related to report, while introducing their own inferential dependence on proxies.¹¹ The sensible conclusion is not “reports are contaminated, so remove them” or “proxies are indirect, so ignore them”. It is to use report and no-report paths as mutual controls, each with a declared bridge.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The coordinate science keeps dropping
  M->>E: The sonar transplant
  E-->>M: compare rival explanations
  E->>D: Put perspective back into the measurement
  D-->>Q: revise scope or act

Put perspective back into the measurement

Do not ask, “Should I trust introspection?” The question is too blunt to guide research. Ask instead: Which property is this response supposed to measure, which transformation should it track, and which rival process could produce the same answer?

If the claim concerns hidden causes, treat the participant’s explanation as a hypothesis. If it concerns confidence, measure calibration separately from accuracy. If it concerns relational appearance, vary the perspective-bearing variable and require the pattern to replicate across response modes. If the report channel changes, model that change rather than silently attributing it to experience.

This decision makes subjectivity neither sovereign nor disposable. It becomes a constrained source of evidence. The first-person character of experience adds the coordinate that objective description tends to remove. Science advances by putting that coordinate back in, then testing whether the proposed transformation survives contact with reports, behaviour, physiology and counterfactual change.

Glossary

First-person character

The way an episode is organised for a situated subject, including its centre, modality, salience and temporal relations.

Meta-consciousness

Explicitly noticing or re-representing an ongoing experience, which may occur later than the experience.

Perspective-bearing variable

A feature such as sensory format, bodily orientation, viewpoint, attention or history that may transform presentation while the world stays fixed.

Measurement bridge

A declared relation connecting reports or choices to a target experiential property, with assumptions and failure conditions.

Rank agreement

How well a model preserves the ordering of reported dissimilarities across conditions.

Normalised stress

The residual geometric mismatch between reported and predicted relations after one overall scale alignment.

References

1. Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450. DOI.
2. Varela, F. J. (1996). Neurophenomenology: A methodological remedy for the hard problem. Journal of Consciousness Studies, 3(4), 330–349. Record.
3. Lutz, A., Lachaux, J.-P., Martinerie, J., & Varela, F. J. (2002). Guiding the study of brain dynamics by using first-person data: Synchrony patterns correlate with ongoing conscious states during a simple visual task. Proceedings of the National Academy of Sciences, 99(3), 1586–1591. DOI; full text.
4. Nisbett, R. E., & Wilson, T. D. (1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84(3), 231–259. DOI.
5. Schooler, J. W. (2002). Re-representing consciousness: Dissociations between experience and meta-consciousness. Trends in Cognitive Sciences, 6(8), 339–344. DOI.
6. Fleming, S. M., Weil, R. S., Nagy, Z., Dolan, R. J., & Rees, G. (2010). Relating introspective accuracy to individual differences in brain structure. Science, 329(5998), 1541–1543. DOI.
7. Maniscalco, B., & Lau, H. (2012). A signal detection theoretic approach for estimating metacognitive sensitivity from confidence ratings. Consciousness and Cognition, 21(1), 422–430. DOI.
8. Petitmengin, C., Remillieux, A., Cahour, B., & Carter-Thomas, S. (2013). A gap in Nisbett and Wilson’s findings? A first-person access to our cognitive processes. Consciousness and Cognition, 22(2), 654–669. DOI.
9. Sparby, T. (2021). First-person access to decision-making using micro-phenomenological self-inquiry. Scandinavian Journal of Psychology, 63(2), 117–127. DOI.
10. Berkovich-Ohana, A., Dor-Ziderman, Y., Trautwein, F.-M., Schweitzer, Y., Nave, O., Fulder, S., & Ataria, Y. (2020). The Hitchhiker’s Guide to neurophenomenology: The case of studying self boundaries with meditators. Frontiers in Psychology, 11, 1680. DOI.
11. Tsuchiya, N., Wilke, M., Frässle, S., & Lamme, V. A. F. (2015). No-report paradigms: Extracting the true neural correlates of consciousness. Trends in Cognitive Sciences, 19(12), 757–770. DOI.
12. Dennett, D. C. (2003). Who’s on first? Heterophenomenology explained. Journal of Consciousness Studies, 10(9–10), 19–30. Author archive.