The stimulus that was not the same

Imagine a rat facing two lights. On some trials the rat's own movement triggers the light. On other trials the same light appears according to a replayed schedule. The photons can be matched, yet the perceptual problem is not necessarily the same.

The familiar intuition says that perception begins when energy reaches a receptor. The eye receives a pattern, the brain processes it, and an experience is produced. Movement may add context, but the sensory input remains the causal centre. This picture hides an assumption: the environment supplies a ready-made stimulus whose meaning is independent of the organism's way of meeting it.

Active sensing experiments disturb that assumption. Concha-Miranda and colleagues trained rats on a visual localisation task under active and passive conditions. Performance improved when the stimulus was coupled to self-initiated movement, and the advantage decayed when the timing between movement and stimulus was displaced.[6] This does not prove enactivism. Neuromodulation, motor preparation, temporal expectation and sensory gain are all live mechanistic explanations. It does show something narrower and decisive: a matched stimulus sequence need not be a matched perceptual situation.

There are three increasingly strong claims hiding inside the phrase “organism and world co-determine perception”. The weakest is historical dependence: previous action changes the organism, so a later input meets a different nervous system. The next is online dependence: current movement changes the sampled signal and the gain with which it is processed. The strongest is constitutive dependence: remove the organised coupling and the cognitive process itself is no longer the same kind of process, even if a sensory sequence is replayed.

Those claims require different evidence. A before-and-after learning study can support historical dependence. A movement-locked perturbation can support online dependence. A yoked control, boundary transfer or closed-loop disruption is needed for the stronger constitutive proposal. Co-determination becomes an empirical claim only when the design states which dependency is at stake and what removal would break it. Without that distinction, “enactive” can mean anything from ordinary feedback to a wholesale metaphysics.

Varela, Thompson and Rosch called this orientation enaction. The term is often weakened into the harmless slogan that cognition is embodied. The stronger proposal is relational: an autonomous organism brings forth a meaningful world through a history of sensorimotor and regulatory engagement.[2] “Brings forth” does not mean inventing mountains or photons. It means that the world encountered as climbable, threatening, edible, distant or graspable depends on the organism's capacities, needs and path through it.

Di Paolo sharpened the biological side of this picture with adaptivity: an organism does more than maintain itself; it regulates its tendencies relative to conditions of viability.[5] A concentration, temperature or sound is not intrinsically good or bad. It acquires a valence within the organisation that can continue or fail. Enaction therefore links perception to a history of self-maintenance, bodily possibility and action.

This formulation also prevents a common mistake. A thermostat is sensitive to temperature, but its norms are assigned by a designer and maintained by a larger technical system. A bacterium's organisation continually recreates the conditions of its own persistence. Whether that difference is sufficient for mind remains contested, but it supplies a principled question: where do the norms that organise sensing and action come from? Neurophenomenology inherits that question when it studies a human participant. Hunger, fatigue, posture, readiness and affect are not nuisance variables by definition. They may be part of the living norm against which the scene is encountered.

The word “world” must also be handled carefully. Enaction does not deny a reality that resists the organism. A cliff can kill regardless of belief. What is enacted is the field of relevance and possibility within that resistant reality. The same ledge can be a seat, an obstacle, a nesting place or nothing salient at all. The environment constrains the organism, while the organism's form of life selects which constraints become a lived world.

This is the first bridge to neurophenomenology. Experience is not an ornamental report placed beside brain data. It belongs to a trajectory in which a living body acts, samples and regulates. The challenge is to make that trajectory experimentally tractable without declaring in advance that one level explains all the others.

flowchart LR
  A["The stimulus that was not the same"] --> B["The response began before the<br/>trial"]
  B --> C["Constraint must be able to<br/>disappoint"]
  C --> D["Where the braid snaps"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#7B4CD8,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

The response began before the trial

Most experiments draw a vertical line at stimulus onset. Before the line lies baseline. After it lies response. Varela's method asks whether the supposed baseline already contains the history that makes the response intelligible.

In a well-known study, Antoine Lutz, Jean-Philippe Lachaux, Jacques Martinerie and Varela trained four participants on a visual illusion and on reporting the cognitive context of each trial.[4] The participants did more than choose a confidence number after the image. Through preparation and repeated inspection, they developed categories for how they arrived at the stimulus, including forms of steady readiness, fragmented readiness and unreadiness.

The categories were then used to partition single trials. Behavioural and electroencephalographic patterns differed across some of those partitions. Prepared trials showed different prestimulus synchrony and subsequent responses from unprepared trials. The categories were not identical across people, and the sample was tiny. The point was not a universal neural signature. It was methodological: what had looked like unexplained trial-to-trial variance became structured when the participant's lived trajectory entered the analysis.

The procedure matters more than the headline. Participants initiated trials themselves, viewed a three-dimensional stereogram and reported how the perceptual episode unfolded. Their descriptions were stabilised into participant-specific phenomenal clusters rather than forced into one universal taxonomy. The reported cluster assignment overlapped reaction-time grouping more often than a chance baseline, and the EEG analysis found distinct patterns before and after stimulation for some experiential states.[4]

That sequence illustrates both the promise and the debt. The first-person work supplied a new way to stratify trials. The physiology then made the categories answerable to an external pattern. Yet the exploratory workflow, four-participant sample and participant-specific categories limit generalisation. A modern replication should freeze the bridge on one block or cohort, test it on another and report how often the phenomenological coding itself can be reproduced. The study is a proof of methodological possibility, not a validated diagnostic instrument.

The phrase first-person data can mislead. An experience is first-personal, but a scientific record is produced through a procedure. The participant notices, stabilises, describes and sometimes revisits an episode. An interviewer may help distinguish sequence, bodily tone, attentional movement and implicit assumptions. The resulting description can be coded, compared across occasions and challenged by other evidence. It is not the experience itself, and it is not automatically accurate.

Three positions are therefore present. The participant has lived access to the episode, including features no camera can directly record. The interviewer has relational access to the participant's language, hesitations and return to the episode. The analyst has access to the coded record, task events and physiology. None occupies a view from nowhere. Reliability comes from coordinating these partial positions while keeping their transformations visible.

This also explains why a questionnaire score and a phenomenological description are not interchangeable. A score compresses variation into a predefined dimension. A description may reveal that the dimension was badly chosen. Conversely, an interview can proliferate distinctions that do not survive repetition. The method should move between discovery and compression: rich elicitation proposes structure, standardised trials test it, and failed structure is removed.

Practical neurophenomenology therefore faces a resolution problem. A thin report such as “seen or unseen” is easy to repeat but may discard the structure that matters. A thick interview may reveal temporal and affective organisation but become harder to standardise, code and align with millisecond physiology. Berkovich-Ohana and colleagues describe this as a continuum rather than a choice between proper science and unstructured testimony.[9]

Neurophenomenology is sometimes presented as adding subjective colour to objective science. That formulation is too weak. The ambition is to change what counts as a good experimental variable. A participant may identify a transition the external protocol did not anticipate. Physiology may then reveal that the description collapses distinct trajectories. Each side can become more precise because the other side resists it.

Constraint must be able to disappoint

A report and an EEG trace can line up beautifully while neither constrains the other. Correlation is a relation between observations. Reciprocal constraint is a rule for revising explanations.

Suppose a study collects a vivid interview and a brain scan. The scan is analysed in many ways until one pattern fits the interview. The interview is then retold using the language suggested by the pattern. The two descriptions converge, but the convergence is guaranteed by the workflow. No possible outcome was permitted to break the bridge.

This motivates a sharper distinction. A domain constrains another only when it reduces the set of explanations that remain admissible before the held-out result is seen. Phenomenology can constrain physiology by specifying a temporal transition, state distinction or intervention-sensitive prediction. Physiology can constrain phenomenology by revealing that one reported category contains two repeatable dynamic regimes, or that a claimed transition occurs too late to explain the event.

From alignment to reciprocal revision

The proposed Reciprocal-Constraint Ladder separates five study designs that are often grouped under one label. Grade 0 collects first-person and third-person material without linking them. Grade 1 reports an association. Grade 2 lets one domain guide selection or interpretation in the other. Grade 3 specifies revision rules in both directions. Grade 4 adds an intervention that changes the live organism-world coupling and tests the bidirectional predictions on held-out data.

A small formal device can clarify what “constraint” means. Let H_N be a finite set of candidate neurophysiological models and H_E a finite set of experiential models. If an independently specified experiential distinction reduces the viable neural models from eight to two, its model-exclusion gain is:

The same bookkeeping can run from physiology to experience. A conservative reciprocal quantity is the weaker direction, R = min(G(E → N), G(N → E)). If either direction excludes nothing, the reciprocal gain is zero. The equation is not a consciousness metric. It assumes a finite, honestly specified candidate set and a held-out test. When those assumptions fail, use the ladder qualitatively rather than manufacturing a number.

The Braided Constraint Protocol

The smallest useful implementation combines six moves. First, familiarise participants with the target episode before recording. Second, derive a compact experiential vocabulary without access to the physiological outcome. Third, manipulate a causal property of the coupling, such as active control versus yoked replay. Fourth, preregister bridge hypotheses that state what each domain would force the other to revise. Fifth, test on held-out trials with independent coding. Sixth, repeat the cycle only after recording the failed predictions.

Each move blocks a different shortcut. Familiarisation reduces the chance that the report vocabulary is invented under scanner pressure. Blind coding prevents neural patterns from steering the experiential labels. The coupling intervention distinguishes a live loop from a replayed sequence. A bridge hypothesis prevents unlimited analysis search. Holdout data separate discovery from confirmation. Symmetric revision prevents physiology from being treated as the final court while experience serves only as annotation.

The workbench tests study logic, not empirical truth. A positive Grade 4 result permits a narrow statement about the configured manipulation, measures and population. A negative result means that the proposed bridge failed under those conditions. The failure may reflect a wrong experiential distinction, an insensitive physiological measure, a weak intervention or an absent coupling. The design must name these alternatives before deciding which model to abandon.

from dataclasses import dataclass
from math import log2

@dataclass(frozen=True)
class Study:
    neural_before: int
    neural_after: int
    experiential_before: int
    experiential_after: int
    predeclared: bool
    independent_coding: bool
    intervention: bool
    holdout: bool
    symmetric_revision: bool


def gain(before: int, after: int) -> float:
    if before < 1 or after < 1 or after > before:
        raise ValueError("Require 1 <= after <= before")
    return log2(before / after)


def assess(study: Study) -> dict:
    e_to_n = gain(study.neural_before, study.neural_after)
    n_to_e = gain(study.experiential_before, study.experiential_after)
    reciprocal = min(e_to_n, n_to_e)

    if e_to_n == 0 and n_to_e == 0:
        grade = 1
    elif not study.predeclared or not study.holdout:
        grade = 1
    elif e_to_n == 0 or n_to_e == 0:
        grade = 2
    elif not study.independent_coding or not study.symmetric_revision:
        grade = 2
    elif not study.intervention:
        grade = 3
    else:
        grade = 4

    return {
        "grade": grade,
        "experience_to_neural_bits": round(e_to_n, 2),
        "neural_to_experience_bits": round(n_to_e, 2),
        "reciprocal_bottleneck_bits": round(reciprocal, 2),
        "cannot_establish": [
            "consciousness",
            "metaphysical enactivism",
            "universal neural signatures",
        ],
    }

stateDiagram-v2
  [*] --> Claim
  state "Two trajectories through one<br/>enacted niche" as Claim
  state "The active trial and its yoked<br/>twin" as Test
  state "One trial is a trajectory, not<br/>an instant" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Where the braid snaps

The elegance of a reciprocal story can become its greatest risk. Once experience, brain and world are placed in one loop, almost any co-variation can be narrated as mutual determination.

The first boundary is causal underdetermination. Active sensing can improve performance because movement predicts stimulus timing, raises arousal, changes sensory gain or supplies an efference copy. These mechanisms are compatible with an enactive description but do not by themselves show that organism-world coupling partly constitutes cognition. The active-yoked contrast locates a dependency. Further interventions must separate the rivals.

One useful sequence is to remove the proposed causes one at a time. Preserve movement but randomise action-outcome timing to test temporal prediction. Preserve timing but remove control to test agency over sampling. Preserve control but clamp sensory gain to test neuromodulatory explanations. Match pupil size and muscle activity to reduce arousal and motor confounds. No single contrast is decisive, but the pattern of failures can locate which part of the loop does the work.

The constitutive claim carries an additional burden. Showing that a process depends on external coupling does not yet show that the external element is part of the cognitive process rather than a crucial cause of it. A dialysis machine is causally necessary for some patients without becoming part of every cognitive act. The stronger enactive inference needs counterfactual integration: changes in the external relation should alter the internal and experiential organisation in the specific, law-like way predicted by the joint model.

The second boundary is report construction. Training may improve discrimination of experience, but it may also change attention, expectation and the experience being studied. Meditators or phenomenologically trained participants are not transparent windows. They are differently trained organisms. This can be an advantage when the target is a trainable mode of awareness, and a limitation when the claim concerns ordinary, untrained experience.

A careful programme should therefore separate at least three questions. Can a participant report the distinction reliably after training? Did training create or amplify the distinction? Does the same organisation occur in untrained participants even when they cannot name it? These are different experiments. Treating expertise as pure access collapses them and risks projecting a learned phenomenological ontology onto everyone.

The third boundary is neural interpretation. Varela's work emphasised large-scale, transient coordination, and synchrony remains an important candidate mechanism for integration.[7] Yet a synchrony difference is not consciousness, and it may reflect motor preparation, attention, signal quality or common drive. The Lutz study showed a way to discover state-conditioned dynamics. It did not establish a universal neural grammar of lived experience.

Formal bridges can help and still overreach

Recent work in computational neurophenomenology proposes a third strand: a formal generative model linking trained descriptions and neurophysiology. Sandved-Smith and colleagues argue that such models may represent not only what is experienced but how an object is given, while explicitly denying that the model alone explains why experience exists.[12] This is a serious proposal because it exposes bridge assumptions that prose can hide.

It is also contested. A model can fit both domains because it is flexible, not because it preserves their structure. Active-inference vocabulary may import representational commitments that some enactivists reject. A latent parameter called “precision” is not automatically felt clarity, neural gain or attentional confidence. Each mapping must state its source structure, target structure, preserved relation, break condition and discriminating experiment.

A useful formal bridge should do more than reproduce both curves. It should predict a cross-domain intervention that was not used to fit the model. For example, changing action contingency should alter a parameter trajectory, a physiological measure and a reported transition in a specified order. If many parameter settings reproduce the same triple, the bridge is underidentified. If the model predicts physiology but not the timing of experience, it is a neural model with phenomenological decoration. If it predicts the report only after fitting participant-specific labels, it has not yet generalised.

The strongest alternative explanation may be plural rather than reductive. Experience, physiology and action could be different measurements of a deeper process that none of them causes. Reciprocal constraint would still improve the model, but the arrows should not be read as direct causation between reports and neurons. The protocol is compatible with that possibility because it tests organised dependency, not a predetermined ontology.

Froese and Sykes argue that much neurophenomenology has developed the pragmatic use of reports more fully than the stronger enactive claims about embodiment and the efficacy of lived experience.[11] Whether one accepts their diagnosis or not, it identifies a useful pressure test: does experience merely label neural data, or does the study allow experiential organisation and embodied action to change the causal model?

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The stimulus that was not the same
  M->>E: Two trajectories through one enacted niche
  E-->>M: compare rival explanations
  E->>D: Make each domain capable of forcing revision
  D-->>Q: revise scope or act

Make each domain capable of forcing revision

Do not decide that a study is neurophenomenological because it contains an interview, a brain measure and the word “embodied”. Decide by asking what each domain was allowed to make false.

For a new study, begin with the coupling rather than the correlate. Specify the organism's relevant regulatory state, the action that changes sampling, the environmental contingency and the experiential transition of interest. Then create at least one intervention that breaks or shifts that relation while preserving the obvious alternatives. Active versus yoked replay is one pattern, not a universal recipe.

This change moves experimental effort upstream. Instead of buying another measurement channel first, invest in task design, phenomenological familiarisation and the counterfactual control. A weak task with more sensors produces a denser correlation map. A strong intervention with fewer measures can reveal which trajectory matters. The practical unit of evidence is not the modality count. It is the number of serious alternatives the configured study can eliminate.

Next, write two revision clauses. The first states which experiential observation would eliminate or alter a physiological model. The second states which physiological observation would eliminate or alter an experiential model. Put both clauses into the preregistration, separate report coding from neural analysis, and reserve data for a real holdout. If neither side can force revision, call the result a multimodal association, not reciprocal constraint.

Finally, match the claim to the grade. A correlation can identify a candidate bridge. One-way constraint can guide an analysis. Bidirectional revision can improve a joint model. An intervention-closed result can support a mechanism claim for the configured trajectory. None of these conclusions, alone, proves a metaphysics or identifies consciousness with a measure.

Varela's enduring contribution is therefore not a promise that experience and neuroscience will meet in one final vocabulary. It is a research discipline for keeping them different enough to resist one another and connected enough to learn. The organism, world and experience co-determine one another scientifically only when the experiment makes that co-determination vulnerable to a changed trajectory.