The report is inside the causal loop

The first mistake is to treat a report as if it arrived from outside the experiment, carrying a timestamped copy of an inner event.

Imagine asking Mira, “When did the image lose its grip?” She must notice a transition, keep it available, choose words or a key, decide that the criterion has been met, and act. Each step has its own dynamics. A report can be careful and informative without being instantaneous, exhaustive or causally inert. The report is therefore neither a transparent window nor useless noise. It is a fallible measurement channel embedded in the system being measured.

This point was already implicit in the original neurophenomenological programme. Varela proposed that disciplined first-person accounts and cognitive neuroscience should constrain one another rather than compete for sole authority.^[1] Lutz and colleagues then used participant-derived experiential categories to examine trial-level brain dynamics in a visual task, showing how first-person distinctions could expose structure that averaging would hide.^[2] Later methodological work developed second-person elicitation, micro-phenomenological interviews and thick-to-thin reporting strategies as ways to make experiential descriptions procedurally inspectable rather than privately certain.^[4][5][7]

No-report paradigms make this logic concrete. Studies have found neural activity that differs between reported and unreported conditions, while also identifying signals that persist when overt reporting is reduced.^[14][15] Yet “no report” does not mean “no post-perceptual processing”. Participants may still monitor, remember or covertly label an event. The control removes one route, not every rival route. This is why a computational bridge needs an explicit causal graph rather than a binary report/no-report label.

What makes a model generative?

A classifier that maps EEG features to “focused” or “distracted” can be useful. It is not yet a generative explanation. A generative model specifies a forward scheme: given a candidate state, context and intervention, what observations should appear, with what variability?^[9] That forward direction matters because it lets us simulate alternatives. Would the same neural trace appear if the report were delayed? Would the same report appear after a sham cue? Would behaviour change if the experiential transition occurred without an overt label?

Recent computational neurophenomenology uses active inference and related generative formalisms to model meta-awareness, attentional control and the “how” of experience, not only its object or content.^[8][10] A 2026 proposal goes further by treating beliefs as a conditional mathematical hub between phenomenology, behaviour and neural dynamics, while making its central assumption explicit: phenomenology must be a function of those beliefs for the derived predictions to hold.^[12] That is a serious, testable hypothesis. It is not a definition of experience.

Content, transition and report are different targets

Mira can report what she sees, how it is given, and that she has noticed a change. These are not interchangeable labels. “Face” names a perceptual content. “The face has lost its grip” describes an organisation of appearance. “I noticed the loss of grip” identifies a metacognitive event. Pressing the face key is an observable action. A study can model any of these, but it should not slide from one to another merely because their timestamps are close.

The distinction matters because each target supports a different intervention. Altering contrast may change perceptual content. Training attention may change the stability or grip of that content. A surprise probe may change meta-awareness. Delaying a button press changes the report process. When all four are compressed into a single class called “conscious state”, a model can score well while remaining causally ambiguous. The state vocabulary must be fine enough to separate the routes the experiment intends to cut, but no finer than the evidence can sustain.

This also clarifies what a computational model contributes. It is not a numerical replacement for an interview. It is a commitment device. The model forces the researcher to say which transition is latent, which observation is expected, which delay is plausible and which alternative generator would mimic the same observation. Formalisation earns its place when it reduces interpretive freedom before the result is known.

flowchart TD
  A["The report is inside the<br/>causal loop"] --> B["One hidden transition, three<br/>visible traces"]
  B --> C["A bridge must survive the<br/>counterfactual"]
  C --> D["Build the model so it can lose"]
  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

One hidden transition, three visible traces

The bridge becomes precise when we stop asking whether a brain pattern “matches” a report and ask what single transition could produce several imperfect traces.

Suppose Mira is doing a breath-attention task rather than binocular rivalry. Her experience moves among three coarse states: focused, drifting and meta-aware, where meta-awareness means that distraction itself has become salient. We do not claim that experience truly comes in three boxes. We choose the smallest state vocabulary that can express the transition of interest and make different predictions.

Let zₜ denote the candidate experiential state at time t. Let rₜ be a report, nₜ a neural or physiological measurement, and bₜ a behavioural outcome. Let xₜ describe the task world, uₜ an intervention intended to alter the experience, and qₜ the report protocol. A compact model is:

The equation compresses four assumptions. First, the chosen state carries enough history for the next prediction. Second, once the state and context are known, the remaining dependencies among the three traces are manageable. Third, the measurement rules are reasonably stable within the configured study. Fourth, the interventions are recorded accurately. If duration or remote history matters, a first-order state is too thin. We then need a semi-Markov model, a recurrent state or an explicit duration variable. Removing an assumption should change the model, not disappear into prose.

Every channel has a different clock and bandwidth

A key press is temporally sharp but semantically thin. A retrospective interview may recover sequence, bodily tone and attentional stance, but its timestamp is reconstructed through memory. EEG is fast but indirect and spatially mixed. Pupil size is easy to record yet affected by luminance, arousal and effort. Behaviour can reveal a consequence of the transition without identifying the experience that produced it. The model should therefore assign a measurement function and delay distribution to each channel rather than treating them as synchronised columns in a spreadsheet.

Suppose a neural marker precedes a key press by 480 milliseconds. That lead does not automatically make it experiential. Motor decisions can begin before movement, and a participant may adopt the response criterion gradually. Conversely, an interview code that appears ten minutes later may concern an earlier transition accurately enough to constrain its order but not its millisecond onset. Temporal precision and experiential specificity are separate properties. A good design combines channels because their errors differ, not because more sensors create truth by accumulation.

The report protocol variable qₜ should therefore include more than “report present”. It can encode whether a probe was expected, whether response was continuous or delayed, whether the participant used a key or language, how much time elapsed before elicitation, and whether the categories came from the participant or the researcher. These details alter both the probability of a report and the physiological cost of producing it.

Manifolds are summaries, not hidden worlds

Researchers often speak of an experiential manifold and a neural manifold. The phrase is useful when it means that many measurements vary along a smaller number of organised directions. It becomes misleading when two low-dimensional plots are visually aligned and the alignment is treated as an explanation. Rotation, scaling, smoothing and flexible nonlinear maps can make unrelated point clouds resemble each other.

The defensible question is operational: does a phenomenological distinction reduce uncertainty about future physiology or behaviour in conditions not used to construct the distinction? Does physiology, in turn, revise the phenomenological model by revealing a transition that interviews had merged? This is the original idea of reciprocal constraint made computational. Each domain must be allowed to make the other model wrong.

Model the person before modelling the average person

Experiential vocabularies rarely align perfectly across participants. One person may distinguish loss of grip from instability; another may use one category for both. Averaging their reports first can manufacture a state no participant actually had. The practical solution is hierarchical modelling. Each participant receives a small set of personal transition and observation parameters, while the group model estimates which relations are shared and how much they vary.

This prevents two opposite errors. Without personal parameters, genuine differences look like noise. Without group structure, every participant can be explained by a bespoke story that never generalises. A hierarchical model asks whether the same causal skeleton survives while thresholds, delays and emission strengths vary. The held-out test can then be demanding: predict a new block for a known participant and a new participant drawn from the population distribution.

Phenomenological training can be represented in the same way. It may improve report reliability, create new discriminations, alter attention itself, or do all three. Treating expertise as a fixed participant label hides those routes. A longitudinal design can instead ask whether observation noise falls, whether the latent transition changes, or whether both occur. Training is an intervention whose target must be declared, not a blanket guarantee of better access.

A bridge must survive the counterfactual

Several models can explain the data we observed. The bridge earns its name only when the models disagree about data we have not yet created.

Take the opening case. Three causal generators fit the first block. The shared-state model says a latent experiential transition drives early physiology, behaviour and the eventual report. The report-demand model says the report operation drives the apparent neural marker and the behavioural pause. The task-only model says stimulus timing and attention demand independently shape all channels, while reports merely track the same schedule.

Fit alone will often leave them nearly tied. Flexible models are especially good at absorbing the past. What we need is a set of interventions designed so that each generator predicts a different future. This paper proposes a practical standard: cross-modal counterfactual closure.

The word closure does not mean final truth. It means that the explanatory loop has been closed tightly enough for the model to carry information across domains under counterfactual change. The bridge predicts not only that two traces co-vary, but how their relationship breaks when one causal route is cut.

A configured study that could close the bridge

Consider a proposed binocular-rivalry study. Before recording, participants learn to discriminate three aspects of a switch: the currently dominant image, the degree of perceptual grip, and the moment at which the instability itself becomes explicit. Training is not used to certify privileged access. It is used to establish a shared task, vocabulary and discrimination procedure. Independent interviewers check whether the distinctions can be recovered without supplying the desired categories.

Write a commitment packet before fitting

The study should begin with a compact packet that another team could criticise. It names the experiential distinction, the elicitation procedure, the candidate states, the predicted transition order, the allowed delays and the observation variables. It also names the rival generators and the result each intervention is meant to separate. Feature families may be broad, but the selection rule and multiplicity control must be specified before the holdout is opened.

The packet should include two revision clauses. The first says what first-person observation would force a change in the physiological model. For example, if participants consistently distinguish loss of grip from perceptual instability, a one-transition neural model may be too coarse. The second says what physiological or behavioural observation would force a change in the experiential model. If two reported states have identical transition consequences across all independent channels, keeping both may add description without explanatory value.

Pre-registration does not freeze the research programme. Exploratory work can discover new categories and features. The discipline is to keep discovery and confirmation on different data. A new state found in one cohort becomes a declared hypothesis for the next cohort. A post-hoc timing window can motivate a replication, but it cannot serve as its own independent test. Generative passage requires an audit trail of how the bridge changed.

The closure test

For each candidate model, calculate predictive log scores on held-out intervention blocks. Do this separately for reports, physiology and behaviour. A single aggregate can hide failure in one channel, so inspect the weakest cross-channel margin. The shared-state model should beat its rivals not merely overall, but on the intervention that targets each rival’s assumed route.

A positive margin is strengthened by calibrated uncertainty, replication and a temporal match. It is weakened when the result depends on one participant, one interviewer, one frequency band selected after inspection, or a state definition rewritten after seeing the physiology. The number itself is not a consciousness score. It is a disciplined summary of a model-comparison obligation.

Closure also requires intervention invariance. The meaning of a latent state should not silently change from one block to another. If “meta-aware” predicts one physiological pattern under continuous report and an unrelated pattern under delayed report, the model may be using the label as a condition-specific container. Partial pooling, posterior predictive checks and state-alignment tests can reveal this drift. Sometimes the correct response is to split the state. Sometimes it is to admit that no stable bridge has been found.

Temporal generalisation provides another pressure test. Train the model on the onset of instability and ask it to predict recovery. Train on short transitions and test longer ones. If the same transition law fails as soon as the timescale changes, the apparent state may be an artefact of the observation window. This does not demand one universal model. It demands that the declared scope be earned rather than implied.

What would strengthen or weaken the proposal?

Cross-modal counterfactual closure would be strengthened by pre-registered predictions that transfer across report protocols, by participant-specific models that improve a hierarchical population model, and by interventions whose effects appear in the declared temporal order. It would be weakened if report-confound or task-only models match the held-out score, if experiential categories are unstable under independent elicitation, or if the bridge requires a new latent state after every condition.

A strong rival explanation is that physiology and behaviour carry enough information to predict reports without any scientifically useful experiential variable. That rival should be built, not dismissed. If it wins, the first-person distinctions may still matter clinically, ethically or descriptively, but they did not add predictive structure in this experiment. Conversely, if first-person distinctions improve held-out prediction, this shows epistemic value without proving that the latent variable is ontologically fundamental.

stateDiagram-v2
  [*] --> Claim
  state "Two trajectories can align<br/>without sharing a cause" as Claim
  state "Where the report enters the<br/>causal graph" as Test
  state "What makes a model generative?" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Build the model so it can lose

The smallest useful implementation is not a giant multimodal foundation model. It is a transparent candidate family, two interventions and a scoring rule that exposes failure.

A three-state hidden Markov model is enough to begin. Give it explicit transition probabilities, separate observation rules for report, physiology and behaviour, and a parameter for report contamination. Build two rivals with different causal commitments. The first explains physiology mainly through report activity. The second explains all channels through task context. Generate synthetic data, withhold one intervention and ask which model predicts it.

This simple design exposes logic that prose can conceal. If the bridge wins only when the report is continuous, it may have learned the reporting circuit. If it wins when the experience intervention is active but fails under the sham cue, expectancy may be missing. If all models tie, the intervention did not separate their predictions. None of these is a software error. Each is a scientific diagnosis.

Choose the formalism by the failure you need to see

A hidden Markov model is useful when discrete states and transition timing are the main hypotheses. A continuous state-space or dynamical-systems model is better when trajectories, attractors and gradual destabilisation matter. A semi-Markov model becomes useful when dwell time carries information. Active-inference models add beliefs, precision, policy and action selection, which can express hypotheses about attention regulation and meta-awareness.^[8][9] None should be chosen because its vocabulary sounds closest to phenomenology.

The selection question is practical: what rival mechanism would the formalism make visible? If the concern is report contamination, the model needs separate report and state routes. If the concern is continuous versus discrete experience, compare models with different state geometry. If the concern is whether action changes perception, include the organism-environment loop. A more expressive model is valuable only when the additional structure creates a discriminating prediction.

Complexity must also be paid for. Deep latent models can absorb participant idiosyncrasy and nonlinear physiology, but they increase identifiability problems. Two very different latent stories may produce the same observations. Simulation-based calibration, parameter-recovery tests and deliberately generated rival data should precede interpretation. If the pipeline cannot recover a known synthetic generator, it has no warrant to identify an unknown human one.

The strongest boundary

Cross-modal closure is still compatible with several explanations. A physicalist may interpret the latent transition as a compressed description of neural causation. A biological naturalist may insist that the relevant organisation depends on living tissue and regulation. A functionalist may focus on the causal role preserved across implementations. An enactivist may argue that the state is not inside the participant at all, but belongs to the organism-task loop. An idealist, neutral monist, panpsychist, dual-aspect theorist or process philosopher may treat neural and experiential trajectories as differently accessible aspects of a more basic process.

The experiment can pressure these views only when they make different predictions. Otherwise, it should remain metaphysically neutral. A successful bridge establishes disciplined co-prediction across perspectives. It does not establish that one perspective produces, reduces to or exhausts the other.