A conventional awareness experiment asks a sensible question in a causally dangerous way: compare trials reported as seen with trials reported as unseen, then attribute the neural difference to seeing.

The comparison is useful, but its interpretation is underdetermined. Before a faint stimulus becomes reportable, attention and baseline state may make its detection more likely. After it is experienced, the observer may evaluate confidence, map the experience to a response category, maintain that category, prepare a movement and monitor whether the movement succeeded. Aru and colleagues named the central distinction: a contrast can mix prerequisites of consciousness, a proper neural correlate of the conscious content, and consequences of conscious processing.[1]

That distinction changes what a brain image means. Suppose a frontal response occurs only on “seen” trials. It might be required for the experience. It might instead reflect the decision to say “seen”. It might be a metacognitive estimate of confidence. It might hold the answer during a delay. A correlation with report cannot decide among these generators because they normally rise together.

Report is a vector of operations, not a binary switch. At minimum it contains content access, metacognitive commitment, decision formation and expression. Speech, a key press and a saccade differ mainly in the last component. Immediate and delayed reports differ in memory demands. Graded visibility ratings add comparison and scale use. A “no-report” block can still contain covert versions of the first three components.

Subtraction is not identification

Pitts and colleagues exploited task relevance to separate early awareness-related activity from later report-related activity. Their work, and later no-report ERP studies, often finds that a late P3 or P3b is strongly affected by task and report, while earlier or intermediate candidates may survive.[2][7] This is an important correction to treating every late widespread response as the experience itself. It is not evidence that one earlier component is a universal marker across contents, modalities and tasks.

flowchart TD
  A["The button press that<br/>swallowed the percept"] --> B["The proxy that inherited the<br/>problem"]
  B --> C["A causal decomposition of report"]
  C --> D["The post-cue decomposition<br/>protocol"]
  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

No-report paradigms solve a real problem by refusing to make every critical trial end in a conscious judgement and overt response. Their hidden cost is that the experimenter still needs a label for what the participant experienced.[5]

That label can come from optokinetic nystagmus, pupil size, fixation, spontaneous facial movement, an involuntary physiological response, a decoder trained in another condition, a surprise memory test, or a post-experiment awareness assessment. Each method removes some report operations. Each also introduces an estimator with its own causal parents.

A revealing success: binocular rivalry

Binocular rivalry provides a favourable test bed. Opposing moving patterns can produce distinct tracking eye movements. Frässle and colleagues combined optokinetic nystagmus and pupil measures with fMRI, allowing perceptual alternations to be estimated without an active button press. Frontal responses associated with reported switches were substantially reduced, supporting the view that some frontal activity reflected introspection and action rather than the perceptual switch itself.[3] Brascamp and colleagues used inconspicuous, unreportable switches and likewise found that frontoparietal BOLD responses were greatly reduced, which they interpreted as evidence that those responses tracked consequences rather than causes of the switch.[4]

Those results broke an important rival generator: overt selection and action. They did not establish that all frontal activity is report-related. Later work decoded internally generated perceptual transitions from macaque prefrontal populations when the animals had not been trained to report the percept.[11] In inferotemporal face patches, Hesse and Tsao used eye movements to infer perception in untrained macaques and found that neurons tracked both the dominant and suppressed stimuli, with different readouts.[8] The studies differ in species, recording scale, stimuli, training history, proxy and analysis. Together they reject a simple equation between frontal activation and overt report.

Five ways a proxy can fail

Proxy error is not only misclassification. A proxy can be accurate overall and still support the wrong causal claim.

Block called this the “bored monkey” problem: the absence of an overt report does not prevent internally generated thought about the stimulus.[6] The inverse problem is equally important. A participant may stop processing a task-irrelevant stimulus deeply enough for the target conscious content to occur. No-report studies therefore need evidence about attention, engagement and actual perceptual rate, rather than only the absence of a response.[9]

What the empirical record actually says

The robust lesson is narrower than either camp’s slogan. Across several EEG studies, late P3-like responses are strongly modulated by task relevance and report. Some earlier or intermediate signals, including variants of the visual awareness negativity or an N2, can remain when overt report is removed.[12][13] Yet the exact surviving component changes with stimulus, analysis and paradigm. A recent no-report study reported a later bilateral occipital positivity for meaningful objects that was less reliable for simple shapes, suggesting that even late candidates can be content-dependent rather than universal.[16]

Spatial claims are also mixed. In one visual masking study, univariate frontal activation for visible stimuli disappeared under no-report, while multivariate patterns still decoded awareness in frontal as well as posterior regions.[10] Kronemer and colleagues found broad cortical and subcortical activity that persisted without report and after analyses designed to remove eye-movement contributions, while also showing that pupil, blink and microsaccade dynamics carried arousal and control signals.[12] An auditory no-report study found that conscious detection of task-irrelevant sounds recruited secondary auditory and attention networks beyond primary auditory responses.[15]

Human intracranial studies add temporal and anatomical precision, but their no-report results remain heterogeneous across contents, regions, tasks and measurement choices. A recent narrative review therefore treats localisation as an open empirical question rather than a settled posterior-versus-frontal division.[17]

The large adversarial test of global neuronal workspace theory and integrated information theory used clearly visible stimuli, task-relevant and task-irrelevant conditions, and multiple recording modalities. It found results that supported some predictions and challenged central predictions of both theories.[14] That outcome is methodologically healthy. Report control is one constraint among several, not a machine that prints a unique theory winner.

The practical move is to stop asking whether a paradigm “has report” and start specifying which causal operations occur, when they occur and how each could reach the neural measurement.

Let the critical conscious content on a trial be C. Let A represent attention and arousal, M metacognitive selection, D a decision state, O overt expression, and P the physiological or behavioural proxy used in a no-report condition. Let N be the measured neural signal.

In a report condition, O is present and M and D are usually high. In a no-report condition, O may be near zero, but A, M, D and P can all change. The candidate content is also unobserved. Instead the analysis conditions on an estimated label:

The proxy model g is learned in a context that includes stimulus statistics, task, species, training and recording quality. If any of these change, calibration may not transport. Treating Ĉ as error-free can attenuate a real NCC, fabricate a contrast through systematic error, or shift timing by aligning neural data to a delayed proxy response.

The proxy-collider trap

Noisy labels are the easy version of proxy failure. A deeper problem appears when the proxy is caused by both the candidate experience and another process that also affects the neural signal. Suppose pupil dilation, P, increases when a face becomes conscious, C, but also when arousal rises, A. Suppose the candidate neural response, N, also increases with arousal. The proxy is now a common effect of two causes. Selecting or weighting trials by P can make C and A statistically dependent even when they were independent before selection.

A crossed numerical example

Now consider a synthetic EEG candidate measured as the seen-minus-unseen amplitude between 300 and 450 milliseconds. The stimulus and fixation task are held constant. Only the timing and object of the later query change. The values below are illustrative effect sizes in arbitrary units.

The convergent component is about 2.8 units. The additional 2.1 to 2.4 units in immediate report are plausibly tied to operations induced by immediate reporting, while the small yoked effect suggests that finger movement alone explains little. This pattern permits a narrow statement: the 300-to-450-millisecond candidate is not wholly generated by content-specific decision or overt response under these conditions.

It still does not identify consciousness. Imagine that an attention-diversion block leaves physical stimulation unchanged, lowers the proxy’s calibration slope and reduces the candidate contrast from 2.7 to 1.4. Three explanations remain live: conscious detection became less frequent, the neural candidate depends on attention as a prerequisite, or proxy error increased. The crossed design has done useful work because it eliminated one family of explanations. It has not manufactured a unique ontology from subtraction. The next intervention must vary attention while preserving enough content evidence to distinguish those remaining worlds.

The report-load vector

A useful protocol records four report dimensions for each condition:

1. Selection load: Must the participant notice and select the critical content rather than merely undergo it?
2. Commitment load: Must the participant settle on a categorical or graded judgement?
3. Maintenance load: Must the judgement persist across a delay or compete with other contents?
4. Expression load: Must it drive speech, a button, gaze, autonomic control or another external channel?

Representing a condition as a vector, rather than “report” or “no-report”, prevents false equivalence. A delayed post-cue has low expression load during the early neural window but high prospective selection load if participants expect a query. A surprise memory test has low online load but adds uncertain retrospective sampling and memory selection. Passive rivalry with eye tracking has low deliberate expression but a non-zero physiological output pathway.

What reports contribute that proxies cannot replace

Reports are more than contaminants. They are evidence about phenomenal character, confidence, ambiguity and temporal structure. A proxy may identify leftward dominance while missing whether the percept was mixed, unstable, effortful or accompanied by uncertainty. In clinical, animal and artificial systems, report may be unavailable, but the evidential loss should remain visible.

This point is compatible with several metaphysical positions. A physicalist may search for neural mechanisms that constitute experience. A biological naturalist may expect substrate-specific causal powers. A functionalist may treat the relevant organisation as multiply realisable. Idealist, neutral-monist, panpsychist and dual-aspect views may reverse or complicate the explanatory direction. No-report methodology does not adjudicate these ontologies. It asks a narrower question: given a candidate experience and a measurement bridge, which neural differences survive changes to the reporting pathway?

stateDiagram-v2
  [*] --> Claim
  state "Subtraction is not identification" as Claim
  state "The report-displacement principle" as Test
  state "A revealing success, binocular<br/>rivalry" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

The proposed contribution is a study architecture that treats reports and proxies as instruments with separable loads, then attacks the remaining rival generators one by one.

Stage 1: calibrate the instrument

Run a report-rich calibration session using the same stimulus family, participant population and recording setup as the critical experiment. Collect immediate graded reports, not only binary labels. Fit a probabilistic model from passive signals to reported content. Evaluate it out of sample. Report calibration curves, ambiguous-state coverage, timing error and failures by stimulus strength.

A positive calibration result permits use of the proxy as an uncertain content estimator within that scope. A negative result means the proxy cannot label critical trials. It does not mean the participant lacked the experience.

Stage 2: randomise the query after the critical window

Present the critical stimulus while holding fixation task, reward and expected effort constant. Only after the pre-registered neural window, draw one of three cues at random: report the critical content, report an orthogonal feature, or make no response. Because the cue is unknown during the early window, immediate decision mapping and motor preparation cannot systematically differ by later cue. Prospective monitoring remains and must be acknowledged.

Stage 3: replay action without judgement

In a yoked motor block, replay each participant’s earlier response timings and button distribution while decoupling the actions from current percepts. For eye or pupil proxies, include physical motion and luminance controls that reproduce the peripheral signal without the target perceptual transition where possible. This estimates the neural footprint of expression and proxy physiology.

Stage 4: inject failures

Stress the proxy rather than merely validating it on easy trials. Include mixed percepts, rapid alternations, low-confidence events, attention diversion, arousal changes and a session boundary that can reveal calibration drift. Add sparse experience-sampling probes about mind-wandering on trials excluded from the main time-locked analysis. The experiment should contain conditions where the proxy is expected to fail. Otherwise its claimed domain has no tested edge.

Executable proxy-assumption ledger

The instrument below implements hard gates rather than an arbitrary total score. Use it before interpreting a no-report result. Select only checks backed by data or a pre-registered control. The output states the strongest claim class currently permitted.

A positive ledger result is permission to make a narrower claim, not a consciousness verdict. The tool cannot establish phenomenal experience, necessity, sufficiency, minimality or metaphysical constitution. It exposes where prose would otherwise hide assumptions.

{
  "proxy": {
    "sameDomainCalibration": true,
    "heldOutValidation": true,
    "uncertaintyPropagated": true
  },
  "causalControls": {
    "attentionAndArousal": true,
    "covertCognition": false,
    "proxyPath": true,
    "temporalSeparation": true,
    "rivalGeneratorIntervention": true
  },
  "triangulation": {
    "crossConditionGeneralisation": true,
    "negativeControls": true
  },
  "permittedClaim": "report-independent NCC candidate",
  "prohibitedClaims": [
    "consciousness established",
    "necessary or sufficient mechanism",
    "universal neural signature"
  ]
}

How the proposal could fail

The protocol would be strengthened if candidate signals remain stable across query probabilities, proxy perturbations and recording modalities, while report-specific signals follow only the post-cue branches. It would be weakened if the candidate disappears whenever active monitoring is reduced, if proxy errors explain the neural effect, or if the signal follows motor or autonomic replay.

A rival explanation could still produce the same evidence. A content-specific attentional process may generalise across all conditions and precede the cue. A theory may identify that process as a prerequisite rather than consciousness. The next discriminating study must then intervene on attention while holding reported content and proxy validity as stable as possible.

The protocol should not be applied as a rigid template to unresponsive patients, infants or artificial systems where calibration reports are unavailable. There, the proxy-assumption ledger becomes even more important, but validation must use convergent physiology, perturbation, theory-derived predictions and explicit missingness rather than pretending a report-trained model transfers unchanged.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: The button press that swallowed the percept
  M->>E: Subtraction is not identification
  E-->>M: compare rival explanations
  E->>D: Replace purification with triangulation
  D-->>Q: revise scope or act

Do not choose between report and no-report paradigms. Choose a portfolio of interventions that decomposes what each instrument adds.

The old decision was methodological branding: use report and risk post-perceptual contamination, or use no-report and reveal a purer neural correlate. The better decision is claim-centred. Begin with the exact inference you want. Name the candidate content, the time window, the system boundary and the rival generators. Then select report, delayed report, proxy, task relevance, motor replay and perturbation conditions that make those generators vary independently enough to matter.

For a new study, the practical rule is:

For an existing result, ask what disappears and what remains when report is removed. If a late signal disappears, conclude that it depends on report context unless another intervention locates the responsible operation. If a signal persists, call it report-independent within the tested scope. Do not promote persistence to constitution. If frontal activity shrinks, do not infer that frontal cortex is irrelevant. If posterior activity survives, do not infer that it alone is sufficient.

For theory comparison, pre-register different predictions for the same intervention. A global-access theory may predict a specific broadcast pattern even without overt expression. A recurrent-processing theory may predict local recurrent activity that survives task irrelevance. A higher-order theory may predict metarepresentational involvement that need not decode detailed content. The experiment earns value when these commitments generate distinct outcomes, not when one anatomical region is assigned to “consciousness” by convention.

The deepest correction is epistemic. Conscious experience is private, but it is not scientifically inaccessible. Reports, behaviour, physiology, neural dynamics and intervention can constrain one another. None is self-interpreting. A report tells us something valuable while producing extra activity. A proxy removes some activity while adding assumptions. A perturbation can break a causal path while changing the state it was meant to preserve.

The decision is therefore to replace purification language with triangulation and typed inference. The best no-report paradigm is not the quietest experiment. It is the experiment that makes its substituted measurement visible, gives rival generators somewhere to fail, and leaves the reader able to tell exactly what the result changes.