Where the eclipse begins

Imagine three questions about a red apple. Is there something it is like for the observer to see red? Can the observer use the colour to compare, remember, choose and reason? Can the observer say “red” or press the red key? Everyday life lets one answer all three at once. The apple is seen, used and reported. That reliable coincidence encourages us to collapse the questions.

Ned Block’s influential distinction separated phenomenal consciousness, the experiential character of a state, from access consciousness, the availability of information for reasoning and the rational guidance of speech and action.[1] The distinction was not a discovery of two brain modules. It was a warning about explanations. A mechanism that makes information available for control may explain access while leaving open why or whether there is something it is like. Conversely, an alleged experience that cannot affect memory, choice or report creates an evidential problem.

There are strong reasons to resist turning this distinction into two independent kinds of consciousness. An access-oriented account can argue that a genuinely subjective content must be available within an active interpretive workspace. Lionel Naccache, for example, treats conscious reports as the end of an inferential chain in which a subject attributes a content to itself, not as a passive broadcast from a private theatre.[4] A recent proposal by Liad Mudrik and colleagues similarly argues that phenomenality and access may be necessary conditions of one phenomenon rather than two detachable consciousnesses.[18] Global neuronal workspace accounts provide a major scientific version of the access-centred alternative by linking conscious access to broad availability for flexible use.[5]

That rival is stronger than “only reports matter”. It says that the lived presence of a content may already involve a form of access, integration or availability. Even if this is correct, overt report still does not become transparent. A thought can be accessible yet unspoken. A motor pathway can fail. A criterion can shift. The conceptual dispute about phenomenality does not remove the engineering problem of the report channel.

Access is always access for something

The word available can hide another collapse. Information may be available to an orienting response but unavailable to verbal comparison. It may guide a forced choice while failing to enter working memory. It may influence a later judgement without being available for deliberate correction. Access is therefore relational: available to which consumer, for which operation, over what delay and under which interference?

Blindsight made this issue vivid. Some people with damage to primary visual cortex can discriminate features in a region they describe as blind, especially when asked to guess. Their performance shows that stimulus information reaches routes capable of guiding constrained behaviour. It does not recreate ordinary seeing, flexible description or confidence. Block used such cases to challenge the inference that any information-guided action reveals the function of phenomenal consciousness.[1] Access-oriented theorists reply that the residual route is too narrow to count as the relevant global availability.

Both sides therefore need an access profile rather than a yes-or-no label. Test immediate discrimination, cross-modal comparison, delayed use, error correction, confidence and transfer to a new response. A content that drives one trained keypress but cannot support any other operation has thin access. A content that can be flexibly recombined has broader access even if the final motor act is blocked. The boundary should be drawn around preserved causal uses, not around the mere existence of an output.

A rich field and a narrow answer

Brief visual displays make the tension intuitive. Flash twelve letters for a fraction of a second, then ask for all of them. People usually report only a few. Cue one row immediately after the display and performance for that row improves. The classic partial-report result suggests that more information was briefly available than could be transferred into a durable report.[2] Block later used related capacity arguments to defend the possibility that phenomenality can overflow cognitive access, a claim that remains contested.[3]

flowchart LR
  A["Where the eclipse begins"] --> B["A report is something the<br/>system does"]
  B --> C["Taking the button away does<br/>not remove the inference"]
  C --> D["An assay for the boundary"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#087A98,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 something the system does

A report sounds like a final readout: observe an experience, convert it into words, send it out. The actual operation is active. The subject must understand the question, select a target, hold it long enough to compare with response options, adopt a criterion, commit to an answer and execute a movement. Even a confidence rating adds a second-order judgement about the first answer.

This matters because those operations recruit neural and cognitive processes of their own. If an experiment contrasts “seen and reported” with “not seen and not reported”, the difference can include the sensory conditions that enabled perception and the processes associated with the candidate experience. It can also include access, decision, answer commitment and the motor act that displayed it. Aru and colleagues called for separating prerequisites, neural correlates proper and consequences of consciousness.[6] A measured contrast is a mixture until the experiment supplies interventions that unmix it.

Formal notation is useful only after the situation is clear. The report probability can be expanded as a chain:

Every term names a possible failure. The stimulus may not produce the candidate content. The content may not be stabilised for access. The subject may adopt a conservative criterion. The chosen answer may not reach the muscle. The expansion does not assert that P exists independently of A; a strict access theory can set them equal. Its value is diagnostic: it prevents the final output from silently absorbing every intermediate process.

Reportability is not the same as reporting

A further distinction prevents a common methodological mistake. A state can be reportable in the counterfactual sense that the subject could answer if asked, while no answer is actually requested. No-report experiments depend on this possibility. They aim to preserve the perceptual state while withholding the commitment and motor operations triggered by a question.

Counterfactual reportability is difficult to verify because asking the question changes the trial. Researchers therefore estimate it across matched blocks. They may show that a stimulus condition supports accurate reports in one block, then infer that a similar state occurs in another block where the question is absent. That inference is reasonable only if the manipulation that removes report does not also remove attention, expectation or the relevant access route.

It helps to keep three levels separate. Potential report is a capacity under a specified query. Report commitment is the internal selection of one answer under a criterion. Overt report is the observable effect through a channel. The first may exist without the second; the second may exist without successful output; the third may be produced by habit or guessing with weak access. Experiments gain precision when they state which level a marker predicts.

The temporal signature of a mixed contrast

The report chain unfolds over time. Early sensory prerequisites precede the candidate conscious content. Access, evidence accumulation and commitment follow. Motor preparation and feedback arrive later. The windows overlap, and recurrent processing can blur any simple serial story, but timing supplies interventions: postpone the report, change its mapping, make the stimulus irrelevant or remove immediate action.

Matched manipulations have produced instructive dissociations. In binocular rivalry, Frässle and colleagues inferred perceptual alternations from eye movements and pupil size. When active report was removed, differential frontal activity was greatly reduced while posterior activity associated with perceptual switches remained.[9] In a separate visual paradigm, Pitts and colleagues found that gamma-band activity and the P3 response depended strongly on task relevance and report requirements, supporting a post-perceptual role for those signals in that design.[10]

Neither result licenses the slogan “the back of the brain is consciousness and the front is only report”. Removing an explicit response also changes attention, task set, expectation and memory. Frontal involvement may differ by content, modality, timescale and theory. Other no-report studies have examined content-related activity in macaque face cells and large-scale human visual networks, extending the method while retaining proxy assumptions.[11][12] The studies show that report demands can explain part of a contrast. They do not settle the complete neural basis of experience.

Taking the button away does not remove the inference

A no-report paradigm avoids asking for an explicit trial-by-trial answer. Instead, the experiment infers perceptual state from a response considered less demanding: reflexive eye movements, pupil changes, physiological signatures or a neural decoder. The method attacks a real confound. It does not create an observation without interpretation.

The proxy must be calibrated against conditions in which the percept is independently known or reported. It must then generalise to the no-report condition. If the proxy is influenced by luminance, surprise, arousal, attention or task context, those causes must be separated. No-report removes one report channel; it does not remove the measurement bridge. Tsuchiya and colleagues therefore argued for using report and no-report paradigms together, not for replacing one dogma with another.[7]

Calibration is a transport claim

Suppose a pupil pattern predicts reported face perception with 90 per cent accuracy in an active block. Using it in a passive block assumes the mapping transports across task contexts. That assumption is stronger than ordinary cross-validation. The no-report block changes motivation, sustained attention, expectation and sometimes eye position. A classifier can remain accurate in the training distribution yet become a detector of task state when the task disappears.

Three tests before a proxy can carry the claim

A proxy should pass three different tests. The first is content validity: within a condition where report is available, does the proxy distinguish the intended percept rather than only stimulus category or response hand? The second is causal specificity: when luminance, surprise, eye position or arousal is manipulated independently, does the proxy continue to track the percept? The third is transport: after the report task is removed, does the calibrated relation survive the changed task ecology?

These tests can fail independently. An eye movement may be an excellent percept classifier in active binocular rivalry yet become ambiguous during mixed dominance. A neural decoder may generalise across response hands but fail when attention is diverted. A pupil signal may retain sensitivity while losing specificity because unexpected sounds create extra positives. Reporting one aggregate accuracy hides these different debts.

The practical record should therefore include a confusion matrix for each context, not only a pooled score. It should name the proxy’s known rival generators, the interventions used against them and the conditions in which the classifier must abstain. A no-report inference is only as strong as the weakest untested transport assumption.

Critics of no-report methods make exactly this point. Overgaard and Fazekas argued that proxies inherit calibration assumptions and may not extract a “true” correlate simply by removing introspective report.[8] Block later proposed stronger designs that compare perceptually matched conditions and make the proxy’s causal role explicit.[13] The disagreement is productive: report paradigms contain report consequences; no-report paradigms contain proxy assumptions. Joint designs can make both visible.

When the body cannot carry the answer

Disorders of consciousness turn the report boundary into a clinical problem. A patient may understand an instruction and form an intention while lacking reliable motor output. Functional magnetic resonance imaging or electroencephalography can substitute a neural response channel. Owen and colleagues demonstrated task-specific imagery in a behaviourally unresponsive patient.[14] Later studies showed willful modulation and, in some cases, yes-or-no communication through such channels.[15] EEG command-following paradigms have also detected covert brain activation in some acutely unresponsive patients.[16]

A large multicentre study led by Yelena Bodien reported task-based fMRI or EEG responses in 60 of 241 participants who did not show an observable response to commands at the bedside.[17] This is compelling evidence that behavioural silence can conceal cognitive motor dissociation. It is not a universal detector. Negative results may reflect fatigue, language impairment, hearing loss, scanner noise, poor imagery, fluctuating arousal or limited statistical sensitivity.

The same boundary appears in artificial self-report

A language model can generate a fluent statement about what it “sees” because a prompt, image representation and learned discourse jointly support that output. The statement is evidence of a configured function: the system can classify, describe and perhaps use the content. It does not reveal, without further intervention, whether the answer was produced by a stable internal representation, a shallow textual cue, a tool result or direct imitation of the prompt. Swapping the output wording, hiding the image after encoding, patching internal state and testing transfer to action can separate some of these routes. None turns a functional dissociation into proof of experience. The report-boundary discipline applies precisely because machine reports can be convincing while their generator remains underdetermined.

stateDiagram-v2
  [*] --> Claim
  state "Access is always access for<br/>something" as Claim
  state "The silent pianist" as Test
  state "A rich field and a narrow answer" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

An assay for the boundary

The usual question, “Which neural signal is consciousness?”, is too early. Begin with a smaller causal question: which signal follows the content when the route to report changes? The paper’s practical contribution is a Triangulated Report Boundary Assay. It treats overt report, delayed report, no-report proxy and substituted output as matched measurement conditions rather than rival schools.

The assay is not a consciousness detector. It is a way to grade evidence. It can identify signals that are likely prerequisites, content-related candidates, access operations, report commitments or output consequences. It can also expose when the study lacks a discriminating intervention.

The four conditions

1. Immediate overt report. The subject reports each trial. This supplies direct task performance and calibrates response timing, but it contains criterion and motor activity.
2. Delayed report. The subject receives the report question after the main content-related window. This reduces immediate preparation but introduces memory and retrospective reconstruction.
3. No-report proxy. The subject receives no trial-wise question. A calibrated eye, physiological or neural proxy estimates content. This removes explicit commitment while importing proxy assumptions.
4. Channel substitution. The informational demand is preserved while the output route changes, such as voice to button, button to gaze, or muscle to fMRI/EEG. This tests whether absence belongs to the content or the channel.

Cross those conditions with a manipulation of content and, where possible, task relevance. Then intervene separately on criterion, access and proxy confounds. The design becomes useful because each manipulation carries a predicted pattern.

Executable lab: how channels distort the same latent event

The simulator below uses synthetic data. It assumes a latent candidate content variable so that channel error can be inspected directly. Real consciousness experiments do not possess that ground truth. This artificial advantage is deliberate: it makes visible how sensitivity, false positives, report contamination and proxy confounds change the conclusions drawn from the same underlying event rate.

How to use the assay

Declare the content manipulation and the candidate markers before collecting data. Calibrate every proxy in an active condition, then reserve a held-out context for transport testing. Change the response mapping without changing the content. Shift the report criterion through payoffs or instructions. Interfere with access at a predeclared time. Perturb each proxy’s known confounds. Record when the design cannot identify the cause.

A positive result does not mean “consciousness found”. It permits a narrower statement: a marker tracks content across specified report routes and resists specified alternative generators. A negative result can mean the marker belongs to report, the proxy failed, the access manipulation changed the content, or the study lacks power. The protocol requires a typed conclusion and an abstention rule for these cases.

from __future__ import annotations

from dataclasses import dataclass
from random import Random
from statistics import mean


@dataclass(frozen=True)
class Assumptions:
    prevalence: float = 0.50
    access_given_content: float = 0.75
    report_sensitivity: float = 0.90
    report_false_positive: float = 0.05
    proxy_sensitivity: float = 0.85
    proxy_confound: float = 0.10
    report_marker_boost: float = 0.60
    trials: int = 4_000
    seed: int = 88


def simulate(a: Assumptions) -> dict[str, float]:
    rng = Random(a.seed)
    rows: list[tuple[bool, bool, bool, float, float]] = []

    for _ in range(a.trials):
        content = rng.random() < a.prevalence
        access = content and rng.random() < a.access_given_content
        report = (
            rng.random() < a.report_sensitivity
            if access
            else rng.random() < a.report_false_positive
        )
        proxy = (
            rng.random() < a.proxy_sensitivity
            if content
            else rng.random() < a.proxy_confound
        )

        # Triangular noise avoids an external numerical dependency.
        noise_a = (rng.random() + rng.random() + rng.random() - 1.5) * 0.35
        noise_b = (rng.random() + rng.random() + rng.random() - 1.5) * 0.35
        report_marker = float(content) + a.report_marker_boost * float(report) + noise_a
        no_report_marker = float(content) + noise_b
        rows.append((content, report, proxy, report_marker, no_report_marker))

    def ratio(num: int, den: int) -> float:
        return num / den if den else float("nan")

    report_pos = [r for r in rows if r[1]]
    proxy_pos = [r for r in rows if r[2]]
    report_neg = [r for r in rows if not r[1]]

    report_ppv = ratio(sum(r[0] for r in report_pos), len(report_pos))
    proxy_ppv = ratio(sum(r[0] for r in proxy_pos), len(proxy_pos))
    missed = ratio(sum(r[0] for r in report_neg), len(report_neg))

    report_effect = mean(r[3] for r in report_pos) - mean(r[3] for r in report_neg)
    proxy_neg = [r for r in rows if not r[2]]
    no_report_effect = mean(r[4] for r in proxy_pos) - mean(r[4] for r in proxy_neg)

    return {
        "report_ppv": report_ppv,
        "proxy_ppv": proxy_ppv,
        "content_among_negative_reports": missed,
        "report_marker_effect": report_effect,
        "no_report_marker_effect": no_report_effect,
        "contrast_inflation": report_effect - no_report_effect,
    }


if __name__ == "__main__":
    for key, value in simulate(Assumptions()).items():
        print(f"{key:38s} {value:.3f}")

The adversarial path

Every proxy needs a mimic. A pupil change can be generated by perceptual alternation, luminance or surprise. A frontoparietal signal can be generated by access, decision or motor preparation. A language model’s statement “I am aware of the image” can be generated by internal access, instruction-following patterns or direct prompt imitation. The experiment should build at least one rival generator that produces the same observation through a different path.

The 2025 COGITATE adversarial collaboration offers a larger lesson. A preregistered, multimethod test challenged key predictions of both global neuronal workspace theory and integrated information theory rather than delivering a ceremonial winner.[19] Report-boundary work should adopt the same discipline. Theories must state where content, access and report enter their predictions, which result would lower confidence, and when the experiment is unable to decide.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: Where the eclipse begins
  M->>E: Access is always access for something
  E-->>M: compare rival explanations
  E->>D: Promote the conclusion by channel invariance
  D-->>Q: revise scope or act

Promote the conclusion by channel invariance

Do not choose between report and no-report as if one were evidence and the other contamination. A report is a rich measurement channel. It carries testimony, discrimination, confidence and the consequences of commitment. It also changes the system being measured. A proxy can reduce that intervention. It also creates a new bridge whose calibration may fail.

The practical decision is to stop asking a single output to bear the whole weight of consciousness. In a laboratory, require matched routes and interventions. In a clinic, treat behavioural silence as a possible output failure and seek validated alternative channels without promising that a negative scan proves absence. In artificial systems, treat self-report as functional evidence about generated statements and possible internal access, not as a verdict on experience.

Promote conclusions by channel invariance, not by rhetorical vividness. A signal that survives report removal, delayed query, output substitution and proxy challenge deserves a stronger content-related status. A signal that follows criterion or response mapping belongs closer to report. A design that cannot separate the two should say so.

This decision is compatible with disagreement about phenomenality. Access may constitute consciousness, accompany it, express it or constrain its manifestation. The assay does not settle that. It does something more immediate and necessary: it prevents the route by which an answer becomes observable from being mistaken for the thing the answer is about.