The scene and the second scene

Start with the smallest useful distinction. A first-order visual state carries information about the grating's orientation. It can drive a left-or-right choice. A higher-order state carries information about the system's relation to that visual state: that it is seeing rather than imagining, that the evidence is weak rather than strong, that the state arose now rather than from memory, or that the representation is trustworthy enough to guide a claim.

Higher-order theories make this relation load-bearing. In a traditional higher-order thought account, a first-order state becomes conscious when the subject has a suitable higher-order representation of itself as being in that state. The higher-order representation need not be verbal, deliberate or itself conscious. Inner-sense variants cast it as a monitoring representation. Perceptual reality-monitoring accounts emphasise whether a sensory state is represented as a reliable description of the present external world. The higher-order state-space model treats awareness as an inferred state in a generative model of perceptual contents, with particular attention to awareness reports. Radical plasticity accounts propose that the brain learns redescriptions of its own activity. These variants disagree about format, learning and what is constituted, but all give the second scene explanatory work. [1] [2] [3] [4]

Why a second state does not create an infinite regress

A natural objection asks who observes the observer. If the first-order state needs a higher-order state to become conscious, must that higher-order state itself be represented by a third state, continuing without end? Higher-order theories answer no. The higher-order state can perform its constitutive or enabling role without itself being conscious. The claim concerns what makes the target state conscious, not what makes every state in the architecture conscious.

This answer is coherent, but it creates a testable burden. The higher-order state cannot be identified by asking the person to introspect it directly, because the theory does not require such access. Researchers must infer it from its target relation and causal consequences. “Higher” therefore names an order of representation, not greater intelligence, verbal sophistication or executive rank.

The content also need not contain an explicit sentence such as “I, Mira, am now seeing a left-tilted grating”. A minimal state could index that a current sensory representation is reliable, externally caused and available for use. This is why infants, animals and non-verbal systems are not excluded by definition. At the same time, removing explicit concepts makes the suitability problem sharper: what distinguishes a genuine representation of being in a state from an ordinary reliability signal? The answer must come from the relation the signal preserves and the interventions it survives.

The attraction of this architecture is clear. It offers a way to explain why two perceptual states with similar first-order task performance can differ in subjective availability. Lau and Passingham's “relative blindsight” paradigm sought exactly this pattern: objective discrimination was matched across conditions while subjective awareness differed. The result is suggestive because it targets the hidden relation between performance and awareness rather than simply comparing strong with weak stimuli. [6]

Yet the pattern does not select a higher-order theory on its own. A recurrent-processing theorist can say local sensory recurrence changed while the chosen discrimination measure failed to reveal it. A global-workspace theorist can say broad availability changed. An attention-schema theorist can say the system's simplified model of attention changed. A decision theorist can say the criterion for saying “seen” moved. The same behavioural dissociation sits downstream of several rival generators.

The empty and inaccurate second scene

Suppose Mira's higher-order system represents her as seeing red when no first-order red state exists. This is the empty higher-order case. A strict higher-order thought view can appear to predict a red experience generated by the higher-order content alone. Now suppose a first-order orange state exists but the higher-order representation says red. Which content is experienced? The first-order state, the higher-order state, or a joint result?

Contemporary higher-order accounts often answer with a joint-determination view: first-order content and higher-order representation contribute different aspects. The first-order state helps determine what appears; the higher-order relation helps determine whether and how it is consciously available. That response avoids making a free-floating higher-order thought sufficient for arbitrary experience, but it also makes the mechanism more demanding. The experiment must now identify the relation, rather than only the presence of activity in a monitoring area. [2]

Higher-order theory becomes scientifically distinctive only where it predicts an intervention fingerprint that a first-order or report-only account does not. The rest of this paper constructs that fingerprint.

flowchart TD
  A["The scene and the second scene"] --> B["What makes a representation<br/>genuinely higher-order"]
  B --> C["The parallax perturbation"]
  C --> D["What the result would and<br/>would not mean"]
  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

What makes a representation genuinely higher-order

A neural network layer transforms the output of an earlier layer. A clinician reads a patient's brain scan. A separate model predicts whether a classifier is correct from its final logits. All three are “about” an earlier process in a loose sense. If every downstream transformation counts as metarepresentation, the concept becomes too cheap to explain consciousness.

Kanai and colleagues identify this triviality pressure and propose that meta-representations should be understood as representations of processes, not simply representations whose input happens to be another representation. The key question becomes what internal relation the second state preserves. Does it track how the first state was produced, or only correlate with its output? [5]

Consider two routes to the same answer. Route S obtains “left” from a weak sensory trace. Route M retrieves “left” from a recently memorised cue. Both are 80 per cent accurate. A state that encodes only the answer cannot distinguish them. A process-sensitive higher-order state can encode that one answer is perceptual and noisy while the other is mnemonic and fluent. That difference can guide whether the system says “I saw it”, “I remember it” or “I guessed”.

Let X be the external stimulus, F the first-order state, D the task decision, H the candidate higher-order state and R the observable report. A minimal architecture is:

Here q names internal process properties, ε_F and ε_H are first-order and monitoring noise, and b is a report or decision bias. The factorisation is a modelling assumption, not a discovered anatomy.

The useful causal question is not whether H predicts R. A response-prediction module can do that by reading the button plan. The question is whether an intervention on H changes the relevant report or control while F, D and the report mapping remain stable enough to rule out simpler paths. In intervention notation, the desired signature is a change in R under do(H = h′) conditional on matched first-order sensitivity and verified report mechanics.

Confidence is a witness with its own incentives

Confidence is useful because it asks whether the system can distinguish its likely correct from likely incorrect decisions. It is not identical to awareness. A person can be highly confident in a guess because the prior odds favour one answer. They can experience a vivid stimulus yet remain uncertain about a difficult category. Higher-order theorists themselves have repeatedly warned against equating consciousness with confidence or sophisticated self-reflection. [2]

Signal detection theory helps separate first-order discrimination from second-order sensitivity. First-order d′ measures how well internal evidence separates stimulus classes, under the model's distributional assumptions. Meta-d′ asks what first-order sensitivity an ideal observer would need to produce the observed pattern of confidence ratings. The ratio meta-d′/d′, often called metacognitive efficiency, asks how much of the available decision information reaches confidence. It is still a model-based behavioural quantity. It is not a unit of consciousness. [9] [10]

Replacement and boundary tests

These substitutions are not philosophical decoration. They define experimental contrasts. A monitor that sees only final logits can be compared with one that receives process traces. An internal monitor can be compared with an external assessor. A live higher-order circuit can be compared with a cached report policy. The target is the preserved relation and its break condition.

The parallax perturbation

The proposed test uses three causal perturbations and at least four measurements. Its purpose is to prevent one ambiguous subjective response from carrying the entire argument.

The first intervention, I_F, changes first-order evidence. Contrast, exposure duration or sensory noise can be altered while keeping response incentives fixed. The second, I_H, targets the candidate monitoring relation. In a biological study this might be time-specific stimulation, lesion-defined comparison or pharmacological manipulation with a validated effect on monitoring. In an artificial system it can be noise injected into the monitor's process fingerprint, a state swap or an ablation that leaves the first-order classifier untouched. The third, I_R, changes report policy without changing the stimulus: base rates, asymmetric payoffs, response mapping or report order.

The minimum measurements are first-order sensitivity, metacognitive sensitivity, response criterion and an appearance measure that does not reuse the same categorical decision. Reproduction, matching or graded visibility can help, but no measure is automatically pure. One recent study found that base rates and payoff manipulations leaked into confidence reports even when they did not uniquely alter subjective experience. A follow-up study went further: asking participants to make a detection judgement could contaminate a subsequent reproduction measure with the same decision biases. [14] [15]

Measurement independence is an experimental achievement, not a property conferred by a different response scale. The design should vary whether the detection decision is requested, counterbalance report order, estimate criterion shifts and pre-register which measure licenses which inference.

The predicted signatures

A trial-level protocol

Each trial begins with a brief stimulus whose difficulty is adjusted independently for each participant. The task response is delayed until a fixed cue so that sensory evidence and motor preparation can be modelled separately. On randomly interleaved trials, the participant either gives a confidence rating, reproduces an appearance property such as contrast, or gives no overt subjective report. Report order is counterbalanced across blocks rather than assumed harmless.

The first-order manipulation changes contrast or masking. The report-policy manipulation changes base rates or asymmetric payoffs but is disclosed before the block, making its effect on criterion estimable. The candidate higher-order manipulation occurs within a time window justified by prior physiology and includes a sham or active-control site. A manipulation check must show that the intended monitoring variable changed without a broad loss of alertness, working memory or motor speed.

The analysis fits task sensitivity and criterion jointly, then estimates metacognitive efficiency with uncertainty carried through from first-order performance. Appearance measures are analysed both when a detection decision precedes them and when it does not. A result counts as selective only if the higher-order intervention changes the pre-registered monitoring or appearance outcome more than the active control, while equivalence bounds support preservation of first-order sensitivity. “No significant change in accuracy” is too weak because an underpowered study can hide a meaningful sensory loss.

This protocol also defines a real negative result. If the monitoring manipulation is demonstrably strong and selective, first-order measures remain within equivalence bounds, and every awareness-related measure remains stable, the tested higher-order mechanism loses support for that task and time window. The theory is not rescued by relabelling the unsuccessful prediction as hidden consciousness.

The decisive comparison is not a single significant effect. It is the vector of changes across measurements. Define a parallax signature for intervention I:

A is a pre-declared appearance measure and c is response criterion. The vector is meaningful only with manipulation checks and uncertainty intervals.

A higher-order-supporting pattern would combine a successful I_H manipulation with approximately preserved d′, reduced metacognitive efficiency and a change in a separately validated appearance or awareness measure. An I_R manipulation should alter criterion or mean confidence without reproducing the full pattern. A first-order intervention should show that the task can still detect genuine sensory degradation.

This standard is stricter than the often-cited transcranial magnetic stimulation result in which dorsolateral prefrontal disruption reduced metacognitive visual awareness while leaving discrimination intact. That study provided an important causal template, but a later replication did not recover the effect and highlighted data stability and analytic sensitivity. Other work reports causal involvement of lateral prefrontal cortex in visual metacognition, while decoding studies show that subjective visibility and confidence signals can overlap yet be dissociated. The evidence supports continued intervention work, not a simple localisation verdict. [7] [8] [11] [12] [13]

The simulator makes one hidden logic visible. Lowering sensory strength reduces first-order performance. Increasing only higher-order noise leaves the decisions unchanged but weakens how well confidence discriminates correct from incorrect trials. Adding a constant report bias raises average confidence and the thresholded “awareness” rate without improving discrimination. The last effect is precisely why a higher mean rating cannot establish richer experience.

from __future__ import annotations

from dataclasses import dataclass
from math import exp, log, pi, sqrt
from random import Random
from typing import Iterable


@dataclass(frozen=True)
class Scenario:
    signal: float
    monitor_noise: float
    report_bias: float
    trials: int = 20_000
    seed: int = 117


def normal(rng: Random) -> float:
    # Box-Muller transform
    u1 = max(rng.random(), 1e-12)
    u2 = rng.random()
    return sqrt(-2.0 * log(u1)) * __import__("math").cos(2.0 * pi * u2)


def sigmoid(value: float) -> float:
    return 1.0 / (1.0 + exp(-value))


def rank_auc(scores: Iterable[tuple[float, int]]) -> float:
    ordered = sorted(scores, key=lambda item: item[0])
    positive_ranks = 0.0
    positives = sum(label for _, label in ordered)
    negatives = len(ordered) - positives
    index = 0

    while index < len(ordered):
        end = index + 1
        while end < len(ordered) and ordered[end][0] == ordered[index][0]:
            end += 1
        average_rank = ((index + 1) + end) / 2.0
        positive_ranks += average_rank * sum(
            ordered[position][1] for position in range(index, end)
        )
        index = end

    if positives == 0 or negatives == 0:
        raise ValueError("AUC requires both correct and error trials")
    return (
        positive_ranks - positives * (positives + 1) / 2.0
    ) / (positives * negatives)


def simulate(scenario: Scenario) -> dict[str, float]:
    rng = Random(scenario.seed)
    records: list[tuple[float, int]] = []
    confidence_total = 0.0
    aware_total = 0
    correct_total = 0

    for _ in range(scenario.trials):
        stimulus = 1 if rng.random() >= 0.5 else -1
        evidence = stimulus * scenario.signal + normal(rng)
        choice = 1 if evidence >= 0 else -1
        correct = int(choice == stimulus)

        monitor = (
            abs(evidence)
            + scenario.monitor_noise * normal(rng)
            + scenario.report_bias
        )
        confidence = sigmoid(2.0 * (monitor - 0.75))
        aware = int(confidence >= 0.60)

        records.append((confidence, correct))
        correct_total += correct
        confidence_total += confidence
        aware_total += aware

    return {
        "accuracy": correct_total / scenario.trials,
        "type2_auc": rank_auc(records),
        "mean_confidence": confidence_total / scenario.trials,
        "awareness_rate": aware_total / scenario.trials,
    }


if __name__ == "__main__":
    scenarios = {
        "baseline": Scenario(1.10, 0.60, 0.00),
        "first_order_degradation": Scenario(0.55, 0.60, 0.00),
        "higher_order_noise": Scenario(1.10, 1.80, 0.00),
        "report_bias": Scenario(1.10, 0.60, 0.80),
    }
    for name, scenario in scenarios.items():
        print(name, simulate(scenario))

Pre-registration and break conditions

The smallest useful study is a within-participant factorial experiment with adaptive staircasing to match first-order performance, separate sessions for the three interventions, a manipulation check for the candidate monitor and a report-order control. The analysis should estimate full posterior uncertainty or confidence intervals for the parallax vector, rather than only test whether one coefficient crosses a threshold.

A result strengthens the proposal when the higher-order intervention is selective, replicable, temporally specific and accompanied by a process-level manipulation check. It weakens the relevant higher-order theory when a strong, selective intervention leaves every awareness and appearance measure unchanged while first-order and report controls remain sensitive. A rival explanation remains possible if the perturbation changes post-decision evidence, attention, working memory or motor preparation.

stateDiagram-v2
  [*] --> Claim
  state "Why a second state does not<br/>create an infinite regress" as Claim
  state "The empty and inaccurate<br/>second scene" as Test
  state "Confidence is a witness with<br/>its own incentives" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

What the result would and would not mean

A configured machine case: Aster-7

If the process-fingerprint monitor identifies route swaps, maintains calibration under report-cost shifts and loses only those capacities when perturbed, it satisfies the suitability contract better than the logit-only monitor. That is a useful engineering result. It supports source monitoring, calibrated abstention and error diagnosis. It also creates a stronger artificial candidate for testing higher-order theory than a language model prompted to describe its “experience”.

A transformer layer attending to another layer is not thereby a higher-order representation. The system needs a declared target, a preserved process relation, an intervention that breaks that relation and a downstream use distinguishable from ordinary task computation. The same caution applies to natural language self-reports. A model can generate “I am uncertain” from training regularities, policy instructions or visible logits without representing the process that produced its answer.

What current evidence permits

Human studies establish that first-order performance and metacognitive performance can dissociate, that prefrontal perturbations sometimes affect metacognitive measures, and that subjective visibility and confidence are not reducible to one variable. They also show that results depend on task design, timing, model choice and replication. No single cortical region or behavioural score has been shown to be the place where experience is manufactured.

No-report work can reduce overt reporting demands, and prefrontal activity related to conscious content has been observed under some no-report designs. This matters because it weakens the simple claim that all prefrontal signatures are just button preparation. It does not remove covert attention, task set, memory or inference about unreported states. A no-report paradigm replaces one measurement bridge with another. [16]

The strongest alternative

The strongest alternative is not that higher-order processes do nothing. It is that they support access, confidence, source monitoring, introspection and report while the phenomenal character of perception depends on first-order recurrent sensory organisation or another mechanism. On this view, Mira's accuracy can remain stable while her ability to know and describe her visual state deteriorates. Her experience may be unchanged, or only the reportable structure around it may change.

Block's criticism presses the ambitious identity claim: empty and misrepresenting higher-order states create counterintuitive predictions, and a modest theory of access does not automatically become a theory of phenomenal consciousness. Higher-order responses argue that the suitable relation is richer than a detached thought and that first-order content participates in the result. The dispute therefore turns on exactly which relation is constitutive, causal or merely evidential. [17]

These interpretations are not extra decorations around the same empirical result. They assign different explanatory jobs to the mechanism. The parallax test can discriminate causal architectures. It cannot, without additional bridge principles, decide whether awareness is produced, constituted, organised, manifested or merely reported by them.

For artificial systems, the same restraint is essential. Butlin and colleagues propose assessing computational properties derived from scientific theories rather than relying on surface behaviour. Aster-7 follows that spirit: it can establish a configured higher-order functional property under intervention. The result should enter a broader candidate dossier as one theory-relative indicator, never as a consciousness score. [18]