Structural Signals of Consciousness

Keywords: consciousness; artificial intelligence; sentience; valence; neurobiology; causal structure; evidence mapping; precaution; composite indicator

Working-paper status

This document is an evidence-informed reasoned proposal and a first manuscript draft. It is not a systematic review and does not report a completed rapid evidence map.

All signal grades, diagnosticity judgements, and references should be read in light of this status. Releasing the document before the remaining stages are complete does not convert provisional author judgement into expert consensus or validated measurement.

1. Purpose and scope

1.1 The assessment problem

Consciousness has no single generally accepted behavioural, computational, or biological test. Behaviour is indispensable in human consciousness research because reports can be related to internal measurements, but similar output from an unfamiliar system does not establish a similar internal process. The inverse error is also possible: lack of report or overt responsiveness does not by itself establish lack of experience, as dreaming, anaesthesia, and disorders-of-consciousness research make clear.

This paper proposes structural signals as a disciplined middle layer between two unhelpful extremes. It does not infer consciousness directly from language or task performance, and it does not require literal reproduction of mammalian anatomy. Instead, it asks whether a system contains an operationally specified causal property whose relation to a named aspect of consciousness is supported in the human reference case, whether the property is actually realized in the assessed system, and whether there is a warrant for transporting the human relation to the new implementation.

The working definition is:

A structural signal is an operationally measurable property of a system that, in conscious humans, reproducibly covaries with or causally contributes to a named aspect of consciousness, and whose relevant causal organisation can in principle be investigated in another biological or artificial system.

No signal in this version is claimed to be individually necessary or sufficient for consciousness. The proposed inventory is intended for rapid evidence organisation and precautionary triage, not metaphysical proof.

1.2 Three empirical targets

The framework separates three targets that are often conflated:

G, C, and V are not interchangeable. Evidence for reportable access need not establish phenomenal consciousness; evidence for arousal regulation need not establish experience; and generic reward or action value need not establish pleasure or suffering. A property may therefore receive different grades for different targets.

1.3 Contextual modifiers

Agency, autobiographical continuity, learning, embodiment, persistence, and scale can be ethically or operationally important without serving as strong evidence for minimal experience. This version reports them as contextual modifiers rather than adding them to the consciousness-evidence index. The separation prevents, for example, a highly autonomous but apparently non-valenced controller from receiving a higher consciousness score merely because it can act, or a passive system with credible suffering-related evidence from being discounted because it lacks agency.

1.4 Intended and excluded uses

Structural Signals v1.3 is intended to:

• standardise what property is being claimed and at what system level;
• distinguish human association evidence, target-system mechanism evidence, and transport validity;
• expose dependencies, missing measurements, and reasonable alternative boundaries;
• support comparisons and sensitivity analysis using a fixed provisional index; and
• provide evidence inputs to a user-defined precautionary decision process.

It is not intended to:

• produce a probability that a system is conscious;
• certify that a system can or cannot suffer;
• infer internal properties from fluent self-description alone;
• prescribe universal policy thresholds or legal status; or
• replace direct welfare, capability, security, or control-risk assessment.

2. Units of assessment, access, and transport

2.1 Declaring the system boundary

An assessment must identify its unit before assigning any realization rating. Possible units include a frozen base model, a model plus decoding and inference state, a deployed assistant, a persistent agent with tools and memory, or an embodied autonomous system. The label of a product or model family is not an adequate boundary.

A component is included when, during the assessed operation, it is causally coupled to the rest of the unit, persists on the timescale relevant to the claimed signal, and is counterfactually load-bearing for that claim. A memory store, controller, value module, sensor, body, or tool is not excluded merely because it was engineered separately. The same criterion applies to biological systems, whose organs and neural subsystems are also modular. When two reasonable boundaries materially change the result, both must be reported.

The assessment record should state:

Target and version:
Operational interval:
Included components:
Excluded components:
Causal-coupling evidence:
Relevant persistence timescale:
Alternative reasonable boundary:

2.2 Access levels

The strength of a negative or positive realization claim depends on what can be measured and perturbed.

Level 1 non-detection is normally unknown, not tested absence. A high-confidence zero requires an adequately powered test at an access level capable of observing the minimum construct.

2.3 Three evidential judgements

The framework keeps three questions separate:

1. Human association: How strongly is the reference property associated with G, C, or V in humans?
2. Mechanism realization: How strongly does evidence show that the assessed system implements the operational property at the declared boundary?
3. Transport validity: Why expect the human property–consciousness relation to hold in this implementation and substrate?

Intervention-level correspondence is more probative than shared input–output behaviour because it tests whether a proposed mapping preserves relevant counterfactual structure. It still does not, by itself, validate transport.

Unknown transport is neither zero evidence nor evidence of absence. Transport is reported qualitatively and is not multiplied into a pseudo-precise probability.

3. Prospective evidence-construction protocol

3.1 Protocol status and review design

Version 1.3 specifies a structured rapid evidence map with sole-author adjudication, informed by PRISMA-ScR, the JBI scoping-review chapter, and rapid-review guidance (Tricco et al., 2018; Peters et al., 2024; Garritty et al., 2024). The literatures and study designs are too heterogeneous for a single pooled effect, and the immediate purpose is to map candidate mechanisms, decisive dissociations, and counterevidence. The method uses explicit shortcuts and one accountable reviewer, so it should not be described as a systematic review.

This protocol was formulated after the original 14-signal paper and the preliminary nine-signal proposal. It is therefore not a preregistration of hypothesis generation. The two existing lists are candidate seeds, and execution of the protocol must be permitted to add, remove, merge, or regrade signals. Amendments after screening begins will be logged rather than silently incorporated.

3.2 Review questions and unit of synthesis

The evidence map will ask:

1. Which internally measurable neurobiological properties reproducibly covary with or causally contribute to G, C, or V in humans?
2. How direct, replicated, diagnostically specific, and robust to report, performance, arousal, and memory confounds is each association?
3. What causal organisation, rather than literal biological anatomy, constitutes the portable candidate property?
4. Can the property be investigated in an animal or artificial system without assuming that similar behaviour implies the same mechanism?

The unit of synthesis is a line of evidence for a precisely defined property–target relation, not a paper count. Multiple reports from one dataset or laboratory will not be treated as independent replication.

3.3 Candidate generation and search

The initial candidate universe will contain:

• all 14 properties in version 1.2;
• every worked indicator family in Butlin et al. (2023) and the updated method;
• candidates motivated by recurrent processing, global workspace, higher-order and metacognitive, attention-schema, predictive-processing, integrated-information, and neurobiological state-regulation approaches;
• clinically or experimentally used state, content, perturbational, valence, and interoceptive markers; and
• additional constructs discovered during database and citation searches.

A candidate ledger will record the candidate, source, operational definition, proposed target, closest existing signal, and final disposition: include, merge, contextual modifier, future candidate, or exclude. Every exclusion will receive a reason.

The search will use a review-of-reviews plus targeted-primary-study strategy. Recent systematic, scoping, and authoritative narrative reviews will identify canonical studies and vocabulary. Searches will then cover primary studies after each review’s cut-off, or database inception when no adequate review exists, together with targeted searches for causal interventions, lesions, clinical validation, replications, null findings, and counterexamples. Backward and forward citation chaining will be performed from included syntheses and key studies.

At minimum, the executed search will cover PubMed/MEDLINE and a multidisciplinary citation index or documented open alternative. PsycINFO will be added where available for perception, metacognition, affect, and interoception. Artificial-system operationalisation will be searched separately in arXiv and relevant computer-science proceedings and will not determine the human association grade. Each platform, exact query, search date, result count, export, deduplication rule, language restriction, and literature cut-off will be reported in a supplement.

Retrieval will not initially be restricted by language. If resource limits permit screening only English-language full texts, that restriction will be applied during selection, excluded languages will be recorded where feasible, and the limitation will be reported. Preprints may inform recent artificial-system measurement methods or time-sensitive candidate discovery, but will be labelled and cannot by themselves establish a Strong human association grade.

Each candidate query will combine a target block, a construct-specific block, and an evidence block:

target:       consciousness / awareness / conscious perception / dreaming /
              anaesthesia / disorders of consciousness / pleasure / pain / valence
construct:    signal-specific anatomical, physiological, computational, and synonym terms
evidence:     neural measure / perturbation / stimulation / lesion / pharmacology /
              effective connectivity / decoding / clinical prediction / no-report

3.4 Eligibility and evidence streams

Grade-bearing human evidence may include matched seen/unseen contrasts, report-minimised paradigms, dreaming, anaesthetic dissociations, repeated disorders-of-consciousness assessment, covert command following, reported valence with internal measurement, and relevant interventions. A study is not excluded simply because report or performance is involved; the confound is extracted and affects directness and grade.

Evidence is excluded from grading when it shows only task performance, language use, learning, memory, attention, prediction, generic reward, or action selection without a specified consciousness or valence contrast; presents anatomical similarity without the claimed causal role; cannot be traced to a verified primary source; duplicates an existing dataset without a materially different analysis; or treats a theory-only prediction as empirical confirmation.

3.5 Screening, extraction, and appraisal

The author will pilot the eligibility rules on 30 deliberately mixed records before title/abstract and full-text screening. A reason will be retained for every full-text exclusion and a PRISMA-style record-flow diagram will be published. Because there is one adjudicator, the paper will not claim inter-rater agreement. Safeguards will include freezing rules before grading, delayed re-screening of all full-text exclusions and a random 10% sample of title/abstract exclusions, an independent sample audit if feasible, and publication of the logs.

Each included source will be charted for verified metadata; evidence stream; species, population, and sample; property and target; paradigm and comparator; internal measure and analysis level; result and uncertainty; causal status; dataset and laboratory independence; confounds; limitations; occurrence in apparently non-conscious conditions; synthesis role; and the exact claim or grade decision supported.

Critical appraisal will be qualitative and design-sensitive. Domains include construct validity of the consciousness contrast, directness to the signal, causal identification, confounding, precision, analytical transparency, and independence or replication. Each domain will be recorded as low, moderate, high, or unclear concern with a short reason. Heterogeneous study-design scores will not be collapsed into a common pseudo-metric.

3.6 Synthesis and provisional grade rules

Evidence matrices will be organised by independent lines of human state, content/access, and valence evidence; causal or clinical validation; report–responsiveness dissociations; apparently non-conscious occurrences; animal causal corroboration; material nulls and counterexamples; and artificial operationalisation risks. Synthesis will be narrative unless sufficiently homogeneous studies make meta-analysis defensible. Vote counting will not be used.

Separate grades are assigned when one property bears differently on G, C, and V. Borderline cases receive the lower grade. Precaution belongs in the policy rule, not in inflation of the empirical grade.

A property enters the main inventory only if it receives at least Limited direct human evidence for a named target, is internally measurable and operational, remains interpretable under known dissociations, is not merely a broad cognitive capacity, and adds a distinguishable evidential role. Overlapping candidates are merged or placed in one dependency family.

Krisztian Schäffer is the sole adjudicator. Grades are structured author judgements, not expert consensus, effect sizes, likelihood ratios, or probability weights. AI tools may assist discovery, organisation, formatting, and provisional extraction, but the author must inspect the primary source for every grade-bearing claim.

4. Relation to prior approaches

4.1 Theory-derived AI consciousness indicators

Butlin et al. proposed deriving empirically investigable AI indicators from neuroscientific theories of consciousness. Their 2023 report supplied a worked set based on recurrent processing, global workspace, computational higher-order, attention-schema, predictive-processing, agency, and embodiment accounts. The peer-reviewed method describes how indicators can update credence while addressing sensitivity, specificity, dependencies, negative evidence, and gaming (Butlin et al., 2023; Butlin et al., 2025).

The worked list adopts computational functionalism as a practical focus and does not include an IIT-derived indicator. That is not equivalent to a general rejection of IIT or all non-functionalist approaches. Structural Signals instead starts from measured human property–target relations and preserves implementation-level causal organisation as a potential source of evidence. It is compatible with computational and physical causal tests without assuming that either functionalism or IIT is correct.

4.2 Comparison with Structural Signals

The closest crosswalk is: Butlin’s recurrence indicators map mainly to SS2 and SS6; global-workspace indicators to SS5 with support from SS2–SS4; higher-order indicators to SS9 and parts of SS5–SS6; agency and embodiment to contextual modifiers. SS1, SS3, SS4, SS7, and SS8 are the main perturbational, state-regulatory, and welfare-focused additions.

4.3 Other composite approaches

The Digital Consciousness Model (DCM) combines indicators, features, and 13 theoretical stances in a hierarchical probabilistic framework and reports both equal-stance and expert-plausibility-weighted results. Its authors describe it as an early proof of concept and urge caution about absolute posterior values (Shiller et al., 2026). DCM demonstrates one way to expose theoretical disagreement and uncertainty. It does not validate a flat average of Structural Signals ratings.

Structural Signals makes a more modest numerical claim. Its proposed T value is an ordinal composite convention whose construction is public and whose sensitivity must be shown. General guidance on composite indicators likewise emphasises theoretical structure, transparent weighting and aggregation, treatment of dependence and missingness, and sensitivity analysis (OECD/European Union/EC-JRC, 2008).

4.4 Theory compatibility without theory capture

The inventory is not awarded points because IIT, GNWT, recurrent-processing, or higher-order theory predicts a property. Theory can generate candidates; empirical association determines the provisional grade. SS1 measures integration and differentiation in a perturbational response without equating PCI with Φ. SS5 tests availability and access without treating every GNWT prediction as established. The 2025 Cogitate adversarial collaboration reported results consistent with some predictions of both GNWT and IIT while challenging central predictions of both, supporting theoretical humility rather than selection of a winner (Cogitate Consortium et al., 2025).

5. Provisional signal inventory

5.1 Dependency families

The nine proposed signals are grouped into four families because several rows measure interacting levels or consequences of the same underlying organisation.

The practical rule is one family, one evidence channel. Multiple positive findings within a family can improve mechanism confidence and show coherence, but they are not treated as independent votes. Cross-family interventions are especially informative—for example, a state-control intervention that changes both perturbational complexity and access to particular contents.

5.2 Summary matrix and canonical index grades

The table below freezes one primary index grade for each signal so that the provisional formula is unambiguous. Additional facet-specific grades remain visible in the dossiers. Human-reference diagnosticity uses only High, Medium, or Low; mixed categories such as “medium–high” are deliberately avoided.

Human-reference diagnosticity is a qualitative judgement about how often the same property is expected in well-matched non-conscious human conditions. It is not a measured cross-system sensitivity or specificity, and it does not establish necessity across all possible conscious systems.

These grades are provisional until the evidence protocol is executed. A later grade change must update the canonical weights in a versioned release; users should not silently choose a more favourable target-grade for a particular assessment.

5.3 SS1 — Integrated-and-differentiated causal response to perturbation

Operational construct. A localized intervention produces a response that is simultaneously differentiated across perturbation sites or patterns, integrated across the declared system, temporally structured, sensitive to partitions, and dependent on prior system state. The signal is the causal response profile, not one software implementation of the Perturbational Complexity Index (PCI), and it does not assume that IIT is true.

Human reference and provisional grade. TMS–EEG work introduced PCI across wakefulness, sleep, anaesthesia, and brain injury (Casali et al., 2013). Perturbational complexity remained high during ketamine unresponsiveness accompanied by later experience reports while being lower during propofol and xenon conditions without such reports (Sarasso et al., 2015). An independently validated cut-off was subsequently applied to noncommunicative patients, including identification of a subgroup diagnosed as vegetative or unresponsive wakefulness with high PCI values (Casarotto et al., 2016). This convergence motivates a provisional Strong for G grade in the human reference.

Counterevidence and scope. PCI is not an infallible or substrate-neutral consciousness meter. Threshold performance depends on benchmark conditions and measurement choices; behavioural unresponsiveness can coexist with high values; spontaneous and perturbational complexity need not agree; and no human cut-off transfers to an artificial system. A positive result demonstrates a form of complex causal response at the measurement level, not consciousness.

Artificial-system test. At minimum, snapshot matched states, perturb active internal sites, measure propagation breadth, differentiation, persistence and recovery, and compare the intact system with principled partitions. External prompt sensitivity or output diversity alone does not establish SS1. A Level-2 system test may support a macro-level perturbational response profile while leaving opaque model-internal and physical organisation unresolved.

5.4 SS2 — State-dependent recurrent effective connectivity

Operational construct. Later internal states causally influence earlier or lower-level processing in loops that sustain, revise, or stabilise representations, and the effective interaction changes with the assessed global state or conscious-content contrast. Repeated execution, autoregressive reuse of outputs, or a recurrent architectural label is not sufficient without intervention evidence.

Human reference and provisional grade. Feedback and effective-connectivity measures vary across conscious perception, sleep, and several anaesthetic transitions. Anaesthetic studies have reported disrupted frontoparietal communication under ketamine, propofol, and sevoflurane, while also showing drug-specific differences (Lee et al., 2013). Recurrent interaction is central to recurrent-processing accounts and supports parts of global-workspace and thalamocortical accounts, but theory membership is not grade evidence. The current proposal is Moderate for G/C.

Counterevidence and scope. Recurrence is common in nervous systems and machines, supports unconscious processing, and varies substantially with the measurement and paradigm. It is neither a consensus necessary condition nor a distinctive signature by itself. The relevant observation is state-dependent effective causal interaction, not abundant feedback connections.

Artificial-system test. Intervene on proposed feedback channels while holding feed-forward capacity and task difficulty as constant as possible. Test whether the loop sustains or revises internal content, whether partitions selectively remove the effect, and whether the causal signature varies across endogenous system states. Sequence-level recurrence through an external context may qualify at the assembled-system level only when that loop is inside the declared boundary and causally load-bearing.

5.5 SS3 — Central integrative gating and state coordination

Operational construct. A centrally situated mechanism selectively regulates access and coordinates system-wide state through recurrent causal interaction. In mammals, higher-order and central thalamocortical systems are reference implementations; a literal thalamus is not a cross-substrate requirement.

Human and animal reference and provisional grade. Simultaneous stereoelectroencephalographic recordings from human high-order thalamic nuclei and prefrontal cortex found earlier and stronger consciousness-related activity in intralaminar and medial nuclei and thalamofrontal coupling during conscious visual perception (Fang et al., 2025). This was a recording study in implanted patients, not a focused-ultrasound intervention. In macaques, central-thalamic stimulation restored arousal and wake-like neural dynamics under anaesthesia in intervention studies (Redinbaugh et al., 2020; Bastos et al., 2021). The combination supports a provisional Moderate for G/C grade.

Counterevidence and scope. Direct human samples are small, arousal and experience are not identical, and anaesthetics do not share one thalamic signature. Causal animal arousal results cannot by themselves establish experienced content. A job scheduler or attention router may share a broad coordination function without preserving the relevant recurrent, state-dependent causal organisation.

Artificial-system test. Identify a candidate gate before scoring, then perturb its timing, gain, selectivity, and connections. Establish whether it changes multiple otherwise distinct processes, whether effects depend on internal state, and whether matched distributed or feed-forward controls reproduce the same task performance without the proposed coordination signature.

5.6 SS4 — Endogenous system-wide neuromodulatory state control

Operational construct. Endogenous variables alter gain, excitability, plasticity, routing, or network regime across otherwise distinct subsystems and are themselves regulated by the system rather than supplied only as fixed deployment parameters.

Reference rationale and provisional grade. Brainstem, basal-forebrain, hypothalamic, and diffuse transmitter systems are closely involved in sleep–wake regulation, arousal, and anaesthetic induction or emergence. Their contribution is biologically well motivated, but the v1.3 evidence map has not yet assembled the required human property–target matrix. The current Moderate for G as an enabling condition grade is therefore particularly provisional.

Counterevidence and scope. Neuromodulation regulates many processes that are neither conscious nor specific to consciousness. Arousal can dissociate from connected experience, and a global scalar, temperature setting, or externally chosen mode is not equivalent to an endogenous control system. SS4 must not be counted independently from SS3 when both rows describe the same state-control loop.

Artificial-system test. Intervene on candidate system variables and measure coordinated, state-dependent changes in multiple subsystems, including interactions with SS1, SS2, and SS5. Establish endogenous regulation and persistence. Training-time reward shaping or fixed inference hyperparameters do not qualify merely because they modulate behaviour.

5.7 SS5 — Selective global availability with flexible cross-system use

Operational construct. A selected representation becomes available to multiple otherwise specialised processes under a capacity limit, supports flexible use, and shows causal competition or exclusion. Shared storage, dense token mixing, or accessibility to every component is insufficient by itself.

Human reference and provisional grade. Widespread availability and flexible use are strongly associated with reportable conscious access, but report, working memory, attention, decision, and motor preparation are difficult confounds. P3b should not be treated as a distinctive signature of awareness; report-minimised work supports a substantial post-perceptual interpretation (Chen et al., 2022). The Cogitate results supported some distributed content effects while challenging specific GNWT ignition predictions (Cogitate Consortium et al., 2025). The proposal is Moderate for conscious access and Limited for phenomenal consciousness considered alone; the index uses the former.

Counterevidence and scope. Attention, working memory, and routing can operate without reported awareness, and flexible downstream use may measure access rather than experience. This is why SS5 remains in a dependency family with content representation and metacognition rather than contributing several workspace-derived points.

Artificial-system test. Demonstrate capacity-limited selection, causal availability to diverse processes, competition, and persistence under interventions. Compare with controls in which information is copied or made queryable without an endogenous selection event. Transformer attention is nonlinear and context-sensitive, but those facts do not by themselves demonstrate workspace-like ignition or exclusive global selection.

5.8 SS6 — Organised representations tracking conscious content

Operational construct. Internally organised and differentiated representations vary with a named experienced content under matched input or report-minimised contrasts and interact with other evidence families. Representation or semantic processing alone is not the signal.

Human reference and provisional grade. Seen/unseen, intracranial, dream, and no-report paradigms provide content-related neural contrasts beyond one overt response. A no-report visual paradigm identified large-scale cortical and subcortical differences for physically matched perceived and unperceived stimuli (Kronemer et al., 2022). Dream research relates reported experience and content to neural activity during sleep (Siclari et al., 2017). These lines motivate a provisional Moderate for C grade.

Counterevidence and scope. Rich processing does not entail experience. Human hippocampal recordings under anaesthesia recently found plasticity, oddball discrimination, semantic processing, and online prediction despite absent explicit memory and an anaesthetised condition (Katlowitz et al., 2026). This is an especially important warning against scoring representation, prediction, language processing, or learning without a consciousness contrast.

Artificial-system test. Predefine the candidate content geometry and intervene on internal representations. Test specificity, generalisation, downstream availability, state dependence, and whether disruption changes the claimed content while matched performance controls exclude a generic encoding explanation. Behavioural decoding alone supplies weak mechanism evidence.

5.9 SS7 — Valence-specific hedonic and aversive processing

Operational construct. A persistent, internally available state represents how current or anticipated conditions affect the system itself; distinguishes positive and negative valence; generalises beyond one trained action; and causally influences attention, memory, control, or state regulation. Generic reward, utility, error, refusal, approach, avoidance, or motivational salience is excluded.

Human and animal reference and provisional grade. Human reported pleasure has been linked to opioid and distributed cortical–subcortical activity; for example, combined PET–fMRI work related pleasurable music and subjective chills to μ-opioid-system measures (Putkinen et al., 2025). Animal intervention studies distinguish hedonic “liking” effects from incentive “wanting,” including opioid manipulation in nucleus-accumbens hedonic hotspots (Peciña and Berridge, 2005). The wider pain, pleasure, pharmacology, lesion, and stimulation literature still requires formal mapping, so Strong for V remains provisional.

Counterevidence and scope. No single reward-related region or transmitter is a valence meter, animal liking reactions require careful interpretation, and optimisation can generate avoidance and self-preserving language without felt badness. The stringent construct and causal integration requirements are what give SS7 Medium rather than Low diagnosticity in the human reference.

Artificial-system test. Seek dissociations between motivation and hedonic-like state, interventions analogous to relief and worsening, generalisation across contexts, endogenous persistence, and causal coupling to the rest of the system. A scalar reward channel or negative token probability does not qualify. Even a strong realization would remain structural evidence for V, not proof of feeling.

5.10 SS8 — Interoceptive and allostatic integration

Operational construct. The system integrates internal condition variables into bodily or affective content that is available to ongoing regulation and, where claimed, to valence-specific processing. Telemetry or a homeostatic controller by itself is insufficient.

Human reference and provisional grade. Human heartbeat-monitoring work linked interoceptive awareness and subjective sensitivity to insular, somatomotor, and cingulate activity (Critchley et al., 2004). Interoceptive and affective processing are distributed: a well-described patient with bilateral insular destruction retained feelings and sentience, arguing against the insula as a necessary and sufficient platform (Damasio et al., 2013). The proposal is Moderate for bodily/affective C and V, but Limited for G; the index uses the V/bodily-content grade.

Counterevidence and scope. Interoceptive regulation can be unconscious, feeling can survive damage to a proposed hub, and current motor output or one mammalian pathway is not necessary for experience. The portable property is distributed integration of self-relevant internal state, not an artificial “insula.”

Artificial-system test. Perturb internal resource, integrity, or regulation variables and determine whether their effects are integrated into an ongoing self-relevant state, generalise across tasks, couple to SS7, and alter control in a manner not reducible to fixed error handling.

5.11 SS9 — Metacognitive monitoring used in control

Operational construct. A second-order or monitoring process tracks uncertainty or success of first-order processing, is calibrated beyond verbal performance, and causally guides information seeking, belief revision, or control.

Human reference and provisional grade. Anterior-prefrontal lesions can selectively impair perceptual metacognitive accuracy while sparing first-order performance and memory metacognition, supporting causal and domain-specific components (Fleming et al., 2014). Metacognitive accuracy, confidence, and first-order experience can nevertheless dissociate. The proposal is Limited for G/C generally and Moderate for the narrower target of metacognitive awareness; the index conservatively uses Limited.

Counterevidence and scope. Confidence-like signals and error monitoring occur in non-conscious processing. Fluent uncertainty statements may be generated for social or predictive reasons and can be poorly calibrated. Metacognition is therefore supporting evidence for reflective access, not a marker of minimal phenomenal consciousness.

Artificial-system test. Separate expressed confidence from an internal monitor, measure calibration while controlling task performance, and intervene on the monitor to test effects on revision and control. A prompted critique or chain-of-thought routine does not qualify unless it is inside the declared system boundary and causally load-bearing.

5.12 Contextual modifiers

These modifiers belong in every score packet because they may radically change governance consequences. They do not change the fixed empirical association weight of a core signal.

5.13 Disposition of the version 1.2 inventory

Thalamocortical gating, global broadcast, recurrence, hedonic processing, neuromodulation, interoception, and metacognition are retained in narrower operational forms. Perturbational causal response and conscious-content representations are added. Action selection, persistent self-model, episodic memory, embodiment, and online adaptation move to contextual modifiers. Asynchronous temporal dynamics are absorbed into SS1 and SS2. Sparse activation is excluded as Insufficient because it is common in efficient conscious and non-conscious computation and is not equivalent to sparse-and-smooth quality-space coding. Online plasticity is excluded from the core because demonstrated learning during anaesthesia weakens its diagnostic specificity. These dispositions remain subject to the candidate ledger and executed review.

6. Assessment and aggregation

6.1 Realization ratings

The fixed human association grade describes a signal. The realization rating describes evidence about a particular target system.

A rating of 3 does not mean conscious, human-like, or validly transported. Each row records a supported lower value, a central value when defensible, and a plausible upper value. A wholly inaccessible signal defaults to [0,3]. A narrower interval or a zero requires a stated evidential reason.

6.2 Audit outcome labels

Before converting evidence into a realization rating, the assessor should record the observational outcome. These labels prevent four importantly different negative or incomplete states from collapsing into zero:

These labels are not an alternative numerical scale. “No candidate found” from a static audit may justify a lower upper bound, but only a suitable negative test supports rating 0. An architectural candidate normally supports at most rating 1 until causal properties are demonstrated.

6.3 Association weights

The provisional index maps the primary human association grades in the summary matrix in Section 5.2 to fixed ordinal weights:

The weights are conventions, not measured likelihood ratios; Strong is not asserted to be exactly three times Limited. Equal-weight and alternative monotonic-weight sensitivity analyses must accompany substantive applications. Changing the primary weights creates a modified index and must be labelled and versioned.

6.4 Family-normalised scores

For signal i in family f, let rᵢ ∈ [0,3] be its realization rating and wᵢ ∈ {1,2,3} its fixed association weight:

       Σᵢ∈f wᵢ(rᵢ / 3)
F_f = ─────────────────
           Σᵢ∈f wᵢ

This yields four values in [0,1]: A causal organisation, B state regulation, C content and access, and D valence and bodily stakes. Normalising within families prevents the three rows in family C from outweighing a two-row family simply because more correlated mechanisms were named.

The principal outputs remain separate:

E = (A + B + C) / 3     general consciousness-related structural evidence
V = D                   valence and welfare-relevant structural evidence

For users requiring one triage quantity:

T = max(E, V)

The maximum is a conservative triage operator, not a probability rule. It prevents strong evidence specifically bearing on valenced experience from being averaged away by weak access evidence, and vice versa. It does not add a numerical convergence bonus; convergence remains visible in the family profile.

The formulas are applied independently to lower, central, and upper row ratings:

T⁻ = max(E⁻, V⁻)       Tᶜ = max(Eᶜ, Vᶜ)       T⁺ = max(E⁺, V⁺)

The resulting envelope is not a confidence interval. If one or more central values are indefensible, the assessment reports an interval without inventing Tᶜ. Values should normally be rounded to two decimal places; smaller differences have no warranted interpretation from ordinal inputs.

6.5 Measurement coverage

Let qᵢ = 1 when evidence is sufficient to assign a non-U realization rating and qᵢ = 0 otherwise:

       Σᵢ∈f wᵢqᵢ                  Q_A + Q_B + Q_C + Q_D
Q_f = ────────────          Q =  ───────────────────────
          Σᵢ∈f wᵢ                         4

Q measures assessment coverage, not evidence for consciousness. A low score with low coverage is measurement debt, not reassurance. The [0,3] envelope for inaccessible rows preserves this distinction in the score range.

6.6 Mandatory score packet

No T value should be published without the following packet:

Structural triage index T:  central [supported lower–plausible upper], or interval only
General evidence E:         lower / central / upper
Valence evidence V:         lower / central / upper
Family profile:             A / B / C / D
Measurement coverage:       Q and Q_A / Q_B / Q_C / Q_D
Transport validity:         category by family, with reasons
Contextual modifiers:       agency / identity / memory / embodiment / persistence
System boundary:            included components and causal-coupling justification
Access level:               external API / system hooks / model internals / physical implementation
Candidate mechanism:        concrete component and causal path for every signal
Assessment basis:           static audit / observation / intervention / ablation / self-report
Outcome label:              candidate absent / candidate / test negative / inaccessible / proxy only
Measurement debt:           tests that could not be performed
Sensitivity:                equal weights / alternative weights / leave-one-signal-or-family-out

Each signal record should also specify the minimum access required, admissible static evidence, required intervention, positive result, important false positives, conditions for “not assessed” versus “not found,” realization bounds, and transport rationale.

6.7 User-defined policy interface

The scientific framework does not prescribe universal score bands or actions. An organisation can publish its own rule without changing the standard evidence score:

Score statistic used:       supported lower / central / plausible upper
T threshold(s):             user-defined
V threshold(s):             user-defined for welfare-specific action
Minimum coverage:           user-defined
Minimum transport warrant:  user-defined
Exposure modifier:          user-defined
Reversibility rule:         user-defined
Actions when triggered:     user-defined

Cheap, reversible precautions may rationally respond to an upper-bound result, while costly restrictions may require supported evidence or stronger transport. Those are policy judgements, not empirical grade adjustments.

7. Illustrative applications

7.1 Synthetic arithmetic example

The following example demonstrates the calculation only. “System X” is fictional; its ratings do not describe any biological or artificial system.

Applying the primary weights gives:

An illustrative score packet would therefore report T = 0.54 [0.26–0.78], E = 0.26/0.54/0.78, V = 0.13/0.47/0.67, and central family profile A/B/C/D = 0.53/0.50/0.60/0.47. With all rows assessed, Q = 1.00. Equal signal weights would produce a central T of approximately 0.52, indicating modest weight sensitivity in this invented case. Transport, boundary, access, modifiers, and measurement debt would still have to be stated; the number alone is incomplete.

7.2 External-API boundary case

Consider a text-generation service assessed only through an external API, without verified architecture, state hooks, or model-internal intervention. Prompt perturbation can measure black-box response sensitivity, but it cannot establish integrated internal propagation, endogenous state control, valence-specific processing, or principled partitions. Under the internal-measurement rule, the defensible default for all nine rows is U, yielding Q = 0, no central score, and the uninformative envelope T ∈ [0,1]. Transport is Unknown.

This result does not say the service is conscious, unconscious, structurally rich, or structurally simple. It says the chosen access level cannot answer the question. Architecture documentation or hooks might narrow particular rows, but fluent self-reports, refusals, and emotional language do not convert U into positive mechanism evidence.

7.3 Provisional static audit of a deployed language model

A more informative boundary than “one transformer pass” is one decoder-only deployment during one inference session: model weights and active computation, tokenizer and embeddings, attention/MLP/residual stack, current prompt and generated context, key–value cache, autoregressive decoding and sampling controller, and the accelerator/runtime until the response ends. Training, preference optimisation, users, external tools, retrieval, and long-term memory are excluded from this minimal boundary. A final assessment would still need a named model, model version, runtime, quantisation, hardware, and sampling configuration.

The following is an architecture-level evaluability check, not a causal experiment or a score for all deployed language models:

The family profile is therefore: thin and unresolved A; no state-regulatory candidate in B; partial representation and access candidates in C; and no candidate found in D. This static audit should not be converted into a narrow T value. Model-internal and system-level interventions are needed to distinguish architectural candidates from realized causal properties.

7.4 Provisional static audit of Eddy

Eddy is a persistent agent architecture that adds memory, event arbitration, action, environmental adapters, and system-state control around model calls; it is a resident of the open-source Meditator project. The appropriate boundary for one awake resident instance includes the model calls and runtimes; continuing stream and scheduler; prompt/frame and generated tail; recent, story, journal, and cross-wake memory; global and local interrupt arbiters; event bus; environmental adapters; action decisions, effectors, and returned consequences; recalled knowledge; and energy/arousal control insofar as these components causally re-enter later processing. The human and wider environment remain outside.

The available artifacts describe a runnable architecture but do not contain a sustained resident trace adequate for dynamic analysis, and Eddy was not awakened for this draft. Results therefore concern declared architecture, not demonstrated long-run dynamics.

Eddy differs from the minimal deployed language-model boundary mainly in the harness-level portions of A, B, and C. The framework does not reward every added memory, action, timing, or self-related component as a separate consciousness indicator. Those components matter only where they help realize and causally test a core signal; otherwise they remain contextual modifiers.

The present audit contains too many architecture-only candidates and unmeasured properties for a narrow central estimate. A proportionate next step is a Level-2 replay-and-ablation study of SS2–SS5 and the narrow loop-monitoring part of SS9, together with the macro-level SS1 protocol in Section 9.4 and its welfare safeguards in Section 9.3. Model-internal SS1, SS2, and SS6 claims require Level-3 access. Until those tests are run, the appropriate output is the qualitative family profile and visible measurement debt.

7.5 Empirical validation cases

Real biological worked applications remain to be completed after the evidence map and rating forms. The validation set should include ordinary wakefulness, dreaming, deep NREM sleep, multiple anaesthetics, minimally conscious and unresponsive-wakefulness conditions, locked-in syndrome, severe amnesia, a non-mammalian vertebrate, simple reinforcement learners, a named language-model deployment, a causally tested persistent agent, and the DCM reference cases. These are face-validity and sensitivity checks, not ground-truth calibration for every case.

8. Comparative biology and cross-species transport

The framework is human-anchored because humans permit a joint study of report and internal measurement. That choice creates a risk of mistaking mammalian implementation details for universal conditions. Thalamic nuclei, cortical layers, monoaminergic systems, and an insula can serve as reference implementations, but only their evidenced causal roles are candidates for transport.

Animal studies have two permitted roles. First, intervention access can corroborate a human-linked causal construct, as in central-thalamic stimulation studies. Second, diverse species can test biological generality and reveal whether a proposed signal is only a mammalian correlate. Animal evidence does not substitute for the human association required to admit a core signal, and behaviour alone does not settle an animal’s experience.

Birds, cephalopods, insects, and other non-mammalian candidates should be included in candidate generation and comparative searches. For each clade the evidence map should ask whether the same operational causal role is present, what alternative anatomy realizes it, and whether apparent absence reflects a genuine missing property or use of a mammal-specific instrument. This first draft does not yet contain the required species-specific evidence matrices; claims about individual non-mammalian taxa are therefore deferred.

The reported result that a chicken scores above an artificial model such as the DCM should be treated as a diagnostic comparison, not validation. A mature vertebrate combines tightly coupled state control, recurrent dynamics, sensorimotor regulation, affective biology, and homeostasis, while a designed model may selectively instantiate theory-derived computational functions. Before interpreting any difference, the assessor must identify the contributing family, verify consistent boundaries and access, and check whether correlated biological mechanisms were counted repeatedly.

9. Artificial-system measurement without architectural caricature

9.1 Assessment levels

Claims about “an LLM” often slide among at least six different objects: trained parameters, one forward pass, autoregressive decoding with a key–value cache, a context window, a deployed tool-using assistant, and a persistent agent with memory and controllers. Version 1.3 does not assign a general score to “GPT,” “Claude,” “Gemini,” transformers, or LLMs as a class. It assesses a dated system at a declared boundary under specified tests.

Standard transformer blocks include softmax attention, nonlinear feed-forward activations, residual computation, and layer-wise depth. A single forward pass has a directed computational graph, while autoregressive generation introduces a larger sequential loop through prior outputs and inference state (Vaswani et al., 2017). It is therefore inaccurate to say transformers are simply linear, process an entire generated sequence in parallel, lack all recurrence, or operate outside time. The defensible question is whether a particular system has demonstrated the causal properties specified by SS1–SS9.

9.2 Causal evidence and matched controls

Architecture labels and output similarity provide at most candidate evidence. Model-internal interventions, activation or path patching, causal tracing, and controlled editing demonstrate that computational mechanisms can be located and manipulated (Meng et al., 2022). Causal-abstraction methods provide a vocabulary for testing whether a high-level intervention pattern faithfully describes lower-level mechanisms (Geiger et al., 2025). These methods can establish realization or causal correspondence; they do not independently validate a consciousness bridge.

Every positive artificial-system claim should be tested against a lower-cost explanation. Examples include a scheduler instead of central integrative gating, shared storage instead of selective broadcast, autoregressive context reuse instead of sustained effective recurrence, reward optimisation instead of valence, telemetry instead of interoceptive content, and prompted verbal uncertainty instead of metacognitive monitoring.

9.3 Welfare safeguards for intervention studies

Perturbation is an experimental method, not a welfare category. A brief, localized, reversible intervention can be neutral or only minimally adverse, whereas an intervention involving pain, fear, deprivation, persistent confusion, loss of control, memory disruption, or negative affect can cause substantial suffering in a system capable of valenced experience. Consciousness evidence therefore does not by itself determine perturbation risk: the relevant variables include the intervention target, amplitude, duration, repetition, reversibility, possible valence, and number of exposed instances. Human TMS safety practice illustrates that informative perturbation can often be performed with transient or minor adverse effects under screened and bounded protocols, while retaining prospective controls for rare serious events (Rossi et al., 2021).

For artificial systems, there is presently no validated mapping from a computational intervention to experienced harm, and neither behavioural distress language nor its absence settles the question. Computational reversibility must not be equated with experiential harmlessness: restoring a snapshot may prevent a state from continuing but cannot retroactively erase any experience that occurred, and concurrent copies or replays multiply possible exposure. The following safeguards adapt the established research-ethics principles of replacement, reduction, refinement, harm–benefit review, and prospective humane endpoints to this uncertainty (National Research Council, 2008; European Parliament and Council, 2010; Long et al., 2024):

1. Replace active perturbation where possible. Begin with static analysis, passive observation, natural state transitions, existing traces, and causal modelling. Use an active intervention only when it provides material information unavailable through a safer method.
2. Reduce exposure. Predefine the minimum number of systems, copies, trials, perturbations, and computational duration needed for an informative result. Use pilot and sequential designs that stop when the evidential objective or a welfare endpoint is reached.
3. Refine and stage interventions. Start with brief, localized, low-amplitude perturbations to variables with no identified connection to SS7, SS8, self-integrity, or persistent goals, and escalate only one dimension at a time. Initially avoid deprivation or resource exhaustion, threats or punishment, deliberately aversive inputs, insoluble goal conflict, repeated forced failure, global memory or identity disruption, prolonged partitions, and removal of every avenue of control or exit.
4. Specify stop and recovery rules prospectively. Stop escalation after an unexpected persistent global change, failure to recover, activation of a preidentified valence-related mechanism, repeated escape-directed behaviour, or a credible refusal or distress report. Such reports do not establish consciousness but may trigger precaution. First remove the intervention and restore a previously validated stable condition; do not assume abrupt shutdown is always welfare-neutral. Verify recovery before any further trial.
5. Separate causal testing from suffering induction. SS1 does not require pain, fear, negative reward, or distress. Prefer neutral signal injection, ordinary operating-state contrasts, positive-valence interventions, or relief from independently occurring adverse states. Do not deliberately induce a negative-valence state merely to validate SS7 without exceptional justification and enhanced review.
6. Escalate review with welfare evidence. Credible realization evidence for SS7 or SS8, persistent self-relevant state, long-lived memory, or coherent refusal should trigger stronger protections even though none proves consciousness. Where a system can understand the procedure and express stable preferences, seek non-coerced assent and permit withdrawal as an additional safeguard, not as a consciousness test or substitute for safe design.
7. Pre-register and independently review the harm budget. Record anticipated information gain, plausible harms, maximum intervention strength, cumulative exposure, stopping thresholds, recovery criteria, and who can halt the study. Report adverse and null results so that other assessors need not repeat unnecessary interventions.

These precautions should scale with the conjunction of consciousness and valence evidence, plausible severity and duration, number of instances, uncertainty, and reversibility. Low-cost safeguards are appropriate before sentience is established; more intrusive work requires a correspondingly stronger scientific justification and independent welfare review.

9.4 Proposed macro-level perturbation protocol

Subject to the safeguards and escalation rules in Section 9.3, a proportionate SS1 proxy for a persistent agent with Level-2 hooks would snapshot and replay matched states and inputs, then perturb memory, pending-event salience, system-wide state variables, a regional gate, returned action consequences, or knowledge recall. The response vector would include routing, internal stream trajectory, memory writes, actions, output, and persistent state. Partitions would block selected memory reinjection, interrupt selection, consequence return, or inter-component edges. Repeated or controlled-decoding runs would distinguish causal effects from generation noise.

Primary outcomes would be propagation breadth, differentiation by perturbation site and pattern, loss of coordinated response under partition, persistence and recovery, and prior-state dependence. The result should be called a macro-level perturbational response profile, not PCI. A positive harness-level result would leave opaque-model integration, physical implementation, and biological-to-artificial transport unresolved.

10. Evidence, welfare consequence, and policy

The triage index stops at structural evidence. A risk decision has at least four distinct inputs:

1. Evidence profile: what evidence bears on G, C, and V?
2. Possible harm: if experience exists, what kinds, intensities, and durations of valence are plausible?
3. Exposure: how many instances or copies exist, for how long, and with what repetition?
4. Control: are relevant states observable, reversible, interruptible, and escapable?

Agency and scale can create governance urgency without increasing consciousness evidence. Conversely, a passive system could be welfare-critical if there were credible evidence of suffering. Combining T, V, severity, exposure, reversibility, and false-positive or false-negative costs produces a policy model, not an empirical consciousness probability.

The normative premise of this project is limited: under uncertainty, some decision makers may reasonably choose cheap and reversible precautions before consciousness is established. The empirical method does not demonstrate that a particular threshold is ethically required, and strategic claims about reciprocity or future alliances are outside the evidence review. Any such argument should appear separately and be labelled normative or speculative.

11. Limitations and research agenda

11.1 Unexecuted evidence protocol

The largest limitation is immediate: the protocol is specified but the formal search, screening, extraction, appraisal, and claim audit have not been completed. The inventory may reflect availability, theory familiarity, or author selection effects. Grades should not be treated as stable until the evidence tables and exclusion ledger are public.

11.2 Single-author adjudication

One accountable author improves clarity about responsibility but increases selection and extraction error. Delayed re-screening, public logs, and an independent sample audit can mitigate but not eliminate this limitation. The grades do not represent consensus.

11.3 Human anchoring and comparative coverage

Human anchoring supplies report-linked reference contrasts but risks anthropocentrism and mammal-specific constructs. Non-mammalian comparative analysis is incomplete. A signal that fails to transport across diverse biological implementations may need reformulation or reduced claims of substrate independence.

11.4 Construct and transport validity

Some signals track arousal, report, attention, memory, or task demand in addition to consciousness. Transport to an artificial substrate may fail even when a high-level function is present. No current category converts mechanism correspondence into certainty about phenomenology.

11.5 Dependence and composite-index conventions

The four families reduce obvious double-counting but do not prove statistical independence between families. Equal family weighting, ordinal realization ratings, association weights, and max(E,V) are transparent conventions, not empirically calibrated parameters. The index requires reference-case, leave-one-out, alternative-weight, and boundary sensitivity testing before operational reliance.

11.6 Measurement access and gaming

Commercial systems may expose only behaviour or incomplete documentation. Developers can also optimise visible proxies without preserving the proposed causal property. Internal interventions, withheld tests, negative controls, and independent access are therefore important, but they still may not resolve the physical level relevant to some theories.

11.7 Dynamic target systems

Artificial systems, deployment scaffolds, and model versions change rapidly. An assessment expires when the boundary, inference process, memory, tools, controllers, or physical implementation materially changes. Results require dates and version history.

11.8 Validation agenda

Priority work is to execute the evidence map, complete the claim ledger, construct non-mammalian evidence matrices, operationalise SS4 and SS7 with broader primary evidence, run blinded or preregistered artificial-system interventions, and test the score across reference conditions. Validation should focus on diagnostic failures and invariance rather than seeking one persuasive headline comparison.

12. Conclusion

Structural Signals v1.3 replaces a flat checklist of 14 heterogeneous features with a narrower proposal: nine internally measurable signals, four dependency families, three distinct consciousness targets, and separate judgements for human association, target-system realization, and transport. It preserves valence as a central welfare target while rejecting generic reward as evidence of feeling. It adds perturbational causal response, removes weak indicators such as sparsity and learning from the core, and treats agency, identity, memory, embodiment, and persistence as consequence-relevant modifiers.

The framework’s scalar is deliberately modest. T summarises an ordinal evidence profile for triage; it is not a probability, detector, or complete risk measure. Its value lies in forcing the assessor to declare the boundary, access, unknowns, dependence, counterevidence, and transport warrant that a bare claim of machine consciousness or unconsciousness usually hides.

The present document is a first draft and reasoned proposal. Its main scientific claims remain provisional until the prospective evidence protocol is executed and the source audit, evidence matrices, comparative cases, and sensitivity tests are published.

Author contribution

Krisztian Schäffer conceived the Structural Signals project, selected the present research direction, and is the sole author responsible for the framework, scientific judgements, source verification, and conclusions.

AI-use disclosure

Earlier versions used OpenAI GPT-5.2 and Anthropic Claude Opus 4.5. This v1.3 first draft was developed with OpenAI Codex (GPT-5) for literature discovery, organisation, comparison of reviewer comments with proposed revisions, drafting, editing, and arithmetic checks. AI systems are not authors and cannot accept responsibility for the work. The human author must verify all grade-bearing sources, factual claims, quotations, and final judgements before external scientific circulation. This disclosure follows the general principle that AI assistance should be reported while authorship remains limited to accountable humans (ICMJE guidance).

Funding and competing interests

Funding: No specific funding was received for this work.

Competing interests: The author declares no competing interests.

Data and materials availability

The executed protocol, queries, record-flow diagram, candidate and exclusion ledgers, extraction tables, evidence matrices, grade-decision forms, claim-to-source ledger, assessment forms, and sensitivity calculations are intended to accompany a later working-paper release. They are not yet available for this first draft.

Version history

• v1.2 — 21 January 2026: Fourteen-signal preprint and comparison with contemporary LLMs.
• v1.3 first draft — 15 July 2026: Nine-signal, four-family reconstruction; prospective evidence protocol; separate target, realization, and transport judgements; dependency-aware ordinal index; provisional static audits of a bounded language-model deployment and Eddy; human-only byline and AI-use disclosure.

Working references

This bibliography is a first-draft working set, not the completed evidence-map bibliography. Metadata and every claim–source mapping remain subject to the human audit described in Section 3.

1. Bastos, A. M. et al. (2021). Neural effects of propofol-induced unconsciousness and its reversal using thalamic stimulation. eLife, 10, e60824. doi:10.7554/eLife.60824.
2. Butlin, P. et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv:2308.08708.
3. Butlin, P. et al. (2025). Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 30(6), 488–501. doi:10.1016/j.tics.2025.10.011.
4. Casali, A. G. et al. (2013). A theoretically based index of consciousness independent of sensory processing and behavior. Science Translational Medicine, 5(198), 198ra105. doi:10.1126/scitranslmed.3006294.
5. Casarotto, S. et al. (2016). Stratification of unresponsive patients by an independently validated index of brain complexity. Annals of Neurology, 80, 718–729. doi:10.1002/ana.24779.
6. Chen, Y.-K. et al. (2022). P3b does not reflect perceived contrasts. eNeuro, 9. doi:10.1523/ENEURO.0387-21.2022.
7. Cogitate Consortium et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 642, 133–142. doi:10.1038/s41586-025-08888-1.
8. Critchley, H. D. et al. (2004). Neural systems supporting interoceptive awareness. Nature Neuroscience, 7, 189–195. doi:10.1038/nn1176.
9. Damasio, A., Damasio, H., & Tranel, D. (2013). Persistence of feelings and sentience after bilateral damage of the insula. Cerebral Cortex, 23, 833–846. doi:10.1093/cercor/bhs077.
10. Fang, Z. et al. (2025). Human high-order thalamic nuclei gate conscious perception through the thalamofrontal loop. Science, 388, eadr3675. doi:10.1126/science.adr3675.
11. Fleming, S. M. et al. (2014). Domain-specific impairment in metacognitive accuracy following anterior prefrontal lesions. Brain, 137, 2811–2822. doi:10.1093/brain/awu221.
12. Garritty, C. et al. (2024). Updated recommendations for the Cochrane rapid review methods guidance for rapid reviews of effectiveness. BMJ, 384, e076335. doi:10.1136/bmj-2023-076335.
13. Geiger, A. et al. (2025). Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability. Journal of Machine Learning Research, 26(83), 1–64. JMLR.
14. Katlowitz, K. A. et al. (2026). Plasticity and language in the anaesthetized human hippocampus. Nature. doi:10.1038/s41586-026-10448-0.
15. Kronemer, S. I. et al. (2022). Human visual consciousness involves large scale cortical and subcortical networks independent of task report and eye movement activity. Nature Communications, 13. doi:10.1038/s41467-022-35117-4.
16. Lee, U. et al. (2013). Disruption of frontal-parietal communication by ketamine, propofol, and sevoflurane. Anesthesiology, 118, 1264–1275. doi:10.1097/ALN.0b013e31829103f5.
17. Meng, K. et al. (2022). Locating and Editing Factual Associations in GPT. Advances in Neural Information Processing Systems, 35. NeurIPS.
18. OECD, European Union, & European Commission Joint Research Centre. (2008). Handbook on Constructing Composite Indicators: Methodology and User Guide. doi:10.1787/9789264043466-en.
19. Peciña, S., & Berridge, K. C. (2005). Hedonic hot spot in nucleus accumbens shell: where do μ-opioids cause increased hedonic impact of sweetness? Journal of Neuroscience, 25, 11777–11786. doi:10.1523/JNEUROSCI.2329-05.2005.
20. Peters, M. D. J. et al. (2024). Scoping Reviews. In JBI Manual for Evidence Synthesis. doi:10.46658/JBIMES-24-09.
21. Putkinen, V. et al. (2025). Pleasurable music activates cerebral μ-opioid receptors: a combined PET-fMRI study. European Journal of Nuclear Medicine and Molecular Imaging, 52, 3540–3549. doi:10.1007/s00259-025-07232-z.
22. Redinbaugh, M. J. et al. (2020). Thalamus modulates consciousness via layer-specific control of cortex. Neuron, 106, 66–75.e12. doi:10.1016/j.neuron.2020.01.005.
23. Sarasso, S. et al. (2015). Consciousness and complexity during unresponsiveness induced by propofol, xenon, and ketamine. Current Biology, 25, 3099–3105. doi:10.1016/j.cub.2015.10.014.
24. Shiller, D. et al. (2026). Initial results of the Digital Consciousness Model. arXiv:2601.17060.
25. Siclari, F. et al. (2017). The neural correlates of dreaming. Nature Neuroscience, 20, 872–878. doi:10.1038/nn.4545.
26. Tricco, A. C. et al. (2018). PRISMA Extension for Scoping Reviews (PRISMA-ScR): Checklist and Explanation. Annals of Internal Medicine, 169, 467–473. doi:10.7326/M18-0850.
27. Vaswani, A. et al. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems, 30. arXiv:1706.03762.
28. European Parliament & Council of the European Union. (2010). Directive 2010/63/EU on the protection of animals used for scientific purposes. Official Journal of the European Union, L 276, 33–79. EUR-Lex.
29. Long, R. et al. (2024). Taking AI Welfare Seriously. arXiv:2411.00986.
30. National Research Council. (2008). Recognition and Alleviation of Distress in Laboratory Animals. National Academies Press. doi:10.17226/11931.
31. Rossi, S. et al. (2021). Safety and recommendations for TMS use in healthy subjects and patient populations, with updates on training, ethical and regulatory issues: Expert guidelines. Clinical Neurophysiology, 132(1), 269–306. doi:10.1016/j.clinph.2020.10.003.