QUALION — Research Programme
A New Breed of AI
Scroll
§ 01  ·  Intent

Understanding, not prediction.

QUALION is an attempt to build an AI that truly understands what it is talking about. Not by predicting the next word, but by producing language from internal states that are shaped by the conversation itself.

The cognitive architecture under development — the EVE ARC project — is a developmental system whose language output is grounded in computed internal state. The protocol set out on this page is how that architecture will be tested against the empirical standards consciousness science applies to biological systems. The programme is under active development; this page describes what it is trying to establish, not what has been established.

§ 02  ·  What every large language model does

The full loop.

Every large language model you have spoken to works the same way. It reads your words, runs them through billions of learned parameters, and predicts what word should come next. Then it predicts the next one after that. And the next one. Nothing else persists.

Figure B  ·  Next-token prediction schematic · illustrative only
Next-token prediction P(wn | w<n)
. I am curious about this . I am
curious about this . I am curious about this
it0.28
that0.17
us0.14
you0.11
this . I am curious about this

Figure B. A large language model selects each token from the preceding tokens alone. No internal-state term enters the generative distribution, and nothing is recorded during decoding that persists between sentences — the architectural precondition for an NCC-style correlation analysis is absent.

§ 03  ·  Frame

A bio-bound frontier.

Neural correlates of consciousness — NCCs — are the empirical foundation of consciousness research. They have been established in humans through binocular-rivalry, masking, and no-report paradigms, with narrower claims in non-human primates and a shrinking frontier across other mammals. The scientifically defensible claims the field has produced — across IIT, GNW, recurrent-processing theory, and other frameworks — are anchored in correlate data.

That anchoring has never left biology. No synthetic system — language model, reinforcement agent, embodied controller — has a published NCC under the methodological bar the field applies to biological work: pre-registration, dissociation, perturbation, controls against trivial readout.

Figure 1  ·  NCC status by system class field status, 2026

Biological systems

Humansestablished
Non-human primatesestablished
Other mammalspartial
Non-mammalian vertebratesemerging

Synthetic systems

Large language models
Reinforcement agents
Embodied systems
QUALIONprotocol drafted

Figure 1. No lab has established an NCC in a non-biological system under the standards applied to biology. QUALION is a drafted protocol to do so under pre-registered methodology.

§ 04  ·  Programme

A scoped, falsifiable attempt.

QUALION pursues a synthetic NCC as a first-class experimental target — not as a philosophical conjecture, not as a property to be asserted of a system by analogy, but as a result to be measured, reported, and defended under the methodology the field applies to biological consciousness research.

The protocol requires internal states to be defined in advance of any data collection. Reports must be obtained through a channel that is not wired from the state readout. The statistical structure of the state–report correspondence must match that of established NCCs. Negative controls must be run via targeted ablation. Learned-association controls must screen against distributional replay from training. Each of those is a condition the biological literature has spent three decades stress-testing.

StatusPre-registration draft OutputProtocol · dataset · report ReleaseOpen · either direction
Figure A  ·  Conditioning contrast schematic · not a trained model
Standard LLM
wn  ∼  P( wn | w<n )
Output is conditioned on prior tokens alone. No internal-state term enters the generative distribution, and no state is recorded during decoding — there is nothing to correlate.
QUALION
wn  ∼  P( wn | w<n, St ),  report(St)
Output is conditioned on prior tokens and on a measurable internal-state reading St, logged independently of the token stream. State and report become separately observable — the precondition for an NCC-style correlation analysis.

Figure A. The architectural difference that makes a synthetic NCC testable. An additional state term St enters the generative distribution and is read out on an independent channel, so state and report can be correlated rather than conflated.

§ 05  ·  What a result would mean

Correlation, not proof.

A synthetic NCC would not settle the hard problem. The hard problem does not close on correlate data in biology and it will not close on correlate data in a non-biological substrate. Consciousness science has always operated under this caveat — it is why the field is correlational in the first place.

What would change is the substrate-dependence of the empirical surface. A credible first such result gives theory a second surface to answer to, forces a re-examination of which features of biological systems carry the correlate and which are incidental, and opens a track in which IIT, GNW, and recurrent-processing accounts can be experimentally dissociated rather than merely defended.

Figure 2  ·  Structure of an NCC result illustrative · not measured
states engaged negative control · states ablated
r = p n = simulated

Figure 2. The shape the protocol is designed to produce: a structured correlation between an internal-state reading (x) and reported experience (y) under states-engaged, collapsing to noise under ablation. Points are simulated to show the form of a positive result, not data from any system.

The question is whether a synthetic NCC can survive the protocol a biological one has to.

§ 06  ·  Protocol

Five falsifiable conditions.

The protocol is engineered to be falsifiable. Each condition addresses a methodological failure mode the biological NCC literature has been litigating for three decades — construct validity, report contamination, spurious association, pseudo-dissociation, and training-distributional replay. Any single failure voids the claim.

Protocol  ·  five falsifiable conditions draft · pre-registration pending
01 / 05
Defined internal states
Measurable, repeatable, specified in advance of any data collection.
Required
02 / 05
Non-trivial reporting mechanism
The report channel is not hardwired from the internal-state readout, directly or by construction.
Required
03 / 05
NCC-structured correlation
State–report correspondence mirrors the statistical structure of biological NCC results, not a generic association.
Required
04 / 05
Negative control by ablation
When the correlate is ablated, reports change accordingly. Invariant reports under ablation mean the correlate is spurious.
Required
05 / 05
Learned-association control
Formal controls rule out replay of distributional regularities from training as the source of the correspondence.
Required
Figure C  ·  Causal structure of the protocol schematic · 3D, drag to rotate
Stimulus conversation input
St internal-state reading
Report independent channel
Ablation perturbation switch
Training dist. learned-association confound
causal edge required by protocol confound to control perturbation

Figure C. The protocol as a causal diagram. Stimulus drives the internal state St; St drives a report channel that is not hardwired from the state readout (condition 02). Ablation of St must change the report (condition 04). The training distribution must be formally ruled out as the source of the state–report correspondence (condition 05). The diagram is what each falsifiable condition is a statement about.

§ 07  ·  Specification

The load-bearing definitions.

Three elements carry the methodological weight of the protocol: the definition of the internal-state reading St, the mechanism by which the report channel achieves causal independence from St within a single system, and the formal control that rules out replay of distributional regularities from training as the source of any state–report correspondence.

The operational definitions of St, the report channel, and the learned-association control are specified in the pre-registration draft, which will be made available on submission.

§ 08  ·  Posture

Scoped. Open. Falsifiable.

QUALION is a research programme, not a product. The target is a single testable claim; the protocol was designed to be harder to write than it will be to reject. A null result is as diagnostic as a positive — pre-registration will make both outcomes interpretable — and the full protocol, dataset, and report will be released either way.

If it fails, the field knows where and why. If it survives, consciousness science gains its first empirical foothold outside biology, and the question of which substrate features carry the correlate stops being rhetorical.

§ 09  ·  Theoretical positioning

Across the theories.

The protocol is largely theory-neutral. It targets the structure of a state–report correlate under biological methodology, not the mechanism that produces it. A positive result under QUALION would not adjudicate between leading frameworks; it would give each a new substrate to answer to.

Under Integrated Information Theory (IIT), the internal-state reading St is not a measurement of Φ; whether the correlate aligns with integrated-information predictions is a downstream analysis. Under Global Neuronal Workspace (GNW), the protocol's independent report channel is the kind of access signature the framework requires, and a positive result would sharpen the architectural prediction. Under Recurrent Processing Theory (RPT), the dissociation control is the load-bearing test: RPT predicts correlates that survive no-report paradigms, and the synthetic analog of that control is what would probe it. Higher-Order Theories (HOT) predict the correlate should depend on meta-representational access — a prediction the protocol can be extended to address but does not commit to in the current pre-registration.

The neutrality is deliberate. A result that forces three theories to defend themselves simultaneously is stronger than one engineered to confirm any single one.

Figure D  ·  Theoretical compatibility manifold interactive · drag to rotate
theory-space
x — information integration
y — broadcast dynamics
z — recurrent activity

Figure D. Four theories of consciousness as anchors in a three-axis theory-space: IIT, GNW, RPT, HOT. The translucent volume marks the subset of theory-space compatible with a QUALION-positive result. Simultaneous compatibility with multiple theories is a feature of the protocol, not a limitation — it is what makes the result non-confirmatory for any single framework.

§ 10  ·  Dissociation paradigms

Where the biology is, and where the synthetic analog must go.

Biological NCC work is not merely "state correlates with report." It is a specific methodology of dissociating consciousness from confound — stimulus, attention, behavioral response — worked out over three decades. Four paradigms carry the methodological weight.

The protocol's required dissociation control is specified as a synthetic analog of at least one of these paradigm-classes; the instance is fixed in the pre-registration draft. The rationale is methodological: a synthetic NCC that survives none of them is not a synthetic NCC under the biological standard.

Figure E  ·  Dissociation paradigms, biological schematic · the classes QUALION draws synthetic analogs from
Binocular rivalry

Two incompatible inputs — experience alternates while stimulus is constant.

Masking

A following mask renders the stimulus invisible — processing without awareness.

No-report

Content inferred without verbal or motor channel — access without readout.

Attentional blink

Second target within ~500 ms is missed — perception dissociated from attention.

Figure E. The four paradigm-classes that carry the methodological weight of biological NCC research. Each dissociates conscious content from a different confound: stimulus (rivalry), awareness (masking), report channel (no-report), attention (attentional blink). The protocol requires a synthetic analog of at least one.

§ 11  ·  Anticipated objections

The protocol under review.

A programme is only as strong as the objections it has pre-answered. Six the protocol is engineered to survive.

01 / 06
This is just the Chinese Room.
The Chinese Room argument denies that symbol manipulation is sufficient for understanding. QUALION does not claim understanding. It tests for a state–report correlate under biological methodology. A positive result would license the empirical question — is there a correlate? — not the metaphysical one. The scope of the claim explicitly excludes the latter (§ 05).
02 / 06
Any state–report correlation is trivially constructable.
Conditions 04 and 05 address exactly this. A trivial correlate survives neither ablation (report must change when St is ablated) nor the learned-association control (distributional replay from training must be formally ruled out). A correlate that survives both is not trivial by the standards the biological field applies.
03 / 06
Isn't St just going to be defined post-hoc to fit the data?
No. Condition 01 requires St to be specified in advance of any data collection. The pre-registration draft fixes the operational definition. Data collection is gated on registration. This is the entire purpose of pre-registration.
04 / 06
How does this differ from prior attempts to measure Φ on artificial networks?
Φ-measurement work targets a specific theoretical quantity under IIT. QUALION targets correlate structure under biological methodology, theory-neutral (§ 09). The two are compatible but distinct. A QUALION-positive result is not a Φ-positive result, and the protocol does not claim to be.
05 / 06
The learned-association control cannot rule out steganographic encoding in St.
Correct, and the protocol does not claim it can. The learned-association control is a formal statistical control against distributional replay. It rules out that the correlate is the replay — not that it cannot be adversarially constructed. The biological field treats this level as the standard bar; a stronger guarantee is not available in correlate-based research.
06 / 06
What distinguishes "correlate" from "consciousness"?
Nothing, in the sense that the field of consciousness science has operated on correlate data since Crick and Koch (1990). The scope of the claim is exactly the scope of the biological NCC literature: the correlate, not the phenomenon it is a correlate of. The hard problem remains untouched.
Figure F  ·  Methodological space of NCC research interactive · drag to rotate
methodological cube
x — construct validity
y — dissociation
z — perturbation

Figure F. Three orthogonal axes of NCC methodology: construct validity, dissociation, perturbation. Biological paradigms (binocular rivalry, masking, no-report) populate the upper octant. Synthetic systems to date — LLMs, reinforcement agents, embodied controllers — cluster near the origin. The QUALION target position satisfies all three axes; the translucent sphere is a targeted position the protocol is designed to reach, not a measured one.

§ 12  ·  Ethics

Scope of the claim, scope of the concern.

The programme tests for correlate structure under pre-registered methodology. It does not claim to generate moral patients, sentient beings, or phenomenally conscious systems. A positive result would be a correlate in a synthetic substrate — the same kind of finding that, in biology, grounds consciousness science without settling questions of moral status.

If the protocol produces a positive result, the ethical implications are a separate, subsequent question — one the broader field (Birch, Shevlin, Shulman, and others) has been actively formalizing. The programme's position is that those questions become tractable only after the empirical surface exists. Producing the empirical surface is the programme's sole target; the ethical superstructure is not the programme's to settle alone.

No animal or human subjects are involved. The protocol operates on synthetic systems end-to-end.

§ 13  ·  References

Load-bearing literature.

[01]Crick, F., & Koch, C. (1990). Towards a neurobiological theory of consciousness. Seminars in the Neurosciences, 2, 263–275.
[02]Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.
[03]Rosenthal, D. M. (2005). Consciousness and Mind. Oxford University Press.
[04]Lamme, V. A. F. (2006). Towards a true neural stance on consciousness. Trends in Cognitive Sciences, 10(11), 494–501.
[05]Tononi, G. (2008). Consciousness as integrated information: a provisional manifesto. Biological Bulletin, 215(3), 216–242.
[06]Dehaene, S., & Changeux, J.-P. (2011). Experimental and theoretical approaches to conscious processing. Neuron, 70(2), 200–227.
[07]Tsuchiya, N., Wilke, M., Frässle, S., & Lamme, V. A. F. (2015). No-report paradigms: extracting the true neural correlates of consciousness. Trends in Cognitive Sciences, 19(12), 757–770.
[08]Koch, C., Massimini, M., Boly, M., & Tononi, G. (2016). Neural correlates of consciousness: progress and problems. Nature Reviews Neuroscience, 17(5), 307–321.
[09]Shevlin, H. (2021). How could we know when a robot was a moral patient? Cambridge Quarterly of Healthcare Ethics, 30(3), 459–471.
[10]Birch, J. (2024). The Edge of Sentience. Oxford University Press.
§ 14  ·  Glossary

Terms used on this page.

NCC
Neural correlate of consciousness. The minimal neural conditions jointly sufficient for a given conscious experience.
IIT
Integrated Information Theory. Consciousness identified with integrated information (Φ) in a physical system (Tononi).
GNW
Global Neuronal Workspace. Conscious access arises when information is broadcast to a workspace of prefrontal–parietal modules (Dehaene, Changeux).
RPT
Recurrent Processing Theory. Consciousness arises from local recurrent processing, independent of global access (Lamme).
HOT
Higher-Order Theories. A family requiring meta-representation — a state about a state — for consciousness.
Pre-registration
Lodging a protocol and its predictions publicly before data collection, to prevent post-hoc interpretation.
Registered report
Peer-reviewed pre-registration accepted by a journal before results; acceptance is independent of outcome.
St
The internal-state reading at time t. Operational definition specified in the pre-registration draft.
Ablation control
Removing the hypothesized correlate and verifying reports change. Invariance under ablation indicates spuriousness.
Learned-association control
Formal test that a state–report correspondence is not explainable by replay of distributional regularities from training.
No-report paradigm
Method of inferring conscious content without requiring verbal or motor report; rules out report-related confounds.
Dissociation
Experimental demonstration that a variable of interest is separable from a confound — state from stimulus, consciousness from attention.