Virtualized Qualia and Causal Architecture

From Refractory Route Substitution to a Predictive, Contextual, and Reportable Internal State

Michal Havelec
Independent Researcher, Czech Republic
havelm3@centrum.cz
October 30, 2026

Abstract

This paper synthesizes the Virtualized Qualia and Causal Architecture (VQCA) program, extending the Dynamical Probabilistic Spiking Hierarchy, the Theory of Refractory Virtualization, and recurrent percept reporting. It asks whether a spiking network can contain an operational state that is neither an input label nor one route's activity, but a distributed, contextual, counterfactually stable structure used for prediction and later report. More than one hundred versioned experiments indicate that neighborhoods may be spatial or spectrotemporal; redundant routes can interfere; virtual capacity depends on route compatibility rather than count; refractoriness and adaptation have dissociable roles; and slow local homeostasis can restore route availability. Unequal delays form a multiscale predictive structure, while local prediction error estimates route relevance and calibrates spatial subcoalitions. Color relations were represented robustly, but random credit and symmetry breaking alone did not create selectivity. Context was learned without labels only when causally identifiable. A live recurrent percept was distinguished from replay, read by a constrained transformer, and its delayed proposal accepted or rejected inside the workspace. Adelson-like and Munker–White-like scenes produced different internal contents under physically identical local input. With constant ambiguous input, spontaneous reversals occurred, and a hidden intervention changed later reports whereas matched replay did not. The results do not establish subjective experience. They support a functional theory of a causally accessible, input-underdetermined recurrent percept and specify when this candidate emerges, fails, or is unidentifiable.

Motivation and research question

A neural network may correctly classify a color, sound, or object without containing a state that should meaningfully be called a percept. Likewise, a language model may produce a plausible sentence about experience without that sentence being caused by a sensory or recurrent representation. Predictive-coding, recurrent-processing, global-workspace, and no-report research motivate different parts of this distinction 2, 3, 4, 5. VQCA therefore does not begin by asking whether a system “has qualia.” It asks a stricter question: what minimal internal architecture would distinguish a live percept from an input label, redundant transmission, memory replay, and a downstream motor response?

The program starts from the hypothesis that neuronal refractoriness is not merely a transmission constraint. If the same functional result can be realized by a different physical route because the previously used route is temporarily unavailable, realization itself can become virtualized. TRV demonstrated limited cases in which percept-related function survived a change of route. VQCA broadens the problem from substituting one pathway to organizing an entire field: how routes compose and interfere, what information their timing carries, how their relevance can be learned, and whether their live recurrent state can causally determine a report.

Vision is the principal experimental modality because it provides spatial variety, neighborhood structure, and contextual illusions in which local physical intensity can be dissociated from appearance. Sound provides a complementary temporal modality. A multiscale transduction converts a waveform into a spectrotemporal graph and demonstrates that relational neighborhood need not be an anatomical surface.

Central theory

A percept as a distributed causal data structure

VQCA proposes that a functional percept is neither a scalar nor a particular neural route. It is a distributed state

$$ P_t = \bigl(Z_t, \mathcal{R}_t, C_t, \Pi_t, L_t\bigr), $$

where $Z_t$ is context-derived content, $\mathcal{R}_t$ is the set of currently available physical realizations, $C_t$ is local and coalition-level confidence, $\Pi_t$ is a predictive profile, and $L_t$ is causal lineage. A state is a perceptual candidate only if all of the following hold:

its content is not fully determined by one local input;
it persists or is reconstructed through live recurrence;
its physical route can change counterfactually while content is preserved;
the state improves prediction, completion, or calibration of a future observation; and
a targeted intervention on the state changes a later constrained report, whereas matched replay or a sham intervention does not.

Virtualization is not redundancy

For routes $r_1, \ldots, r_n$, route count is not a measure of virtualization. Let $Q(P)$ be a content readout and $I$ a set of admissible interventions. The routes virtually realize a percept when, after an intervention on the active route, another live lineage $r_j$ exists such that

$$ Q(P^{I(r_i)}) \approx Q(P), \qquad L(P^{I(r_i)}) \neq L(P), $$

while cutting all relevant return routes abolishes the content. A fixed replay may reproduce the same spike trace, but without the ability to respond to a new intervention it fails the second condition.

Neighborhood is relational rather than merely geometric

Two elements are neighbors when a small change in a latent cause transforms one smoothly into the other or when a local predictive relation connects them. A neighborhood may therefore be spatial, temporal, frequency-based, receptor-based, or generally graph-defined. In audition it is produced by a multiscale filter bank: low frequencies require longer windows and high frequencies shorter ones. One global Fourier transform is insufficient because it destroys frequency-dependent temporal resolution.

Two regulatory quantities

The experiments distinguish the relevance of an individual route from the confidence of the whole coalition. Route precision controls the contribution of a temporal or spatial scale. Coalition confidence calibrates the joint output of correlated routes. Suppressing an obsolete route need not improve the Brier score if the remaining routes become overconfident. At least two slow local control loops are therefore required.

Experimental program

All decision-bearing experiments were versioned and preregistered. Development, validation, and held-out partitions were separated. Experimental branches were forked from a common snapshot, and randomness was frozen whenever required by the causal contrast. Depending on the experiment, matched controls preserved neuron and edge counts, polarity, weight and delay multisets, activity, input information, or complete spike traces. Observers had no write path into network state. Negative, invalid, and inconclusive results were retained rather than overwritten by later interpretations.

The production substrate was DPSH: asynchronous events in integer model time, refractoriness, adaptation, excitatory and inhibitory edges, local prediction-error channels, precision, plasticity, chemistry, snapshot/replay, and causal event lineage. The transformer was a separate engine connected through typed orchestration and had no write path into the perceptual state.

Discoveries I: relational transduction and temporal topology

O1 — A symmetric frequency edge is not dynamically symmetric

In multiscale auditory transduction, ascending and descending chirps did not behave as mirror-image trajectories. A peak-aligned control restored completion in only one direction. After oriented delays were introduced, the correct topology outperformed degree-, weight-, and delay-matched rewiring in 24 of 32 held-out cases and was never worse. The functional delay shifted from approximately 10–14 ms toward 30 ms as chirp velocity changed. A relational neighborhood must therefore include transduction group delay and direction, not frequency distance alone. This is consistent with the importance of spatiotemporal structure in recurrent spiking computation 6.

O2 — Parallel routes can interfere destructively

A delay bank containing several individually functional routes failed for slow traversals despite high activity. An early route changed network state so that a later, correct route could no longer act. The same principle reappeared in the spatial experiments: a slow-only route preserved content, but adding a medium route destroyed it. More routes are not monotonically beneficial; a virtual repertoire requires coordinated availability.

O3 — Refractoriness and adaptation have dissociable roles

Route-set turnover arose from branch refractoriness combined with relay adaptation. Short refractoriness increased reuse of previously active fast routes, whereas removing relay adaptation rescued all eight temporal-gap cases. Refractoriness controls immediate branch availability; adaptation controls the slower availability of the relay node supporting that branch.

Discoveries II: capacity and topology of virtualized routes

O4 — Virtual capacity is compatibility-limited

Of all 128 masks over seven slow routes, 32 configurations worked but produced only two causal-lineage signatures. Two older routes formed a minimal blocking set. The Pareto map exhibited a finite, context-dependent frontier: a three-route topology could jointly provide high delivery, lineage diversity, and removal robustness, whereas additional routes did not improve performance and could degrade it. Effective capacity is the number of compatible realizations, not the number of edges.

O5 — Virtualization has a local cost

In a 5×5 opponent field, an intact center survived with two supporting neighbors, whereas live rerouting required three. Refractory route substitution therefore consumed one degree of local relational support. An isolated point remained a systematic limit. The result supports the view that a rich spatial percept is a property of a neighborhood-supported coalition rather than of an individual neuron.

O6 — The coordinated coalition, not the first threshold lineage, is causal

The first event to cross threshold need not contain the complete causal history because subthreshold contributions from another route may be decisive. Counterfactual ablation showed that a slow route was sufficient even when trigger provenance named the medium route. Functional causality therefore requires event lineage together with sufficiency and necessity interventions.

Discoveries III: homeostasis, chemistry, and state availability

O7 — Relay adaptation is the homeostatic target

Of ten reset candidates, only resetting relay adaptation increased delivery and virtualization without losses. On new profiles, partial relative relief transferred the benefit at lower spike cost than a full reset. A continuous readout localized a sharp adaptation threshold between approximately $\Sigma u = 0.937$ and $0.985$. The missing resource was not output gain but relay-state availability.

O8 — State-dependent clearance is a genuine local rule

A fixed time constant and state-dependent clearance were functionally similar in a small test, but a load-by-gap factor identified a nonlinear law: relief produced by clearance grew with adaptation load, whereas a fixed time constant remained constant. Across 72 new profiles, a Pareto frontier emerged among rescue, loss, and spike cost. The rule was subsequently implemented as a general neuronal property with CPU/GPU parity and declarative lowering.

O9 — An endogenous neuromodulator can control clearance but does not automatically solve credit assignment

A spike-triggered local modulator with identified amplitude and decay restored routes without host intervention. Spatial chemical reach, by contrast, was not supported: a local field did not outperform an autocrine trace, and receptor topologies were functionally indistinguishable on held-out data. Chemically gated STDP and outcome-gated eligibility were real, replay-exact mechanisms, but in this task they produced a trade-off: exact lineage attribution lost to a shuffled control, and chemical plasticity closed more routes than it rescued. The discovery is an effective slow state and its timing, not a universal chemical credit mechanism.

O10 — Route availability requires hysteresis

Slow homeostasis benefited from separate opening and closing thresholds. Without hysteresis, a route oscillated around its boundary and repeatedly destabilized the coalition. With hysteresis, recovery remained local and autonomous while avoiding rapid reopen–close cycles. This provides a mechanism by which chemical or adaptation state can change global operating mode without directly choosing perceptual content.

Discoveries IV: color fields and autonomous specialization

O11 — Color is represented by a relational opponent field

Color experiments showed that a single local value was insufficient. Selectivity arose from opponent channels combined with neighborhood support on a surface. Multiple target positions were necessary to separate a genuine local rule from boundary effects and coordinate memorization. The stable signal was a relation among a center and its neighbors rather than an intrinsic “color value” assigned to one neuron.

O12 — Random credit and symmetry breaking do not create semantic selectivity

Strong initial credit, all-active starts, ignition pulses, and stochastic symmetry breaking could start activity, but they did not reliably assign color roles. Random credit also damaged virtualized routes by changing their balance before causal evidence existed. Stochastic exploration is useful for escaping a silent state, but semantic specialization requires an observable contribution to prediction.

O13 — Centers specialize by relative local error and marginal contribution

Competition alone created diversity without correct routing; marginal error routing improved mean squared error without learning a new factor. Combining them with capacity-weighted credit produced specialization and free capacity in 12/12 runs and acquired a new relation in 10–12/12, while controls remained at 0/12. A functional center is therefore defined by its own relative error and marginal contribution, rather than by a global error rate or a predetermined anatomical label.

Discoveries V: the self-predictive percept

O14 — A live loop can be spike-identical to replay

Production closure produced 502 live returns. Forced replay reproduced the entire spike trace bit for bit but lacked live closure lineage. Workspace-only prediction was identical in live, replay, and open branches. Instantaneous state alone therefore cannot identify causal organization; the difference must appear under a future intervention or change in route availability.

O15 — Unequal delays turn time into a data structure

A 2-ms fast route and a 22-ms slow route at 20-ms cadence exposed the current and previous state to one instantaneous readout. The Brier score fell from 0.282916 to 0.182852, and the temporal pair won in 16/16 seeds; an equal-size duplicate-current control did not help. The useful horizon was finite: lags 1–3 helped, lag 4 failed the gain criterion, and lags 5–8 were harmful. A route is therefore not only a transmission path but a hypothesis about a relevant temporal scale.

O16 — The network can evaluate temporal routes autonomously but must calibrate confidence separately

Local prediction errors correctly identified lag 6 as obsolete, and an autonomous homeostat changed uniform precision into high values for lags 1 and 2 and a minimum for lag 6. Nevertheless, overall Brier score did not significantly improve across 64 seeds. The reason was not the obsolete route but excessive confidence in the two correlated surviving routes. An independent temperature $T = 0.7$ reduced Brier from 0.171116 to 0.144443 and expected calibration error from 0.163070 to 0.017239 without changing accuracy.

O17 — Confidence belongs to slow spatial subcoalitions

Pixel-level updates learned marginal noise and degraded performance, while trial-level updates overshot. Slow frame-level homeostasis found $T = 0.720854$, close to the oracle value 0.7, reducing Brier by more than 10% throughout a neighborhood of parameter values and ECE from approximately 0.158 to 0.023. Separate local confidence for two subcoalitions yielded a stable 4.2–4.5% gain over a global scalar across 128 seeds; on spatial partitions the gain was 5.2–6.2% and generalized after rotating the mask. Perceptual confidence is thus a field of slow local variables, not one global state.

Discoveries VI: latent context and learned illusions

O18 — A silent state can be a causal data structure

After 21, 81, and 321 ms of silence, a uniform probe recovered the same spatial mode. Erasing the latent state abolished it completely without removing any spike, and feedback ablation changed the later response. This separates latent storage from in-flight events and post-input reverberation. It is not evidence of qualia, but it is an internal data structure accessible only through future dynamics.

O19 — One latent/re-entry apparatus carries different contextual relations

Adelson-like lightness 1 and Munker–White-like chromaticity emerged in one graph with a double dissociation: surround ablation removed only the checker-shadow effect, whereas collinear ablation removed only chromatic assimilation. The same local input therefore acquired different internal meaning through relational context rather than absolute pixel value.

O20 — Context can be learned from local error only when causally identifiable

The first positive relation-learning result was invalidated because the orchestrator read a hidden relation label. Once that shortcut was removed, the network failed on conflicting scenes; a Bayes oracle demonstrated irreducible ambiguity in the original generator. A new temporal Markov generator added observable history. Relative-error centers then reached 0.998 dominance, captured all transitions, and reduced MSE to 0.0308; confirmation passed 64/64 runs and 8/8 parameter points. Functional-center identity can emerge from local error, but its time constant must match the rate at which the world changes.

O21 — Prediction error selects a distributed coalition of complete routes

Dense random mixtures of spatial modes achieved 100% diversity but covered 0/12 roles. Sparse mixtures of one to three modes covered 12/12 with mean minimum cosine 0.978. Local error selection enriched the correct routes in 48/48 cases, yet top-8 ablation was causal in only 23/48. Ablating 16 routes increased median error 51-fold and ablating 32 increased it 974-fold, whereas random ablation remained near one. Production DPSH achieved autonomous MSE 0.014, compared with 0.340 for uniform and 0.375 for crossed controls. The carrier is neither a winner neuron nor a small top-k set but a broad, prediction-selected coalition.

Discoveries VII: causal percept reporting

O22 — Contextual appearance is separable from physical value

Slow self-supervised learning produced contextual separation 0.231, versus a maximum of 0.043 across eight shuffle controls and zero after ablation. An Adelson-like local predictor mapped physically identical targets of 0.4 to 0.386682 and 0.765043; their difference of 0.378361 fell to 0.0106854 after shuffling. Munker–White-like scenes produced red–green separation 0.392, compared with 0.00059 under shuffling. The contextual error therefore arose in the learned internal representation before language reporting.

O23 — Content transfers automatically into a live carrier

Predictor output passed through a blind population encoder into a fast/slow loop without a hand-written color label. Cutting the fast route changed trajectory while preserving the red/green code through 21/21 slow returns; cutting all routes abolished it, and replay did not respond to intervention. Sparse red and green interventions produced a double dissociation. Across 180 states, 178 were strong, two were borderline, and none showed a confident inversion. Carrier capacities 8, 16, 32, and 64 distinguished 69, 135, 175, and 180 states, respectively; a matched repetition of the 16-unit code remained at 135.

O24 — Completion is not hidden-pixel reconstruction

Two attempts at pixel-perfect filling-in lost to shuffled or scene-mean controls. A weaker and more precise hypothesis succeeded: contextually coherent internal completion reduced boundary error from 0.501125 to 0.113931, preserved contextual separation 0.24288, and transferred 16/16 fragments through the live carrier. A percept may complete consistent meaning without reconstructing the physically correct unobserved pixel.

O25 — A constrained report tracks content rather than scene or route

The transformer received only carrier state. On the boundary map it emitted 135 reports, 45 abstentions, and no opposite token. A content-changing cut reversed the report, while a content-preserving reroute did not. Pixel-only and shuffled controls abstained, whereas a Munker workspace produced the correct relative chromatic report. Sensory versus internal origin could not be inferred from content alone; a truthful source report required physical lineage.

O26 — A linguistic proposal can be verified again by the recurrent percept

A delayed proposal was returned to DPSH, where pair-specific populations produced endorse, reject, or stale. In the full Adelson test, three physical control branches were correct; only the physically equal illusion branch produced appear-different. A workspace-faithful proposal was endorsed and a swapped proposal rejected. Zero-shot composition of an omitted rule had previously failed; the successful version applied a learned abstract rule rather than discovering new logic at report time.

O27 — A private bistable state can be read causally under constant input

Constant ambiguous input produced 20 clear windows out of 24, both percepts, and ten spontaneous reversals. All 20 clear states were reported and 20/20 were recurrently confirmed. A hidden coalition intervention changed four subsequent live reports, while fixed replay changed none. This combination separates report from stimulus change and is the strongest functional evidence obtained for a private, input-underdetermined recurrent state in this particular apparatus. A later transfer through the public DPSH path passed three of four held-out seeds, but content-specific and temporal generalizations of the intervention each produced zero positive seeds out of eight. O27 therefore establishes one reproducible causal case rather than general identifiability of private state across realizations. The paradigm is mechanistically narrower than biological binocular rivalry, but uses its key separation between constant stimulation and changing perceptual organization 7.

Summary map of discoveries

Layer	Main discovery	Architectural consequence
Transduction	Oriented spectrotemporal neighborhood	Neighborhood must include latency and direction, not distance alone.
Routes	Destructive interference among redundant paths	Route availability must be state-coordinated.
Capacity	Compatibility-limited virtual repertoire	More edges do not imply more realizations; capacity has a contextual Pareto frontier.
Dynamics	Dissociable roles of refractoriness and adaptation	Refractoriness switches branches; adaptation controls slower relay availability.
Homeostasis	Clearance and hysteresis	The network can restore and reopen routes locally without a boundary reset.
Chemistry	Timed modulation works; spatial credit does not	Chemistry is a slow state regulator, not an automatic source of correct attribution.
Color	Relational fields and capacity-weighted credit	Semantic specialization requires neighborhood, normalization, and relative local error.
Prediction	Delay as a temporal data structure	Routes explicitly represent distinct lags and have a finite useful horizon.
Confidence	Route precision ≠ coalition confidence	Route relevance and coalition calibration require separate circuits.
Latency	Silent causal trace	Content can be stored in state and expressed only by a future probe.
Learning	Center identity from local error	A specialized center need not be anatomically predetermined.
Coalition	Broad prediction-selected carrier	A top-k neuron set is not the percept; causal content is distributed.
Illusions	Same pixel, different contextual state	Internal appearance can be separated from local physical measurement.
Report	Content-sensitive, route-invariant, rejectable report	The transformer verbalizes a percept; it does not create it.
Bistability	Reversals and hidden perturbation under constant input	Report is causally downstream of a private recurrent state.

Reproducibility map

The table below maps each synthesis claim to the immutable experiment families in the research archive. Each family has a preregistration and a result report containing the exact engine path, configuration, development/validation/held-out split, seeds, gates, control outcomes, and machine-readable artifact names. The ranges include failed and invalid intermediate versions; the synthesis above uses only the terminal status recorded by each report.

Claims	Archive identifiers	Principal causal or matched controls
O1–O3	VQ-AUD-03, K11–K24	Peak alignment; degree/weight/delay-matched rewiring; adaptation × refractoriness factorial; lineage completion.
O4	VQ-AUD-03, K25–K31	Exhaustive route masks; topology Pareto sweep; held-out transfer; causal-eligibility and counterfactual searches.
O5–O6	VQ-COL-01/02, K65–K93	Closed-loop known truth; full-cut and rerouting interventions; opponent and minimum-neighbor controls; route-composition ablations.
O7–O10	VQ-AUD-03, K34–K60; VQ-COL-20/21, K151–K152	Absolute/relative and route-local homeostasis; load × gap identification; clamped versus endogenous modulation; shuffled chemical credit; hysteresis and drift.
O11–O13	VQ-LRN-10–21, K125–K136; VQ-COL-06–21, K137–K152	Pixel-only and shuffled-neighborhood controls; matched capacity; random-credit, ignition, stochastic-exploration, and learned-kernel failures; held-out spatial confirmation.
O14–O17	VQ-SPP-02–13, K95–K107	Bit-identical replay; open-loop and equal-delay controls; obsolete-lag ablation; frozen precision; global versus local confidence; rotated spatial masks.
O18–O20	VQ-SPP-14–21, K108–K115; VQ-LRN-01–09, K116–K124	Latent-state erasure; feedback ablation; surround/collinear double dissociation; hidden-label audit; Bayes identifiability oracle; temporal held-out generator.
O21	VQ-COL-22–27, K153a–K154a.2	Dense versus sparse generators; uniform and crossed controls; ranked versus random ablation; autonomous production-DPSH selection.
O22–O24	RPR-04A–04I and RPR-05A–05E	Pixel-only, shuffled, context-edge, all-route-cut, replay, repetition-capacity, scene-mean, and hidden-pixel reconstruction controls.
O25–O27	RPR-06A–06K	Carrier-only transformer input; content-changing cut versus content-preserving reroute; swapped/stale proposal; physical control branches; constant-input bistability; hidden intervention versus fixed replay.

What was falsified

Negative results are part of the theory because they delimit the mechanism more precisely than positive scores alone:

One Fourier transform or a symmetric frequency topology is insufficient for multiple temporal scales.
Parallel redundancy is not virtualization and can degrade performance.
Route count is not a measure of virtual capacity.
Orchestrator resets, global gain, and random credit are not autonomous homeostasis.
Spatial chemical reach and receptor geometry were not identified in the tested task.
Exact past lineage was not the best learning-credit rule.
Repeating a spike state through a live loop did not by itself add predictive content.
A global E/I spatial mode found in development did not transfer to new seeds.
Neither a static closed loop nor chemical state alone maintained a post-input percept.
Symmetry breaking, ignition, and stochasticity alone did not create color selectivity.
Pixel-perfect filling-in was unsupported; contextual completion was supported.
Content alone did not reveal whether it originated in a sensor or an internal workspace.
The transformer did not compose an unseen conflict rule zero-shot.
None of these tests directly observes subjective experience.

Synthetic VQCA theory

The combined evidence supports the following working theory:

A functional percept arises when transduction creates a relational field, local prediction error selects a distributed coalition of compatible routes within that field, and recurrent dynamics maintain the coalition as a counterfactually stable data structure. Refractoriness, adaptation, and slow homeostasis continually change its physical realization; content remains stable only when it has sufficient neighborhood support and a separate opponent representation. Each route expresses a hypothesis about a spatial or temporal relation, while slower confidence represents trust in the whole subcoalition. A percept is functionally private when it is not fully determined by instantaneous input or spike replay, yet can be causally revealed through future prediction, completion, intervention, and constrained reporting. A linguistic report is credible only when it follows content changes, ignores mere realization changes, and is verified again by the live recurrent workspace.

This theory unifies robustness, prediction, context, completion, confidence, and report in one mechanistic framework. It does not claim that every recurrent predictive network has qualia. It claims that this causal profile is a stronger candidate than classification, activity, recurrence, redundancy, or language report considered separately.

Falsifiable predictions of the unified theory

With topology frozen, a content-changing intervention must change the report, a content-preserving reroute must not, and an all-return-path cut must abolish it.
A spike trace created by replay must diverge from a live loop under a future intervention or input change.
Adding an incompatible route may degrade a percept despite greater activity; a local gate or homeostat must restore performance without the orchestrator selecting the result.
The useful temporal horizon must be finite and vary with environmental dynamics; excessively old routes must produce systematic error.
Route precision and coalition confidence must be doubly dissociable: the former changes scale weighting, whereas the latter changes calibration without changing the decision.
A new center may specialize only when observable causal evidence exists; with a hidden independent relation label, even an optimal label-blind network must fail.
Contextual illusions must be learnable from prediction error and generalize to unused scenes; pixel-only, shuffled, and context-edge-ablation controls must abolish the effect.
A bistable report under constant input must be predictable from internal state before querying and controllable by a hidden perturbation, not by the reporter prompt.

Epistemic boundary

The work supports a functional candidate: a causally accessible, input-underdetermined, contextual, and recurrent percept. It does not show that the system biologically reproduces human cortex, that a transformer's sentence is introspection in the human sense, or that phenomenal consciousness has been demonstrated. The term qualia in the title denotes the target of a falsifiable research program, not a completed ontological conclusion.

The strongest current evidence is not one illusion score. It is a closed causal chain: a physically identical or constant input can produce different internal states; a live, substitutable recurrent coalition carries the state; a hidden intervention on the coalition changes a future report whereas replay does not; and a delayed linguistic proposal is accepted or rejected according to the current workspace.

Next decisive experiments

The next phase should turn the present synthesis from a collection of known-truth experiments into a small set of prospective, blindly evaluated tests:

freeze a generator of novel Adelson and Munker scenes after all thresholds are locked, then open one held-out package;
predict a spontaneous reversal from the internal coalition before reporting, and intervene within a preregistered temporal window;
combine learned context, route homeostasis, and transformer reporting in one long run without a host reset;
replace the current reporter with an independently trained model and perform a blind replication;
test a multimodal conflict with separate physical input lineages; and
derive biologically measurable predictions linking refractoriness, adaptation threshold, local confidence, and perceptual reversals.

Conclusion

VQCA does not yield the simple conclusion that consciousness appears when recurrence, redundancy, chemistry, or a transformer is added. It yields a narrower architectural picture. A perceptual candidate is a distributed coalition of relations that must be learnable from local prediction error, possess finite compatible capacity, preserve content across changing physical realizations, regulate its own availability and confidence over multiple time scales, and have a causally faithful but separate route to report. Many intuitive shortcuts were falsified. Those failures make the resulting theory more precise, testable, and valuable: it specifies not only what a successful system requires, but also what is insufficient.

Availability and reproducibility

Source code, preregistrations, results including negative versions, and experimental artifacts are maintained in the private Cognia repository. The URL https://github.com/havelm3/cognia may return “404 Not Found” without authorization. Research access or a reproduction package may be requested from the author at havelm3@centrum.cz; availability is subject to repository permissions and licensing. No human or animal participants and no personal data were used.

Use of generative tools

Generative language tools assisted with implementation, experimental record keeping, and editorial synthesis under the author's direction. Hypotheses, preregistrations, interpretations, and responsibility for all claims remain with the author.

References

E. H. Adelson, “Perceptual organization and the judgment of brightness,” Science, vol. 262, no. 5142, pp. 2042–2044, 1993. doi:10.1126/science.8266102.
R. P. N. Rao and D. H. Ballard, “Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects,” Nature Neuroscience, vol. 2, no. 1, pp. 79–87, 1999. doi:10.1038/4580.
V. A. F. Lamme, “Towards a true neural stance on consciousness,” Trends in Cognitive Sciences, vol. 10, no. 11, pp. 494–501, 2006. doi:10.1016/j.tics.2006.09.001.
S. Dehaene and L. Naccache, “Towards a cognitive neuroscience of consciousness: basic evidence and a workspace framework,” Cognition, vol. 79, nos. 1–2, pp. 1–37, 2001. doi:10.1016/S0010-0277(00)00123-2.
N. Tsuchiya, M. Wilke, S. Frässle, and V. A. F. Lamme, “No-report paradigms: extracting the true neural correlates of consciousness,” Trends in Cognitive Sciences, vol. 19, no. 12, pp. 757–770, 2015. doi:10.1016/j.tics.2015.10.002.
R. Pyle and R. Rosenbaum, “Spatiotemporal dynamics and reliable computations in recurrent spiking neural networks,” arXiv:1611.01557, 2016.
N. K. Logothetis, D. A. Leopold, and D. L. Sheinberg, “What is rivalling during binocular rivalry?” Nature, vol. 380, pp. 621–624, 1996. doi:10.1038/380621a0.