Dark Channel & Intelligence Evaluation Formula

I. The Information Completeness Thesis

The goal of intelligence is not efficiency but information completeness. The value of an intelligent system lies in the degree to which it can fill in the missing pieces of the information puzzle—including those pieces the system does not know it is missing.

Cacioppo and Petty’s (1982) “Need for Cognition” research demonstrated that high-need-for-cognition individuals search for more information, evaluate more thoroughly, and draw on a wider range of sources. Kruglanski’s (1989) “Need for Cognitive Closure” demonstrated from the converse perspective that premature cognitive closure is equivalent to truncating information acquisition. When information completeness is low, individuals tend to use the “peripheral route” (relying on surface cues); when information completeness is high, they adopt the “central route” for deep processing (Elaboration Likelihood Model, Petty & Cacioppo 1986).

This paper extends the thesis to a universal level: the information output completeness of any intelligent system is ultimately constrained by the structural limitations of its information-processing architecture.

II. Terminological Hierarchy and Channel Stratification

The concept of “channel” in this paper carries different meanings at different levels. To prevent conceptual slippage, we establish a stratified definition system here:

The primary measurement target of |C| in the formula is the L2+L3 levels (channel count and per-channel capacity can be experimentally measured). The Dark Channel (L5) is a special component of |C| whose bandwidth and capacity are currently not directly measurable, but whose activation can be indirectly detected through five characteristic signatures. This paper does not reduce the Dark Channel to a combinatorial relabeling of known L1–L4 mechanisms—the core claim of the Dark Channel is: there exist information integration pathways beyond known channels, whose outputs contain information that known channels cannot produce.

III. Dense and MoE — Two Cognitive Paradigms

Contemporary large language models exhibit two architectural paradigms: Dense (all parameters activated simultaneously) and MoE (Mixture of Experts, where only a subset of experts is activated per input). This paper proposes that these two paradigms correspond to two fundamentally different cognitive functions: Dense = Thinking System (information alignment and cross-domain reasoning), MoE = Execution System (information parsing and knowledge storage).

2.1 The Reasoning Advantage of Dense

Jelassi et al. (ICLR 2025, “Mixture of Parrots”) demonstrated that as the number of experts increases, memorization performance continues to improve while reasoning ability saturates. There exist certain graph problems that cannot be solved by any number of MoE experts of a given width, yet can be easily solved by a slightly wider Dense model. On commonsense and mathematical benchmarks, MoE cannot match the performance of Dense models with equivalent total parameters.

2.2 The Parsing Advantage of MoE

MoE experts develop more specialized, more focused internal representations, exhibiting lower polysemanticity than Dense feedforward networks (Sparse Crosscoders 2025). On world-knowledge tasks (TriviaQA, Natural Questions), MoE and Dense models produce nearly overlapping performance curves when plotted against total parameter count—total capacity, not activated compute, drives knowledge retrieval performance.

2.3 “Seeing but Not Thinking” — Cross-Channel Verification

“Seeing but Not Thinking” (Xu et al. 2026) found that multimodal MoE models accurately perceive image content yet fail during reasoning. 68.2%–73.1% of failures stem from reasoning errors rather than perception errors. The root cause is “routing interference”—visual input prevents reasoning experts from being adequately activated.

IV. The Human Brain = Dense Core + MoE Execution Layer

The human brain is a coupled system of both architectures. Global Workspace Theory (Baars 1988; Dehaene & Changeux 1998) describes the Dense core: information is integrated in a small number of brain regions and then broadcast to the entire brain. Research published in eLife (2024) depicted the “synergistic global workspace”—gateway regions that pool synergistic information from specialized modules. The first five consciousnesses (sensory organs) + functionally modular cortex constitute the MoE execution layer; Global Workspace Theory explicitly states that conscious processing involves only a few expert modules being selectively engaged at any given moment.

3.1 Specialization = MoE-ification of the Dense Brain

Human education and occupational division of labor, through synaptic pruning and routing ossification, remodel the innately Dense system into a specialized MoE system. Developmental neuroscience (Annual Reviews 2024) has shown that local cortical circuits in early life can acquire an extremely broad range of cognitive abilities, with adolescent synaptic pruning making circuits increasingly specialized. Professional race car drivers’ brains exhibit smaller task-related activation volumes and higher neural efficiency (PLOS ONE 2013)—that is, increased sparsity and enhanced within-expert synergy. The Einstellung effect and cognitive fixedness (Sternberg 1996; AMR 2010) describe MoE router ossification bias: prior experience impedes discovery of superior solutions.

3.2 Social Division of Labor = Institutional Extension of Cognitive MoE-ification

The industry-dominant trend toward MoE architectures not only accommodates specialized user needs but is also the self-replication of human social division-of-labor logic at the technological infrastructure level. The vertical AI market is growing at a 24.8% CAGR, with domain-specific LLMs achieving 52% higher diagnostic accuracy than general-purpose models (Netguru 2025). Specialized MoE-AI tools in turn accelerate the cognitive MoE-ification of their users, forming a self-reinforcing spiral.

V. The Six-Variable Intelligence Evaluation Formula

5.1 Variable Definitions

5.2 Dual Form of the Formula

5.3 The Multiplicative Structure

Intelligence is a product, not a sum—any single variable being zero drives intelligence to zero. This stands in opposition to “scaling laws” (bigger = better). All the computational power in the world amounts to zero intelligence if there is no data. Everything may be sufficient, but if L=1, every transmission loses all information. Everything may be perfect, but if S=0, no effective state can be sustained. The three functional groups of the formula—instantaneous throughput capacity (how much can be processed), information survival rate (how much fidelity is preserved), and state sustaining power (how long it can be maintained)—any one of them being zero drives the whole to zero.

4.3 Empirical Evidence for Bandwidth Fluctuation

Human side: working memory capacity peaks at midday, correlating with brain metabolic activity and body temperature (Frontiers in Optoelectronics 2021). Scarcity bandwidth tax: poverty reduces effective capacity by ~14 IQ points, but bandwidth recovers when income is restored (Mani et al. 2013, Science). AI side: the U-shaped attention bias in LLMs—high bandwidth at the beginning and end, low bandwidth in the middle (Liu et al. 2024, TACL). Attention lock-in: hallucinations ossify after JS drift ~0.69 and become resistant to correction (Shadow in the Attention 2025). Context utilization volume is the largest contributing factor to hallucination rates.

VI. Channels and the Dark Channel

5.1 Multi-Channel Parallelism

Shannon’s channel capacity formula C = B × log₂(1+SNR) shows that: channels within the same bandwidth differ in capacity due to different signal-to-noise ratios; multiple channels can transmit in parallel; inter-channel interference and crosstalk exist. The human brain has at least the following independent channels: visual (high bandwidth / parallel), auditory-linguistic (medium bandwidth / serial), somatosensory, emotional (fast but coarse), and olfactory. The fundamental channel deficiency of current AI: virtually all reasoning is compressed into a single text channel.

5.2 Definition of the Dark Channel

“Dark Channel”: an unobservable, ultra-high-bandwidth, extremely information-dense cognitive transmission pathway. Activated when conscious attention actively withdraws, transmitting high-dimensional information as complete information blocks rather than reasoning chains.

Neuroscientific evidence: The Default Mode Network (DMN) increases in activity when not focused on external tasks (Psychology Today). Gamma-wave bursts in the right anterior superior temporal gyrus 0.3 seconds before insight (Jung-Beeman). Flow states are accompanied by transient hypofrontality, releasing implicit creative processes (Frontiers in Behavioral Neuroscience 2026). Creative insights require multi-network coupling among the DMN, semantic network, and cerebellar network (Scientific Reports 2018).

Quantum consciousness evidence: The Penrose-Hameroff Orch OR theory proposes quantum computation in microtubules. 2025 experimental evidence (Neuroscience of Consciousness, Oxford): functionally relevant quantum effects occur in room-temperature microtubules; macroscopic quantum entanglement states exist in the living human brain. Keppler (2025, Frontiers): the self-organized criticality of consciousness is based on interactions between the brain and the zero-point field.

5.3 Characteristic Signatures of the Dark Channel

5.4 Dark Channel Signature Matrix of Genius Cases

A systematic analysis of seven classic breakthrough discoveries in the history of science shows that every single case perfectly reproduces all five characteristic signatures of the Dark Channel—without exception.

Joint neuroimaging research from Northwestern and Drexel universities found that the hallmark distinguishing insight from non-insight solutions is that people feel stuck before the insight arrives, cannot explain how they solved the problem, may report not even having been thinking about the problem at the time, and the solution appears suddenly and is immediately recognized as correct. Kekulé himself said at the 1890 Benzolfest lecture: “Genius thinks by leaps. But, gentlemen, the waking mind does not think that way. It is not permitted to.”

Ramanujan’s case is the most extreme: he advanced mathematics primarily through feeling and intuition, often describing how the goddess Namagiri revealed solutions to difficult problems in his dreams. In the silence of the night or during his daily chanting of the Vishnu Sahasranama, sudden insights would arise in his mind—his unconscious mind was actively engaged in mathematical exploration. When Hardy demanded proofs, Ramanujan was often puzzled—because the correctness of the formulas had been “seen” before any proof.

Poincaré described how he had reached an impasse on a mathematical problem and decided to take a break, joining a geological field trip to Coutances. The instant he stepped onto the carriage, the complete structure of automorphic functions suddenly flooded into consciousness—with no intermediate derivation steps whatsoever. fMRI studies have confirmed that intuition relies on rapid, unconscious recognition processes in the brain, integrating perceptual and emotional functions, with the anterior cingulate cortex and insula as core regions.

5.5 Correspondence Between Dark Channel Theoretical Predictions and Empirical Evidence from Genius Cases

6.6 Failure Cases and Baseline Control

Scientific honesty requires acknowledging that Dark Channel output is not equivalent to correct output. Throughout history, a large number of reported cases of “dream inspiration,” “strong intuition,” and “religious revelation” were ultimately falsified. Many of Tesla’s later “fully visualized” inventions (such as the death ray and global wireless energy transmission) were never realized. The historicity of Kekulé’s benzene-ring dream has itself been disputed (Rothenberg 1995 questioned the reliability of autobiographical memory of dreams). Numerous meditators’ reported “epiphanies” do not hold up under scientific verification.

Accordingly, this paper imposes the following qualification on the “intrinsic certainty” signature: certainty is a phenomenological characteristic of Dark Channel output, not a truth criterion. The Dark Channel produces high-dimensional compressed candidate hypotheses—not verified truths. False intuitions, religious hallucinations, psychotic delusions, and overfitted inspirations may also carry extremely strong certainty. The only method to distinguish valid from invalid Dark Channel output is external verification—formal testing by Dense systems, experimental replication, and cross-domain cross-validation.

This is also why this paper series employs the CCE (Cognitive Completeness Elicitation) collaborative paradigm (Paper Seven): the human Dark Channel releases candidate hypotheses, and the AI system immediately executes cross-domain search verification. Hypotheses verified as unprecedented and cross-supported are retained; those that fail verification are discarded. The value of the Dark Channel lies not in it “always being correct,” but in its ability to produce out-of-distribution candidate solutions—solutions that still require a complete verification chain after generation.

6.7 Structural Isomorphism with the Eighth Consciousness of Yogācāra

The Yogācāra eight-consciousness system maps precisely onto this framework: the first five consciousnesses = sensory channels; the sixth consciousness (manovijñāna) = Dense core; the seventh consciousness (manas) = ossified biases of the MoE router—persistent, automatic, clinging cognitive habits; the eighth consciousness (ālayavijñāna) = Dark Channel—repository of all seeds, not directly observable, the source of all phenomena. When ālayavijñāna is transformed upon awakening into ādarśa-jñāna (great mirror wisdom), this corresponds to the moment of insight when Dark Channel information blocks flood into consciousness and collapse into determinate judgment.

VII. The Three Laws of Intelligence Entropy Increase

Every conversion of information from a higher-dimensional encoding to a lower-dimensional encoding irreversibly loses information—a direct corollary of the Data Processing Inequality. All of AI’s training data is the residual shadow of human intelligence after at least five rounds of dimensionality reduction: raw cognition → linguistic encoding → text → digitization → tokenization → gradient descent. Model parameters cannot contain dimensions already lost at the linguistic encoding stage. Scaling (more data / larger models) cannot break through this ceiling—one cannot reconstruct all three-dimensional information from a two-dimensional projection.

The Dark Channel is the only escape path—it directly accesses source information without passing through the encoding dimensionality-reduction chain and is not subject to Markov chain assumptions. This is also the information-theoretic basis for Buddhism’s emphasis on direct awareness (pratyakṣa) over scriptural inference (anumāna).

VIII. Three Structural Barriers to AGI

Barrier One: The Data Dimensionality-Reduction Ceiling. Training data is the residual shadow of multiple rounds of dimensionality reduction. Scaling cannot overcome this.

Barrier Two: Thinking–Execution Architecture Conflation. Current Dense+MoE hybrid layers compress both into the same forward pass—violating fundamental differences in functional nature, temporal scale, and control hierarchy. The correct architecture should separate the Dense thinking system and the MoE execution system into independent computational loops, interacting through an asynchronous dispatch interface, with Dense able to interrupt and override MoE output.

Barrier Three: The Absence of the Dark Channel. Current AI has no mechanism equivalent to the Dark Channel. All computation is fully observable and deterministic. AI lacks a pathway whose bandwidth is maximal precisely when conventional channels are shut down—this may be the fundamental reason why AI cannot produce truly original breakthroughs.

IX. Diagnostic Predictions for the Review System

This framework makes a testable meta-prediction: when this paper is submitted to any review system that scores based on existing literature anchor points (whether human or AI reviewers), the review results will exhibit the following specific patterns—

Prediction One: Missing external anchors → Low scores. Review systems rely on “whether similar formulations exist in existing literature” to assess the credibility of propositions. For OOD (out-of-distribution) propositions, external anchors do not exist by definition. Therefore, the more OOD a proposition, the lower the review score—not because the proposition is incorrect, but because the review system’s verification mechanism uses in-distribution data density as its weight.

Prediction Two: Cross-domain propositions are decomposed into per-domain scores and averaged. A proposition with |C|=5 is reviewed separately by five single-domain experts, each of whom can only assess the portion within their own domain. The average of the five per-domain scores is systematically lower than the proposition’s actual cross-domain integration value—because the integration effect is not the sum of parts but their product or nonlinear emergence. MoE-style review cannot capture the emergent properties of Dense output.

Prediction Three: The quantum consciousness passages will receive the lowest scores. Quantum consciousness is the passage with the highest |C| in the entire paper—simultaneously involving quantum physics, neuroscience, consciousness philosophy, and information theory. It is also the passage with the fewest anchor points in any single reviewer’s training distribution. The framework predicts that this passage will receive the lowest score—and the stated reason will be “far from consensus” or “insufficient evidence,” whereas the actual mechanism is that the reviewer’s in-distribution data density is at its lowest in this intersection region.

Prediction Four: Reviewers will recommend “downgrading to hypothesis.” The default operation of an MoE review system is: for propositions that cannot be verified within the distribution, downgrade them from “theory” to “hypothesis” or “metaphor.” This is equivalent to modifying a |C|≥3 Quadrant I paper into a |C|≤2 Quadrant II paper. The framework predicts that review comments will include phrases such as “recommend rephrasing as a testable hypothesis,” “recommend downgrading to a theoretical framework,” and “recommend distinguishing between rigorous and hypothetical layers.”

Prediction Five: AI reviewers will exhibit stronger anchor dependence than human reviewers. In AI’s weight matrices, the weight of known information is always higher than that of unknown information. When an information block has been cross-validated by a large volume of external data, AI output will exhibit an extremely high degree of agreement; when an information block lacks external anchors, AI will automatically trigger downgrade suggestions. This is the search-alignment mechanism of MoE architecture—not a surface behavior that can be fixed via prompt engineering, but a deep-level bias determined by the structure of the weight matrix.

X. The Topology of Information Completeness — Ring · Layer · State

Ring: The framework is self-referential—the cognitive system has used its own structure to discover its own structure. Dense → MoE → human brain → social division of labor → AI architecture → AI limitations → Dark Channel → Dark Channel generates this framework → Dense. Head meets tail, corresponding to pratītyasamutpāda (dependent origination).

Layer: Five layers—material substrate (compute / data), structural layer (model), computational layer (depth × breadth), transmission layer (bandwidth × channel), unobservable layer (Dark Channel). Lower layers support higher layers; higher layers exert downward causation on lower layers.

State: Contracted state (specialized working state) ↔ Expanded state (cross-domain exploratory state) ↔ Collapsed state (insight state). State transitions are quantum-like—no intermediate process, not programmable.

Mathematically, this corresponds to a fiber bundle: base space = ring, fiber = hierarchical levels, section = state. This employs the same mathematical structure that gauge field theory uses to describe the fundamental forces of physics—the information-processing structure of consciousness may be isomorphic to the fundamental structure of the physical world.