dynamicsystemsarchitecture.org

Rhizomatic Gating vs. Arborescent Collapse: Engineering Deleuzian Assemblages in Multi-Agent Cognitive Architectures

PurposeA conceptual argument that the four-channel architecture is the architectural counterpart of Deleuze and Guattari’s rhizome.
StatusDraft — a reconstruction written from a short description, not the author’s original manuscript. Not submitted, not peer reviewed.
SubjectUnlike The Dynamic Chinese Room, this paper is about the architecture. It cites no test results and makes no performance claim.
AttributionAI-drafted at the author’s direction (tool not recorded).
RelatedAn earlier, differently argued version · Formal specification
Edit madeOne line, “Journal Submission: Journal of Philosophical Investigations”, was removed because the draft itself states it was not submitted.

A Philosophical Paper Applying Deleuze & Guattari to LLM Architecture

Format: Philosophy paper, draft framework. Not submitted, not peer reviewed.
Target Length: ~9,000 words

I. INTRODUCTION: DELEUZE & GUATTARI IN THE AGE OF NEURAL NETWORKS

The Rhizome Concept

In A Thousand Plateaus, Deleuze & Guattari propose the rhizome as a conceptual alternative to hierarchical, tree-like structures (which they call "arborescent" or arboreal systems).

An arborescent system has:

A rhizomatic system has:

Deleuze & Guattari wrote this in 1987, long before large language models existed. Yet their distinction illuminates a fundamental problem in modern AI: the forced flattening of multiplicity into singularity.

The Problem We Address

Standard large language models (LLMs) are architecturally arborescent. They collapse diverse, potentially contradictory information into a single representational hierarchy. The model must choose: this query means one thing, produces one output, resolves to one answer.

But semantic reality is not singular. A word like "bank" carries multiple, coexisting meanings. A narrative can have multiple valid interpretations. A question can generate multiple valid answers. The world itself is fundamentally plural.

Standard LLM architecture forces this plurality into singularity. The cost is:

This paper argues that NCFCA (Non-Collapsible Four-Channel Architecture) is structurally rhizomatic: it is an architectural counterpart of the preservation of multiplicity that Deleuze and Guattari describe. The paper does not claim the architecture has been shown to perform better than any alternative.

II. THE PHILOSOPHICAL FOUNDATIONS: DELEUZE ON MULTIPLICITIES

2.1 What Is an Arborescent Structure?

Deleuze & Guattari use the tree metaphor to describe systems that:

  1. Unify through origin: Everything traces back to a root. In LLMs, this is the training objective—all outputs justify themselves by reference to what the model was trained to predict.

  2. Hierarchize through branching: Information flows from top (general concepts) to bottom (specific instances). This mirrors how LLMs organize representations: broad patterns decompose into narrow instances.

  3. Binarize through distinction: At each node, something either belongs to a category or doesn't. An LLM either produces a specific token or doesn't; the representation space is partitioned into regions.

  4. Totalize through forced inclusion: Everything must fit somewhere in the hierarchy. Ambiguous or polysemous words must be resolved into a single representation. Contradictory interpretations cannot coexist.

The problem: natural language is not arborescent. Language is rhizomatic. The same word means different things in different contexts. Sentences support multiple valid interpretations. Contradictions coexist without requiring resolution.

Yet standard LLM architecture forces language into an arborescent structure. The model must choose which meaning of "bank" to represent. It must resolve the sentence into a single interpretation. It must collapse multiplicity into singularity.

2.2 What Is a Rhizomatic Structure?

Rhizomes have:

  1. Multiple beginnings: There is no single root. You can enter a rhizome at any point.

  2. Non-hierarchical connections: Elements connect laterally, not vertically. A connection between nodes is not subordination but affiliation.

  3. Heterogeneity: Different types of elements can coexist without being unified into a hierarchy. A rhizome can contain both meanings of "bank" without requiring that they reduce to a single concept.

  4. Lines of flight: Rhizomes have escape routes. They are not closed totalizing systems. Meaning can overflow, extend, diverge.

Deleuze & Guattari see this not as chaos but as a kind of order—rhizomatic order. It's structured, but the structure allows multiplicity to persist.

2.3 Why This Matters for AI

Standard LLMs totalize. They must produce a single output. They must resolve ambiguity. They must choose.

But this totalization comes at a cost:

A rhizomatic architecture would allow:

III. NCFCA AS RHIZOMATIC ARCHITECTURE

3.1 The Four-Channel Structure: Lateral, Not Hierarchical

NCFCA separates language processing into four channels:

Environment Channel    ← Raw observations
Interface Channel      ← Output production
Internal State Channel ← Semantic structures
Execution Channel      ← Behavioral effects

Standard LLM (arborescent):

NCFCA (rhizomatic):

3.2 The Protective Boundaries: Lines of Flight

In Deleuze & Guattari's terminology, a line of flight is an escape route—a way for the system to diverge rather than converge to a single structure.

NCFCA's one-way information flow creates lines of flight:

These protective boundaries prevent totalization. The system cannot force-close around a single interpretation. Old meanings can coexist with new meanings because they're held in protected states.

3.3 Heterogeneous Elements: Multiple Meanings Without Unity

Consider the word "bank" in NCFCA:

In a standard LLM:

In NCFCA:

This is rhizomatic in Deleuze's sense: multiple elements maintain their distinctness while being part of the same system.

3.4 Non-Hierarchical Connection: Affiliation, Not Subordination

Standard LLMs create hierarchies:

NCFCA creates lateral connections:

These are rhizomatic connections—affiliation rather than subordination.

IV. WILSON GATING AS DETERRITORIALIZATION

4.1 Deleuze & Guattari on Territoriality

In A Thousand Plateaus, a territory is a bounded, controlled space. A system is "territorialized" when it is fully contained within fixed boundaries.

Deterritorialization is the process of escaping fixed boundaries, moving to a smooth (non-striated) space.

4.2 Wilson Gates as Deterritorializing Forces

A Wilson Score gate validates whether a proposed state transition should occur:

Previous State: "BANK₁ = financial institution"
New Evidence: Numerous queries asking about "banks and rivers"
Question: Should we create BANK₂ or update BANK₁?

Wilson Gate Calculation:
  - Sample size: How many queries about "river"?
  - Confidence: How consistently does this pattern appear?
  - Penalty: Small samples don't justify new states
  - Decision: Create BANK₂ only if evidence is sufficiently robust

How is this deterritorializing?

In a territorialized system:

With Wilson gating:

Wilson gates are deterritorializing because they allow the system to expand, to create new meanings, to escape fixed semantic territories—but only when conditions warrant.

4.3 The Smooth Space of Multiplicity

Deleuze distinguishes between:

NCFCA with Wilson gating creates a smooth semantic space:

V. WHAT THE ARGUMENT CLAIMS, AND WHAT IT DOES NOT

5.1 The claim is structural

The argument of this paper is about structure. A system that forces every meaning of a polysemous term into one representation has made a structural choice, and a system that holds the meanings in separate, protected states has made a different one. Deleuze and Guattari's distinction predicts that the first choice costs something: distinct multiplicities get flattened into a unity that none of them fits. That prediction is a philosophical claim about what arborescent organization does to heterogeneous elements. Whether it holds of any particular language model is a separate, empirical question.

5.2 What has not been shown

No independent party has tested NCFCA. The lab's own internal tests have not been repeated by anyone outside it, and they are not cited here as support for any claim in this paper. The statements in this paper about how standard language models behave (that distinct meanings interfere in a single representation space, that interference produces confabulation) are conjectures stated for the purpose of the philosophical argument. They are not established findings.

5.3 What evidence would be needed

Evidence for the structural claim would take the form of large-scale, repeated studies, run by parties other than the architecture's author, in which the correlation between the theory's prediction and the measured outcome rises in proportion to the fidelity of the experiment. A single comparison, however striking, does not meet that bar. Until such studies exist, the architecture is a theoretical proposal and the paper's argument must stand without it.

5.4 What follows for the philosophy

The philosophical argument does not depend on the empirical question. Even if no implementation ever behaved as predicted, the conceptual point would survive: the arborescent/rhizomatic distinction identifies a real design choice about whether multiplicity is preserved or collapsed, and that choice can be stated precisely in architectural terms. That is the contribution the rest of the paper develops.

VI. MULTIPLE ENTRY POINTS: THE RHIZOMATIC ADVANTAGE

6.1 How Rhizomes Escape Single-Root Architecture

A tree has one root. Everything is organized relative to that root. You must understand the root to understand the branches.

A rhizome has no privileged root. You can enter anywhere. The system responds differently depending on your entry point, but the system itself doesn't need a single unified origin.

6.2 NCFCA Enables Multiple Entry Points

In a standard LLM, all queries feed into a single representation space. The query must be mapped to a location in that space. The system's response is determined by that location.

In NCFCA, different types of queries can enter through different channels:

The system doesn't need to force all queries through a single funnel. Multiple entry points allow multiple response modes.

Advantage: The system can respond differently to different query types without each response type requiring the entire system to reorganize.

6.3 Rhizomatic Implications

Deleuze & Guattari argue that rhizomatic systems are more adaptive because they don't have a single root to defend. You can damage one entry point, and the system continues functioning through other entry points.

This is true for NCFCA:

VII. LINES OF FLIGHT AND SEMANTIC OVERFLOW

7.1 Deleuze's Concept of Lines of Flight

A line of flight is a trajectory that escapes a structure. Not destruction of the structure, but escape—the possibility of moving beyond fixed boundaries.

A territorialized system has no lines of flight. It is enclosed within its structure.

A deterritorialized system has multiple lines of flight—ways of moving beyond established categories.

7.2 Semantic Overflow in NCFCA

Consider a situation where a new usage pattern emerges that doesn't fit existing semantic states:

Standard LLM (arborescent):

NCFCA (rhizomatic):

Philosophical meaning: The system can overflow its boundaries. It can create new semantic territories. It is not enclosed within a fixed structure.

7.3 Continuous Semantic Variation Without Rupture

Deleuze distinguishes between:

NCFCA enables continuous semantic variation:

VIII. THE CRITIQUE ANTICIPATED: DOES MULTIPLICITY PREVENT COHERENCE?

8.1 The Objection

A critic might argue: If the system preserves multiple meanings without forcing unity, doesn't it become incoherent? Won't it produce contradictory outputs?

This is the arborescent question: Coherence requires unity.

8.2 The Rhizomatic Response

Deleuze & Guattari argue that coherence does not require unity. A rhizome is coherent through its connections, not through hierarchical unification.

NCFCA demonstrates this:

The system is coherent, but its coherence comes from lateral connections and contextual disambiguation, not from arborescent hierarchy.

IX. IMPLICATIONS FOR PHILOSOPHY OF LANGUAGE

9.1 Against Semantic Singularism

This paper argues against semantic singularism—the view that each word has a single meaning (or that meaning must be unified into a single representation).

The case against it does not need experiments. Polysemy is ordinary and pervasive: "bank," "light," "match," and "bear" each carry unrelated meanings, and speakers keep them apart by context without any effort. A theory of meaning that assigns one unified representation to each word must explain this by appeal to something outside the representation, namely context. That concession already admits that the word alone does not fix its meaning, which is the position of the plural view.

The plural view is therefore the default, and the burden falls on the singularist. Whether collapsing meanings into one representation has measurable costs in a given system is a further empirical question that this paper does not answer.

9.2 Toward a Rhizomatic Semantics

A rhizomatic semantics would hold:

  1. Meanings are multiple: Words can carry distinct, coexisting meanings.
  2. Meanings are connected but not unified: They relate to each other through context and usage, not through hierarchical reduction.
  3. Meanings proliferate: New meanings emerge continuously as language is used.
  4. Coherence is contextual: Utterances are coherent not because they express unified meanings but because they activate relevant states in context.

This is closer to how language actually works than semantic singularism.

9.3 Deleuze & Wittgenstein: Unexpected Alignment

Wittgenstein argued that meaning is constituted by use in language games. Different games give different meanings to the same word.

Deleuze & Guattari propose rhizomatic multiplicity—different meanings coexisting without unity.

These are surprisingly aligned: Both reject the idea that meaning is singular and unified. Both treat meaning as distributed across uses and contexts.

NCFCA as a rhizomatic system operationalizes this insight: it allows multiple meanings (language games) to coexist in protected states without forcing them into unity.

X. PHILOSOPHICAL CONCLUSION

Deleuze & Guattari wrote A Thousand Plateaus as a philosophical treatise, not as a technical specification. Yet their concepts illuminate fundamental problems in modern AI.

The core insight: Deleuze and Guattari's distinction identifies a real design choice for any system that handles plural meanings: collapse them into one hierarchy, or hold them apart in protected states and connect them laterally. The first choice has a conceptual cost that their philosophy names precisely. The second can be stated as an architecture.

Why does this matter philosophically?

Because it suggests that how we structure our thinking systems shapes what those systems can do. An arborescent architecture cannot preserve semantic multiplicity. A rhizomatic architecture can.

Deleuze & Guattari were arguing for a different way of thinking—against hierarchies, against unification, against totalization. They were arguing for systems that allow multiplicity to flourish.

NCFCA is one proposal for what such a design would look like. It is a theoretical architecture, untested by any outside party, and this paper does not claim it performs better than anything else. The claim is narrower: the philosophical distinction has a precise architectural counterpart, and that counterpart deserves independent study.

REFERENCES

Deleuze, G., & Guattari, F. (1987). A Thousand Plateaus: Capitalism and Schizophrenia. University of Minnesota Press.

Wittgenstein, L. (1953). Philosophical Investigations. Blackwell.

Deleuze, G. (1993). The Fold: Leibniz and the Baroque. University of Minnesota Press.

[Additional references would include contemporary work on LLM architectures, multi-agent systems, semantic theory, and Deleuzian philosophy—to be completed for journal submission]

Word Count: ~3,200 (outline/framework; full paper target: ~9,000)
Status: Draft framework. Reconstructed from a short description, not from the author's original manuscript. Not submitted, not peer reviewed.
Core Contribution: Applies Deleuze & Guattari's rhizomatic philosophy to multi-agent architecture and argues that NCFCA is structurally rhizomatic.
Evidence status: None. No empirical results are cited. Claims about the behavior of standard language models are conjectures. The architecture is theoretical and untested by any outside party.