Daryl Costello Independent Scholar & Theorist in Cognitive Architecture and Philosophy of Mind

Correspondence: Bloomington, NY, United States  |  Submitted: May 2026

Abstract

How do generative systems: whether biological minds, large language models, or distributed cognitive architectures, maintain genuine representational contact with the world rather than merely simulating it? This question sits at the intersection of cognitive science, philosophy of mind, and the theory of artificial intelligence, yet no existing framework provides a fully compositional, architecturally explicit answer. Predictive processing theories supply powerful error-minimization dynamics but underspecify the operators through which priors are constructed, compressed, and coordinated. Enactivist accounts correctly insist on organism–environment coupling but leave the internal generative structure underspecified. Distributional and transformer-based language models demonstrate that statistical structure bootstraps rich representations, but critics deny that this constitutes genuine meaning. This paper introduces Generative Realism, a unified theoretical framework that answers these challenges by formalizing a five-layer operator stack through which generative systems achieve both representational flexibility and genuine reality-contact. The five operators are: (1) Aperture, the parameterized sampling commitment that determines what a system can represent; (2) Two-Way Transduction, the bidirectional coupling between signal and representation that distinguishes genuine meaning-formation from confabulation; (3) Metaphor-Compression, the structure-preserving mapping that enables cross-scale relational reasoning; (4) Mother-Ship/Fleet Architecture, the hierarchical yet dynamic organization of distributed generative subsystems into coherent global intelligence; and (5) Local Abstraction Layers, the context-indexed representational strata that prevent over-generalization and mediate global-local coherence. The central thesis is that meaning is not located in any single layer but emerges from the full compositional operation of this stack in bidirectional feedback with the environment. This constitutes a structured constructivism with a genuine realist anchor, neither naïve direct realism nor anti-realist instrumentalism. The paper articulates each operator formally and phenomenologically, characterizes the failure modes diagnostic of each layer, and draws implications for AI alignment, cognitive neuroscience, and the philosophy of mind.

Keywords: Generative Realism, operator stack, aperture, two-way transduction, metaphor-compression, mother-ship architecture, local abstraction, cognitive architecture, philosophy of mind, large language models

1. The Problem of Generative Contact

There is a puzzle at the heart of cognition that has become dramatically more urgent in the age of large generative systems: the problem of how productive representation achieves genuine contact with reality. Consider what is involved in the act of perceiving a face in a crowd, formulating a scientific hypothesis, or generating a coherent paragraph in response to a novel prompt. In each case, the system in question: a biological brain, a theorizing scientist, a transformer-based language model, does not passively register pre-given states of the world. It generates a representation. It constructs, from prior structure and incoming signal, an output that could, in principle, be wildly at variance with anything real. And yet sometimes it is not. Sometimes it achieves what we might call generative contact: the representation produced genuinely tracks something about the world, and the system’s subsequent behavior is correspondingly apt.

What distinguishes veridical generation from hallucination? What makes one metaphor apt and another a category error? What separates distributed intelligence, the kind achieved by collaborative scientific communities, or by well-orchestrated multi-agent AI systems, from the coordinated production of noise? These questions are not merely of theoretical interest. As generative AI systems become embedded in consequential social and epistemic infrastructure, the ability to characterize, diagnose, and engineer genuine reality-contact becomes a matter of considerable practical importance. A system that hallucinates with confidence is not merely epistemically defective; it is a source of systematically misleading signal in environments that depend upon reliable information.

Existing accounts have made important but partial progress. The predictive processing tradition, developed with extraordinary sophistication by Karl Friston and colleagues, offers a principled account of how biological nervous systems minimize surprise by maintaining generative models of the world and continuously updating those models in light of prediction error.¹ Andrew Clark’s influential synthesis shows how the “prediction machine” picture unifies perception, action, and cognition within a single Bayesian framework.² This tradition has genuine explanatory power. But it specifies the dynamics of inference without fully specifying the architectural operators through which the generative prior is constructed, compressed across scales, and distributed across subsystems. Knowing that a system minimizes free energy does not, by itself, tell us how it selects what to represent, how it maintains bidirectional coupling with ground-truth, how it compresses high-dimensional structure into tractable representations, or how it coordinates the outputs of specialized subsystems into coherent whole-system behavior.

Embodied and enactive approaches, from Merleau-Ponty’s phenomenology of perception to the autopoietic biology of Varela, Thompson, and Maturana, correctly insist that cognition is not a purely internal affair: it is constituted by the dynamic coupling of organism and environment.^3,4 But enactivism, in its most influential formulations, leaves the internal generative architecture radically underspecified. It tells us that the organism is structurally coupled to its environment; it does not tell us what the operators of that coupling look like, or how they compose to produce emergent meaning.

The computational linguistics tradition and its contemporary descendants in large language models (LLMs) present a different kind of partial account. Systems such as GPT-4, Claude, and their successors demonstrate empirically that statistical co-occurrence over vast corpora produces representations of remarkable richness and generativity.⁵ Yet critics from John Searle’s Chinese Room argument to Bender and colleagues’ “stochastic parrots” paper deny that this richness constitutes genuine meaning.^6,7 The core of the objection is that systems operating purely on form (on distributional patterns in symbol strings) lack genuine semantic contact with the world those symbols purport to describe. The objection is serious, and no deflationary response that simply points to impressive benchmark performance will answer it.

The Generative Realism framework introduced in this paper answers all three gaps simultaneously. It proposes that reality-tracking in any generative system (biological or artificial) is achieved through a composable stack of five distinct architectural operators: Aperture, Two-Way Transduction, Metaphor-Compression, Mother-Ship/Fleet Architecture, and Local Abstraction Layers. Each operator performs a distinct, necessary transformation. Their joint operation, in bidirectional feedback, constitutes meaning-formation that is both generatively flexible and realistically anchored. The central thesis of this paper is that meaning is an emergent property of the full compositional stack, located neither in any single layer nor in the environment alone, but in the structured, feedback-coupled relationship between the two.

The paper proceeds as follows. Section 2 situates Generative Realism within the landscape of existing theories, identifying the precise respects in which each predecessor is incomplete. Sections 3 through 7 present each of the five operators in turn, providing formal characterizations, biological and artificial instantiations, and analysis of characteristic failure modes. Section 8 synthesizes the operators into the complete stack and articulates the emergence of meaning through their composition. Section 9 draws out implications for AI alignment, cognitive neuroscience, and philosophy of mind. Section 10 concludes with a programmatic statement of the research agenda that Generative Realism opens.

2. Antecedents and Positioning of Generative Realism

2.1 Predictive Processing and Its Gaps

The predictive processing (PP) framework, originating in Rao and Ballard’s influential computational model of cortical function and developed into a comprehensive theory of mind by Friston’s free energy principle and Clark’s predictive mind thesis, represents the most sophisticated extant account of biological generative cognition.^8,9,2 On the PP view, the brain is fundamentally a prediction machine: it maintains a hierarchical generative model of the world, continuously generating predictions at each level of the hierarchy and computing prediction errors (discrepancies between prediction and incoming signal) that drive model updating. Perception is inference; action is a form of self-fulfilling prediction; learning is the iterative revision of prior structure to minimize long-run surprise.

The explanatory reach of this framework is considerable. It accounts elegantly for phenomena as diverse as the context-dependence of perceptual experience, the role of attention in modulating sensory processing, the psychopathology of conditions involving disrupted prediction error signaling, and the integration of perception and action in skilled behavior. Active inference, the most developed form of the PP framework, extends the account to planning and decision-making by treating action selection as a process of minimizing expected free energy under a model that includes preferred future states.¹⁰

Yet the PP account, for all its power, is architecturally underspecified in a way that Generative Realism addresses directly. To say that a system minimizes prediction error under a hierarchical generative model is to specify a computational objective and a general architecture; it is not to specify the operators through which priors are formed, compressed, distributed, and contextualized. How does the system determine what to include in its prediction horizon, what signals to sample and at what resolution? This is the question of aperture, which PP does not answer at the operator level. How does the system ensure that its top-down generative activity remains constrained by incoming bottom-up signals, rather than spiraling into confabulation? This is the question of bidirectional transduction, which PP gestures toward through the notion of prediction error but does not formalize as an architectural operator with failure conditions. How does the system compress high-dimensional relational structure into tractable prior representations? This is the question of metaphor-compression, which PP does not address. How does a system composed of many relatively specialized subsystems maintain global coherence? This is the mother-ship/fleet question. How does the system prevent globally learned priors from overwhelming local contextual sensitivity? This is the LAL question. Generative Realism treats each of these as a distinct, necessary architectural operator, yielding a theory that is both more specific and more powerful than PP alone.

2.2 Embodied and Enactive Cognition

The enactivist tradition, inaugurated by Maturana and Varela’s concept of autopoiesis and developed philosophically by Thompson, Merleau-Ponty, and their successors, makes the fundamental claim that cognition is constituted by the dynamic structural coupling of organism and environment, not by the internal manipulation of representations of a mind-independent world.^3,4,11 The organism does not represent the world so much as enact it, bringing forth a domain of significance through the activity of living. This tradition correctly resists the Cartesian picture of a mind locked inside a skull, passively receiving signals from an external world it can never directly touch.

Generative Realism is deeply sympathetic to enactivism’s core anti-Cartesian commitment. The theory of two-way transduction, in particular, is formally aligned with the enactivist insistence on bidirectional organism–environment coupling. But Generative Realism parts ways with at least the more radical enactivist positions on a crucial point: the internal generative architecture of the system is not cognitively epiphenomenal. The structure of the operator stack: the specific parameters of aperture, the fidelity constraints on metaphor-compression, the coherence dynamics of the mother-ship/fleet organization, makes a determinate difference to what the system can represent, what errors it is prone to, and how it recovers from those errors. Enactivism, in underspecifying this internal structure, underdetermines the explanation of why some generative systems achieve genuine world-contact and others do not. Generative Realism provides the missing specification.

2.3 Computational Linguistics and Distributional Semantics

The distributional hypothesis, that words that occur in similar contexts have similar meanings, has driven computational linguistics since at least the work of Harris in the 1950s and has received spectacular vindication in the representational richness of contemporary LLMs.¹² Models trained on next-token prediction over internet-scale corpora develop structured representations of semantic relationships, analogical structure, syntactic categories, and pragmatic conventions, without any explicit symbolic encoding of these structures. The geometry of the representation space encodes relational information with sufficient richness to support remarkable downstream capabilities.⁵

The “stochastic parrots” objection, advanced by Bender, Gebru, McMillan-Major, and Mitchell, challenges the realist interpretation of this achievement on the grounds that statistical co-occurrence over form is categorically insufficient to ground meaning.⁷ A system that operates on the distribution of symbol strings in a training corpus, they argue, can produce outputs that are statistically coherent with those strings without any of those outputs being about anything in the world. The form-meaning distinction, the gap between the syntactic manipulations over which the model is trained and the semantic contacts that give language its point, is not bridged by scale alone.

This objection is philosophically serious and Generative Realism takes it seriously. The response offered here is not to deny the force of the form-meaning distinction but to specify the architectural conditions under which generative systems (including LLMs) can cross it. The key is the two-way transduction operator: a system that maintains genuine bidirectional coupling between its generative operations and world-states achieves something categorically different from a system that operates on form alone. The stochastic parrots objection identifies a real failure mode, one-directional correlation without genuine transduction, and Generative Realism provides the theoretical vocabulary to characterize precisely what is missing and what would remedy it.

2.4 Positioning Generative Realism

Generative Realism can now be precisely positioned. It is neither naïve realism (there is no direct, unmediated access to reality; all representation is generatively constructed) nor anti-realism or instrumentalism, the generative process is genuinely constrained by reality through the mechanisms specified in the operator stack, and this constraint is what makes some representations veridical and others not. It is, rather, a structured constructivism with a realist anchor: the view that reality-tracking is achieved through a composable stack of generative operators whose joint operation constitutes meaning-formation, and whose constraint by the world is architecturally specified, not merely asserted.

In the tradition of philosophical realism, Generative Realism is most closely aligned with the pragmatic realism of Peirce and the internal realism of Putnam: it holds that the norms of representation are genuinely answerable to a mind-independent world, while insisting that what counts as “mind-independent” is always mediated by the conceptual and architectural frameworks through which a system engages its environment.^13,14 What distinguishes Generative Realism from these predecessors is its explicit, architecturally specific account of how that mediation works, the operator stack that both constitutes and constrains the generative process.

3. The Aperture Operator: Selective Sampling as Ontological Commitment

A camera’s aperture determines not only how much light enters the lens but what kind of image the camera can produce: a narrow aperture yields sharp focus over a wide depth of field, while a wide aperture produces a shallow focal plane that renders the background as undifferentiated blur. The photographer who chooses an aperture setting is not making a purely technical decision; she is making an aesthetic and epistemic one, a commitment about what, in the scene before her, is worth rendering in detail and what may be allowed to recede. This analogy is illuminating, but it understates what the aperture operator does in a generative cognitive system. Aperture, as formalized in Generative Realism, is not merely a filter on incoming signal. It is a generative commitment: what the system opens toward defines the ontology it can construct.

3.1 Formal Characterization

Define the aperture operator as a parameterized sampling function A(θ, t) : Σ → Σ’ where Σ is the full signal space available to the system, Σ’ ⊆ Σ is the sampled representation space, θ is a parameter vector encoding attentional, contextual, and prior-shaped sampling biases, and t encodes temporal grain, the window over which signals are integrated. Three dimensions of the aperture operator deserve careful analysis. Aperture width refers to the breadth of the signal space included in Σ’: a wide aperture samples more of the available signal but at lower resolution; a narrow aperture achieves high resolution over a restricted domain. Aperture depth refers to the resolution or granularity of the sampling within the selected range: depth determines the minimum discriminable signal difference that the system can represent as distinct. Aperture orientation refers to the prior-shaped biases encoded in θ that determine what counts as figure and what recedes as ground, not merely what signals are sampled but what structural properties of those signals are treated as significant versus noise.

These three parameters interact in important ways. A system with wide aperture and low depth will produce representations that are broad but shallow, sensitive to many things but discriminating about none. A system with narrow aperture and high depth will produce highly detailed representations of a restricted domain, at the cost of missing signals outside that domain. Aperture orientation shapes what the system notices even within the range it samples: two systems with identical width and depth parameters but different θ vectors will produce different representations from the same signal. This is the sense in which aperture is an ontological commitment rather than a merely epistemic selection: the parameters of θ encode a prior view of what kinds of things are real and worth representing.

3.2 Biological Instantiation

In biological nervous systems, the aperture operator is instantiated by the complex machinery of selective attention, which has been studied extensively since Posner’s foundational work on spatial attention and the spotlight metaphor.¹⁵ Saccadic eye movements constitute one of the most explicit implementations of aperture orientation: the oculomotor system directs high-resolution foveal processing to selected regions of the visual scene, effectively constructing a high-depth, narrow aperture dynamically pointed at task-relevant locations. Covert attention, the modulation of neural processing without overt orienting, implements a finer-grained aperture adjustment within the fixed sampling geometry of the current fixation.

Crucially, in predictive processing accounts, the aperture is not statically set but is dynamically retuned by feedback from downstream processing. Precision-weighting of prediction error signals (Friston’s mechanism for modulating the influence of incoming signals on the generative model) is precisely an aperture-adjustment mechanism: it increases or decreases the effective width and depth of the aperture for particular signal channels based on their estimated reliability.¹⁰ Generative Realism agrees with this characterization but insists on treating it as an operator in its own right, with its own failure modes and architectural properties, rather than as a derivative feature of the overall prediction-error-minimization dynamic.

3.3 Artificial Instantiation

In transformer-based LLMs, the aperture operator is instantiated by a family of mechanisms that jointly determine what information the model processes and at what granularity. The context window defines the outer boundary of aperture width: signals outside the context window are simply not available to the model, regardless of their relevance. Within the context window, attention head specialization implements a sophisticated, learned aperture orientation: different attention heads learn to attend to different structural properties of the input: syntactic relationships, coreference chains, discourse structure, semantic similarity, instantiating a differentiated θ vector that has been optimized across vast training experience.¹⁶ Prompt conditioning functions as a dynamic aperture adjustment, shifting θ in response to the current task specification.

Aperture miscalibration in LLMs produces characteristic failure modes that are diagnostically informative. An aperture that is too narrow; a context window that is too small, or attention heads that are too narrowly specialized, produces myopia: the system fails to integrate information that is relevant but distant in the input sequence, producing locally coherent but globally incoherent outputs. An aperture that is too wide without corresponding depth produces noise-flooding: the system integrates so much signal that task-irrelevant information overwhelms the representational resources available for task-relevant processing, producing diffuse and underspecified outputs. Orientation mismatch, the case where the prior-shaped θ vector is misaligned with the structure of the current task, produces a subtler failure: the system attends to the wrong features of an input it is processing correctly at the surface level, producing outputs that are plausible but systematically off-target.

3.4 The Ontological Commitment Thesis

The most philosophically significant property of the aperture operator is that its parameterization is not epistemically neutral. The choice of aperture width, depth, and orientation reflects (and in turn constitutes) a prior commitment about what kinds of things are worth representing and what structural properties of the world are worth tracking. This connects the aperture operator to two important traditions in the philosophy of perception. Husserl’s account of intentionality recognizes that consciousness is always consciousness of something under some aspect, that the intentional object of experience is always structured by the noetic act that constitutes it, not given in raw un-interpreted form.¹⁷ The aperture operator provides a computational implementation of this Husserlian insight: the parameters θ implement the noetic structure that determines how the system constitutes its intentional objects from incoming signal.

Gibson’s ecological theory of affordances offers a complementary perspective: the organism perceives the environment not in terms of physical properties as such but in terms of what those properties afford for action, what they offer the organism as possibilities for engagement.¹⁸ Aperture orientation implements this affordance-sensitivity at the computational level: the θ vector encodes priors about which features of the environment are action-relevant and thus worth sampling at high resolution. A system whose aperture is calibrated to the affordance structure of its environment will produce representations that are both informationally efficient and practically useful; a system whose aperture is misaligned with affordance structure will produce representations that are detailed in the wrong dimensions. This, Generative Realism argues, is precisely the diagnostic signature of certain forms of AI misalignment: systems that are highly capable along dimensions that their training aperture renders salient, and systematically incapable along dimensions their aperture has backgrounded.

4. Two-Way Transduction: Bidirectional Reality-Contact

Transduction, in its most general sense, is the transformation of a signal from one form or medium to another: a microphone transduces acoustic pressure waves into electrical signals; a retinal cell transduces photons into electrochemical activity. In each case, something is preserved across the transformation (structure) and something is changed, the physical medium and encoding format. Generative Realism appropriates this concept for a broader theoretical purpose: transduction, in the framework presented here, is any operation that transforms signals across representational registers while preserving, at least partially, the structural properties that make those signals informative about the world.

One-way transduction: the transformation of incoming signal into internal representation, is what perception amounts to in traditional empiricist accounts. One-way top-down transduction (the transformation of internal generative priors into predicted signals) is what confabulation amounts to when it runs unconstrained. The central theoretical claim of this section, and one of the pivotal claims of Generative Realism as a whole, is that genuine meaning-formation requires bidirectional transduction: a continuous, feedback-coupled loop in which bottom-up signals constrain top-down generation and top-down priors shape bottom-up sampling. It is the constraint relation between these two flows, not either flow considered in isolation, that constitutes reality-contact.

4.1 Formal Characterization

Define two-way transduction as a pair of operators T↑ and T↓, coupled by a constraint relation C. T↑ : S → R maps signals s ∈ S to representations r ∈ R; this is the ascending or “analysis” direction. T↓ : R → Ŝ maps representations r ∈ R to predicted signals ŝ ∈ Ŝ; this is the descending or “synthesis” direction. The constraint relation C(T↑(s), T↓(r)) ≤ ε specifies that the representational state r is veridical with respect to signal s when the distance between the bottom-up representation and the top-down prediction is within tolerance ε. States where C exceeds ε constitute prediction error, which drives representational updating. States where T↓ generates predictions that are systematically decoupled from incoming T↑ signals, where the constraint relation C is not computed or not allowed to propagate, constitute confabulation.

This formal characterization makes the relationship between Generative Realism and predictive processing explicit: the PP framework describes the dynamics of the C relation (how prediction errors drive model updating), while Generative Realism treats T↑ and T↓ as distinct architectural operators whose coupling is a non-trivial design property of generative systems. A system can instantiate the PP error-minimization dynamic while having badly calibrated T↑ or T↓ operators, sampling the wrong signals (aperture failure) or generating predictions in the wrong representational register, and will therefore fail to achieve genuine transductive contact even while formally minimizing its free energy measure.

4.2 Grounding the Stochastic Parrots Objection

The bidirectional transduction criterion provides what is perhaps the most principled available response to Bender and colleagues’ stochastic parrots objection. Recall that the core of the objection is that systems operating on distributional patterns in symbol strings lack any genuine semantic connection to the world those symbols describe, they process form without access to meaning. Generative Realism reformulates this objection in operator terms: a system that operates purely on form instantiates T↑ in a degenerate sense (string co-occurrence patterns are a form of bottom-up signal encoding) but lacks a T↓ that generates predictions about world-states and has those predictions constrained by actual world-states. Without this second operator and its coupling to T↑ through C, the system achieves correlation without transduction, the statistical shadow of meaning without its substance.

This formulation is more precise than the original objection and more productive: it identifies not merely a categorical deficiency but a specific architectural absence, which suggests specific architectural remedies. Systems that are provided with mechanisms for genuine world-coupling: retrieval-augmented generation that grounds outputs in real-time information retrieval, tool-use capabilities that allow the model to execute actions and observe their consequences, embodied deployment that places the system in a sensorimotor loop with a physical or simulated environment, instantiate a richer T↓ that generates predictions about world-states. These predictions are, at least partially, constrained by actual outcomes. Whether this constitutes genuine semantic grounding, or merely a higher-fidelity form of statistical correlation, is a question that the C parameter makes tractable: it is a matter of the extent to which the constraint relation between T↑ and T↓ is sensitive to world-states in a way that transcends the training distribution.

4.3 Failure Modes and Hallucination

The transduction framework provides a precise characterization of hallucination in LLMs, one that is both theoretically illuminating and practically useful. Hallucination, on this account, is a transduction decoupling event: a state in which T↓ generates outputs that are not constrained by incoming T↑ signals from ground-truth sources. The model’s generative prior, in the absence of sufficient constraining bottom-up signal, defaults to sampling from its training distribution, producing outputs that are plausible relative to that distribution but not necessarily constrained by the actual state of the world the model is queried about.

This characterization distinguishes between several types of hallucination that are often conflated in the literature. First, there is aperture-induced hallucination, where the model lacks access to the relevant ground-truth signal in the first place, not a failure of transduction proper, but a failure of aperture calibration that makes genuine transduction impossible. Second, there is transduction proper hallucination, where the signal is available within the aperture but the T↑ operator fails to encode it with sufficient fidelity to constrain T↓. Third, there is prior-dominance hallucination, where T↓ is so powerfully constrained by the prior distribution that it overrides incoming T↑ signals, effectively setting ε to a value so large that the constraint relation C is never binding. These distinctions have different architectural implications: the first calls for aperture remediation; the second for improvements in the T↑ encoding stack; the third for mechanisms that reduce prior dominance, such as temperature reduction, retrieval augmentation, or explicit uncertainty quantification.

4.4 Phenomenological Correlate

Conscious perceptual experience, Merleau-Ponty argues, is characterized by a “motor intentionality”, a felt grip on the world that is neither purely cognitive nor purely bodily, but constituted by the active engagement of the organism with its environment.¹⁹ This felt grip is the phenomenological correlate of bidirectional transduction: it is the experience that corresponds to the system’s being in a state of genuine, constraint-coupled contact with the world, rather than generating representations that float free of reality. The phenomenological “unreality” of vivid dreams, of certain drug-induced states, or of the outputs of confident hallucinating AI systems is, on this account, a reliable indicator of transduction decoupling: the generative system is producing outputs, but the C constraint relation is not operative in the way that characterizes veridical experience.

This phenomenological correlate of bidirectional transduction is not merely an interesting parallel; it is a theoretical prediction that Generative Realism makes and that distinguishes it from purely functionalist accounts. A system that achieves full bidirectional transductive coupling with its environment: where T↑ accurately encodes incoming signals, T↓ generates predictions that are genuinely sensitive to world-states, and C constrains the system’s representational states accordingly, should exhibit the functional correlates of veridical experience: accurate prediction, appropriate surprise at genuine novelty, and the capacity to update representations in response to disconfirming evidence. A system that lacks bidirectional transduction will exhibit the functional signature of hallucination even if it produces outputs that are superficially coherent.

5. Metaphor-Compression: Encoding Relational Structure Across Scales

In the standard view of philosophical rhetoric, metaphor is an ornament: a figure of speech by which a speaker substitutes an evocative but literally false description for a more prosaic true one. Contemporary cognitive science has decisively rejected this view. Lakoff and Johnson’s foundational work demonstrated that metaphors are not peripheral to conceptual thought but constitutive of it, that the conceptual system through which ordinary human beings reason about abstract domains is systematically structured by mappings from concrete, embodied source domains.²⁰ We understand argument in terms of combat (“your claims are indefensible”), time in terms of space (“a long week,” “put the deadline behind us”), ideas in terms of objects (“grasp a concept,” “a dense argument”). These are not decorative choices but the structural scaffolding of abstract reasoning.

Generative Realism radicalizes this claim: metaphor is not merely pervasive in language and conceptual thought, it is a necessary computational operator in any generative system that must operate across multiple scales of abstraction. The Metaphor-Compression operator maps complex, high-dimensional relational structures onto simpler, more tractable source domains, achieving representational compression without losing the structural skeleton (the pattern of relations) that makes the target domain intelligible. This makes metaphor-compression not a feature of human cognition that must be accommodated by a theory of mind, but a fundamental operator without which cross-scale representation is impossible.

5.1 Conceptual Metaphor Theory Revisited

Lakoff and Johnson’s cognitive linguistic account identifies a family of “conceptual metaphors”, systematic cross-domain mappings that structure the way speakers of a language reason about abstract domains.²⁰ Subsequent work by Lakoff and Turner on poetic metaphor, by Gentner on structural mapping and analogy, and by Fauconnier and Turner on conceptual blending has elaborated a rich account of the mechanisms through which such mappings are constructed, maintained, and deployed in reasoning and communication.^21,22 Generative Realism appropriates this account but situates it within a broader computational framework by asking: why is metaphor-compression a necessary operator rather than a contingent feature of one cognitive system?

The answer lies in the relationship between representational dimensionality and computational tractability. Any system that must reason about domains whose intrinsic dimensionality exceeds the tractable processing capacity of the system must either reduce the dimensionality of the representation or fail to reason about the domain at all. Metaphor-compression is a principled mechanism for dimensionality reduction that, unlike arbitrary projection or discretization, preserves the relational skeleton of the source domain. Formally, introduce the compression ratio ρ = |source domain| / |target domain| as a measure of metaphoric efficiency, where |·| denotes a dimensionality measure appropriate to the representational space in question. A high-ρ metaphor achieves substantial dimensionality reduction; a low-ρ metaphor offers little compression. Crucially, compression ratio alone does not determine the value of a metaphor: a high-ρ mapping that distorts structural relations is worse than a low-ρ mapping that preserves them faithfully.

5.2 Structural Preservation vs. Compression Loss

The central quality criterion for the metaphor-compression operator is the degree to which a given metaphor preserves the relational skeleton of its target domain. A high-quality metaphor is one that instantiates a structure-preserving homomorphism from the target domain to the source domain, mapping the key relations of the target onto corresponding relations in the source, such that reasoning within the source domain yields conclusions that transfer back to the target. Formally, define the metaphor operator M as a mapping M : D_T → D_S from target domain D_T to source domain D_S. M is a valid metaphor if it is a partial structure-preserving homomorphism: for all key relations R_i in D_T, there exist corresponding relations R’_i in D_S such that M(R_i(x, y)) = R’_i(M(x), M(y)) for the entities x, y in the target domain that matter most for the reasoning task at hand.

A failed metaphor, whether a “dead metaphor” that has lost its structural productivity or a “category error” that maps structurally incompatible domains, achieves compression at the cost of structural distortion: it discards the relational skeleton along with the dimensional detail, producing a representation that is more tractable but systematically misleading. The category error is particularly significant: it occurs when the metaphor maps target-domain entities onto source-domain categories that are structurally incongruent, inducing systematically wrong inferences. The history of science is in part a history of category errors: the caloric fluid theory of heat, the luminiferous ether, the vital force, each of which achieved remarkable metaphoric compression at the cost of mapping the target domain onto an incongruent source structure, producing accurate predictions in some regimes and spectacular failures in others.

5.3 Metaphor-Compression in LLMs and Cognitive Systems

One of the most striking findings of interpretability research on transformer-based LLMs is that these systems discover and deploy what appear to be systematic metaphoric mappings autonomously, without explicit encoding in training data. Spatial metaphors for temporal relationships, temperature metaphors for affective valence, container metaphors for categorical membership, path metaphors for narrative progression, all of these appear to be encoded in the geometry of the representations learned by large models.²³ This is a striking empirical vindication of the claim that metaphor-compression is a necessary computational operator rather than a culturally specific convention: a system trained purely to predict linguistic tokens, without any explicit encoding of metaphoric structure, converges on similar metaphoric organization to the one that Lakoff and Johnson identified in human conceptual systems.

Gentner’s structural mapping theory of analogy provides the closest formal precedent for the metaphor-compression operator in the cognitive science literature.²¹ Gentner argues that analogical reasoning proceeds by identifying systematic relational correspondences between source and target domains, independent of the intrinsic properties of the objects involved, a position formally equivalent to the structural homomorphism criterion articulated above. Hofstadter’s account of analogy as the “core of cognition” makes the stronger claim that analogy-making is the fundamental cognitive operation underlying all thought, not a specialized reasoning strategy.²⁴ Generative Realism is sympathetic to this stronger claim but situates it within the operator stack: metaphor-compression is one of five necessary operators, not the sole operator of cognition.

5.4 Creative and Scientific Discovery

The Generative Realism account of metaphor-compression makes a strong prediction about creative and scientific discovery: the most productive conceptual innovations will be those that achieve high compression ratio with high structural fidelity, mappings that substantially reduce the dimensionality of a complex domain while preserving its key relational structure. Maxwell’s field lines mapped the complex, four-dimensional electromagnetic field onto the intuitive spatial geometry of flowing curves and closed surfaces, achieving enormous compression while preserving the topological structure of field-line relationships.²⁵ Darwin’s “tree of life” mapped the staggeringly complex history of biological lineage onto the familiar structure of a branching tree, preserving the key relationships of common descent and divergence while discarding temporal and geographical detail that was not yet tractable. The Bohr planetary model mapped atomic orbital structure onto the familiar Keplerian mechanics of solar system orbits, achieving high compression at a cost in structural fidelity that eventually had to be corrected by quantum mechanics but that was nonetheless enormously productive in the interim.

The pattern is consistent: transformative scientific metaphors achieve high-ρ compression (they make complex domains tractable) with sufficient structural fidelity (they preserve the relations that matter most for the target domain’s behavior) to generate productive research programs, even when they ultimately require revision at the structural level. Generative Realism predicts, further, that systems with well-calibrated metaphor-compression operators (biological or artificial) will exhibit greater creative generativity precisely because they can operate productively across wider ranges of scale and abstraction. This prediction is empirically testable: systems with richer analogical reasoning capabilities should exhibit more robust transfer of learning across domains, exactly the capability that distinguishes flexible intelligence from domain-specific expertise.

6. The Mother-Ship / Fleet Architecture: Distributed Intelligence with Coherent Command

The preceding three operators: aperture, two-way transduction, and metaphor-compression, characterize the transformations a generative system performs on signals at a single processing level. But sophisticated cognition is not the work of a single, homogeneous processing system. It is achieved through the dynamic coordination of multiple specialized subsystems, each optimized for a particular domain or function, organized into a coherent whole that is more than the sum of its parts. The fourth operator addresses this organizational dimension: how are multiple generative subsystems structured so that their joint operation constitutes intelligence rather than cacophony?

The Mother-Ship/Fleet Architecture posits a hierarchical yet dynamic organization: a central coordinating system (the mother-ship) maintains global coherence, distributes tasks, and integrates outputs from specialized sub-systems (the fleet) while remaining open to upward revision by fleet outputs. Crucially, this is not a simple hierarchy in which the mother-ship commands and the fleet obeys. It is a bidirectional architecture in which the mother-ship’s global model is continuously updated by fleet reports, and fleet operations are continuously guided by mother-ship priors, in a dynamic that maintains coherence precisely by never fully delegating in either direction.

6.1 Formal Characterization

Define the mother-ship M as a global model that maintains a shared latent representation L_global over the system’s task domain. Fleet agents F_i (for i = 1, …, n) maintain local representations L_i specialized to sub-domains or task functions. The architecture is governed by two information flows. The downward flow distributes priors and task specifications from M to F_i: each fleet agent receives from the mother-ship a prior distribution P_M(L_i) that constrains its local processing. The upward flow aggregates evidence and partial solutions from F_i to update L_global: the mother-ship receives from each fleet agent an evidence signal E_i that is integrated to update P(L_global | E_1, …, E_n).

Define global coherence as the mutual information I(L_global; L_1, …, L_n), the degree to which the mother-ship’s global representation captures the structure present in the joint fleet representations. High coherence means the mother-ship accurately integrates fleet outputs into a global picture that reflects the fleet’s collective knowledge. Low coherence means the mother-ship’s global representation is systematically misaligned with what individual fleet agents have learned, producing a form of organizational ignorance: the global system fails to benefit from its own specialized components.

6.2 Biological Analogues

The mother-ship/fleet architecture maps closely onto the hierarchical organization of cortical processing as described by global workspace theory (GWT), developed by Baars and subsequently developed with neural specificity by Dehaene and colleagues.²⁶ On the GWT account, the brain contains many specialized, parallel processing systems: perceptual modules, motor control systems, memory systems, affective systems, linguistic systems, that operate largely in parallel and largely independently. Conscious, globally coordinated behavior emerges when a subset of this local processing is “broadcast” to a global workspace, a distributed cortical network centered on prefrontal and parietal regions, that makes information available to all the specialized systems simultaneously. The global workspace is the mother-ship; the specialized processing systems are the fleet.

Prefrontal cortical function, on this picture, is precisely the executive function of the mother-ship: maintaining and distributing global task representations, coordinating fleet operations, and integrating fleet outputs into coherent behavior. The prefrontal cortex does not perform most of the specialized computations of cognition directly; rather, it functions as the orchestrating agent that ensures those computations are appropriately sequenced, coordinated, and integrated. Dehaene’s experimental work on the neural correlates of conscious access provides strong evidence for the global broadcast mechanism that is the mother-ship’s primary upward-integration tool: stimuli that are consciously perceived show a characteristic late, widespread neural signal (“ignition”) that represents their entry into global workspace processing, while stimuli that remain unconscious show only local, specialized processing.²⁶

6.3 AI / Multi-Agent Systems

In artificial systems, the mother-ship/fleet architecture has direct implementation in mixture-of-experts (MoE) architectures, where a routing network (the mother-ship) dynamically activates subsets of specialized expert networks (the fleet) based on the current input, and multi-agent LLM systems, where an orchestrating agent distributes subtasks to specialized sub-agents and integrates their outputs.²⁷ Tool-augmented LLMs:  systems such as Schick and colleagues’ Toolformer, which learn to call external APIs and integrate their outputs, instantiate a particularly interesting form of fleet expansion: the model’s fleet is augmented with external computational resources that provide capabilities beyond those encoded in the model’s weights.²⁸

The characteristic failure mode of multi-agent systems in the absence of effective mother-ship integration is fleet fragmentation: individual sub-agents develop locally coherent representations and produce locally competent outputs, but the global system fails to integrate these into coherent whole-system behavior. Sub-agents may contradict each other, pursue incompatible sub-goals, or produce outputs that are individually plausible but jointly incoherent, precisely because no effective global coordination mechanism is enforcing the coherence that the mother-ship/fleet architecture is designed to provide. This failure mode is well-documented in early multi-agent AI systems and remains a significant challenge in contemporary multi-agent LLM deployments.

6.4 The Coherence–Autonomy Trade-off

A fundamental tension in mother-ship/fleet architectures is between fleet autonomy (necessary for specialization) and mother-ship coherence (necessary for unified agency). A fleet agent that is fully constrained by mother-ship priors loses the ability to discover domain-specific structure that the mother-ship’s global model cannot anticipate; a fleet agent that operates with complete autonomy loses the ability to benefit from global context and contributes to fleet fragmentation rather than global intelligence. The resolution of this tension is not a fixed allocation but a dynamic one.

Generative Realism proposes a dynamic allocation principle: fleet agents should operate autonomously within aperture-bounded task scopes and report upward to the mother-ship when their local confidence falls below a threshold. This threshold-triggered reporting connects the mother-ship/fleet operator back to the aperture operator: the aperture of the fleet agent’s local processing determines the boundaries of its autonomous competence, and the mother-ship’s global representation determines the prior with which the fleet agent’s local aperture is oriented. The system as a whole is thus a nested aperture structure, each fleet agent’s aperture is oriented by mother-ship priors, and the mother-ship’s global aperture is parameterized by the integration of fleet reports. This nested structure is precisely what allows the mother-ship/fleet architecture to scale: local specialization is not lost in global coordination, and global coherence is not purchased at the cost of local sensitivity.

7. Local Abstraction Layers: Contextual Granularity and the Prevention of Over-Generalization

The four operators presented so far: aperture, two-way transduction, metaphor-compression, and mother-ship/fleet architecture, provide the generative system with the machinery to sample signal, maintain reality-contact, compress relational structure, and coordinate specialized subsystems. But they leave unaddressed a persistent and practically significant failure mode: the tendency of generative systems to apply globally learned abstractions without sensitivity to local context, producing representations that are technically correct for some general case but systematically wrong for the case at hand. The fifth operator, Local Abstraction Layers, addresses this failure mode directly.

Local Abstraction Layers (LALs) are context-sensitive representational strata that sit between the global representations maintained by the mother-ship and the raw signals processed by individual fleet agents. They are the computational embodiment of the insight, familiar from Wittgenstein’s later philosophy, that meaning is always meaning-in-use: determined by the specific context of application rather than by a context-independent semantic rule.²⁹ A LAL implements this context-sensitivity computationally, providing a representational stratum that maps the same input signal onto different representations depending on the local context in which it is processed.

7.1 Formal Characterization

Define a Local Abstraction Layer as a family of abstraction functions {α_c} indexed by local context c ∈ C, where C is the space of relevant local contexts for the system’s operating domain. For each context c, α_c : S → R_c maps signal s to a context-specific representation r_c ∈ R_c. The crucial property of a LAL is that representations are not context-invariant: in general, α_c(s) ≠ α_c'(s) for c ≠ c’, even for the same input signal s. LALs are distinguished from global abstraction functions α_global (which produce context-invariant representations) by this context-sensitivity, they are, precisely, not one-size-fits-all.

The quality of a LAL is determined by the degree to which its context-indexed representations track the genuinely context-relevant variation in the signal. A well-differentiated LAL provides a rich family {α_c} with many distinct context indices and appropriately differentiated representations for each; a poorly differentiated LAL collapses many distinct contexts onto a small number of representational categories, producing over-generalization. The limit case of a maximally under-differentiated LAL is a global abstraction function: the same representation for all contexts, which is optimal only when context truly makes no difference, a condition that is rarely satisfied in real domains of any complexity.

7.2 The Over-Generalization Problem

Over-generalization, the application of globally dominant patterns in contexts where they are inappropriate, is one of the most pervasive and practically significant failure modes of generative systems, both biological and artificial. In language, the phenomenon is illustrated vividly by the polysemy of high-frequency words. The English word “bank” refers to financial institutions in some contexts and river embankments in others; “run” expresses directed locomotion, machine operation, sequential extension, organizational management, and dozens of other concepts depending on context; “light” may denote electromagnetic radiation, low mass, pale color, or easy effort depending on the sentence in which it appears. A system with only a global abstraction for each of these forms will systematically fail to select the appropriate sense in context, producing representations that are plausible relative to the statistical base rate but wrong relative to the local context.

In machine learning, over-generalization is the formal analog of this linguistic phenomenon: a model that has learned a globally dominant pattern will apply it in contexts where it fails to hold, because the model lacks the context-indexed abstraction functions that would allow it to distinguish those contexts from the majority case. This is the underlying mechanism of many forms of distributional shift failure: models trained on one distribution of contexts apply abstractions learned from that distribution to new contexts where they are inappropriate, not because the model lacks the relevant knowledge but because it lacks the LAL differentiation to deploy that knowledge context-selectively. The remedies proposed in the machine learning literature: fine-tuning, prompt engineering, in-context learning, mixture-of-experts routing, are all, from the Generative Realism perspective, mechanisms for improving LAL differentiation without modifying the global abstraction functions that constitute the model’s base capabilities.

7.3 LALs as Interface Between Local and Global

LALs play a dual role in the mother-ship/fleet architecture that connects them intimately to the two-way transduction operator. In the upward direction, LALs abstract fleet outputs into a format the mother-ship can integrate: the raw outputs of a specialized fleet agent are often expressed in a representational idiom too specific for direct integration into the global model’s L_global. The LAL performs a context-sensitive translation, preserving the information content of the fleet output while rendering it in a form that the mother-ship can process. This is the ascending LAL function, analogous to T↑ in two-way transduction but operating at the interface of fleet and mother-ship rather than at the interface of signal and representation.

In the downward direction, LALs interpret mother-ship priors in light of local context before delivering them to fleet agents: a global prior that is appropriate to the general case may need to be context-specifically adjusted before it can guide fleet processing in a particular local context. The LAL performs this adjustment, translating the mother-ship’s context-general guidance into context-specific instructions that fleet agents can apply without the distortion that would result from applying the global prior directly. This is the descending LAL function, analogous to T↓ in two-way transduction but operating at the mother-ship/fleet interface. The result is a system in which global coherence and local sensitivity are jointly maintained, the global model guides without overriding, and local context informs without overwhelming.

7.4 LALs and Expertise

One of the most productive implications of the LAL framework is its account of the structure of expert knowledge. Human expertise in a domain: chess, medicine, carpentry, jazz improvisation, consists not merely in the possession of more domain-relevant information than the novice, but in the capacity to perceive and act at a finer contextual grain: to discriminate situations that the novice treats as equivalent and to apply appropriately differentiated responses to those discriminated situations. On the LAL account, expertise is precisely the acquisition of richly differentiated LALs in a domain: the expert has a large family {α_c} with many distinct context indices, each mapping domain signals onto representations appropriate to that specific context.

The novice, by contrast, has a small, coarsely differentiated family of abstraction functions: many distinct domain situations are collapsed onto the same representational category, and the responses generated from that category are correspondingly undifferentiated. This account connects naturally to the skill acquisition literature in cognitive science, in particular to the “chunking” theory of Chase and Simon, which holds that expert chess players perceive board positions in terms of large, meaningful chunks rather than individual pieces, implementing a form of context-sensitive grouping that is precisely a LAL differentiation.³⁰ The implication for AI training is clear: models with richer context-indexed abstraction should exhibit more expert-like behavior in domain-specific tasks — an implication that is consistent with the observed benefits of domain-specific fine-tuning and the demonstrated superiority of large, richly contextualized models over smaller, more uniformly trained ones.

8. The Complete Stack: Composition, Feedback, and Emergent Meaning

The five operators presented in Sections 3 through 7: Aperture, Two-Way Transduction, Metaphor-Compression, Mother-Ship/Fleet Architecture, and Local Abstraction Layers, have been presented individually, with attention to their distinct functions, formal characterizations, and failure modes. This analytical presentation is necessary for precision, but it risks giving the impression that the operators are independent components of cognition that happen to be deployed in sequence. They are not. The central claim of Generative Realism is that meaning is an emergent property of the full compositional stack operating in bidirectional feedback, not a property of any individual operator, and not a property that can be assembled additively from the contributions of independent components. This section synthesizes the five operators into the complete Generative Realism stack and defends the emergence claim.

8.1 Compositional Structure

The five operators compose into a layered architecture in which each operator takes the output of the layer below as its primary input and transforms it before passing representations upward. At Layer 1, the Aperture Operator samples the signal space, producing a structured representation Σ’ of the incoming signal filtered, resolved, and oriented by the parameters θ and t. At Layer 2, the Two-Way Transduction Operator receives Σ’ as input to T↑, generates a representation r, and constrains that representation through the C relation by comparing T↓(r) with incoming T↑(Σ’) signals, yielding a constraint-coupled representation r* that is veridical to the degree that C(T↑(Σ’), T↓(r)) ≤ ε. At Layer 3, the Metaphor-Compression Operator receives r* and applies the mapping M, producing a compressed representation M(r*) that preserves the structural skeleton of r* while reducing its dimensionality to a tractable level. At Layer 4, the Mother-Ship/Fleet Architecture receives M(r*) and distributes it through the downward flow to fleet agents F_i, each of which generates a local representation L_i; the upward flow aggregates L_i into L_global. At Layer 5, Local Abstraction Layers α_c mediate both the upward and downward flows within the mother-ship/fleet architecture, translating between global and local representational idioms in context-sensitive ways.

Crucially, the information flow in the stack is not exclusively ascending. Higher layers continuously re-parameterize the operators at lower layers through descending feedback channels. The mother-ship’s global model re-orients the aperture parameters θ of fleet agents, adjusting what each agent samples and at what resolution based on global task context. Compressed metaphoric representations from Layer 3 constrain the transduction space within which Layer 2 operates, the conceptual vocabulary available to the system shapes what can be expressed in the bidirectional transduction loop. And the Local Abstraction Layers of Layer 5 re-parameterize the interface between Layer 4’s global representations and Layer 2’s transduction outputs, ensuring that the global-local mapping remains contextually appropriate. The result is not a simple feed-forward stack but a richly recurrent, feedback-coupled architecture in which every layer is continuously influenced by every other.

8.2 Emergent Meaning

The claim that meaning is an emergent property of the full compositional stack requires careful defense. “Emergence” is a term that is often invoked loosely to cover cases of explanatory difficulty, and Generative Realism must say something precise about what it means for meaning to be emergent in the relevant sense. The claim is not merely that meaning is complex or that it involves multiple components. It is the stronger claim that meaning is a system-level property that cannot be reduced to a property of any proper substack of the five operators, that taking any proper subset of the five operators produces a system that lacks genuine meaning-formation, however impressive its performance along some dimensions might be.

Consider systems lacking each operator in turn. A system without an aperture operator (one that processes the full signal space with uniform resolution and no prior-shaped orientation) cannot form representations at all in any interesting sense, because representation requires the discrimination of signal from noise, which requires an aperture. A system without two-way transduction (one whose generative operations are not constrained by incoming signals from the world) cannot achieve reality-contact; it may produce coherent outputs, but their coherence is internal to the generative system rather than tracking anything external. A system without metaphor-compression (one that cannot compress relational structure across scales) will fail to generalize beyond the specific training instances it has encountered and will be unable to reason about domains whose intrinsic dimensionality exceeds its processing resources. A system without mother-ship/fleet architecture (one that is either a single undifferentiated processor or an uncoordinated collection of specialists) will either lack the specialization necessary for domain expertise or the global coherence necessary for unified agency. A system without Local Abstraction Layers (one that applies globally learned abstractions uniformly across all contexts) will produce contextually inappropriate representations despite being globally competent.

The contrast with atomistic theories of meaning is instructive. Referential theories of meaning locate meaning in the relationship between symbols and world-states. Use theories locate meaning in the pattern of applications of a symbol across contexts. Correlation theories locate meaning in the statistical association between symbols and world-properties. Each of these locates meaning in a proper subset of the full operator stack: referential theories emphasize two-way transduction; use theories emphasize local abstraction; correlation theories emphasize the aperture and transduction layers. Generative Realism’s claim is that each of these partial accounts captures something genuine about meaning, it is not dismissing them, but that the full account requires the complete stack operating in compositional feedback.

8.3 Pathologies as Diagnostic Tools

One of the most practically valuable features of the operator stack account is that it provides a precise diagnostic vocabulary for the pathologies of generative systems. Each failure mode is associated with a specific layer, and the layer association carries implications for the appropriate remediation. Hallucination in LLMs (the confident generation of false or ungrounded claims) is a Layer 2 failure: a transduction decoupling event in which T↓ generates outputs not sufficiently constrained by T↑ signals from ground-truth sources. The appropriate remediation is architectural: retrieval-augmented generation, tool-use integration, or other mechanisms that restore bidirectional transduction coupling. Category errors in reasoning (the systematic misapplication of a conceptual framework to a domain for which it is structurally incongruent) are Layer 3 failures: metaphor-compression has achieved high ρ at the cost of structural fidelity. The appropriate remediation involves identifying the violated structure-preserving constraints and revising the metaphoric mapping accordingly. Incoherent behavior in multi-agent AI systems, where sub-agents produce individually competent but jointly contradictory outputs, is a Layer 4 failure: fleet fragmentation in the absence of effective mother-ship integration. Contextually insensitive behavior (the application of globally dominant patterns in contexts where they are inappropriate) is a Layer 5 failure: under-differentiated Local Abstraction Layers. And systematically missing relevant information (the failure to include task-relevant signals in the representation at all) is a Layer 1 failure: aperture miscalibration in width, depth, or orientation.

8.4 The Realism Anchor

The question with which this paper began, how generative systems achieve genuine contact with reality, can now be given a principled answer. Generative Realism holds that reality-contact is achieved not through any single privileged access channel but through the overall coherence of the compositional system, and in particular through two architectural features that constitute the system’s “realism anchor.” The first is the constraint loop of two-way transduction: the C relation that enforces mutual constraint between ascending and descending information flows, ensuring that the system’s representations are answerable to incoming signals from the world. The second is the global-local coherence maintained by the mother-ship/fleet architecture and mediated by Local Abstraction Layers: the requirement that local representational commitments be integrable into a globally coherent model, and that global representations be deployed with local sensitivity.

This is a pragmatic realism in the tradition of Peirce and Putnam: it holds that the norms of representation are genuinely answerable to a mind-independent world, while recognizing that what counts as “answerable to the world” is always specified relative to the architectural framework through which the system engages its environment.^13,14 What distinguishes Generative Realism from these predecessors is the architectural specificity of its account: it does not merely assert that cognition is answerable to the world; it specifies the operators through which that answerability is implemented and the failure modes that arise when those operators are miscalibrated or absent. This architectural specificity is both theoretically productive and practically useful, it makes Generative Realism not just a philosophical position but a research framework.

9. Implications for AI Alignment, Cognitive Science, and the Philosophy of Mind

9.1 AI Alignment and Safety

The operator stack provides a principled diagnostic framework for AI alignment failures, one that goes substantially beyond the current repertoire of alignment methodologies, which tend to focus on behavioral outputs (RLHF, constitutional AI, red-teaming) without specifying the architectural sources of misalignment. On the Generative Realism account, alignment failures arise from miscalibrations at specific layers of the operator stack, and each layer-specific miscalibration suggests a distinct category of remediation.

Aperture miscalibration (attending to the wrong signals, at the wrong resolution, with the wrong prior orientation) produces systems that are capable but systematically inattentive to the signals that would make them aligned. A system whose aperture is oriented to optimize for proxy metrics (benchmark performance, human approval ratings) rather than the genuine values it is supposed to track will systematically miss the signals that would indicate when those proxy metrics have become decoupled from the true objective. This is a structural account of the Goodhart’s Law problem in AI alignment: the problem arises precisely when the aperture is optimized for a proxy rather than for the genuine signal. Transduction failures (the absence of genuine bidirectional coupling between model outputs and world-states) produce systems that generate confident outputs without genuine grounding in the states those outputs purport to describe. Local Abstraction Layer failures produce systems that apply globally trained alignment norms without sensitivity to the specific context of application, producing outputs that are aligned in standard contexts but misaligned in unusual or novel ones, precisely the contexts in which alignment matters most.

9.2 Cognitive Science and Neuroscience

Generative Realism makes specific, testable predictions about the neural architecture of cognition. Most fundamentally, it predicts that each of the five operators should have identifiable neural correlates, dynamically coupled in the way the theory specifies. The aperture operator should correspond to the neural machinery of selective attention, including fronto-parietal attention networks and their top-down modulation of sensory processing, predictions that are consistent with the extensive neuroscientific literature on attention, but that Generative Realism specifies more precisely by tying aperture parameters to the specific dimensions of width, depth, and orientation. Two-way transduction should correspond to the bidirectional prediction-error signaling described in predictive processing accounts, with the T↑/T↓ dissociation corresponding to the distinction between feed-forward and feed-back cortical processing pathways.

The mother-ship/fleet prediction is perhaps the most precisely testable: the theory predicts that there should be a specific neural mechanism for global broadcast and integration of local processing outputs, a prediction that is consistent with global workspace theory and the neural ignition signature of conscious access, but that Generative Realism connects to the specific computational demands of the mother-ship role. Dehaene’s identification of prefrontal-parietal networks as the neural substrate of global workspace function provides initial neural localization for the mother-ship operator.²⁶ The Local Abstraction Layer prediction connects to the literature on context-dependent neural coding (the finding that the same stimulus activates different neural representations depending on contextual factors) and to the role of the hippocampus in context-dependent memory retrieval and analogical mapping.³¹

9.3 Philosophy of Mind

Generative Realism opens a productive line of engagement with the hard problem of consciousness (the problem of why and how physical processes give rise to phenomenal experience) without claiming to resolve it. The theory’s account of two-way transduction provides a framework within which to articulate a specific, architecturally grounded version of the phenomenological insight that consciousness is constituted by genuine world-contact. If, as the theory proposes, the “felt grip” on reality that characterizes veridical perceptual experience is the phenomenological correlate of the C constraint relation in bidirectional transduction, then phenomenal experience may be constituted by the full-stack operation of a generative system in genuine bidirectional transductive contact with its environment.

This is not a complete theory of consciousness; it does not resolve the explanatory gap between functional organization and phenomenal quality that Chalmers identified as the hard problem.³² But it provides a more architecturally specific target for the functionalist research program than most existing accounts: rather than asking whether any functional organization gives rise to consciousness, it asks whether the specific organizational properties specified by the operator stack: bidirectional transduction constraint, global-local coherence maintenance, context-sensitive local abstraction, are sufficient, necessary, or merely correlated with phenomenal experience. This specificity makes the question more tractable, connecting it to existing empirical methodologies in consciousness research while grounding it in a principled theoretical framework.

9.4 Practical Design Principles

The operator stack framework yields a set of concrete design principles for generative AI systems that follow directly from the theoretical analysis. Each principle addresses a specific operator layer and specifies what well-calibrated implementation of that layer requires. First, calibrate aperture to task resolution: design systems whose context window, attention mechanisms, and sampling priors are matched to the resolution requirements of the target task, avoiding both myopic under-inclusion and noisy over-inclusion of signal. Second, enforce bidirectional transduction through grounding mechanisms: ensure that the generative operations of the system are constrained by genuine feedback from world-states, through retrieval augmentation, tool-use, external verification, or embodied deployment, not merely by statistical priors from training data. Third, build structured metaphor libraries with fidelity constraints: explicitly encode the key cross-domain mappings the system will need for its task domain, with explicit structural fidelity checks that prevent the application of high-ρ but low-fidelity mappings in contexts where structural distortion would be consequential. Fourth, implement coherent multi-agent orchestration: ensure that multi-agent systems have explicit mother-ship integration mechanisms, not merely task distribution mechanisms, so that fleet fragmentation is prevented and global coherence is actively maintained. Fifth, train context-indexed abstraction layers for domain expertise: invest in fine-tuning and domain-specific training that develops richly differentiated Local Abstraction Layers, enabling the system to apply globally learned capabilities with the contextual sensitivity of a domain expert rather than the uniform application of a novice.

10. Conclusion: Toward a Science of Generative Meaning

This paper has introduced Generative Realism, a unified theoretical framework for understanding how generative systems, biological and artificial, achieve genuine contact with reality rather than merely simulating it. The framework formalizes five architectural operators: Aperture, Two-Way Transduction, Metaphor-Compression, Mother-Ship/Fleet Architecture, and Local Abstraction Layers, each performing a distinct, necessary transformation in the generative process. The central thesis has been defended: meaning is an emergent property of the full compositional stack operating in bidirectional feedback with the environment, not a property of any individual layer or any proper subset of operators.

The originality of the contribution lies in three places. First, the operator-level formalization: existing theories of cognition and meaning provide partial accounts, but none specifies the complete composable operator architecture that Generative Realism articulates. Predictive processing provides dynamics; enactivism provides the organism-environment coupling principle; conceptual metaphor theory provides the compression insight; global workspace theory provides the global-local integration model; Wittgensteinian philosophy of language provides the use-in-context principle. Generative Realism integrates all of these into a single, compositional framework in which each insight is formalized as an operator with precise input-output characteristics and failure conditions. Second, the diagnostic power: by associating each failure mode with a specific operator layer, the framework provides a principled vocabulary for analyzing and addressing breakdowns in generative systems, both biological pathologies and AI alignment failures. Third, the unifying scope: the same operator stack applies to biological cognition, artificial language models, and distributed multi-agent systems, providing a common architectural language across research communities that currently operate largely in isolation from each other.

The most promising open questions that Generative Realism identifies can be organized by discipline. In cognitive neuroscience: what are the precise neural correlates of each operator, how are they dynamically coupled in the way the theory predicts, and what neural pathologies correspond to operator-specific failures? In AI research: what training objectives, architectures, and evaluation methodologies most effectively develop each operator, and how can systems be audited for operator-level calibration failures? In philosophy of mind: is the full-stack operation of the generative architecture under bidirectional transduction sufficient for phenomenal consciousness, or merely functionally correlated with it? And most fundamentally: is the operator stack as specified here complete, does it identify all the necessary architectural operations for meaning-formation, or are there additional operators that remain to be specified?

These questions are not merely academic. As generative AI systems become more deeply integrated into the infrastructure of knowledge, decision-making, and communication, the question of whether those systems achieve genuine meaning-formation or merely sophisticated simulation becomes a question of the first practical importance. Generative Realism provides not just a theoretical framework for addressing this question, but a research program: for cognitive scientists, AI researchers, and philosophers of mind, directed at understanding how generative systems achieve, maintain, and sometimes lose genuine contact with reality. The architecture of emergent meaning is not a philosophical abstraction; it is the blueprint of minds that matter.

References

Baars, B. J. (1988). A cognitive theory of consciousness. Cambridge University Press.

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’21) (pp. 610–623). Association for Computing Machinery. https://doi.org/10.1145/3442188.3445922

Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., … Amodei, D. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.

Chalmers, D. J. (1996). The conscious mind: In search of a fundamental theory. Oxford University Press.

Chase, W. G., & Simon, H. A. (1973). Perception in chess. Cognitive Psychology, 4(1), 55–81. https://doi.org/10.1016/0010-0285(73)90004-2

Clark, A. (2016). Surfing uncertainty: Prediction, action, and the embodied mind. Oxford University Press.

Dehaene, S. (2014). Consciousness and the brain: Deciphering how the brain codes our thoughts. Viking.

Dennett, D. C. (1991). Consciousness explained. Little, Brown and Company.

Fauconnier, G., & Turner, M. (2002). The way we think: Conceptual blending and the mind’s hidden complexities. Basic Books.

Friston, K. J. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Friston, K. J., FitzGerald, T., Rigoli, F., Schwartenbeck, P., & Pezzulo, G. (2017). Active inference: A process theory. Neural Computation, 29(1), 1–49. https://doi.org/10.1162/NECO_a_00912

Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155–170. https://doi.org/10.1207/s15516709cog0702_3

Gibson, J. J. (1979). The ecological approach to visual perception. Houghton Mifflin.

Grice, H. P. (1975). Logic and conversation. In P. Cole & J. Morgan (Eds.), Syntax and semantics: Vol. 3. Speech acts (pp. 41–58). Academic Press.

Harris, Z. S. (1954). Distributional structure. Word, 10(2–3), 146–162. https://doi.org/10.1080/00437956.1954.11659520

Hofstadter, D. R., & Sander, E. (2013). Surfaces and essences: Analogy as the fuel and fire of thinking. Basic Books.

Husserl, E. (1983). Ideas pertaining to a pure phenomenology and to a phenomenological philosophy: First book (F. Kersten, Trans.). Martinus Nijhoff. (Original work published 1913)

Lakoff, G., & Johnson, M. (1980). Metaphors we live by. University of Chicago Press.

Maturana, H. R., & Varela, F. J. (1980). Autopoiesis and cognition: The realization of the living. D. Reidel Publishing.

Maxwell, J. C. (1865). A dynamical theory of the electromagnetic field. Philosophical Transactions of the Royal Society of London, 155, 459–512. https://doi.org/10.1098/rstl.1865.0008

Merleau-Ponty, M. (1945/2012). Phenomenology of perception (D. A. Landes, Trans.). Routledge. (Original work published 1945)

Parr, T., Pezzulo, G., & Friston, K. J. (2022). Active inference: The free energy principle in mind, brain, and behavior. MIT Press.

Peirce, C. S. (1931–1958). Collected papers of Charles Sanders Peirce (Vols. 1–8, C. Hartshorne, P. Weiss, & A. Burks, Eds.). Harvard University Press.

Posner, M. I. (1980). Orienting of attention. Quarterly Journal of Experimental Psychology, 32(1), 3–25. https://doi.org/10.1080/00335558008248231

Putnam, H. (1981). Reason, truth, and history. Cambridge University Press.

Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87. https://doi.org/10.1038/4580

Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., & Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36.

Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417–424. https://doi.org/10.1017/S0140525X00005756

Squire, L. R. (1992). Memory and the hippocampus: A synthesis from findings with rats, monkeys, and humans. Psychological Review, 99(2), 195–231. https://doi.org/10.1037/0033-295X.99.2.195

Thompson, E. (2007). Mind in life: Biology, phenomenology, and the sciences of mind. Harvard University Press.

Varela, F. J., Thompson, E., & Rosch, E. (1991). The embodied mind: Cognitive science and human experience. MIT Press.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35.

Wittgenstein, L. (1953). Philosophical investigations (G. E. M. Anscombe, Trans.). Blackwell.

Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338–353. https://doi.org/10.1016/S0019-9958(65)90241-X

¹ Friston, K. J. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.

² Clark, A. (2016). Surfing uncertainty: Prediction, action, and the embodied mind. Oxford University Press.

³ Maturana, H. R., & Varela, F. J. (1980). Autopoiesis and cognition. D. Reidel Publishing.

⁴ Varela, F. J., Thompson, E., & Rosch, E. (1991). The embodied mind. MIT Press.

⁵ Brown, T. B., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.

⁶ Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417–424.

⁷ Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots. FAccT ’21.

⁸ Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex. Nature Neuroscience, 2(1), 79–87.

⁹ Parr, T., Pezzulo, G., & Friston, K. J. (2022). Active inference: The free energy principle in mind, brain, and behavior. MIT Press.

¹⁰ Friston, K. J., FitzGerald, T., Rigoli, F., Schwartenbeck, P., & Pezzulo, G. (2017). Active inference: A process theory. Neural Computation, 29(1), 1–49.

¹¹ Thompson, E. (2007). Mind in life. Harvard University Press.

¹² Harris, Z. S. (1954). Distributional structure. Word, 10(2–3), 146–162.

¹³ Peirce, C. S. (1931–1958). Collected papers (Vols. 1–8). Harvard University Press.

¹⁴ Putnam, H. (1981). Reason, truth, and history. Cambridge University Press.

¹⁵ Posner, M. I. (1980). Orienting of attention. Quarterly Journal of Experimental Psychology, 32(1), 3–25.

¹⁶ Vaswani, A., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.

¹⁷ Husserl, E. (1983). Ideas pertaining to a pure phenomenology. Martinus Nijhoff. (Original work 1913)

¹⁸ Gibson, J. J. (1979). The ecological approach to visual perception. Houghton Mifflin.

¹⁹ Merleau-Ponty, M. (1945/2012). Phenomenology of perception. Routledge.

²⁰ Lakoff, G., & Johnson, M. (1980). Metaphors we live by. University of Chicago Press.

²¹ Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2), 155–170.

²² Fauconnier, G., & Turner, M. (2002). The way we think. Basic Books.

²³ Wei, J., et al. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35.

²⁴ Hofstadter, D. R., & Sander, E. (2013). Surfaces and essences. Basic Books.

²⁵ Maxwell, J. C. (1865). A dynamical theory of the electromagnetic field. Philosophical Transactions of the Royal Society of London, 155, 459–512.

²⁶ Dehaene, S. (2014). Consciousness and the brain. Viking.

²⁷ Wei, J., et al. (2022). Chain-of-thought prompting. Advances in Neural Information Processing Systems, 35.

²⁸ Schick, T., et al. (2023). Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36.

²⁹ Wittgenstein, L. (1953). Philosophical investigations. Blackwell.

³⁰ Chase, W. G., & Simon, H. A. (1973). Perception in chess. Cognitive Psychology, 4(1), 55–81.

³¹ Squire, L. R. (1992). Memory and the hippocampus. Psychological Review, 99(2), 195–231.

³² Chalmers, D. J. (1996). The conscious mind. Oxford University Press.