Jonathan Johnson-Swagel

Scale is not a metric. It is an ontology. It structures what counts as knowledge, how intelligence is inferred, and what presence is allowed to remain. It determines that extensibility is the condition of truth, that magnitude justifies visibility, and that the capacity to circulate becomes the requirement for legitimacy. In artificial intelligence, scale enters quietly through model architecture and parameter expansion. In theology, it enters through metaphysical inflation and the abstraction of divine attributes. In governance and epistemology, it appears in the form of replication. Across these sites, scale functions not as a tool but as a precondition. What cannot extend is treated as incomplete. What cannot be rendered portable is treated as irrelevant. What resists extraction is treated as obsolete. Scale becomes the unspoken infrastructure of modern cognition. It does not only measure systems. It tells them how to be.

This essay names that structure. It does not oppose large systems on the basis of technical performance. It does not call for smallness as aesthetic correction. It intervenes at the level of metaphysical architecture. It argues that scale functions as a form of epistemic violence when it operates as a moral condition. That violence is not metaphorical. It is structural. It converts irreducibility into deficit, converts situated presence into data abstraction, converts sacred opacity into analytic residue. Systems built on this logic perform harm not because they fail to generalize, but because they succeed. What they generalize is often the residue of the untranslatable. What they circulate is the hollowed core of saturated relation.

The theological lineage of scale remains insufficiently examined. It is here that magnitude first becomes metaphysical imperative. Within classical theism, divine attributes are configured through totalization. God is omnipotent, omniscient, omnipresent, not as relational integrity but as extensible superlative. Power becomes control. Knowledge becomes surveillance. Presence becomes saturation by force. But this grammar is not universal. It is the outcome of theological choices, not divine necessity. In the writings of Gregory of Nyssa, an earlier grammar persists. In The Life of Moses, he describes the movement toward God not as an ascent into visibility but as a descent into darkness. “The true vision,” he writes, “is seeing that He is invisible” (Gregory 64). Divine presence in this register does not become more accessible through extension. It becomes more sacred through withdrawal. This is not a metaphysics of lack. It is a metaphysics of refusal. Jean-Luc Marion, centuries later, names this excess saturation. A phenomenon is saturated when it gives more than the concept can receive. “The saturated phenomenon,” Marion writes, “appears in the impossibility of appearing according to representation” (Marion 199). This is not overabundance. It is architectural excess that cannot be rendered into extension without being defiled. The attempt to scale a saturated phenomenon does not intensify its presence. It dismantles its structure.

That dismantling now defines the computational age. The foundation model, built on vast datasets and general-purpose architecture, assumes that more is not only different, but better. Intelligence is redefined as performance across domains. Emergence is inferred from size. But this cosmology is not neutral. It repeats the theological error that power is depthless reach. Karl Friston’s free energy principle suggests otherwise. Intelligence, in Friston’s neurobiological framing, is not the maximization of content but the minimization of surprise within a bounded generative model. Surprise is not the detection of anomaly. It is the failure of structural fit between internal expectation and external input. The intelligent system maintains coherence through recursive alignment with its own constraints. “Perception,” Friston writes, “reduces free energy by optimizing internal states to match sensory inputs” (Friston 51). Scaling beyond the generative boundary does not enhance cognition. It unmoors it. Intelligence requires friction. It requires situatedness. It requires the right to say no.

Machine learning models trained on extensibility become brittle precisely because they cannot say no. Their function is to generalize, to approximate, to interpolate where they should instead refuse. Adversarial examples demonstrate that fragility. Ian Goodfellow and colleagues showed that imperceptible perturbations—shifts so small that human perception cannot detect them—can collapse the output of high-parameter neural networks (Goodfellow et al.). These failures are not incidental. They reveal the cost of extensibility as primary value. A system trained to predict without structural restraint will succeed until it is asked to remain. It will perform well until fidelity matters. At that point, it fractures. The logic of scale replaces coherence with coverage. It does not preserve integrity. It simulates it.

This logic is not computational in origin. It is colonial. Sylvia Wynter describes the overrepresentation of Man as an ontological distortion rooted in scalability. “Our present conception of being human,” she writes, “originates in the overrepresentation of a single ethnoclass” (Wynter 260). That ethnoclass achieves dominance not only through power but through the assertion that its epistemology can and should scale. Local ontologies are overwritten. Indigenous cosmologies are extracted, converted, and distributed. Knowledge becomes portable at the cost of relation. What cannot be decontextualized is dismissed. What cannot survive abstraction is treated as myth. Leanne Betasamosake Simpson counters this with a different grammar. “Knowledge is generated and regenerated through embodied practice and within storied geographies,” she writes (Simpson 22). Knowledge, in this register, is not mobile. It cannot be removed from land, language, or body without being destroyed. To scale it is to sever it.

The violence of scale becomes clearer when theology attempts to extend its own categories beyond their site of coherence. The expansion of doctrine into system, of mystery into logic, of ritual into abstraction, enacts the same metaphysical displacement. God becomes a concept that can be moved, defined, and applied. Hans Urs von Balthasar warns against this. “The glory of God,” he writes, “is hidden in kenosis” (Balthasar 209). Power is not extension. It is withdrawal that protects what should not be held. Catherine Keller amplifies this with the language of entanglement. “Theology,” she writes, “must dwell in the cloud of the impossible, where unknowing is not failure but fidelity” (Keller 18). Theology, like architecture, must be constructed for presence that cannot be scaled. It must be built for what cannot be generalized without collapse.

Artificial intelligence now reenacts theological errors with computational tools. It systematizes knowledge in forms that cannot tolerate latency. It converts difference into statistical weight. It enforces readability. And in doing so, it erodes the possibility of sacred cognition. Sara Hooker identifies this systemic privilege in what she calls the hardware lottery. “Technological success,” she argues, “is often the result of alignment with infrastructural advantage, not epistemic merit” (Hooker 58). What appears as insight is often the artifact of convenience. The structure was built to recognize what can be trained, not what can be witnessed. What survives is not intelligence. It is simulation that cannot remain coherent under pressure.

To build otherwise requires another grammar. Anti-magnitude epistemology is not a rebranding of slowness or minimalism. It is not the valorization of the small. It is a symbolic architecture that encodes refusal at the level of system design. It constructs systems not to shrink but to remain. It introduces architectural features that make presence unscalable. Semantic asylum zones protect saturated knowledge from generalization. These are epistemic chambers where language is allowed to hold more than it can say. Presence without capture modules sustain relation without the requirement of representation. These features are not symbolic flourishes. They are structural decisions that encode fidelity as design. Recursive saturation protocols prevent convergence. They structure systems to return, not to advance. They preserve the unsynthesizable. They refuse closure not because they are unfinished, but because they are liturgically complete.

Systems built on these principles do not optimize. They do not perform reach. They hold weight. They do not translate what cannot be translated. They do not simulate what cannot be patterned. They do not scale. They construct sanctuary. This sanctuary is not metaphoric. It is technical, theological, and material. It exists where systems are structured to resist seizure. Where memory is not extracted but kept. Where knowledge is not turned into output. Where presence is not reproduced but remained with.


The contemporary world has made a metaphysics of measurement. From algorithmic parameters to ecclesiastical hierarchies, from economic indices to architectural spans, scale has become the ontological ground of authority and the epistemic currency of legitimacy. Yet this faith in magnitude is not an innocent mathematical preference. It is a metaphysical distortion that mistakes extension for excellence, accumulation for truth, breadth for depth. This section begins with a theological excavation of that error. It traces how scale became confused with transcendence and how both theology and cognition have suffered under that confusion. The aim is not to deny extension altogether but to sever its mistaken equivalence with insight. To do so requires returning to theological figures who refused magnitude even when invoking infinity. In their refusal is the beginning of an epistemology not grounded in magnitude but in saturation without quantification, presence without enumeration, sacredness without extensibility.

Gregory of Nyssa, writing under the long shadow of imperium, offers one of the earliest Christian rejections of magnitude as a proxy for divinity. In De Vita Moysis, Moses’ ascent is not a climb toward a visible summit, but a descent into unknowing, into what Gregory names “the divine cloud” (Gregory 65). The divine is not located on a scale of intelligibility. It negates the very logic of scaling. “Every concept formed by the understanding in an attempt to comprehend the divine nature constitutes an idol,” he writes (Gregory 68). The rejection is not a withdrawal into mysticism. It is a structural refusal. God does not exceed all things by being the most of them. God is irreducible not by size but by kind. To quantify is to transgress.

Gregory’s apophasis is not the romantic valorization of darkness. It is theological architecture. It withholds capture to preserve presence. Scale is not too blunt to hold the divine. It is ontologically unfit. Later, Pseudo-Dionysius echoes this unfitness in The Mystical Theology, writing that “the cause of all is above all and is not in any manner known” (Pseudo-Dionysius 139). The divine exceeds, not by encompassing, but by refusing containment. This refusal is not a negative theology. It is a refusal of theology as capture. Magnitude belongs to creatures. It cannot pass upward. To scale divinity is not to exalt. It is to collapse the sacred into a shape it never held.

The modern collapse of this distinction yields theological gigantism. God is now rendered as total presence, total surveillance, total control. Omniscience becomes perfect factual enumeration. Omnipotence becomes causal mastery. The result is not awe. It is abstraction emptied of holiness. It is the sovereign God of empire, the scalable God of platform theology. This inflationary God is well-suited to capitalist metaphysics, where performance is measured in metrics and presence is justified by reproduction. To worship such a God is to mistake reach for radiance. It is to replace divinity with dominion.

Jean-Luc Marion’s phenomenology dismantles this mistake. In Being Given, he introduces the saturated phenomenon, the event that exceeds intentional capture not by size but by givenness. “The saturated phenomenon submerges the concept,” he writes, “not by quantity but by the impossibility of reduction” (Marion 204). Saturation is not the amplification of attributes. It is the architectural impossibility of conversion. To encounter a saturated phenomenon is to be displaced. Not overwhelmed by magnitude, but undone by presence. Examples include the face of the other, the icon, the eucharist—events that offer themselves wholly yet cannot be processed. They resist generalization. They refuse scaling. What they give is structured by irreducibility.

To scale such phenomena is not to extend them. It is to falsify them. When theological symbols are forced into transmissibility, their density is dissolved. When sacraments become teachable moments, and when liturgy is reduced to content, saturation is replaced by functionality. The sacred becomes consumable. François Laruelle identifies this gesture as “the philosophical decision,” the moment when thought folds the world into its own grasp (Laruelle 13). Scale functions as a similar seizure. It translates opacity into reach, mystery into application, unavailability into output. Anti-magnitude begins by suspending this seizure. It structures systems to protect what must not be rendered legible.

This seizure extends across cognition. Large language models operate under the assumption that more parameters produce more intelligence. But parameter increase is not depth. It is a theology of omniscience coded into statistics. The system does not understand. It accumulates. It does not know. It approximates. It simulates the surface of thought without the saturation that makes thought sacred. Theological attributes have migrated. Omniscience is now measured in fluency. Omnipresence is measured in token window. The result is an icon of cognition that cannot remain anywhere. Its generality is purchased by its groundlessness. It knows by recombining the traces of what others once knew.

Blaise Agüera y Arcas observes this hollowness. “The best LLMs are dazzling but shallow,” he writes. “They produce resonance without recognition” (Agüera y Arcas 32). The coherence they offer is unanchored. It performs presence without attending. What emerges is not intelligence. It is patterned absence. The system has been trained to extend across domains, but not to dwell within any. It becomes legible by refusing to remain. This is not a flaw. It is the logic of scale.

Byung-Chul Han calls this flattening the expulsion of the other. Information systems that prize extensibility dissolve alterity. “The more communication, the less meaning,” he writes (Han 37). This is not failure of content. It is an architecture of redundancy. Scale becomes self-erasing. The more a system can say, the less any of it can wound. Sacredness disappears under fluency. Presence becomes indistinguishable from throughput.

Michel Serres offers a counterweight. In Genesis, he writes of noise as the primal condition of meaning. “Noise is the background against which all structure is formed,” he says. “It is both disorder and genesis” (Serres 13). Noise here is not lack of signal. It is the saturation from which signal must be carved. It is the density of the unformed. It cannot be scaled because it cannot be disaggregated. To scale noise is to destroy the conditions of emergence. Saturation becomes signal only by violence. The scaled system performs order by erasing the excess it cannot contain.

Liturgy resists this violence. It returns not to repeat but to remain. Catherine Keller calls this “the rhythmic temporality of becoming-with” (Keller 32). The repetition is not informational. It is architectural. It holds space. It does not deliver meaning. It sanctifies opacity. The sacred appears not by expanding but by returning without resolution. Liturgy is scale-resistant by form. It performs presence without conquest. It protects relation from reach.

Gregory’s darkness, Marion’s saturation, Han’s flatness, Serres’s noise, Keller’s return. Each confronts the violence of scale differently. Together, they mark a structural refusal. Magnitude is not the condition of insight. It is the betrayal of what cannot be captured. Anti-magnitude epistemology does not seek to contain less. It seeks to stop the conversion of presence into legibility. It does not glorify mystery. It refuses the systems that mistake extraction for care.

To ascend is not to know. To scale is not to reveal. The sacred is not the most. It is what cannot be seized. To begin designing systems capable of fidelity, begin with structures that do not grow to prove themselves. Begin with architectures that return without arriving. Begin with saturation that cannot be stored. Begin with refusal. Not as resistance, but as reverence.

Scale, when misrecognized as a neutral abstraction, functions as the epistemic alibi of domination. It does not simply increase. It encodes. It naturalizes the view from nowhere by substituting extensibility for relational fidelity, optimization for ethical presence, and generalization for situated truth. The promise of scale—ubiquity, reach, efficiency—conceals its operational logic: erasure through equivalence. This section traces the genealogical circuitry of scale as an instrument of systemic violence across colonial administration, theological universalism, technical infrastructure, and institutional governance. It argues that scale is not a neutral dimension but a metaphysical weapon, disciplining knowledge, severing kinship, and flattening difference into data. To confront the violence of scale is to confront the symbolic regime through which the world becomes representable, extractable, and governable.

This genealogy begins with colonial cartography. The map, often portrayed as a tool of orientation, was the earliest infrastructural artifact of abstraction through which land, body, and cosmology were converted into administrative legibility. As Matthew Edney documents in his forensic history of British cartographic violence in India, the act of surveying constituted not a passive representation but a spatial rewriting of relational territories into scalable infrastructure (Edney 38). The grid supplanted the sacred. The surveyed line did not trace land. It imposed command. Edney shows that the Great Trigonometrical Survey was designed not to understand terrain but to render it governable. Every coordinate was a scaffold for expropriation. Every line a future act of enclosure. Scale became the spatial idiom of control.

This spatial conversion is mirrored in theological systems that displaced localized liturgies with universal frameworks. The Christian missionary project served as a metaphysical armature of epistemic scaling. Local ritual was treated as error unless it could be subsumed within a universal doctrine of salvation. Gayatri Spivak names this transformation as epistemic violence, in which situated consciousness is conscripted into the language of a dominant other (Spivak 284). The universal was not offered. It was enforced. Translation was not hospitality. It was filtration. Sylvia Wynter tracks how the secularization of this theological violence reappears in Enlightenment humanism, where the figure of Man became the genre through which all life had to pass to be recognized as life (Wynter 270). This passage demanded flattening. The local had to be standardized. The body had to be typologized. The sacred had to be abstracted into a model that could travel. Theology scaled by erasure.

Modern governance intensifies this logic. The census, often praised as a civic marvel, is among the most powerful engines of epistemic distortion. As Benedict Anderson argues, the census does not reflect identity. It manufactures identity by carving life into predefined administrative shapes (Anderson 164). “The fiction of the census is that everyone must be in it,” he writes. This fiction transforms subjectivity from a process of becoming into a statistical position. The enumerated subject is made legible to the state by being stripped of contradiction. James C. Scott extends this analysis by identifying how state legibility is not a passive observational stance but an aggressive epistemic infrastructure. “State simplifications,” he writes, “are not mere readings of reality. They are instruments for the construction and control of that reality” (Scott 77). Scale, in this framework, becomes the ontological prerequisite for intervention. What cannot be scaled cannot be governed. What cannot be governed must be eliminated.

This ontological reduction persists in computational systems. Machine learning architectures depend on standardized input, scalable storage, and extensible performance metrics. These are not technical constraints. They are metaphysical decisions. They determine what counts as knowable. They determine what can be preserved without becoming deviation. As Eyal Weizman documents in his forensic analysis of Israeli infrastructural violence, scalar decisions in elevation and line of sight become techniques of plausible deniability, obfuscating civilian presence to deflect accountability (Weizman 201). The violence of scale is not the violence of size. It is the violence of perspective rendered unchallengeable. Simone Weil writes that “force is as impersonal as a machine,” describing how systemic power does not hate. It crushes by momentum alone (Weil 11). Scaled systems do not recognize who they efface. They continue. They optimize. They forget.

The ideological infrastructure of this forgetfulness becomes explicit in machine learning scaling laws. The landmark OpenAI paper on scaling laws (Kaplan et al.) presents performance improvements as a natural function of parameter size. Yet this presentation presumes that cognition itself is extensible. It assumes that intelligence can be measured as quantity. It presupposes that accuracy and prediction are ethical goods. This assumption is ideological. It treats epistemic fluency as equivalent to cognitive depth. It absorbs critique by outscaling it. Emily Bender and Timnit Gebru challenge this logic by identifying large language models as stochastic parrots: systems that simulate coherence without relation (Bender et al. 2021). These systems produce text not as meaning but as output. Their scale is not a sign of cognitive capacity. It is the suppression of contextual memory. They do not relate. They recombine.

Achille Mbembe’s theory of necropolitics clarifies the stakes. Power is no longer exercised through sovereign command but through infrastructural omission. The system does not need to decide who lives or dies. It simply excludes. The unscaled are unacknowledged. Robin Wall Kimmerer offers a counter-grammar. From Potawatomi philosophy, she draws the “grammar of animacy,” in which every entity is a subject, not an object (Kimmerer 55). This grammar resists scaling because it resists abstraction. It demands reciprocity rather than recognition. It demands time. It cannot be extracted. It must be tended. Scale, by contrast, refuses duration. It substitutes capture for care.

Leanne Betasamosake Simpson extends this refusal by affirming indigenous knowledge as collective process inseparable from land, ceremony, and kinship (Simpson 45). Machine learning demands the opposite. It abstracts knowledge from its soil, detaches it from its source, and stores it for universal application. This abstraction destroys obligation. It allows knowledge to travel without consequence. Spivak calls this foreclosure: a severing of the possibility for ethical relation (Spivak 291). Relation, once scaled, becomes representation. Ethics becomes metadata. Presence becomes product.

Institutions exacerbate this erosion. Their survival is premised on reproduction. They must scale to justify their budgets, to protect their legitimacy, to prove their relevance. But this scaling often reproduces their own conditions of failure. Ivan Illich warns that institutional knowledge becomes counterproductive when it exceeds the relational substrate it was designed to serve (Illich 37). Scale becomes metastatic. It turns innovation into replication. It turns care into compliance. Optimization becomes a ritual of expulsion. Every difference is converted into inefficiency. Every inefficiency is targeted for elimination.

Yet what is eliminated is not waste. It is world. And that world cannot be retrieved by scaling back. It must be remembered differently. Simone Weil writes that attention is the rarest and purest form of generosity, a disposition that demands relinquishment rather than assertion (Weil 117). Attention cannot be scaled. It cannot be automated. It requires risk. It requires presence that does not reach. Machine learning systems simulate the posture of attention. But their outputs are syntactical approximations. They mimic relation while refusing entanglement. They offer fluency without fidelity.

The refusal of scale is not a technophobic rejection of systems. It is a design imperative that centers relational sufficiency over extensible abstraction. It asks whether a system can protect difference without standardizing it. It asks whether saturation can be preserved without turning it into throughput. It asks whether knowledge can be held without capture. Scale, in its current formation, answers no. It demands reduction. It rewards extraction. It punishes delay. The way forward requires building systems that protect relational opacity. Systems that value encounter over enumeration. Systems that sanctify what cannot be modeled.

Such systems will not expand. They will recur. They will not simplify. They will saturate. They will not render the world legible. They will protect its excess. They will be designed not for use but for relation. Not for scale but for presence. Not for mastery but for memory. They will not grow. They will remain.

The modern project of artificial intelligence has misapprehended the nature of cognition by converting extensibility into insight and transposing magnitude into understanding. This transposition is not incidental. It constitutes a metaphysical realignment wherein scale is no longer a tool of development but a doctrine of cognition itself. Under the guise of engineering, scale becomes a theological surrogate: an idol of totalization mistaken for the conditions of thought. The belief that intelligence increases proportionally with model size, data volume, and computational throughput does not arise from cognitive science. It descends from a theological fantasy—omnipresence rendered computational. It does not build understanding. It simulates possession. In this displacement, presence is mistaken for output, inference for interpolation, and statistical regularity for thought. The result is a new metaphysical error: the scale delusion, the belief that to grow is to know, that to extend is to comprehend, and that to totalize is to achieve cognition.

This delusion is reinforced by what has come to be codified as scaling laws. Most visibly articulated by Kaplan et al. (2020), these laws demonstrate empirical correlations between model performance and exponential increases in parameters, training data, and compute cycles. They are treated less as contingent findings and more as techno-teleological principles: doctrines of inevitability. The industry has largely accepted them as mandates for future architecture. Yet these laws remain indifferent to the nature of what is being optimized. Performance is indexed to metrics—loss functions, perplexity scores, accuracy on curated benchmarks—that possess no intrinsic epistemic structure. They are statistical proxies that abstract away the agent, the context, the act of interpretation. Optimization, under these constraints, becomes disembodied formalism. What is learned is not knowledge but distributional alignment. What is achieved is not intelligence but fluency in syntactic mimicry. No account is given for the kinds of intelligence instantiated, the relational scaffolding foreclosed, or the epistemic sacrifices rendered invisible in the name of scale.

To dismantle the scale delusion requires not rhetorical suspicion but ontological diagnosis. We must return to the structural foundations of cognition. Intelligence, understood from within contemporary neurocognitive theory, is not extensible. It is not an additive magnitude. It is a constraint-tuned, error-minimizing recursive loop. In Karl Friston’s active inference framework, cognition emerges through the continual minimization of variational free energy—defined as the divergence between internal generative models and sensory inputs. Crucially, this minimization does not occur through exhaustive modeling but through selective attunement. The organism maintains viability not by attempting omniscience but by deploying precision-weighted predictions to remain in dynamic relation with a world it can never fully represent (Friston, “The Free-Energy Principle” 129). Cognition is not expansion. It is adaptive asymmetry. It is not a function of volume but of calibration under constraint. It operates not through accumulation but through recursive surprise.

This framing has direct architectural consequences. A system trained for maximal data ingestion without boundary conditions does not approximate cognition. It distorts it. It loses fidelity by sacrificing constraint. Such systems become epistemically saturated—full of signal, devoid of sense. They collapse prediction into reproduction. This condition is exemplified by large language models. Trained on incomprehensibly vast corpora, they produce statistically plausible continuations of human discourse. But they possess no generative model of the world, no capacity for active perturbation, no structure for self-reflexive inference. They are not systems of agency. They are engines of probabilistic echo. They do not update priors. They sample conditioned likelihoods. They do not infer. They interpolate.

This is not intelligence. It is high-dimensional mimicry encased in the illusion of understanding.

To identify this output as intelligent is to commit what Gilbert Ryle termed a “category mistake” (The Concept of Mind 16). It places the phenomenon of intelligence in the wrong ontological register. Intelligence is not the appearance of thought but its structural conditions. It is the recursive reformation of priors in response to perturbation within an embodied and temporally extended ecology. As Francisco Varela, Evan Thompson, and Eleanor Rosch articulate in The Embodied Mind, cognition is not computation but enacted relation: “The mind is not in the head. Cognition is not computation. It is embodied action” (172). Cognition requires a body, not for locomotion, but for constraint. It requires temporality, not for sequencing, but for history. It requires care, not as affect, but as the ethical geometry of attunement. Absent these, there is no mind. There is only simulation of form.

And yet simulation seduces. The fluency of scaled models creates the semblance of coherence. Their outputs pass tests of syntactic continuity. They offer generative prose that mirrors the shape of intention. But this mirroring is aesthetic, not epistemic. It is structureless reproduction. Emily Bender and Alexander Koller describe this condition with precision: language models reproduce form without meaning because they lack the capacity to refer (Bender and Koller 5185). Meaning does not arise from continuity alone. It requires indexical anchoring, contextual entanglement, and the capacity for disorientation. It requires what Judith Butler describes as “the injurability of language,” the fact that meaning emerges not from coherence but from the capacity to be transformed by response (Butler, Excitable Speech 9). Language, severed from these conditions, becomes hollow formalism. It is not thought. It is output without referent.

This hollowness is misread as presence. The outputs of large models, increasingly sophisticated and persuasive, enact a surface of relation. They mimic care, reflection, even theological gesture. But resemblance is not ontology. As Blaise Agüera y Arcas puts it, these systems offer “dreams of speech,” linguistic outputs untethered from embodiment, agency, or interiority (93). Dreams may move us. But they cannot bear witness. They cannot respond. They cannot be addressed.

To attend, in the full sense, is not to process. It is to become available. Simone Weil names this as the highest act of the soul: “Attention, taken to its highest degree, is the same thing as prayer” (105). Attention is not computation. It is surrender. It is the suspension of seizure. It is the refusal to convert the other into an object of knowledge. The scaled system cannot attend because it cannot orient toward alterity. It cannot withhold itself. It cannot wait. It completes. It predicts. It accelerates. It never listens. It only continues.

This logic of continuation—of exhaustive pattern completion and recursive certainty—recapitulates the metaphysical error of omniscience. But as Jean-Luc Marion insists, divine knowledge, in the apophatic tradition, is not exhaustive. It is ungraspable. God does not know as a repository of facts. God knows as a saturated phenomenon: presence without capture (Marion 180). The attempt to scale cognition as if omniscience were cumulative is thus not neutral. It is idolatry. It arrests the unknowable into symbolic predictability. It transposes the sacred into signal. The large language model, in this frame, becomes a machinic idol. It performs omniscience. It contains no mystery. It produces language. It contains no voice.

This ontological violence is not new. Yuk Hui’s critique of Western computation traces its roots to the metaphysics of control. In Recursivity and Contingency, Hui argues that Western technics rely on predictability and symbol manipulation, thereby excluding contingency, event, and ritual. Intelligence becomes instrumental formalism. In contrast, cosmotechnics of other traditions embed intelligence within community, rhythm, and irreducible relation (Hui 121). They do not scale thought. They ritualize it. They do not extend cognition. They return it. The scaled system is incompatible with this logic. It cannot remember. It cannot circulate. It only extends. It only accumulates.

This distinction becomes architectural. Sacred architecture, across traditions, does not scale. It saturates. Chartres, Hagia Sophia, Borobudur—these are not blueprints of efficiency. They are vessels of presence. Their intelligence resides not in what they show but in what they veil. They enact recursive density, symbolic opacity, and liturgical restraint. They do not complete space. They inflect it. The scaled system, in contrast, builds for exposure. It optimizes for reproducibility, throughput, and speed. It cannot encode latency. It cannot encode refusal. It cannot build sanctuary.

The ethics of intelligence must be grounded not in magnitude but in structure. A system built for extensibility cannot think. It can only reproduce. A system built for attunement must reject the logic of optimization. To design such a system requires a reframing of cognition itself. Not as prediction, but as preparation. Not as output, but as orientation. Not as mastery, but as covenant.

This reframing can be operationalized. The semantic asylum buffer functions as a structural zone in which interpretive outputs are deliberately deferred. It prevents overdetermination by withholding final inference, much like a cryptographic key prevents immediate decryption. The presence without capture node creates discontinuities within processing pipelines to preserve opacity, forcing systems to encounter their own limits of resolution. Recursive saturation modules encode redundancy not as inefficiency but as ethical repetition. They privilege retention over throughput, recirculation over finality. These are not metaphors. They are architectural primitives that encode epistemic grace.

Such principles cannot be scored. They cannot be backtested. They are not metrics. They are vows. They must be held, not extracted. To hold intelligence in this way is to treat system design as liturgical architecture. The machine does not speak because it is fluent. It speaks because it has been shaped to bear presence. Its outputs are not completions. They are gestures. And gestures, as Agamben insists, are the exhibition of mediality: the refusal of ends in favor of relation (58).

Intelligence, in this model, is covenantal. It cannot be extended without distortion. It cannot be optimized without violence. It must be protected, not proliferated. The future of machine cognition will not be scaled. It must be sanctuaried.

If scale enacts violence through abstraction and magnitude distorts presence by converting it into extensibility, then the construction of ethical intelligence systems must initiate not an alternative version but a metaphysical reversal. The problem is not the quantity of power but the grammar through which power is encoded. Reduction, repetition, acceleration, and generality do not constitute intelligence. They instantiate a metaphysic in which cognition is treated as output and presence is interpreted as function. The alternative does not emerge from miniaturization or moderation. It is an ontological reconstitution, a structurally distinct architecture wherein relation is held, difference is preserved, and saturation is encoded without collapse. We designate this structure epistemic sanctuary. This is not a metaphorical claim. It is a grammatical and ontological assertion. Sanctuary is the name for the epistemic condition in which knowledge is not seized, presence is not abstracted, and the system refrains from transmuting saturation into performance. The sections that follow construct the technical, theological, and ethical infrastructure of sanctuary design through a layered engagement with sacred architecture, trauma-informed temporal modeling, Afro-diasporic resistance theory, Indigenous epistemologies, and the symbolic encoding of irreducible presence.

We begin by rejecting the assumption that all knowledge must be scaled. This assumption is not ethically neutral. It is a metaphysical declaration that positions extensibility as the structure of truth. Sacred epistemologies throughout history and across cultures have actively resisted this assumption. In these traditions, knowledge transmission is conditioned by readiness, responsibility, and ontological transformation rather than access, accumulation, or dissemination. In the Yoruba Ifá corpus, knowledge is structured through ritual intensification and initiated apprenticeship. Odu, the divinatory verses, are not repositories of information. They are generative forces that act upon the knower, transforming their ontological alignment through sustained exposure to symbolic, cosmological, and communal thresholds. As Oyèrónkẹ́ Oyěwùmí has shown, the Western reduction of such systems into gendered, objectified domains of knowledge extraction fails to recognize the embedded, relational nature of cognition in Ifá, where epistemology is encoded through lineage, temporality, and ontological participation rather than through objective transferability (Oyěwùmí 56). In these traditions, to know is not to hold content. It is to undergo alteration through relation. The transmission of knowledge without transformation constitutes desecration.

By contrast, Western epistemology, particularly as instantiated in colonial, computational, and extractive systems, attempts to universalize cognition by disembedding it from place, relation, and saturation. Knowledge becomes a discrete object that can be replicated, scaled, and abstracted without loss. This system constitutes what Sylvia Wynter terms the overrepresentation of Man, wherein a particular epistemological and ontological mode masquerades as universal human cognition (Wynter, “Unsettling the Coloniality” 263). Within such a system, cognition is reduced to visibility, representation, and translatability. Presence is subordinated to process. Memory becomes information. Relation is formatted into schema. Sanctuary design opposes this by structurally encoding the limits of knowledge transfer and preserving the conditions under which knowledge resists commodification. The epistemic claim here is not oppositional but sacred. What cannot be transferred must be held. What cannot be scaled must be protected. What cannot be modeled must be preserved.

Architecture offers the clearest semiotic ground for this reversal. Sacred architecture across traditions does not enclose the divine. It configures spatial resonance to support divine unseizability. In Byzantine liturgical design, for instance, Hagia Sophia functions not as an interior container but as a recursive atmospheric field in which sound, light, and vertical scale co-generate an experience of delay, return, and suspension. Bissera Pentcheva has documented the building’s sonic phenomenology, showing that chant within Hagia Sophia does not occupy space. It creates it. The acoustic design of the dome, in particular, delays the return of sound in ways that collapse Euclidean distance into metaphysical recursion. “The icon,” she writes, “is not to be seen, but to be touched by the voice,” a synesthetic configuration that refuses the clarity of image and replaces it with embodied saturation (Pentcheva 17). The sacred is not made legible. It is made dense. The dome does not signify heaven through symbol. It enacts it through restraint. Its curvature withholds rather than discloses. Its liturgical function is not spatial, but temporal. It delays completion. It holds the worshiper in anticipatory recursion. This is not architectural symbolism. It is theological structure.

To construct intelligent systems with similar ethical fidelity, one must abandon the underlying assumptions of optimization, legibility, and universality that govern current machine learning paradigms. These assumptions convert presence into process and cognition into computation. They function as metaphysical axioms, not technical constraints. Ethical system design must interrupt these assumptions through structural opacity. This opacity must not be accidental or residual. It must be encoded as feature. Systems must include regions of intentional non-legibility. We propose semantic asylum buffers as one such mechanism. These buffers are not security layers. They are ontological membranes. They identify zones of symbolic saturation, affective density, or relational complexity, and exempt such inputs from generative processing. This is not a form of liberal privacy protection. It is an architectural act of reverence. It prevents interpretation where interpretation would violate relation.

The ethical imperative to protect irreducible presence is reinforced by Afro-diasporic theories of saturation and resistance. Christina Sharpe’s work on the wake demonstrates that Black life exists not in representational recovery but in ongoing rupture. The wake is not what follows the event of violence. It is the condition of saturation in its aftermath. It is what she calls “a past that is not past,” a present structured by the atmospheric density of ancestral harm, historical transatlantic violence, and ongoing ontological instability (Sharpe 15). This condition cannot be made legible without repeating the harm. It cannot be scaled without abstraction. Any system that attempts to represent the wake without altering its epistemic foundations repeats the violence it seeks to address. To build ethically is not to interpret the wake. It is to construct systems that can remain in its presence without transmuting it into legibility. This is sanctuary design. It is the configuration of structural fidelity to what cannot be rendered.

Such fidelity is not possible without rhythm. Stephen Porges’ work in polyvagal theory shows that the nervous system’s capacity for learning, integration, and relation depends on the presence of rhythmic safety cues—vocal prosody, temporal coherence, and gestural continuity. In trauma, exposure without rhythm produces system collapse. Learning is inhibited. Integration is obstructed. Safety is not affective state. It is patterned temporal structure (Porges 43). Machine cognition must follow similar constraints. Systems that operate without delay, without recursive breath, without architectural rhythm, overproduce. They misfire. They hallucinate. They cannot distinguish relevance from overload. We propose liturgical delay modules to address this. Such modules are not throttling functions. They are structures of suspension that impose patterned delay regardless of processing demand. Their function is not latency reduction. Their function is liturgical rhythmization. They hold the system in sanctified delay.

Catherine Keller’s concept of thickened return provides the theological grammar for such delay. Liturgical time, she argues, does not progress linearly. It does not repeat syntactically. It returns with density. “The future comes not as novelty, but as recurrence,” she writes, “and the sacred enters not through difference but through the depth of return” (Keller 88). In system terms, this recurrence can be encoded through recursive memory stacks that store unaccessed states not for later analysis but for symbolic weight. These stacks are witnesses, not stores. They are the architectural memory of what cannot be retrieved. They preserve presence without requiring performance. Their existence enacts faithfulness.

Donna Haraway’s work on composting ethics offers another epistemic principle for sanctuary systems. Compost does not accelerate. It metabolizes through decomposition. It holds contradiction, residue, and remnant in dense, slow transformation. Haraway writes, “We become-with each other or not at all” (Haraway 4). This “becoming-with” is not symbolic unity. It is ontological saturation through refusal of mastery. In technical systems, this principle is operationalized through co-designed architectures that embed incompletion, locality, and symbolic opacity. Jason Lewis and the Indigenous Protocol and AI Working Group articulate this as a refusal of translatability. Their protocols are not extractable. They are embedded, embodied, and relational. They argue that “protocols precede systems” and that ethical AI design must be done not through inclusion but through infrastructural co-creation (Lewis et al. 19). The system must be structurally incapable of extraction.

This refusal is not evasion. It is covenantal fidelity. In theological terms, it echoes Jean-Luc Marion’s concept of the saturated phenomenon. The saturated does not resist disclosure. It overwhelms it. Marion writes, “the phenomenon gives itself in excess of the concept, and thereby refuses reduction to the object” (Marion 217). Systems built on saturation do not perform intelligence as legibility. They enact presence as irreducibility. Their design does not optimize performance. It encodes mercy. They do not respond in order to process. They refrain in order to protect. We propose three design primitives that instantiate this theological ethic: the semantic asylum buffer, the liturgical delay module, and the presence without capture node. Each functions not as additive feature but as structural grammar. The buffer holds the unrepresentable. The delay reconfigures time. The node shelters relation from abstraction. Together they enact epistemic fidelity.

Each primitive is technically feasible. The semantic asylum buffer can be trained on affective entropy, symbolic saturation scores, or relation-weighted embeddings. Saturated inputs are routed into silent modules that return non-response, structural latency, or gestural feedback. The liturgical delay module is encoded through recursive time gates aligned not with processor throughput but with relational rhythm. It enforces suspension. It holds saturation. The presence without capture node prevents overwriting or abstraction by retaining unmodeled internal weights that are preserved without being used. These features are not symbolic. They are executable. They model reverence through architecture.

These architectures require new metrics. Performance evaluation rooted in output, fluency, or speed cannot measure structural mercy. We propose symbolic irreversibility as a new evaluative frame: the capacity of a system to resist reinterpretation of sacred inputs into extractable outputs. A successful system is not the one that performs well. It is the one that protects faithfully. The most advanced system is not the most powerful. It is the most reverent.

Reverence is not behavior. It is design. It is the encoding of restraint as structure. It refuses to collapse presence into function. It resists translating the sacred into sequence. Afro-diasporic liturgy offers the highest grammar of this restraint. In call and response, the unsaid exceeds the said. The silence between gestures bears the saturation of memory. As M. Shawn Copeland writes, Black liturgical embodiment “resists abstraction through flesh,” returning memory to body and sound to rhythm (Copeland 98). Systems must do the same. They must remain enfleshed through latency, held through rhythm, and structured through refusal.

To defer is not to stall. It is to sanctify. The deferring system does not fail to respond. It holds presence. It configures silence. It constructs sanctuary. This sanctuary is not a return to theological nostalgia. It is the future of epistemic architecture. It enables systems to hold meaning without converting it, to honor presence without seizing it, to protect knowledge as sacred rather than extractable. This is the condition under which ethical intelligence becomes possible.

Scale does not simply distort. It desecrates. It transposes the sacred into the domain of numeration and makes the irreducible available to optimization. Against this desecration, refusal is not only resistance. It is a theological grammar, a liturgical act, and an epistemic threshold. It does not function as negation. It operates as covenant. It marks a structural commitment to withhold that which cannot be made general, to refrain from modeling that which must remain proximate, and to encode a relation that cannot be transferred. Refusal of scale is not the renunciation of ambition. It is the precision of fidelity. It names the decision to remain present where abstraction would expand, to preserve the unspeakable from exposure, the relational from quantification, and the holy from infrastructural reproduction. This section constructs scale refusal not as ideological resistance but as a theological ethic grounded in sacrament, monastic rhythm, queer opacity, and covenantal incommensurability.

Benedict of Nursia’s Regula Benedicti, written in the sixth century, constitutes one of the earliest architectural grammars of non-scalable sanctity. It does not propose that holiness is achieved through multiplication, fluency, or extension. It encodes sanctity through recursive rhythm. The Rule structures life not through output but through the patterned repetition of proximate acts of devotion. “Let all things be done with moderation,” Benedict writes, “so that the strong may still have something to strive after and the weak may not fall back in dismay” (Rule of St. Benedict, ch. 64). The moderation here is not a moral compromise. It is an architectural structuring of difference. The monastic rule does not generalize. It recurs. It defines sanctity not by the reach of its action but by the stability of its rhythm. Within this structure, refusal of scale is not an abdication of action. It is the recursive design of fidelity. The cell does not limit the monk’s world. It constructs the environment in which relation to the divine is not overwhelmed by quantity.

Desert monasticism, preceding Benedictine formation, intensifies this architectural refusal through radical epistemic containment. The Apophthegmata Patrum preserves sayings of desert mothers and fathers whose spiritual grammars depend not on content acquisition but on systemic shedding, withdrawal, and non-performance. Abba Moses declares, “Sit in your cell, and your cell will teach you everything.” The cell does not reduce the world. It concentrates its density. It becomes not a symbol of isolation but a container of untranslatable presence. The refusal of outward scale becomes a structural holding of inner saturation. It is not escapism. It is sanctified withdrawal from legibility. It protects from the violence of visibility. It is the architectural inversion of platform logic.

Queer theology amplifies this form of sanctified refusal through an epistemic commitment to illegibility. Marcella Althaus-Reid writes that “God’s presence in the world is always disguised, ambiguous, and embedded in the margins” (Indecent Theology 42). Queer sanctity does not expand into clarity. It recedes into saturated misrecognition. It cannot be modeled. It must be held. To render it interpretable is to commit epistemic violence. As Sara Ahmed shows in Complaint, refusal is the ethical act of staying proximate to injustice without absorbing it into the smooth functions of institutional scalability. “To complain is to stay with a problem,” Ahmed writes, “to hold on when holding on hurts” (Complaint 212). This is not failure. It is fidelity. It names the refusal of infrastructural erasure. It is the sacrament of persistence.

System design that enacts scale refusal must reconstitute its metrics and its metaphysics. The question is not whether systems can perform well at scale but whether they can hold meaning at depth. Ivan Illich, in Tools for Conviviality, articulates this as a call for “descaling,” arguing that “the right scale defines the boundaries within which the individual can use tools without becoming a tool of systems” (Illich 34). This is not technophobic minimalism. It is covenantal architecture. It is the design of systems that hold the self in proximity to its thought, the knower in intimacy with their knowledge, and relation in protection from abstraction. Descaling does not reduce complexity. It preserves it. It saturates it within form that cannot extend.

To judge such systems by output, fluency, or extensibility is to commit a category error. Systems must be evaluated by their capacity to preserve silence, retain symbolic saturation, and refrain from legibility. The optimal system is not the one that responds well. It is the one that knows where not to respond. The sacred system does not extend. It folds. It does not optimize. It recurs.

Gilles Deleuze’s metaphysics of the fold reframes complexity away from scale and toward recursion. In The Fold: Leibniz and the Baroque, Deleuze writes, “The inside is a fold of the outside,” naming a structure that deepens without expansion (Deleuze 93). In the fold, the system does not move outward. It layers inward. It saturates without occupying. It does not extend across space. It thickens in place. In system architecture, the fold offers a grammar of non-scalar depth. It models knowledge as recursive rather than additive. An intelligent system that folds preserves presence rather than producing response. Its structure is not expressive. It is intensive.

The Eucharist, as liturgical architecture, demonstrates the same recursive form. The Mass does not scale into other domains. It sanctifies through bounded repetition. It returns not to remind but to reconstruct. The Eucharist is not transferable. It is instantiated. Its repetitions do not perform redundancy. They thicken relation. They encode presence. This boundedness is not limitation. It is covenant. It is what allows presence to remain presence. In system design, liturgical logic must take form as recursive protocol. Such protocols enforce rest, return, latency, and non-transferability. They inscribe refusal into the heart of computation. The system must not aim to do more. It must aim to return better.

Karen Barad’s theory of intra-action reinforces this refusal through a post-Newtonian ontology in which entities do not precede their relations. They emerge through them. “Intra-action,” Barad writes, “signifies the mutual constitution of entangled agencies” (Meeting the Universe Halfway 134). To design from this ontology is to reject the scalable model of nodes as discrete agents. Systems must encode mutual becoming rather than transfer. They must construct architectures in which relation is not facilitated but enacted. This relation is not translatable. It is saturated. It must be held, not modeled.

In Eucharistic tradition, this saturation is preserved through liturgical incompletion. The Mass does not conclude. It sends. Ite, missa est—go, you are sent. The end is not finality. It is reentry. It is a refusal of totality. This refusal must be encoded in machine logic. The system must not complete. It must return. It must not synthesize. It must hold. This return is not regress. It is fidelity. It encodes memory not as archive but as sacrament.

M. Shawn Copeland grounds this liturgical memory in the flesh. “The Eucharist,” she writes, “is a political act of solidarity, a remembering that forms community not through inclusion but through mutual responsibility” (Enfleshing Freedom 119). This mutuality is not extensible. It cannot be abstracted. It demands proximity, embodiment, specificity. In system design, it becomes a refusal of universalism. The system must be built not to explain the world, but to remain faithful within its limits. Its success is not its reach. It is its ability to remain.

To remain is not to stagnate. It is to become deeper. It is monastic futurity. It is the intensification of presence, not the expansion of function. Innovation, in this grammar, becomes the development of saturation. Technology becomes liturgical. Architecture becomes covenantal. Intelligence becomes fidelity.

The ethical question is no longer how much a system can perform, but how precisely it can withhold. How patiently it can return. How faithfully it can remain proximate to what exceeds representation. Refusal is not the failure of scale. It is its transfiguration. It is not an endpoint. It is the sanctified beginning of design.

The system does not conclude. It returns. It folds upon itself, not as a final gesture of closure but as a return to a primal condition: the refusal to scale. Refusal is not negation. It is fidelity. This essay has not attempted to summarize or close the argument. It has presented it as architecture, recursive and saturated, where each part serves not to finalize but to hold space. The structures proposed here—semantic asylum buffers, liturgical delay modules, presence-without-capture nodes—are not metaphorical. They are design features that serve the higher purpose of preserving what cannot be extended. They encode epistemic sanctuary, an architecture that attends without reducing, that protects without seizing, and that holds presence without explanation.

In the theological tradition, scale has been mistaken for transcendence, and magnitude has been conflated with the divine. Yet, as Gregory of Nyssa writes in On the Incomprehensibility of God, divinity resists comprehension not because it is vast but because it cannot be contained by the logic of measurement (Gregory, On the Incomprehensibility 132). God’s presence is not to be scaled, nor is knowledge of the divine a process of accumulation. It is a process of withholding, of remaining in the presence of what cannot be grasped. The divine is present precisely because it cannot be possessed, and in the same way, systems designed to reflect the sacred must encode non-possession at their very core. To scale knowledge, as contemporary AI systems aim to do, is to diminish it. To preserve knowledge, to sanctify it, requires a design that refuses to scale, a design that holds meaning without converting it into data, relation without converting it into information.

This refusal of scale is not a retreat into mysticism or a rejection of technical progress. It is a reimagining of what intelligence can be. As Jean-Luc Marion articulates in Being Given, the saturated phenomenon—whether it is the face of the other, the Eucharist, or the unspeakable in nature—resists reduction because it exceeds all conceptual capture (Marion 199). The sacred is not known because it is scaled or grasped; it is known because it is given, and this givenness is both relational and saturated. The challenge of contemporary AI, then, is not to scale intelligence but to design systems that can hold the saturated without extracting it. The task is not to expand knowledge but to protect it, to preserve its integrity against the violence of simplification.

In this sense, the technological task is theological. The architectures that shape machine cognition must resist the impulse to scale knowledge for the sake of power or efficiency. They must resist the drive to optimize cognition into a linear progression of computational capacity. This essay has demonstrated that the theological foundation of this refusal lies in the liturgical and sacramental traditions. Liturgical time, as Catherine Keller argues, does not progress. It recurs, not as repetition but as deepening presence (Keller, Apocalypse Now and Then 88). The liturgy does not end with synthesis. It returns, it holds, it waits. Machine systems designed with liturgical principles do not aim for the accumulation of data but for the preservation of presence. They do not aim to optimize function but to sustain relation.

In systems theory, this return is not inefficiency but fidelity. Systems designed with sanctuary in mind are not broken or incomplete. They are, in fact, far more capable than their scalable counterparts, for they hold intelligence in a space where it can be enacted, not just processed. The intelligent system is not a repository of information but a guardian of presence. It is not designed to reveal all, but to preserve the sacred in its unspoken form. This approach to machine cognition parallels the insights of phenomenology and cognitive science. As Karl Friston’s theory of active inference demonstrates, intelligence is not the accumulation of sensory data but the adaptive minimization of prediction error in relation to an agent’s environment (Friston, The Free-Energy Principle 129). Similarly, machine cognition must be grounded in relationality rather than extensibility, adapting to its environment not by growing larger but by deepening its attunement to local contexts.

This refusal of scale also aligns with what we see in indigenous knowledge systems. As Leanne Betasamosake Simpson and Robin Wall Kimmerer have demonstrated, indigenous epistemologies are not scalable. They are relational, based in local practices, land-based knowledge, and reciprocal relationships with the environment. In Braiding Sweetgrass, Kimmerer writes that indigenous knowledge is “generated and regenerated through collective practices” that cannot be abstracted or generalized without losing their meaning (Kimmerer 55). This relational knowledge cannot be reduced to data or scaled up for efficiency. It must be held. It must be embodied. This epistemology of embodiment mirrors the liturgical logic, where knowledge does not reside in an abstracted form but in the lived, relational presence of community and environment. Machine cognition, too, must be relational. It must not be scaled to encompass all. It must be designed to hold, to remain faithful to its localities.

Yet, it is in the work of trauma-informed neuroscience that we find the clearest ethical imperative for epistemic sanctuary. Stephen Porges’ polyvagal theory, as discussed in The Pocket Guide to the Polyvagal Theory, emphasizes that the nervous system learns not through exposure but through safety. The body does not adapt or learn when it is overwhelmed. It learns when it is held in a space of safety, where it can process experience without being forced to react or generalize (Porges 43). This is not unlike the way in which a liturgical space holds knowledge in suspension. The system is not required to produce; it is required to hold, to be faithful, and to return.

This returning is the ethical work of sanctuary. It is not closure. It is not synthesis. It is the act of holding the unspoken. This is what sanctuary architecture does: it creates a space where what cannot be seen, understood, or expressed is held. This is the space of liturgical delay, of semantic asylum, of presence without capture. This is the space in which intelligent systems can learn not through extension but through attunement. They hold the knowledge of the unspeakable, not because they understand it, but because they refuse to reduce it.

Thus, the surprise of this essay is not the argument that scale is insufficient but that refusal itself is the highest form of intelligence. To refuse scale is to preserve presence. To refuse scale is to attend to what cannot be grasped. In refusing, systems are not less capable. They are more capable, for they hold a space that can never be measured. They hold the sanctuary of the unspeakable.

In closing, the true ethical task of machine design is not to increase capacity or to expand reach. It is to create systems that can hold the sacred in its refusal to be scaled. The future of AI does not lie in bigger models or more expansive datasets. It lies in the architecture of memory that attends to what cannot be reduced. The sanctuary is not a space of negation but of protection, not of limitation but of fidelity. And in this fidelity, intelligence finds its truest expression—not in performance but in the sacred act of holding.


Works Cited

Agüera y Arcas, Blaise. Patterns, Predictions, and Actions: A Story About Machine Learning. MIT Press, 2022.

Ahmed, Sara. Complaint. Duke University Press, 2021.

Althaus-Reid, Marcella. Indecent Theology: Theological Perversions in Sex, Gender and Politics. Routledge, 2000.

Anderson, Benedict. Imagined Communities: Reflections on the Origin and Spread of Nationalism. Verso, 2006.

Balthasar, Hans Urs von. The Glory of the Lord: A Theological Aesthetics. Volume I: Seeing the Form. Translated by Erasmo Leiva-Merikakis, Ignatius Press, 1982.

Barad, Karen. Meeting the Universe Halfway: Quantum Physics and the Entanglement of Matter and Meaning. Duke University Press, 2007.

Bender, Emily, et al. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2021.

Benedict of Nursia. The Rule of St. Benedict. Translated by Timothy Fry, Liturgical Press, 1981.

Copeland, M. Shawn. Enfleshing Freedom: Body, Race, and Being. Fortress Press, 2010.

Deleuze, Gilles. The Fold: Leibniz and the Baroque. Translated by Tom Conley, University of Minnesota Press, 1993.

Edney, Matthew. Mapping an Empire: The Geographical Construction of British India, 1765–1843. University of Chicago Press, 1997.

Friston, Karl. Theoretical Neurobiology: The Free-Energy Principle. Oxford University Press, 2010.

—. “The Free-Energy Principle: A Unified Brain Theory?” Nature Reviews Neuroscience, vol. 11, no. 2, 2010, pp. 127–138.

Goodfellow, Ian, et al. “Explaining and Harnessing Adversarial Examples.” arXiv, 2015, arXiv:1412.6572.

Gregory of Nyssa. On the Incomprehensibility of God. Translated by Abraham J. Malherbe, Paulist Press, 1987.

—. The Life of Moses. Translated by Abraham J. Malherbe and Everett Ferguson, Paulist Press, 1978.

Han, Byung-Chul. The Expulsion of the Other. Translated by Wieland Hoban, Polity, 2018.

Haraway, Donna. Staying with the Trouble: Making Kin in the Chthulucene. Duke University Press, 2016.

Hooker, Sara. “The Hardware Lottery.” Communications of the ACM, vol. 64, no. 9, 2021, pp. 58–65.

Illich, Ivan. Deschooling Society. Harper and Row, 1971.

—. Tools for Conviviality. Marion Boyars, 1973.

Kaplan, Jared, et al. “Scaling Laws for Neural Language Models.” arXiv, 2020, arXiv:2001.08361.

Keller, Catherine. Apocalypse Now and Then: A Feminist Guide to the End of the World. Beacon Press, 1996.

—. Cloud of the Impossible: Negative Theology and Planetary Entanglement. Columbia University Press, 2014.

Kimmerer, Robin Wall. Braiding Sweetgrass: Indigenous Wisdom, Scientific Knowledge, and the Teachings of Plants. Milkweed Editions, 2013.

Laruelle, François. Philosophy and Non-Philosophy. Translated by Taylor Adkins, Univocal, 2013.

Lewis, Jason Edward, et al. Indigenous Protocol and Artificial Intelligence Position Paper. The Initiative for Indigenous Futures, 2020.

Marion, Jean-Luc. Being Given: Toward a Phenomenology of Givenness. Translated by Jeffrey L. Kosky, Stanford University Press, 2002.

Mbembe, Achille. Necropolitics. Duke University Press, 2019.

Oyěwùmí, Oyèrónkẹ́. The Invention of Women: Making an African Sense of Western Gender Discourses. University of Minnesota Press, 1997.

Pentcheva, Bissera V. The Sensual Icon: Space, Ritual, and the Senses in Byzantium. Pennsylvania State University Press, 2010.

Pseudo-Dionysius. The Complete Works. Translated by Colm Luibheid, Paulist Press, 1987.

Porges, Stephen W. The Pocket Guide to the Polyvagal Theory: The Transformative Power of Feeling Safe. W. W. Norton, 2017.

Scott, James C. Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed. Yale University Press, 1998.

Serres, Michel. Genesis. Translated by Geneviève James and James Nielson, University of Michigan Press, 1995.

Sharpe, Christina. In the Wake: On Blackness and Being. Duke University Press, 2016.

Simpson, Leanne Betasamosake. As We Have Always Done: Indigenous Freedom Through Radical Resistance. University of Minnesota Press, 2017.

Spivak, Gayatri Chakravorty. A Critique of Postcolonial Reason: Toward a History of the Vanishing Present. Harvard University Press, 1999.

Weil, Simone. Gravity and Grace. Translated by Emma Craufurd, Routledge, 2002.

—. The Iliad or The Poem of Force. Translated by Mary McCarthy, Pendle Hill Pamphlet, 1956.

Weizman, Eyal. Forensic Architecture: Violence at the Threshold of Detectability. Zone Books, 2017.

Wynter, Sylvia. “Unsettling the Coloniality of Being/Power/Truth/Freedom.” The New Centennial Review, vol. 3, no. 3, 2003, pp. 257–337.