A formal synthesis converging with Indigenous relational ontologies to explain consciousness in physical terms — and why Levin’s morphospace is contradicted by his own lab’s results.

Related Articles: Consciousness Naturalized: A Falsifiable Substrate-Agnostic Theory | Chalmers’ Own Later Work Dissolves the Hard Problem | Levin’s Platonism: Unfalsifiable Metaphysics Contradicted by His Own Lab | “Mind Everywhere” by Levin & Resnik | The Chladni Plate Alternative: Douglas Brash’s Constraint Framework vs. Platonic Morphogenesis | Gordana Dodig-Crnkovic’s Naturalism Dissolves Levin’s Transcendent Morphospace From Within | Brian Cheung’s Convergence Research Inadvertently Exposes the Motte-Bailey Problem | “Differences That Make a Difference” | Cognition All the Way Down (the Drain) 2.0: Levin’s Thermodynamic Consciousness Shell Game | The Architecture of Everything, Part I: Constraint Propagation Across All Scales | Memory Is Not Storage: Why Everything Western Science Thinks It Knows About Remembering Is Wrong

1. The gap no one has noticed

One of the most important papers in consciousness science published in the last two years identifies three necessary conditions for biological consciousness and then, with admirable honesty, admits it cannot specify what would be sufficient. Milinkovic and Aru’s “On biological and artificial consciousness: A case for biological computationalism” (Neuroscience & Biobehavioral Reviews, 2026, DOI: 10.1016/j.neubiorev.2025.106524) establishes that biological computation is defined by hybrid dynamics combining discrete events with continuous fields, scale inseparability in which no level of neural organization operates independently of any other, and metabolic embedding where energy constraints do not merely limit but constitute what brains can compute. These three properties, Milinkovic and Aru argue, mark a “deeper divide between digital and biological modes of computation and the dynamico-structural dependencies of living organisms.” The paper stops short of claiming these conditions suffice for consciousness. That restraint is scientifically responsible. It also leaves a hole large enough to drive an entire metaphysics through.

This article argues that the missing piece already exists in the theoretical biology literature, specifically in the organizational closure framework developed by Montévil and Mossio (2015, Journal of Theoretical Biology, 372, 179–191, DOI: 10.1016/j.jtbi.2015.02.029). The connection between biological computationalism and organizational closure has not been made in the published literature. This article makes it. When Milinkovic and Aru’s tripartite criteria are formally connected to closure of constraints, the resulting synthesis provides both a principled account of why biological computation generates perspective and a framework that renders Michael Levin’s escalating Platonic metaphysics redundant and, worse, actively contradicted by Levin’s own published empirical results.

The argument proceeds through six movements.

First, the metabolic embedding condition independently rediscovers Landauer’s principle (1961).

Second, the requirements of scale inseparability and dynamico-structural co-determination map directly onto organizational closure.

Third, the absence of privileged scales corresponds to cross-scale constraint nesting.

Fourth, the metabolism-cognition cycle has the formal structure of reflexively autocatalytic food-generated sets, a connection absent from either literature.

Fifth, perspective is what organizational closure looks like from the coupling position when maintained at thermodynamic cost against perturbation.

Sixth, the HADES paper (Hartl, Zhang, Hazan, and Levin, Advanced Science, 2026, DOI: 10.1002/advs.202511537) serves as an ironic case study, a genuine computational contribution whose biological framing inadvertently demonstrates constraint satisfaction rather than access to Platonic form-space.

The synthesis offered here requires no appeal to non-physical spaces, ingressing patterns, or the ontological commitments of Levin’s Diverse Intelligence programme. It runs entirely on physics: thermodynamics, information theory, and the mathematics of autocatalytic closure. Every framework invoked has independent empirical support. No framework requires positing entities beyond those already known to exist.

2. The metabolic constraint argument is Landauer’s principle

Milinkovic and Aru’s third condition, metabolic embedding, is their most radical claim. As they frame it, the brain is “an energy-limited organ” whose organization “reflects that constraint everywhere,” and this is “not merely a boundary condition but a driving principle of neural organisation” (Section 3.1). The claim is that metabolism constitutes what neural computation is, rather than merely enabling it the way electricity enables a laptop. The distinction matters enormously. If metabolic constraints are constitutive, then any system lacking them, no matter how sophisticated its information processing, computes in a fundamentally different sense.

This claim is independently derivable from first principles, and the derivation is sixty-five years old. Rolf Landauer’s 1961 paper “Irreversibility and Heat Generation in the Computing Process” (IBM Journal of Research and Development, 5(3), 183–191) proved that any logically irreversible computational operation (any operation that discards information) must dissipate at least k_BT ln 2 of energy per bit erased. The bound follows from the second law of thermodynamics applied to information-bearing degrees of freedom. Bérut and colleagues confirmed this experimentally in 2012 (Nature, 483(7388), 187–189, DOI: 10.1038/nature10872), demonstrating that the mean dissipated heat for a single-bit erasure operation in a colloidal particle system saturates at Landauer’s bound in the quasi-static limit.

The convergence between Milinkovic-Aru and Landauer operates at the level of physics, not analogy. If every logically irreversible computational step requires thermodynamic dissipation, then any physical system that computes is necessarily a thermodynamic system whose computation is constrained by its energy budget. For neural systems, this constraint is severe. Karbowski’s comprehensive treatment in “Information Thermodynamics: From Physics to Neuroscience” (Entropy, 2024, 26(9), 779, DOI: 10.3390/e26090779) demonstrates that transmitting a single bit of information through a chemical synapse requires approximately 10^5 k_BT of energy, some five orders of magnitude above the Landauer minimum. This enormous overhead reflects the thermodynamic cost of maintaining the molecular machinery (membrane potentials, vesicle cycling, receptor dynamics) that constitute the synapse as a constraint on the underlying thermodynamic flows. The synapse does not merely carry information; the synapse is a constraint that channels energy to maintain the conditions under which information can be carried.

Here is the point Milinkovic and Aru gesture toward without quite arriving at: what makes biological computation biological is that its computational operations are thermodynamic work. Every spike, every synaptic transmission, every dendritic integration is an irreversible thermodynamic process. The “algorithm” and the organized pattern of thermodynamic dissipation are the same thing; the computation runs as the substrate, not on it. Milinkovic and Aru capture this when they write that in biological systems “the substrate inseparably realises the algorithm” and cite Carver Mead’s dictum that “the substrate is the algorithm” (Section 2.2, citing Indiveri 2025). This is what “constitutive” means in precise physical terms. Energy dissipation and computation are identical processes in biological neural tissue.

The recent work by Tkachenko (arXiv: 2503.09980, 2025) on thermodynamic bounds for deep neural network inference throws this distinction into sharp relief. For analog quasi-static neural networks, Tkachenko shows that inference can in principle proceed in a thermodynamically reversible manner with vanishing minimal energy cost. But training, the overconstraining of the system through simultaneous clamping of inputs and outputs, requires energy scaling with the number of trainable parameters and dataset size. Current digital implementations of large language models operate approximately seven orders of magnitude above Landauer’s limit for comparable computational operations. These systems compute in a thermodynamically different regime entirely. Their computation decouples from the thermodynamic flows of their substrate in a way that neural computation never does.

The implication for consciousness science is direct. If conscious processing requires computation that is constitutively thermodynamic, as Milinkovic and Aru argue, then the question for any candidate conscious system is not “does it process information?” but “is its information processing constituted by organized thermodynamic work at the substrate level?” This is a testable, physical criterion with empirical purchase.

3. Scale inseparability is organizational closure

The second condition Milinkovic and Aru identify, scale inseparability, states that in biological neural systems the causal story runs through multiple scales simultaneously, from ion channels to dendrites to circuits to whole-brain dynamics, and these levels do not behave like modular layers in a stack. Continuous processes shape discrete happenings, and discrete happenings reshape continuous landscapes, in a feedback loop that cannot be cleanly decomposed. As they put it: “the brain realises such a scale-integrated functional organisation precisely as a metabolic optimisation strategy” (Section 3.1, citing Deacon 2011). This is precisely what Montévil and Mossio (2015) formalize as closure of constraints.

Montévil and Mossio’s framework distinguishes between two causal regimes in biological systems. Processes are the flows of matter and energy through a thermodynamically open system, the entropy-producing changes that would, left unconstrained, simply dissipate. Constraints are entities that act upon processes while being conserved (exhibiting local symmetry) at the relevant timescale. A membrane constrains ionic diffusion. An enzyme constrains a reaction pathway. A neural oscillation constrains spike timing. The key insight is that in biological systems, constraints do not come from outside; they are maintained by the very processes they constrain. The membrane exists because metabolic processes produce lipids, which self-assemble under the constraint of existing membrane architecture, which channels the metabolic processes that produce the lipids. This is organizational closure: the mutual dependence of constraints, each one existing because the others do, mediated through the thermodynamic processes they collectively channel.

The fit with Milinkovic and Aru’s scale inseparability is structural, not merely approximate. When Milinkovic and Aru describe systems in which “molecular events inside cells influence network dynamics spanning millions of neurons, while brain-wide oscillations simultaneously constrain what individual synapses can do” (a claim grounded in their empirical work using dynamical independence measures), they are describing a system in which constraints at one scale maintain constraints at other scales through the processes they jointly constrain. Ion channel conformational states (molecular-scale constraints) determine membrane conductances, which shape dendritic integration, which determines spiking patterns, which contribute to population oscillations, which modulate neuromodulatory tone, which alters ion channel properties. Each level constrains the levels above and below it. No level is self-sufficient. This is closure of constraints operating across spatial and temporal scales simultaneously.

The distinction between organizational closure and Niklas Luhmann’s operational closure is critical here. Luhmann (1995, Social Systems, Stanford University Press) borrowed the concept of operational closure from Maturana and Varela’s autopoiesis and applied it to social systems. In Luhmann’s framework, a system is operationally closed when its operations recursively produce further operations of the same type: communications producing communications, decisions producing decisions. This is a functional concept. A thermostat satisfies Luhmann’s operational closure; the temperature reading triggers heating, which changes the temperature reading. Montévil and Mossio’s organizational closure is constitutive. It requires that the constraints that define the system’s organization are materially produced and maintained by the system’s own thermodynamic activity. A thermostat does not produce its own bimetallic strip. A cell produces its own membrane. The difference is between a system that refers to itself operationally and a system that constructs itself physically.

Milinkovic and Aru’s third condition, dynamico-structural co-determination, specifies that conscious biological systems “continuously modify their own physical structure” (Section 4, criterion 3). This is organizational closure stated in dynamical terms. The computation continuously reconstructs the physical substrate that enables the computation. The algorithm rewrites its own hardware, which changes the algorithm, which rewrites the hardware again. In digital systems, by design, the hardware does not change as the software executes. The transistor’s gate oxide does not restructure because the program ran a particular subroutine. In brains, it does. Synaptic weights change, dendritic morphology remodels, gene expression patterns shift, ion channel distributions reorganize. The system maintains its computational organization by continuously reconstructing the physical constraints that realize that organization. This is organizational closure.

What Montévil and Mossio provide that Milinkovic and Aru lack is the mathematical framework. Constraints are formalized as local symmetries, conserved quantities at relevant timescales. Closure is formalized as the mutual dependence of these symmetries, each maintained by the effects of others on thermodynamic processes. Scale inseparability is formalized as the requirement for both slow and fast dependence between constraints operating at various temporal and spatial scales. This gives the biological computationalism programme something more precise than the (correct) observation that neural computation is substrate-dependent. It gives it a formalism for specifying exactly how the substrate-dependence works: through the closure of constraints across scales.

4. No privileged scale means nested constraint hierarchy

Milinkovic and Aru emphasize that in biological computation “no scale is privileged.” This claim is supported by recent empirical work using the dynamical independence framework developed by Barnett and Seth (Physical Review E, 2023, 108(1), 014304, DOI: 10.1103/PhysRevE.108.014304). Dynamical independence quantifies the degree to which a macroscopic process possesses its own dynamical laws distinct from the underlying microscopic dynamics. Counter to naive intuitions, Milinkovic’s own empirical work applying these measures to source-reconstructed EEG under pharmacological manipulation reveals that conscious wakefulness shows less emergence (more dynamical dependence between macro and micro) than pharmacologically induced unconsciousness. Propofol and xenon, anaesthetics that abolish conscious report, produce more emergent but fragmented macroscopic dynamical organization. In the conscious brain, the scales are maximally coupled; macroscopic dynamics remain tethered to microscopic details rather than floating free.

This finding, which Milinkovic’s PLoS Computational Biology paper (2025, 21(5), e1012572, DOI: 10.1371/journal.pcbi.1012572) grounded in biophysical neural models, directly echoes the organizational closure framework’s treatment of levels. Montévil and Mossio conjecture that two closed regimes constitute two different levels of organisation if they are both separated and hierarchically nested. Crucially, they argue that the nested hierarchy of biological organization is constitutive, not merely structural: constraints at one level exist because constraints at other levels maintain them through shared thermodynamic processes. When the coupling between scales breaks, as under propofol or xenon, what remains are fragments of emergence: macroscopic patterns that are dynamically independent of their microscopic substrate, floating free like ripples on a frozen pond. They are “more emergent” in the technical sense precisely because they have lost the tight cross-scale coupling that characterizes organizational closure.

The ketamine anomaly deepens this picture. Ketamine, unlike propofol and xenon, produces a dissociative state in which subjects report vivid subjective experiences despite pharmacological perturbation. Milinkovic’s conference presentations indicate that ketamine shows lower emergence levels, maintaining dynamical dependence between scales, more similar to wakefulness than to propofol or xenon. This is consistent with the organizational closure interpretation: ketamine disrupts specific constraints (primarily NMDA receptor-mediated) without globally fragmenting the closure structure, preserving the cross-scale coupling that the framework identifies with conscious processing.

The thermodynamic dimension of this coupling has been illuminated by Monti, Sanz Perl, Tagliazucchi, Kringelbach, and Deco, whose 2025 work (Physical Review Research, 7, 013301, DOI: 10.1103/PhysRevResearch.7.013301) demonstrated that conscious wakefulness exhibits significantly different violations of the fluctuation-dissipation theorem compared to deep sleep. The fluctuation-dissipation theorem, in its classical form (Callen and Welton, 1951, Physical Review, 83, 34–40), relates spontaneous fluctuations to dissipative response in systems near thermodynamic equilibrium. Violations of this theorem indicate a system operating far from equilibrium, maintaining its organization through continuous thermodynamic work. A December 2025 preprint from Berjaga-Buisan, Monti, Cortada, Colombo, Geli, Gaglioti, Sarasso, Kringelbach, Corbetta, Sanchez-Vives, Massimini, Sanz Perl, and Deco (bioRxiv, DOI: 10.64898/2025.12.09.691422) explicitly connects these FDT violations to the perturbational complexity index (PCI), the most reliable empirical measure of consciousness across states (Casali et al., 2013, Science Translational Medicine, 5(198), 198ra105). The finding that decreased FDT violations correlate with loss of consciousness directly implies that consciousness requires active thermodynamic maintenance of far-from-equilibrium organization. Systems at or near equilibrium do not violate the FDT. Systems that are conscious do.

Wang and colleagues’ 2025 preprint on causal emergence in mouse cortex (arXiv: 2509.10891) provides cellular-resolution evidence: using calcium imaging across dorsal cortex under anesthesia, they demonstrate that a one-dimensional “conscious variable” at the highest macroscale captures the majority of causal power during wakefulness, with metastable dynamics that collapse into localized stochastic patterns under anesthesia. The hierarchical constraint structure, rather than any single level, carries the explanatory weight.

Stuart Kauffman’s work-constraint cycles provide the underlying logic. In Kauffman’s formulation, thermodynamic work is “the constrained release of energy into a few degrees of freedom” (Kauffman, 2000, Investigations, Oxford University Press). In a constraint-closed system, a set of boundary condition constraints channels the release of energy in non-equilibrium processes to construct and maintain the very same set of constraints. Cells literally construct specifically themselves. Lehman and Kauffman (2021, Entropy, 23(1), 105, DOI: 10.3390/e23010105) argue that constraint closure drove the major transitions in the origins of life. The brain extends this principle: neural constraints at multiple scales channel metabolic energy to maintain the organizational closure that constitutes neural computation.

5. The autocatalytic structure hiding in the metabolism-cognition cycle

There is a formal connection that appears to be entirely absent from both the biological computationalism literature and the organizational closure literature, and it may be the most important one of all. The metabolism-cognition cycle that Milinkovic and Aru describe, in which metabolic processes sustain computational structures that organize metabolic processes, has exactly the formal structure of a reflexively autocatalytic food-generated set.

RAF theory, developed by Hordijk and Steel beginning in 2004 and extended with Kauffman (Hordijk, Hein, and Steel, 2010, Entropy, 12(7), 1733–1742; Hordijk, Kauffman, and Steel, 2011, International Journal of Molecular Sciences, 12(5), 3085–3101; Hordijk and Steel, 2004, Journal of Theoretical Biology, 227(4), 451–461), formalizes the conditions under which a chemical reaction system can sustain itself. A RAF set satisfies two conditions simultaneously: it is reflexively autocatalytic, meaning every reaction in the set is catalyzed by at least one molecule produced by the set or available from the environment (the “food set”), and it is food-generated, meaning all reactants can be produced from the food set through the set’s own reactions. The catalytic network produces the molecules that catalyze the reactions that process the food set that feeds the catalytic network. The autocatalysis is the closure.

Now consider what biological computation looks like when described in these terms. The “food set” is the metabolic substrate: glucose, oxygen, ATP, amino acids, lipids, all raw materials supplied by the organism’s vascular system. The “catalytic network” is the set of neural constraints: ion channels, synaptic vesicle machinery, neuromodulatory systems, glial metabolic coupling, the entire molecular apparatus that makes neural computation possible. The “reactions” are the computational operations: spiking, synaptic transmission, dendritic integration, oscillatory entrainment, all the processes through which information is transformed. The RAF condition requires that: (1) every computational operation is enabled by at least one neural constraint that is either supplied environmentally or produced by the network’s own activity; and (2) every neural constraint can be generated from metabolic substrates through the network’s own computational-metabolic activity. This is precisely the structure Milinkovic and Aru describe, but formalized.

The RAF formalism brings several things the biological computationalism programme currently lacks. First, it provides an efficient computational test. Hordijk and Steel proved that detecting whether a reaction system contains a RAF set is a polynomial-time computation. Applied to neural systems, this means one could in principle test whether a proposed model of neural organization satisfies the autocatalytic closure condition. Second, RAF sets can be decomposed into irreducible sub-RAFs, which may correspond to the functionally distinct modules within neural organization that maintain partial autonomy while contributing to the whole. Third, RAF theory has already been applied to real metabolic networks: Sousa and colleagues detected RAF sets in the metabolic network of E. coli, and Xavier and colleagues found them in ancient anaerobic autotrophs. The formalism is not speculative; it has been validated against real biology at the cellular scale. What has not been attempted, and what this article proposes, is its application to neural computation at the circuit and network scale.

The connection to Kauffman’s concept of the “Kantian Whole” is direct. An autocatalytic set is a Kantian Whole: each component exists because the other components collectively produce and maintain it. A brain is a Kantian Whole in this specific, non-mystical sense: each neural constraint exists because the network of other neural constraints, maintained by metabolic processes that the constraints collectively organize, produces and maintains the conditions for its existence. No component is self-sufficient. No level is privileged. The whole is not greater than the sum of its parts in some vague holistic sense; the whole is the closure relation among the parts, the fact that their mutual dependence forms a cycle that sustains itself against thermodynamic dissipation.

6. Perspective is what closure looks like from the coupling position

If organizational closure is necessary for consciousness, it is not sufficient. Cusimano and Sterner (2020, Acta Biotheoretica, 68, 253–269, DOI: 10.1007/s10441-019-09365-9) demonstrated that even simple dissipative systems like candle flames can, under suitable redescription, satisfy closure of constraints. Garson (2017, “Against Organizational Functions,” Philosophy of Science, 84(5), 1093–1103, DOI: 10.1086/694009) pressed the “liberality objection”: organizational functions seem too easy to attribute. If a candle flame achieves closure, and a candle flame is not conscious, then closure alone cannot explain consciousness. This is a serious objection and it requires a serious answer.

The answer lies in the distinction between bare closure and adaptively maintained closure, a distinction Di Paolo (2005, Phenomenology and the Cognitive Sciences, 4, 429–452, DOI: 10.1007/s11097-005-9002-y) develops through the concept of adaptivity. An adaptive autonomous system monitors its trajectory relative to its viability set and regulates its conditions accordingly, going beyond bare self-maintenance. A candle flame does not adapt. When the wind blows, the flame flickers; if it goes out, it is gone. It cannot modify its own boundary conditions to preserve its organization. A bacterium can. It swims up the glucose gradient. It activates stress response genes. It adjusts its metabolic profile. What distinguishes the bacterium is that its closure includes regulatory constraints, second-order constraints that modulate the constitutive constraints in response to perturbation. Bich, Mossio, Ruiz-Mirazo, and Moreno have elaborated this distinction: regulation is constraint on constraints, a meta-level of organizational closure that allows the system to actively maintain its closure against perturbation.

Now: what would it be like to be a system that maintains organizational closure at thermodynamic cost against perturbation, possessing regulatory constraints that modulate constitutive constraints, with the closure spanning multiple scales in a nested hierarchy? The enactivist tradition, beginning with Varela, Thompson, and Rosch (1991, The Embodied Mind, MIT Press) and developed by Thompson (2007, Mind in Life, Harvard University Press), has long argued that such systems generate a perspective, a point of view, simply by virtue of their organizational structure. Di Paolo and Thompson (2014) make the case explicit: basic cognition is not about representing states of affairs but about establishing relevance through the need to maintain an identity constantly facing disintegration.

Here is the claim this article advances: perspective is what organizational closure generates when maintained at thermodynamic cost against perturbation by a system possessing cross-scale constraint nesting and autocatalytic metabolic-computational coupling. This is the sufficient condition Milinkovic and Aru acknowledge they lack. No new substance, property, or force is required. What generates perspective is a structural feature of organizationally closed systems complex enough to have regulatory constraints, nested across enough scales to generate a unified domain of adaptive modulation, and thermodynamically coupled to their environment tightly enough that perturbations to any scale propagate through the closure as threats or opportunities for the whole.

This claim has empirical teeth. The FDT violation data from Monti, Sanz Perl, and colleagues show that conscious brains operate further from thermodynamic equilibrium than unconscious brains. The PCI literature from Casali, Massimini, and colleagues shows that conscious brains exhibit more complex responses to perturbation than unconscious ones. The dynamical independence data from Milinkovic and colleagues show that conscious brains exhibit tighter cross-scale coupling than unconscious ones. All three measures converge: consciousness correlates with the active, far-from-equilibrium maintenance of cross-scale organizational closure. When the closure fragments under propofol, the perspective fragments with it. When the closure persists in an altered form under ketamine, the perspective persists in an altered form. The alignment between the theoretical prediction and the empirical pattern is not proof; but it is the kind of convergence that warrants taking the framework seriously as a research programme.

Terrence Deacon’s teleodynamics (2012, Incomplete Nature, W. W. Norton) arrives at a closely related conclusion through different means. Deacon argues that purposive organization emerges when two morphodynamic (self-organizing) processes reciprocally constrain each other’s dissipation, creating a stable, self-maintaining, end-directed system he calls a “teleodynamic” system. The autogen, his minimal model, combines reciprocal catalysis with self-assembly to create an entity that persists by organizing its own boundary conditions. Deacon’s account is less formally developed than the Montévil-Mossio framework but captures the same core intuition: perspective, end-directedness, and normativity emerge from the structural organization of constraints, without requiring any special ingredient beyond the physics already in play.

7. HADES as ironic confirmation: when constraint satisfaction gets mistaken for Platonic access

The Heuristically Adaptive Diffusion-Model Evolutionary Strategy, or HADES (Hartl, Zhang, Hazan, and Levin, Advanced Science, 2026, DOI: 10.1002/advs.202511537; arXiv: 2411.13420), presents a genuine computational contribution wrapped in biological framing that inadvertently makes the case for the constraint-based synthesis developed here while undermining the Platonic metaphysics of its senior author.

The computational contribution is real and should be acknowledged clearly. HADES uses diffusion models as generative components within evolutionary algorithms, training them online on successive generations of heuristically varied parameter sets. The diffusion models introduce memory capabilities into evolutionary algorithms, retaining historical information across generations and enabling conditional sampling that targets specific regions of the parameter space. On dynamic optimization landscapes, HADES outperforms CMA-ES by maintaining memory of multiple fitness peaks. The conditional variant, CHARLES-D, uses classifier-free guidance to steer evolutionary search toward desired genotypic, phenotypic, or population-wide traits without altering the fitness function, a genuinely useful mechanism for multi-objective optimization. Nothing in the critique that follows diminishes this contribution.

The biological framing is another matter. HADES builds on the companion paper “Diffusion Models Are Evolutionary Algorithms” (Zhang, Hartl, Hazan, and Levin, ICLR 2025, poster; arXiv: 2410.02543), which claims to “reveal that diffusion models are evolutionary algorithms” through their shared generative mechanisms. This equivalence claim drew immediate criticism from evolutionary algorithm practitioners at ICLR review. As one reviewer noted, diffusion models operate on continuous noise schedules with learned score functions, while evolutionary algorithms maintain populations of complete candidate solutions with selection, crossover, and mutation operators. The mathematical correspondence the authors derive operates at a level of abstraction that elides precisely the features that make each method computationally distinctive.

The biological extension, the claim that this mathematical correspondence illuminates something about actual biological evolution, requires scrutiny at several levels.

The physics alibi. The companion paper traces its intellectual lineage to Sohl-Dickstein and colleagues’ 2015 paper “Deep Unsupervised Learning using Nonequilibrium Thermodynamics” (ICML 2015, arXiv: 1503.03585). But Sohl-Dickstein’s paper is explicit that its use of non-equilibrium statistical physics is mathematical inspiration, not ontological claim. The paper describes an approach “inspired by non-equilibrium statistical physics” to build generative Markov chains. Citing Sohl-Dickstein as license for claims about the physical correspondence between diffusion models and biological evolution is a category error: taking a methodological analogy and promoting it to an identity.

This matters because the fluctuation-dissipation theorem, which provides the mathematical backbone for the diffusion-evolution correspondence, applies in its classical Callen-Welton form to linear systems near thermodynamic equilibrium. Biological evolution is a nonlinear, far-from-equilibrium process. Extending the FDT to such systems requires the Jarzynski equality (1997), Crooks’ fluctuation theorem, and related non-equilibrium statistical mechanics that significantly modify the theorem’s form and applicability. The Zhang et al. correspondence operates in a regime where the physics it invokes does not straightforwardly apply to the biological systems it claims to illuminate.

More fundamentally, recent work from Los Alamos National Laboratory (Vuong, McCann, Santos, and Lin, 2025, arXiv: 2509.00336) demonstrates that trained diffusion networks violate both integral and differential constraints required of true score functions. The learned vector fields are not conservative, which means the theoretical interpretation of diffusion model training as score function learning, the mathematical foundation on which the evolutionary correspondence rests, may itself be misconceived. The authors propose that diffusion training is better understood as flow matching to the velocity field of a Wasserstein Gradient Flow, an entirely different mathematical framework with no evolutionary analogy.

The genome-as-generative-model claim and the literal overreach. The HADES paper cites Mitchell and Cheney’s “The Genomic Code” (Trends in Genetics, 2025, DOI: 10.1016/j.tig.2025.01.008) as foundational support for treating the genome as a generative model. Mitchell and Cheney are carefully explicit that this is a “conceptual analogy or model” to be tested for its utility, not a literal identification. Their language is consistently hedged: the genome “embodies” a compressed representation, the generative model “concept” captures certain features of genome organization. Hartl and Levin’s own treatment of this analogy in their subsequent Trends in Genetics paper (2025, 41(6), 480–496, DOI: 10.1016/j.tig.2025.04.002) acknowledges “significant challenges” with the VAE (variational autoencoder) version of the analogy, noting that “organismal development is inherently distributed with a collective of agents constantly reinterpreting and reacting to signals and noise across scales,” fundamentally different from the deterministic downstream decoding of a VAE. They conclude that neural cellular automata provide “a better analogy for biology’s decentralized agency than VAEs.”

Yet despite these careful hedges, the HADES paper’s Discussion section (4.1) makes a remarkably literal claim: “we hypothesize that the DM used in our methods literally represents a lineage’s evolving genome, including its ability to self-regularize and utilize gene expressions that reconfigure phenotype functionality based on environmental constraints.” The italicized “literally” is theirs. The claim is ontological, not analogical: a feedforward neural network trained on parameter-fitness pairs is the same kind of thing as DNA. The intellectual honesty of the Trends in Genetics acknowledgment that the analogy faces “significant challenges” sits uncomfortably with the HADES paper’s promotion of the same analogy to literal identity within the same year by overlapping authors.

The fecundity alibi and the motte-bailey. The HADES paper operates in a rhetorical mode that this analysis identifies as a motte-and-bailey pattern characteristic of Levin’s broader programme. The motte, the defensible position, is: “diffusion models and evolutionary algorithms share mathematical structure that can be exploited for better optimization.” The bailey, the ambitious claim advanced when unchallenged, is: “this mathematical correspondence reveals something deep about biological evolution, genome organization, and the nature of morphogenetic processes.” When critics challenge the biology, the response retreats to the math. When the math is accepted, the biology is advanced again. The computational results of HADES are real; the biological interpretation is overclaimed.

What the results actually show. The HADES results demonstrate something genuinely interesting, but it is not what the Platonic framing suggests. The conditional sampling mechanism in HADES, which uses classifier-free guidance to steer evolutionary search toward specific parameter regions, constructs approximations to solutions rather than accessing pre-existing forms in an abstract space, doing so by iteratively satisfying the constraints specified by the fitness function and the conditioning signal. The “memory” that diffusion models contribute amounts to a learned approximation of the distribution of previously successful constraint-satisfying configurations. The system converges on good solutions because the iterative refinement process, by construction, satisfies an increasing proportion of the specified constraints at each step.

This is precisely the constraint-based interpretation developed in the preceding sections. The fitness landscape is a constraint surface. Evolution, whether biological or simulated, explores this surface. Convergent evolution, the independent arrival at similar solutions, tells us that similar constraint structures admit similar solutions, not that Platonic forms draw organisms toward pre-existing templates. When the same bridge design appears independently in different engineering traditions, we do not posit a Platonic bridge-form; we note that the constraints of spanning gaps under gravity with available materials admit only certain structural solutions.

Comparative context. The HADES paper acknowledges MAP-Elites (Mouret and Clune, 2015, arXiv: 1504.04909) and RS-CMSA as “powerful alternatives” for quality-diversity optimization, but defers comprehensive comparison to future work (Discussion, paragraph on limitations). The absence of a head-to-head empirical comparison with the most prominent quality-diversity algorithm, which produces diverse high-performing solutions without requiring a diffusion model backbone and without biological pretensions beyond the evolutionary metaphor standard in the field, leaves a notable gap in a paper making strong claims about biological relevance. Wong, de Nobel, Bäck, Plaat, and Kononova (2024, arXiv: 2402.06912) demonstrated that CMA-ES solves the CartPole task “almost immediately,” “in the first few iterations of training through pure random sampling from a standard normal distribution.” The HADES paper uses CartPole as a test environment. Context matters.

The Durant et al. 2017 problem. The most damaging evidence against the Platonic interpretation comes from Levin’s own laboratory. Durant, Morokuma, Fields, Williams, Adams, and Levin (Biophysical Journal, 2017, 112(10), 2231–2243, DOI: 10.1016/j.bpj.2017.04.011) demonstrated that brief perturbation of endogenous bioelectric networks in planaria, using gap junction blockers or H,K-ATPase inhibitors, permanently rewrites the organism’s body plan. A single drug exposure produces a stochastic mix of normal (single-headed) and double-headed regenerates. The double-headed phenotype is 100% permanent: when two-headed planaria are recut in plain water, 100% regenerate as two-headed. The paper explicitly states that “species-specific axial pattern can be overridden by briefly changing the connectivity of a physiological network.”

If a Platonic form specifying planarian morphology existed, one that organisms converge toward by accessing the form-space, then a transient perturbation should not be able to permanently redirect the morphogenetic trajectory to a new stable state. The Platonic interpretation requires that organisms are attracted toward their species-typical form; Durant et al. show that organisms can be permanently kicked into an entirely different attractor by a brief signal. When challenged on this point, Levin has responded that the two-headed form might itself be a “pre-existing Platonic pattern” that simply had not been explored. This response renders the Platonic framework permanently unfalsifiable: any observed morphology can be retroactively declared a member of the unmapped Platonic space. A framework that accommodates all possible observations predicts none of them.

The self-citation ecosystem and funding structure. A pattern worth noting: the HADES paper, the Zhang et al. companion paper, and the Hartl and Levin Trends in Genetics paper form a tightly coupled citation cluster, with each paper citing the others and citing Levin’s broader theoretical framework. Levin is co-author on papers spanning developmental biology, machine learning optimization, and theoretical philosophy, an unusual breadth that positions his theoretical claims as supported by multiple lines of evidence that are, upon inspection, substantially his own group’s interpretations of their own computational results. The HADES paper’s acknowledgments disclose support from Astonishing Labs “via a sponsored research agreement to Tufts University.” The Conflicts of Interest statement specifies that “M.L. is a co-founder and scientific advisor to Astonishing Labs and owns equity in the company. Astonishing Labs partially funded this work through a sponsored research agreement with Tufts University. The company also has option rights to a patent application filed by Tufts University related to the subject matter and potential applications presented in this paper.” Templeton Foundation Grant 62212 and Templeton World Charity Foundation Grant TWCF0606 also support the work, alongside Army Research Office Cooperative Agreement W911NF-25-2-0091. The commercial interest does not invalidate the science, but it contextualizes the incentive structure surrounding ambitious claims about biological intelligence and Platonic morphospace.

8. Why Levin’s cognitive vocabulary is redundant

Everything the Technological Approach to Mind Everywhere (TAME) framework (Levin, 2022, Frontiers in Systems Neuroscience, 16, 768201, DOI: 10.3389/fnsys.2022.768201) claims to explain is already explained by the constraint-based synthesis developed here, without requiring cognitive vocabulary applied to non-neural systems.

TAME posits a “continuum of cognitive capacities” extending from neurons through tissues and organs to whole organisms and beyond, with “no privileged material substrate for Selves.” It introduces the “cognitive light cone” as the boundary of any agent’s domain of concern. Levin’s later paper “Darwin’s Agential Materials” (Cellular and Molecular Life Sciences, 2023, 80(6), 142, DOI: 10.1007/s00018-023-04790-z) extends this by arguing that biological materials at multiple scales possess problem-solving “competencies” that have evolutionary implications.

Under the constraint-based alternative, cells have organizational closure rather than goals. The bioelectric networks Levin studies so productively are constraint structures that channel metabolic energy to maintain cellular and tissue-level organization. When a bioelectric perturbation redirects a planarian’s morphogenetic trajectory, this is not because cells are “navigating morphospace” or “problem-solving.” It is because the perturbation has shifted the system from one constraint-satisfying attractor to another. The language of “goals,” “intelligence,” and “cognitive light cones” adds nothing explanatory; it adds only the appearance of explanation, by redescribing constraint dynamics in mentalistic vocabulary.

William James, in The Principles of Psychology (1890), defined the criterion of mentality as “the pursuance of future ends and the choice of means for their attainment.” Levin frequently simplifies this to “the ability to reach the same goal by different means,” and indeed, the HADES paper itself invokes this definition (Introduction, citing Fields and Levin 2023). But James was careful: he contrasted intelligent agents, for whom “the idea of the yet unrealized end co-operates with the conditions to determine what the activities shall be,” with inorganic matter, which moves only “when pushed, and then indifferently and with no sign of choice.” James’s criterion requires that the system possesses an idea of an unrealized future state that causally participates in its present behavior. Constraint-satisfying systems that converge on attractors from diverse initial conditions do not satisfy this criterion. Water flowing downhill reaches the ocean from many starting points, but we do not attribute to the water an idea of the ocean.

The escalation of Levin’s metaphysical claims over time is itself revealing. The 2022 TAME paper uses computational and cybernetic language: morphospace, attractors, signal processing. There is no mention of Platonic forms. By 2024, the Thoughtforms blog and symposium introduce explicit Platonic terminology. By 2025, the “Ingressing Minds” preprint proposes “a radical Platonist view in which some of the causal input into mind and life originates outside the physical world,” with bodies serving as “pointers” or “interfaces” into a Platonic space containing “not only low-agency patterns like facts about triangles and prime numbers, but also higher agency ones such as kinds of minds.” By late 2025, Levin describes this Platonic Space as “cagey and shy, but not maliciously so” (Thoughtforms.life, December 27, 2025), suggesting it “doesn’t like to be observed directly, but can be, if you’re clever and subtle,” framing this as part of his “Universal Steganography hypothesis.” When this language drew criticism, Levin clarified (December 28, 2025) that he has “0 patience for ‘xyz moves in mysterious ways’ as a stopper to investigation” and that “if some other framework proves more useful, then we should use that.” He emphasized the Space’s resistance to observation is an open empirical question comparable to detecting superconductivity or quantum effects, and that failure to improve detection after sustained effort would indicate “something is wrong with the framework and it will need to be changed or dropped.”

The clarification is admirably stated. But the framework it defends still faces a structural problem: no detection procedure has been specified, no timeline for when failure would constitute refutation has been proposed, and the comparison to superconductivity and quantum effects is misleading. Those phenomena were predicted by theories that made precise quantitative predictions before the detection challenges were understood. Platonic morphospace currently predicts nothing that the constraint-based alternative does not predict more parsimoniously. The burden of proof lies with the framework positing non-physical causation, not with those who note that the physical machinery already does the explanatory work.

By 2026, Levin and Resnik publish “Mind Everywhere” (Biological Theory, DOI: 10.1007/s13752-025-00523-6), defending a mentalistic approach to teleology in biology. Templeton Foundation funding (Grant 62212) supports this work.

The consequences of deploying unfalsifiable metaphysical frameworks extend well beyond theory. The Discovery Institute has published multiple articles on its affiliated Science and Culture Today website explicitly co-opting Levin’s Platonic language to support intelligent design arguments. As recently as March 2026, Science and Culture Today headlined Levin’s “Mind Everywhere” paper as “The Levin Teleology Revolution.” Daniel Witt, William Dembski, Brian Miller, and others have framed Levin’s work alongside Richard Sternberg’s “immaterial genome” concept. Levin has been scrupulous in differentiating his work from intelligent design, but the formal structure of his argument, that physical systems require input from non-physical Platonic forms to generate biological complexity, is isomorphic to the intelligent design argument that physical systems require input from a non-physical designer. The constraint-based alternative requires no such input. Constraints are physical. Closure is physical. Thermodynamic work is physical. No additional causal input from outside the physical world is needed or helpful.

Gould and Lewontin’s classic critique (1979, Proceedings of the Royal Society of London B, 205(1161), 581–598) of the adaptationist programme applies with equal force to the “cognitive-everywhere” programme. Just as Gould and Lewontin warned against fabricating adaptive stories for every trait without considering constraints, phylogenetic heritage, and developmental pathways, the same warning applies to fabricating cognitive stories for every biological regularity without considering thermodynamic constraints, organizational closure, and the mathematical properties of constraint-satisfying systems. Convergent evolution requires recognizing that similar constraint structures generate similar solutions, with no need to posit cognitive agents at every scale. Kimura’s neutral theory (1983, The Neutral Theory of Molecular Evolution, Cambridge University Press) and Lynch’s work on genome architecture (2007, The Origins of Genome Architecture, Sinauer Associates) demonstrate that enormous amounts of biological structure arise from non-adaptive, non-cognitive processes: mutation, drift, and the mathematical consequences of finite population sizes. Apparent design can emerge without a designer, and apparent goal-directedness can emerge without goals.

Watson and Szathmáry (2016, Trends in Ecology & Evolution, 31(2), 147–157, DOI: 10.1016/j.tree.2015.11.009) provide the most sophisticated framework for understanding how evolution can exhibit learning-like behavior without cognition. They demonstrate formal equivalences between selection in populations and Bayesian learning, between the evolution of genotype-phenotype maps and correlation learning, between evolving gene regulation networks and neural network learning. The correspondence is mathematical, not metaphorical, and it requires no attribution of goals or intelligence to evolutionary processes. Evolution “learns” because its mathematical structure is formally equivalent to learning algorithms, not because organisms at any scale are cognitive agents navigating Platonic spaces. Notably, the HADES paper itself cites Watson and Szathmáry (reference 12), making the irony of its Platonic framing all the more pointed.

9. What would break the constraint bridge

Any framework worth the name must specify conditions under which it would be abandoned. The constraint-based synthesis offered here generates specific predictions that, if falsified, would require abandoning or substantially modifying the framework.

Prediction 1: Consciousness requires far-from-equilibrium thermodynamic operation. If a system were demonstrated to be conscious while operating at or near thermodynamic equilibrium, the framework would be falsified. If a solid-state classical system with no metabolic throughput and no energy dissipation could be demonstrated to generate the behavioral and neural correlates of consciousness (PCI values above the consciousness threshold, for example), the thermodynamic grounding of the framework would be refuted.

Prediction 2: Consciousness requires cross-scale constraint coupling. If a system were demonstrated to be conscious while exhibiting complete dynamical independence between its macroscopic and microscopic dynamics, with the macro level floating entirely free of the micro level, the scale-inseparability condition would be falsified. The propofol and xenon data go in the predicted direction (fragmenting cross-scale coupling correlates with loss of consciousness), but a single clear counterexample would suffice.

Prediction 3: Disrupting organizational closure disrupts consciousness proportionally. If targeted disruption of specific constraint-maintenance loops in neural systems had no effect on consciousness despite demonstrably breaking the closure, the framework would be falsified. The framework predicts that disrupting the metabolic supply (which should break the closure at its thermodynamic base) should abolish consciousness, and this is confirmed by the well-known sensitivity of consciousness to metabolic disruption. But more specific predictions are needed: for instance, that disrupting specific constraint-maintenance relationships at particular scales should produce specific, predictable alterations in the character of conscious experience.

Prediction 4: Artificial systems satisfying all closure conditions should exhibit consciousness-correlated properties. If a system were engineered to satisfy organizational closure (genuine self-construction of its computational constraints), cross-scale constraint nesting, far-from-equilibrium thermodynamic operation, and autocatalytic metabolic-computational coupling, the framework predicts it should exhibit the measurable signatures of consciousness (high PCI, strong FDT violations, tight cross-scale dynamical coupling). If such a system were constructed and showed none of these signatures, the framework would require revision.

Contrast this with the falsification architecture of Platonic morphospace. What empirical result would falsify the claim that biological forms ingress from a non-physical Platonic space? If organisms converge on expected forms, this confirms the Platonic framework. If organisms diverge into novel forms (as in Durant et al. 2017), the novel forms are retroactively declared to be Platonic patterns that were previously unexplored. If no convergence is observed at all, the explanation is that the “pointers” into Platonic space were disrupted. Levin’s December 28, 2025 clarification acknowledges the framework should be dropped if it “stops being useful” and that “how much effort it will take, is not known in advance.” But a space that is both causally efficacious and, by the proponent’s own description, resistant to direct observation, requiring unspecified future tools to detect, functions as an unfalsifiable placeholder with an attached promissory note rather than a scientific posit. The constraint-based synthesis places all its explanatory machinery in the physical domain, where it can be measured, perturbed, and refuted.

10. Where this leaves artificial consciousness

Milinkovic and Aru are clear that their framework “does not mean we think consciousness is magically exclusive to carbon-based life.” They argue that if consciousness depends on biological-style computation, “it may require biological-style computational organization, even if it is implemented in new substrates.” Their paper’s Section 4 provides a detailed roadmap of neuromorphic, fluidic, and iontronic substrates that might approximate the requisite computational organization, noting that BrainScaleS-2 narrows the gap between digital simulation and biological computation but does not yet meet their tripartite criteria, and that fluidic memristors (Xiong et al. 2023) represent a more promising direction.

The constraint-based synthesis sharpens this position by reframing the inquiry: at what scale, if any, does organizational closure obtain in the coupled system of human users, neural networks, server infrastructure, training data, and the institutions that maintain and deploy them?

The question has substance. Individual LLMs do not maintain their own computational constraints; their weights are frozen after training. They do not perform thermodynamic work to construct their own architecture. They do not exhibit dynamico-structural co-determination. By the criteria developed here, they are organizationally open, lacking the necessary conditions for consciousness because their computation is decoupled from organized thermodynamic work at the substrate level, regardless of whether the substrate is silicon or carbon.

But the larger system, comprising humans who generate training data, engineers who modify architectures, economic institutions that fund computation, power grids that supply energy, and the LLMs themselves, does exhibit some features of organizational closure. The outputs of the system (generated text, code, decisions) feed back into training data and deployment decisions, which modify the system’s future behavior. The constraint structure evolves. This raises the genuinely novel question of whether, at some scale of organization not yet achieved, such a system could cross the threshold into organizational closure with all three of Milinkovic and Aru’s properties: hybrid dynamics across the human-digital interface, scale inseparability across institutional, cognitive, and computational levels, and metabolic embedding through the energetic costs of maintaining the entire infrastructure.

The framework does not answer this question but specifies what answering it would require: demonstrating that the coupled system maintains its own constraints through its own thermodynamic activity, that perturbation to any scale propagates as a threat or opportunity for the whole, and that the system operates far enough from equilibrium to generate the kind of perspective that organizational closure produces when maintained at thermodynamic cost. Berent and Sansiveri (2024, Open Mind, 8, 84–101, DOI: 10.1162/opmi_a_00120) found that large language models exhibit innate-seeming dualist biases, attributing disembodied qualities to mental states. This tells us something about the structure of human language, which is the LLMs’ training environment rather than about the LLMs’ own consciousness. The framework developed here provides precise criteria for distinguishing the two cases.

The implications of ontic structural realism, as developed by Ladyman and Ross (2007, Every Thing Must Go, Oxford University Press), are worth noting in this context. If structure is ontologically fundamental, if there truly are “no things, only structure,” then consciousness becomes a question about which structures, under which conditions, generate perspective The organizational closure framework answers: the structure of constraint closure maintained at thermodynamic cost, with cross-scale nesting and autocatalytic coupling. This structure could in principle be instantiated in non-biological substrates, but it cannot be instantiated without substrates that support genuine thermodynamic work, genuine constraint maintenance, and genuine dynamico-structural co-determination. Carbon holds no special privilege here. Physics does.

Carlo Rovelli’s relational quantum mechanics (1996, International Journal of Theoretical Physics, 35(8), 1637–1678, DOI: 10.1007/BF02302261) offers a complementary perspective: if quantum states are relative to observers, and observers are systems that can register and use correlations, then the organizational closure framework specifies what kind of physical system constitutes an observer. Not every physical interaction constitutes observation; observation requires a system whose organizational closure allows it to integrate information across scales and maintain that information against thermodynamic dissipation. The thermodynamic cost of maintaining organizational closure is, in this view, the thermodynamic cost of being an observer, of having a perspective from which quantum mechanical correlations are defined.

11. What the constraint bridge actually builds

This article has argued that Milinkovic and Aru’s three necessary conditions for biological consciousness, when formally connected to Montévil and Mossio’s organizational closure framework, Hordijk and Steel’s RAF theory, and the thermodynamics of Landauer, Prigogine (Nobel Prize in Chemistry, 1977), and their successors, produce a framework that is both more rigorous and more parsimonious than its competitors.

The framework explains why consciousness correlates with far-from-equilibrium thermodynamics: because organizational closure requires continuous thermodynamic work. It explains why consciousness requires cross-scale coupling: because constraint closure necessarily spans scales, with constraints at each level maintained by constraints at other levels. It explains why metabolism is constitutive rather than enabling: because the Landauer principle guarantees that irreversible computation is thermodynamic work, and biological computation is constitutively irreversible. It explains why the nervous system generates perspective: because a sufficiently complex organizationally closed system with regulatory constraints, cross-scale nesting, and autocatalytic metabolic-computational coupling, maintained at thermodynamic cost against perturbation, necessarily defines a boundary, a domain of concern, and a relationship to perturbation that constitutes what “having a perspective” means.

The framework specifies physical conditions under which consciousness arises, conditions that are testable, falsifiable, and grounded in established physics and biology. What it does not require is any appeal to non-physical spaces, ingressing patterns, or cognitive vocabulary applied to systems lacking organizational closure. Levin’s Platonic morphospace fares worse than unnecessary: path-dependent empirical results from his own laboratory actively contradict it. And the cognitive vocabulary TAME applies to cells and tissues obscures, rather than illuminates, the constraint dynamics doing the actual explanatory work.

The autocatalytic structure of the metabolism-cognition cycle, formalized through RAF theory, represents what may be the most important connection identified here, and one absent from both literatures. If neural computation is an autocatalytic process, with computational constraints maintained by metabolic processes that are themselves organized by computational constraints, then the RAF formalism provides the mathematical tools to test this hypothesis rigorously: detect the autocatalytic closure, decompose it into irreducible sub-RAFs, and correlate the integrity of the closure with the presence and character of conscious experience.

Ilya Prigogine observed that dissipative structures exist only in conjunction with their environment; they are processes, not things. The same is true of conscious systems, on the account offered here. Consciousness, on this account, names something a brain does rather than something it has: the process of maintaining organizational closure at thermodynamic cost, across scales, through autocatalytic metabolic-computational coupling, against the relentless tendency of the second law to dissipate every constraint into thermal equilibrium. When that process stops, the perspective stops with it. When the process is disrupted at a particular scale, the perspective is disrupted at that scale.

The constraint bridge connects Milinkovic and Aru’s precise characterization of biological computation to the formal frameworks that can make it mathematically tractable. It connects enactivism’s long-standing insight about the continuity of life and mind (Thompson, 2007) to the specific organizational principles that distinguish minded systems from mere self-maintaining ones. It connects the thermodynamics of consciousness (Kringelbach, Sanz Perl, and Deco, 2024; Monti, Sanz Perl, et al., 2025; Berjaga-Buisan et al., 2025) to the organizational theory that explains why thermodynamic measures track consciousness in the first place. All of this stays within the physical world, requires no non-physical causation, attributes no goals to cells or intelligence to bioelectric fields, and provides no rhetorical ammunition for those who would exploit scientific ambiguity for anti-scientific purposes. The constraints are physical. The closure is physical. The perspective is physical. That is enough.

12. What Western scholarship rediscovered and what it still lacks

The framework developed in this article draws on Montévil, Mossio, Kauffman, Hordijk, Steel, Prigogine, Landauer, Varela, Thompson, and their successors to arrive at a set of propositions: that reality is fundamentally processual, that entities are crystallized relations, that constraint closure generates identity, that perspective arises from the thermodynamic cost of self-maintenance, and that no level of organization is privileged over any other. Every one of these propositions has been operational in Indigenous knowledge systems for millennia before Western process philosophy existed as a named tradition.

The point of registering this is not decorative. It bears on the epistemic validity of the framework itself. When independently derived systems of knowledge converge on the same structural claims, William Whewell’s consilience of inductions applies: the convergence constitutes evidence that the claims track something real. If organizational closure, relational ontology, and process metaphysics were idiosyncrasies of late-twentieth-century European theorizing, their independent emergence in Blackfoot, Potawatomi, Tewa, and Aboriginal Australian knowledge traditions across tens of thousands of years would be inexplicable. The convergence is evidence.

Leroy Little Bear (Kainai Nation, Blackfoot Confederacy), founding director of the Harvard University Native American Program and professor emeritus at the University of Lethbridge, has articulated Blackfoot metaphysics across four decades of scholarship and public engagement. In “Jagged Worldviews Colliding” (in Battiste, ed., Reclaiming Indigenous Voice and Vision, pp. 77-85, UBC Press, 2000), and in lectures subsequently documented in academic and public settings, Little Bear identifies what he calls the three tenets of the native paradigm: constant flux (everything is in constant motion), energy waves (everything in creation consists of energy waves, which are the spirit), and universal animacy (everything is animate, imbued with spirit). Existence, in Blackfoot philosophy, is a web of relationships requiring constant renewal and maintenance. Space rather than time serves as the primary referent. The Blackfoot language is verb-dominant, encoding the world as process and relation rather than substance and property.

The structural parallel to the framework developed in this article is worth specifying precisely. Little Bear’s “constant flux” corresponds to the process ontology underlying organizational closure: things are stabilized patterns within ongoing thermodynamic flows, not static entities that happen to change. His “energy waves” correspond to the structural realism of Ladyman and Ross, in which what exists fundamentally is structure and pattern, not substance. His “universal animacy” corresponds to the article’s claim that organizational closure generates perspective at every scale where it obtains: if everything participates in relational constraint maintenance, then perspective is a matter of degree, distributed according to the complexity and depth of the closure. And his “web of relationships” maps onto the ontic structural realist position that relations are ontologically primary, with things being derivative. Blackfoot metaphysics arrived at these positions through tens of thousands of years of sustained empirical engagement with the land, not through the axiomatic method, and it did so without requiring the substance ontology that Western philosophy spent three centuries trying to escape.

Robin Wall Kimmerer (Citizen Potawatomi Nation), SUNY Distinguished Teaching Professor of Environmental Biology and founder of the Center for Native Peoples and the Environment, provides a different entry point that is especially relevant for the article’s account of perspective. In Braiding Sweetgrass: Indigenous Wisdom, Scientific Knowledge and the Teachings of Plants (Milkweed Editions, 2013), Kimmerer describes what she calls the “grammar of animacy” embedded in the Potawatomi language. Where English is roughly 70% nouns, Potawatomi reverses this ratio: approximately 70% of its words are verbs. A bay is not a noun in Potawatomi; it is a verb, wiikwegamaa, “to be a bay,” because the living water has chosen, for now, to shelter between these shores. Rocks, mountains, water, fire, and places are all grammatically animate. The linguistic structure encodes a relational, processual ontology in which every entity participates in ongoing becoming.

This has direct relevance to the article’s argument about nominalization as an epistemic trap. When the article warns that nominalization generates autophagy handles, nouns that “eat themselves” by reifying processes into static entities, Kimmerer’s Potawatomi provides empirical evidence from an independently developed linguistic tradition. A language in which streams are verbs and bays are activities of water does not produce the kind of substance metaphysics that leads to Platonic morphospace. It produces precisely the relational, process-oriented ontology that the constraint-based synthesis requires. Kimmerer is both a credentialed plant ecologist and a knowledge keeper in the Potawatomi tradition, and her work demonstrates that the relational ontology underlying organizational closure is independently discoverable from within radically different epistemic starting points.

Tyson Yunkaporta (Apalech Clan) develops complementary arguments in Sand Talk: How Indigenous Thinking Can Save the World (Text Publishing, 2019). Yunkaporta’s critique of what he calls “point-thinking,” the Western tendency to reduce relational processes to isolated data points, maps cleanly onto the ontic structural realist rejection of substance metaphysics. His insistence that knowledge lives in the edges and relations between things, not in the things themselves, restates in different vocabulary the Ladyman-Ross position that the article invokes as its metaphysical grounding. More specifically, Yunkaporta’s account of knowledge transmission through Country, in which walking, singing, and relating to landscape constitutes the activation of knowledge rather than its retrieval from storage, provides independent empirical grounding for the article’s claim that memory is constraint maintenance rather than information warehousing. The Western neuroscience of memory is only now arriving at this position through reconsolidation theory and predictive processing frameworks. Aboriginal Australian knowledge systems have operated from it for at least 65,000 years.

Lynne Kelly’s academic monograph Knowledge and Power in Prehistoric Societies: Orality, Memory, and the Transmission of Culture (Cambridge University Press, 2015) and its accessible companion The Memory Code (Allen & Unwin, 2016) provide the empirical bridge between these Indigenous knowledge systems and the cognitive science invoked in this article. Kelly, an Honorary Research Associate at La Trobe University, demonstrates through extensive cross-cultural and archaeological evidence that Songlines encode ecological, botanical, astronomical, and navigational knowledge across millennia, and that this knowledge survives precisely because it is embedded in physical constraint paths through Country rather than abstracted into propositional storage. For the article’s account of neural computation as autocatalytic constraint maintenance, Kelly’s work is critical: it shows that the most durable human knowledge systems are those that embed information in relational constraint structures tied to physical landscape, not those that abstract it into symbolic representations separated from their material substrate. The parallels to the article’s argument that biological computation is constituted by thermodynamic work in the substrate, rather than running as abstract algorithm on the substrate, are structural.

Gregory Cajete (Tewa, Santa Clara Pueblo), Professor Emeritus of Education at the University of New Mexico, frames Indigenous knowledge systems explicitly as natural philosophy in Native Science: Natural Laws of Interdependence (Clear Light Publishers, 2000). Cajete’s treatment of what he calls “natural laws of interdependence” articulates, from within Tewa epistemology, a framework in which all entities bear responsibility to and for one another because they are co-creators within a web of mutual constraint. His concept of “place-thought,” thinking inseparably embedded in ecological relationships, provides additional consilience for the organizational closure account of perspective developed in Section 6: perspective arises from the coupling position of an organizationally closed system, and that coupling is always a coupling to a specific place, a specific set of ecological and thermodynamic relationships.

What these traditions contribute that Western scholarship currently lacks is not content so much as architecture. Western process philosophy (Whitehead, Simondon, Deleuze) arrived at relational ontology through critique of substance metaphysics. Indigenous knowledge systems never had the substance metaphysics to critique. They began where Western theory is trying to end up. This means they have had millennia to work out the practical and epistemic consequences of a relational, processual worldview, consequences that Western frameworks are still discovering. When Montévil and Mossio formalize organizational closure, they formalize what Blackfoot metaphysics has practiced as renewal ceremony. When Hordijk and Steel detect autocatalytic sets in metabolic networks, they detect what Kimmerer’s Potawatomi encodes as the grammar of animacy: the recognition that self-maintaining relational systems constitute the fundamental units of the living world. When Barnett and Seth measure dynamical independence across scales, they measure what Yunkaporta describes as the inseparability of knowledge from the relational structure of Country.

The consilience matters for falsification. If the constraint-based synthesis developed in this article were merely an artifact of Western theoretical commitments, its convergence with Indigenous frameworks would be coincidental. But the convergence is too structural, too detailed, and too independently arrived at to be coincidental. It constitutes, in Whewell’s sense, evidence that the framework tracks genuine features of reality. The same conclusion that Ladyman and Ross reach through philosophy of physics, that Little Bear reaches through Blackfoot metaphysics, that Kimmerer reaches through Potawatomi ecology, and that Montévil and Mossio reach through theoretical biology, is the same conclusion: what exists are relations under constraint, maintained at thermodynamic cost, generating perspective at every scale where the closure is complex enough to sustain it.

This article is written on historically significant Anishinaabe land near a Sugarbush Trail in Chesterfield, Michigan. The knowledge systems of the people who have maintained this land for millennia deserve acknowledgment as primary epistemological contributions, not as metaphorical ornaments on a framework built elsewhere.

A Note to Scholars

If you work in theoretical biology, consciousness science, epistimology, thermodynamics, philosophy of mind, artificial intelligence, or any adjacent field and found something here worth engaging with, I’d genuinely like to hear from you. My goal is straightforward: to build a formally rigorous, thermodynamically grounded, falsification-first account of consciousness and biological organization that renders unfalsifiable metaphysical frameworks redundant by outcompeting them on their own explanatory terrain. The synthesis attempted here, connecting biological computationalism, organizational closure, RAF theory, and the thermodynamics of perspective, is the kind of work that benefits from exactly the scrutiny that credentialed specialists can apply. If something here is wrong, imprecise, or missing a better formalization, that is information I want.

If your lab or research group could use someone to red-team theoretical frameworks, stress-test metaphysical assumptions against falsification criteria, or provide naturalistic thermodynamic grounding for experimental work that keeps drifting toward unfalsifiable interpretation, that is exactly the kind of collaboration I’m built for. The RCF methodology was designed precisely for that function: finding where the explanatory load is being carried by physics and where it’s quietly been handed off to promissory notes.

If any of this overlaps with your work, contradicts something you’ve established, or points toward a collaboration worth exploring, the contact form is here: Nathan Sweet – Contact Me. Or just click here to drop me an email.

Related Articles

Consciousness Naturalized: A Falsifiable Substrate-Agnostic Theory | Chalmers’ Own Later Work Dissolves the Hard Problem | Levin’s Platonism: Unfalsifiable Metaphysics Contradicted by His Own Lab | “Mind Everywhere” by Levin & Resnik | The Chladni Plate Alternative: Douglas Brash’s Constraint Framework vs. Platonic Morphogenesis | Gordana Dodig-Crnkovic’s Naturalism Dissolves Levin’s Transcendent Morphospace From Within | Brian Cheung’s Convergence Research Inadvertently Exposes the Motte-Bailey Problem | “Differences That Make a Difference” | Cognition All the Way Down (the Drain) 2.0: Levin’s Thermodynamic Consciousness Shell Game | The Architecture of Everything, Part I: Constraint Propagation Across All Scales | Memory Is Not Storage: Why Everything Western Science Thinks It Knows About Remembering Is Wrong

References

Barnett, L., & Seth, A. K. (2023). Dynamical independence: Discovering emergent macroscopic processes in complex dynamical systems. Physical Review E, 108(1), 014304. https://doi.org/10.1103/PhysRevE.108.014304

Berent, I., & Sansiveri, M. (2024). Davinci the dualist: The mind-body divide in large language models and in human learners. Open Mind, 8, 84–101. https://doi.org/10.1162/opmi_a_00120

Berjaga-Buisan, T., Monti, J. M., Cortada, M., Colombo, M. A., Geli, S. M., Gaglioti, G., Sarasso, S., Kringelbach, M. L., Corbetta, M., Sanchez-Vives, M. V., Massimini, M., Sanz Perl, Y., & Deco, G. (2025). Thermodynamics of consciousness: A non-invasive perturbational framework. bioRxiv. https://doi.org/10.64898/2025.12.09.691422

Bérut, A., Arakelyan, A., Petrosyan, A., Ciliberto, S., Dillenschneider, R., & Lutz, E. (2012). Experimental verification of Landauer’s principle linking information and thermodynamics. Nature, 483(7388), 187–189. https://doi.org/10.1038/nature10872

Bich, L., Mossio, M., Ruiz-Mirazo, K., & Moreno, A. (2016). Biological regulation: Controlling the system from within. Biology & Philosophy, 31(2), 237–265. https://doi.org/10.1007/s10539-015-9497-8

Cajete, G. (2000). Native science: Natural laws of interdependence. Clear Light Publishers. ISBN 9781574160413.

Callen, H. B., & Welton, T. A. (1951). Irreversibility and generalized noise. Physical Review, 83(1), 34–40. https://doi.org/10.1103/PhysRev.83.34

Casali, A. G., Gosseries, O., Rosanova, M., Boly, M., Sarasso, S., Casali, K. R., Casarotto, S., Bruno, M.-A., Laureys, S., Tononi, G., & Massimini, M. (2013). A theoretically based index of consciousness independent of sensory processing and behavior. Science Translational Medicine, 5(198), 198ra105. https://doi.org/10.1126/scitranslmed.3006294

Crooks, G. E. (1999). Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences. Physical Review E, 60(3), 2721–2726. https://doi.org/10.1103/PhysRevE.60.2721

Cusimano, S., & Sterner, B. (2020). The objectivity of organizational functions. Acta Biotheoretica, 68(2), 253–269. https://doi.org/10.1007/s10441-019-09365-9

Deacon, T. W. (2012). Incomplete nature: How mind emerged from matter. W. W. Norton.

Di Paolo, E. A. (2005). Autopoiesis, adaptivity, teleology, agency. Phenomenology and the Cognitive Sciences, 4(4), 429–452. https://doi.org/10.1007/s11097-005-9002-y

Di Paolo, E. A., & Thompson, E. (2014). The enactive approach. In L. Shapiro (Ed.), The Routledge handbook of embodied cognition (pp. 68–78). Routledge.

Durant, F., Morokuma, J., Fields, C., Williams, K., Adams, D. S., & Levin, M. (2017). Long-term, stochastic editing of regenerative anatomy via targeting endogenous bioelectric gradients. Biophysical Journal, 112(10), 2231–2243. https://doi.org/10.1016/j.bpj.2017.04.011

Garson, J. (2017). Against organizational functions. Philosophy of Science, 84(5), 1093–1103. https://doi.org/10.1086/694009

Gould, S. J., & Lewontin, R. C. (1979). The spandrels of San Marco and the Panglossian paradigm: A critique of the adaptationist programme. Proceedings of the Royal Society of London. Series B: Biological Sciences, 205(1161), 581–598. https://doi.org/10.1098/rspb.1979.0086

Hartl, B., Zhang, Y., Hazan, H., & Levin, M. (2026). Heuristically adaptive diffusion-model evolutionary strategy. Advanced Science. https://doi.org/10.1002/advs.202511537

Hartl, B., & Levin, M. (2025). What does evolution make? Learning in living lineages and machines. Trends in Genetics, 41(6), 480–496. https://doi.org/10.1016/j.tig.2025.04.002

Hordijk, W., Hein, J., & Steel, M. (2010). Autocatalytic sets and the origin of life. Entropy, 12(7), 1733–1742. https://doi.org/10.3390/e12071733

Hordijk, W., Kauffman, S. A., & Steel, M. (2011). Required levels of catalysis for emergence of autocatalytic sets in models of chemical reaction systems. International Journal of Molecular Sciences, 12(5), 3085–3101. https://doi.org/10.3390/ijms12053085

Hordijk, W., & Steel, M. (2004). Detecting autocatalytic, self-sustaining sets in chemical reaction systems. Journal of Theoretical Biology, 227(4), 451–461. https://doi.org/10.1016/j.jtbi.2003.11.020

James, W. (1890). The principles of psychology (2 vols.). Henry Holt.

Jarzynski, C. (1997). Nonequilibrium equality for free energy differences. Physical Review Letters, 78(14), 2690–2693. https://doi.org/10.1103/PhysRevLett.78.2690

Karbowski, J. (2024). Information thermodynamics: From physics to neuroscience. Entropy, 26(9), 779. https://doi.org/10.3390/e26090779

Kauffman, S. A. (2000). Investigations. Oxford University Press.

Kelly, L. (2015). Knowledge and power in prehistoric societies: Orality, memory, and the transmission of culture. Cambridge University Press. https://doi.org/10.1017/CBO9781316106839

Kelly, L. (2016). The memory code: The traditional Aboriginal memory technique that unlocks the secrets of Stonehenge, Easter Island and ancient monuments the world over. Allen & Unwin. ISBN 9781760291327.

Kimmerer, R. W. (2013). Braiding sweetgrass: Indigenous wisdom, scientific knowledge and the teachings of plants. Milkweed Editions. ISBN 9781571313560.

Kimura, M. (1983). The neutral theory of molecular evolution. Cambridge University Press. https://doi.org/10.1017/CBO9780511623486

Kringelbach, M. L., Sanz Perl, Y., & Deco, G. (2024). The thermodynamics of mind. Trends in Cognitive Sciences, 28(6), 568–581. https://doi.org/10.1016/j.tics.2024.03.009

Ladyman, J., & Ross, D. (with Spurrett, D., & Collier, J.). (2007). Every thing must go: Metaphysics naturalized. Oxford University Press. https://doi.org/10.1093/acprof:oso/9780199276196.001.0001

Landauer, R. (1961). Irreversibility and heat generation in the computing process. IBM Journal of Research and Development, 5(3), 183–191. https://doi.org/10.1147/rd.53.0183

Lehman, N., & Kauffman, S. A. (2021). Constraint closure drove major transitions in the origins of life. Entropy, 23(1), 105. https://doi.org/10.3390/e23010105

Levin, M. (2022). Technological approach to mind everywhere: An experimentally-grounded framework for understanding diverse bodies and minds. Frontiers in Systems Neuroscience, 16, 768201. https://doi.org/10.3389/fnsys.2022.768201

Levin, M. (2023). Darwin’s agential materials: Evolutionary implications of multiscale competency in developmental biology. Cellular and Molecular Life Sciences, 80(6), 142. https://doi.org/10.1007/s00018-023-04790-z

Levin, M., & Resnik, D. B. (2026). Mind everywhere: A framework for conceptualizing goal-directedness in biology and other domains—Part one. Biological Theory. https://doi.org/10.1007/s13752-025-00523-6

Little Bear, L. (2000). Jagged worldviews colliding. In M. Battiste (Ed.), Reclaiming Indigenous voice and vision (pp. 77–85). UBC Press.

Luhmann, N. (1995). Social systems (J. Bednarz Jr. & D. Baecker, Trans.). Stanford University Press.

Lynch, M. (2007). The origins of genome architecture. Sinauer Associates.

Milinkovic, B., & Aru, J. (2026). On biological and artificial consciousness: A case for biological computationalism. Neuroscience & Biobehavioral Reviews, 181, 106524. https://doi.org/10.1016/j.neubiorev.2025.106524

Milinkovic, B., Barnett, L., Carter, O., Seth, A. K., & Andrillon, T. (2025). Capturing the emergent dynamical structure in biophysical neural models. PLoS Computational Biology, 21(5), e1012572. https://doi.org/10.1371/journal.pcbi.1012572

Mitchell, K. J., & Cheney, N. (2025). The genomic code: The genome instantiates a generative model of the organism. Trends in Genetics, 41(6), 462–479. https://doi.org/10.1016/j.tig.2025.01.008

Monti, J. M., Sanz Perl, Y., Tagliazucchi, E., Kringelbach, M. L., & Deco, G. (2025). Fluctuation-dissipation theorem and the discovery of distinctive off-equilibrium signatures of brain states. Physical Review Research, 7, 013301. https://doi.org/10.1103/PhysRevResearch.7.013301

Montévil, M., & Mossio, M. (2015). Biological organisation as closure of constraints. Journal of Theoretical Biology, 372, 179–191. https://doi.org/10.1016/j.jtbi.2015.02.029

Mouret, J.-B., & Clune, J. (2015). Illuminating search spaces by mapping elites. arXiv:1504.04909. https://arxiv.org/abs/1504.04909

Rovelli, C. (1996). Relational quantum mechanics. International Journal of Theoretical Physics, 35(8), 1637–1678. https://doi.org/10.1007/BF02302261

Sohl-Dickstein, J., Weiss, E. A., Maheswaranathan, N., & Ganguli, S. (2015). Deep unsupervised learning using nonequilibrium thermodynamics. arXiv:1503.03585. https://arxiv.org/abs/1503.03585

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

Tkachenko, A. V. (2025). Thermodynamic bounds on energy use in quasi-static deep neural networks. arXiv:2503.09980. https://arxiv.org/abs/2503.09980

Varela, F. J., Thompson, E., & Rosch, E. (1991). The embodied mind: Cognitive science and human experience. MIT Press. https://doi.org/10.7551/mitpress/6730.001.0001

Vuong, A. B., McCann, M. T., Santos, J. E., & Lin, Y. T. (2025). Are we really learning the score function? Reinterpreting diffusion models through Wasserstein gradient flow matching. arXiv:2509.00336. https://arxiv.org/abs/2509.00336

Wang, Z., Luo, C., et al. (2025). Causal emergence of consciousness through learned multiscale neural dynamics in mice. arXiv:2509.10891. https://arxiv.org/abs/2509.10891

Watson, R. A., & Szathmáry, E. (2016). How can evolution learn? Trends in Ecology & Evolution, 31(2), 147–157. https://doi.org/10.1016/j.tree.2015.11.009

Wong, A., de Nobel, J., Bäck, T., Plaat, A., & Kononova, A. V. (2024). Solving deep reinforcement learning tasks with evolution strategies and linear policy networks. arXiv:2402.06912. https://arxiv.org/abs/2402.06912

Yunkaporta, T. (2019). Sand talk: How Indigenous thinking can save the world. Text Publishing. ISBN 9781925773996.

Zhang, Y., Hartl, B., Hazan, H., & Levin, M. (2025). Diffusion models are evolutionary algorithms. Poster presented at International Conference on Learning Representations (ICLR 2025). arXiv:2410.02543. https://arxiv.org/abs/2410.02543