A difference only “makes a difference” if it can persist long enough to constrain what comes next. Everything else is noise that thermodynamics erases.
Prologue: The Question That Would Not Stay Answered
In 1979, Gregory Bateson posed a question that would haunt systems theory, cybernetics, and philosophy of mind for nearly half a century: What is the pattern which connects all patterns? The question appeared in Mind and Nature: A Necessary Unity, his final major work. Bateson circled the answer beautifully, suggesting it was “a metapattern,” that the pattern which connects is itself a pattern of patterns. He offered his famous formulation, “a difference which makes a difference,” as the elementary unit of information, the minimal act of distinction that allows anything to be known at all. But he never specified the constraint structure that would make such an answer operational. He died three years later, the question still open.
The candidates accumulated across decades. Mind, said Bateson himself, tentatively. Information, said the cyberneticians following Shannon and Wiener. Recursion, proposed Douglas Hofstadter in Gödel, Escher, Bach. Process, argued the Whiteheadians. Mathematics, claimed Max Tegmark. Consciousness, insisted the panpsychists. Self-organization, offered Stuart Kauffman. Free energy minimization, suggested Karl Friston. The list is effectively unbounded. Any concept claiming universality becomes a candidate. And that is precisely the problem: without criteria for distinguishing correct answers from merely plausible ones, the question generates discourse without resolution.
This essay reports a different approach to Bateson’s question, one grounded in thermodynamics, constraint theory, and falsification-first methodology. The answer that emerges is not a new primitive to add to the ontological inventory. It is a selection principle that explains why any pattern persists at all: constraint satisfaction under thermodynamic bounds. Differences that persist are differences that make a difference; if they did not constrain what comes next, they would be erased by noise, dissipation, or indifference. The pattern which connects is not information, not mind, not structure, not process. It is the invariant condition under which any of those could exist long enough to be observed.
This principle does not compete with other frameworks. It subsumes them. It is not an ontology describing what exists, not an epistemology describing how we know, not a cosmology describing why there is something rather than nothing. It is a selection rule on what can remain available to explanation at all. And it is falsifiable: find a pattern that persists without thermodynamic cost, find a connection mechanism that does not reduce to constraint satisfaction, and the principle fails.
What follows is an attempt to make that claim rigorous, to show why recent metaphysical interpretations of biological efficiency fail where constraint-based explanations succeed, and to demonstrate that the real work is already being done by frameworks that are experimentally grounded, mechanistically specified, and willing to lose. The question of what connects is also, it turns out, the question of what survives.
Part I: The Structural Problem (Efficiency Is Not a Mechanism)
The Recurring Argument Pattern
There is a recurring temptation in the sciences of mind, life, and complexity: when confronted with systems that are efficient, robust, adaptive, or astonishingly competent, we reach for a special noun. Intelligence. Cognition. Purpose. Agency. Consciousness. The word changes, but the move stays the same. We observe an impressive outcome, compare it to a weak or inappropriate null model, declare the gap too large to be bridged by ordinary mechanisms, and then install a new primitive to do the explanatory work. The noun becomes the answer. The question quietly disappears.
Consider the structure more carefully. We observe a biological system solving a problem efficiently: regeneration, error correction, morphological repair, goal-directed behavior. We compare that performance to a null model, typically random search or maximal entropy sampling. The gap is large. Sometimes astronomically large. We then infer that something special must be doing the work. Intelligence. Cognition. A problem-space being searched. A goal being represented. This inferential move should immediately set off alarms, because it conflates two fundamentally different claims: that the system is impressive, and that a cognitive primitive explains why.
Efficiency is an output, not a mechanism. It is the result of a process, not an explanation of one. If we define intelligence as efficient problem-solving, measure efficiency, and conclude intelligence, we have not discovered anything. We have defined our conclusion into existence. The circularity is not subtle. It is the same logical structure that plagued “irreducible complexity” arguments in intelligent design discourse: define a property such that it cannot arise gradually, find systems matching the definition, and announce that design is required. The conclusion was baked into the premise. The only difference is vocabulary.
The null model matters enormously here. Comparing an evolved, developmentally constrained, historically filtered system to random assembly is not illuminating; it is misleading. Evolution is the search process. Development is constraint propagation. The present phenotype embodies compressed history: billions of years of selection, developmental canalization, and constraint satisfaction have shaped every molecular interaction. Measuring the endpoint against an incoherent starting distribution and calling the gap “cognition” mistakes the record of selection for a new causal ingredient.
William Wimsatt’s work on robustness analysis (DOI: 10.1086/288763) provides a crucial methodological corrective. Wimsatt showed that what we take to be “real” in scientific practice is typically what remains invariant across multiple independent methods of detection, derivation, or measurement. Robustness is not a property of things; it is a property of how things survive different forms of probing. Applied to the efficiency-cognition inference, this means asking: does the “intelligence” we detect remain invariant when we change our detection method? Or does it dissolve when we stop comparing to inappropriate null models? The answer, repeatedly, is the latter. What remains robust is constraint satisfaction. What dissolves is the cognitive gloss.
Why This Matters Beyond Academic Dispute
This is not a merely academic dispute about terminology. Unfalsifiable frameworks do harm, even when wrapped around real data. When a theory cannot fail, it cannot learn. When every possible outcome is retroactively compatible with “basal cognition,” the concept ceases to constrain explanation. Worse, it creates an attack surface for genuinely anti-scientific reasoning by normalizing inference from efficiency gaps to special causal substances. That is not a hypothetical risk. It is a pattern that has already played out repeatedly in public discourse, from vitalism to intelligent design to the current moment.
The stakes are epistemic and ethical. Epistemic, because frameworks that insulate themselves from falsification do not accumulate knowledge; they accumulate interpretations. Ethical, because generations of scientists, thinkers, and experimenters spent their lifetimes building constraint-based frameworks precisely because intuition and metaphor failed. Every advance in thermodynamics, information theory, developmental biology, and neuroscience came from refusing to stop at vibes. People froze, starved, died of infection, got exiled, or were ignored for insisting that explanations cash out in mechanisms, constraints, and testable consequences. To install an unfalsifiable primitive after that labor is not humility. It is a quiet betrayal of the discipline that made knowledge possible.
Part II: Information Is Physical (The Thermodynamic Ground Floor)
Landauer’s Principle and the Cost of Distinction
The foundation of constraint-based explanation is not philosophical preference but physical law. In 1961, Rolf Landauer published a paper that would fundamentally reshape our understanding of the relationship between information and physics: “Irreversibility and Heat Generation in the Computing Process” (IBM Journal of Research and Development, DOI: 10.1147/rd.53.0183). Landauer demonstrated that logically irreversible operations, operations that destroy information, such as erasing a bit, necessarily produce heat. The minimum energy dissipated per bit erased is kT ln 2, where k is Boltzmann’s constant and T is temperature. At room temperature (300K), this works out to approximately 2.87 × 10⁻²¹ joules per bit.
This is not an engineering limitation to be overcome with better technology. It is a fundamental constraint arising from the second law of thermodynamics. Information is physical. Maintaining a distinction between 0 and 1 requires physical substrate. Erasing that distinction releases energy. There is no free information, no costless computation, no disembodied memory. Any explanation that treats information or cognition as causally efficacious while exempting it from energetic accounting is incomplete by definition, not wrong in some subtle philosophical sense, but literally missing the physics.
Charles Bennett extended Landauer’s work in his 1973 paper “Logical Reversibility of Computation” (IBM Journal of Research and Development, DOI: 10.1147/rd.176.0525). Bennett demonstrated that while computation itself can be thermodynamically reversible (dissipating arbitrarily close to zero energy per step), maintaining a result against thermal noise cannot be. The essential irreversible step is not measurement or computation but erasure, the forgetting of intermediate states. This insight resolved the century-old puzzle of Maxwell’s Demon: the demon appears to violate the second law by selectively allowing fast molecules through a gate, but the demon must remember which molecules it allowed. Eventually, the demon’s memory fills up and must be erased, and that erasure produces exactly the entropy the demon seemed to avoid.
In 2012, Bérut and colleagues experimentally verified Landauer’s principle directly (Nature, DOI: 10.1038/nature10872). Using a colloidal particle trapped in a double-well potential created by optical tweezers, they measured the heat dissipated during bit erasure and found it saturated at the Landauer bound in the limit of slow erasure cycles. The experimental precision was remarkable: the measured heat dissipation matched the theoretical minimum to within a few percent. This is not a thought experiment. It is laboratory physics. Information erasure costs energy, and the cost has been measured.
The implications ripple outward. Every distinction maintained, every boundary, every memory, every pattern that persists, pays thermodynamic rent. Bennett’s 1982 review (International Journal of Theoretical Physics, DOI: 10.1007/BF02084158) surveyed biological molecular machines and found they typically dissipate 20-100 kT per step, far above the Landauer minimum but still operating within thermodynamic constraints. The gap between biological dissipation and the theoretical minimum represents room for optimization, not exemption from physics. Life is not transcending thermodynamics; it is navigating within thermodynamic constraints with extraordinary skill accumulated over billions of years of selection.
The Death of Disembodied Information
Why does this matter for the cognition debate? Because it eliminates an entire class of explanatory moves. Once we recognize that information processing has irreducible physical cost, we cannot invoke “information” as a causal agent without specifying where the energy comes from and where the entropy goes. We cannot appeal to “patterns” causing effects without explaining how those patterns are maintained against thermal degradation. We cannot treat cognition as a primitive without asking what physical substrate implements it and what thermodynamic transactions sustain it.
The Jarzynski equality (Physical Review Letters, 1997, DOI: 10.1103/PhysRevLett.78.2690) and subsequent work on fluctuation theorems have made this even more precise. Jarzynski showed that the relationship between work and free energy holds not just on average but through a precise equality involving exponentials of work: ⟨e^(-βW)⟩ = e^(-βΔF). This allows extraction of equilibrium free energy differences from non-equilibrium measurements and provides a rigorous framework for understanding how far systems can deviate from minimum dissipation. Crooks’ fluctuation theorem (Physical Review E, 1999, DOI: 10.1103/PhysRevE.60.2721) extended this, relating the probability of observing a given entropy production to the probability of observing its time-reverse. Seifert’s comprehensive review (Reports on Progress in Physics, 2012, DOI: 10.1088/0034-4885/75/12/126001) synthesized decades of work into what is now called stochastic thermodynamics, a complete framework for understanding thermodynamic processes at the scale where fluctuations matter.
The upshot is that physics provides not just constraints but accounting. We can audit any proposed mechanism for thermodynamic consistency. We can calculate the minimum cost of maintaining distinctions. We can identify where dissipation occurs and measure whether it matches predictions. This converts metaphysical speculation into empirical science. The question is no longer “does cognition exist?” but “does the proposed cognitive process pay its thermodynamic rent?” If the answer is no, the proposal is not merely philosophically suspect; it violates physics.
Part III: Constraint Closure (How Organization Persists Without Goals)
The Montevil-Mossio Framework
If thermodynamics provides the accounting, constraint theory provides the mechanism. The question is not merely how distinctions are maintained but how entire organizations persist, how living systems maintain themselves far from equilibrium, how they reproduce, repair, and adapt, all without invoking goals, purposes, or cognitive primitives. The answer emerging from theoretical biology is constraint closure.
Maël Montévil and Matteo Mossio’s 2015 paper “Biological organisation as closure of constraints” (Journal of Theoretical Biology, DOI: 10.1016/j.jtbi.2015.02.029) provides the most rigorous formalization. They distinguish between two fundamentally different causal roles: processes and constraints. Processes are thermodynamic changes, chemical reactions, flows, transformations. Constraints are entities that act upon processes while remaining conserved (at the relevant time scale) by those processes. A catalyst enables a reaction without being consumed by it. A membrane channels flow without being destroyed by it. A developmental boundary shapes morphogenesis without being erased by it.
The crucial insight is that biological organization arises when constraints exhibit closure: “A set of constraints achieves closure if, and only if, in the set, each constraint realises mutual dependence such that they both depend on and contribute to maintaining each other.” This is not circular causation in the vicious sense. It is organizational closure, the constraints that enable the system’s behavior are themselves produced and maintained by that behavior. Break the closure, and the system collapses. Maintain it, and robust function emerges without any need for external supervision, goals, or cognitive direction.
Moreno and Mossio’s book Biological Autonomy (Springer, 2015, DOI: 10.1007/978-94-017-9837-2) develops this framework into a full theory of biological organization. They argue that autonomy, the capacity of living systems to self-maintain, arises precisely from constraint closure. An autonomous system is not merely self-organizing (many physical systems self-organize). It is organizationally closed: its constraints depend on one another in a way that constitutes a self-maintaining network. This is what distinguishes a cell from a crystal, an organism from a whirlpool. Both involve self-organization; only the former involves constraint closure.
The connection to thermodynamics is immediate. Maintaining constraint closure requires continuous energy throughput. Living systems are dissipative structures in Prigogine’s sense (DOI: 10.1007/978-3-663-01511-6_10), they maintain organization by exporting entropy to their environment. But they are not merely dissipative structures; they are organizationally closed dissipative structures, which is why they persist, reproduce, and evolve in ways that mere dissipative patterns (like Bénard cells or hurricanes) do not.
RAF Theory and the Origin of Closure
Stuart Kauffman anticipated much of this in his work on autocatalytic sets, beginning with his 1986 paper “Autocatalytic sets of proteins” (Journal of Theoretical Biology, DOI: 10.1016/S0022-5193(86)80047-9). Kauffman asked: how could self-sustaining organization arise in a pre-biotic world without enzymes, without genetic information, without design? His answer was that reaction networks can close on themselves. If molecule A catalyzes the formation of molecule B, and B catalyzes C, and C catalyzes A, the network is autocatalytic; it produces the very catalysts it needs to function.
Wim Hordijk, Mike Steel, and Kauffman formalized this intuition in RAF (Reflexively Autocatalytic Food-generated) theory. Their 2004 paper (Journal of Theoretical Biology, DOI: 10.1016/j.jtbi.2003.11.020) established the mathematical foundations, and subsequent work (DOI: 10.1007/s10441-012-9165-1; DOI: 10.3390/ijms12053085) determined the conditions under which RAFs emerge. The key finding: autocatalytic closure is not rare. With realistic assumptions about catalytic promiscuity, where each molecule catalyzes only 1-2 reactions on average, well within what biochemistry allows, RAFs emerge with high probability once reaction networks reach modest size.
Xavier and colleagues (Proceedings of the Royal Society B, 2020, DOI: 10.1098/rspb.2019.2377) brought this into contact with actual biochemistry. They identified RAF structures within the metabolic networks of ancient anaerobic autotrophs, showing that autocatalytic closure is not merely theoretically possible but actually present in the deep branches of the tree of life. Hordijk’s historical review (Biological Theory, 2019, DOI: 10.1007/s13752-019-00330-w) traces the development of these ideas and their implications: “Alone, each molecular species is dead. Jointly, once catalytic closure among them is achieved, the collective system is alive.”
The implications for the cognition debate are direct. If organizational closure explains how living systems persist, repair, and adapt, then we do not need to invoke cognition as a primitive. Constraint closure is the mechanism. What looks like goal-directed behavior is the statistical signature of systems whose internal constraints maintain themselves more effectively than external perturbations destroy them. Purpose is not an inner property; it is an externally legible outcome of constraint closure under selection.
Part IV: Memory Without Representation (The Embedded Memory Framework)
Attractor Landscapes and Persistent States
If constraint closure explains how organizations persist, what explains memory? The traditional answer involves storage, patterns encoded in stable substrates and retrieved on demand. But this model faces deep problems when applied to biological systems. Neural memory is not like computer memory; synaptic weights change constantly, neurons die and are replaced, and memories persist through upheavals that would destroy any conventional storage system. Morphological memory is even stranger: planarian worms retain body-plan information through complete cellular turnover, and organisms regenerate complex structures without consulting any identifiable blueprint.
The emerging alternative treats memory not as storage but as constraint closure applied to dynamical systems. John Hopfield’s landmark 1982 paper (PNAS, DOI: 10.1073/pnas.79.8.2554) showed that neural networks could store memories as stable fixed points of an energy landscape. Memory is not written into a particular location; it is a basin of attraction, a region of state space that the system tends to enter and remain in. Recall is not retrieval; it is relaxation to an attractor.
This insight extends far beyond neural networks. Miller’s review (F1000Research, 2016, DOI: 10.12688/f1000research.7698.1) surveys how attractor dynamics explain working memory, decision-making, and cognitive flexibility. But the framework applies equally to bioelectric memory. Pezzulo, LaPalme, Durant, and Levin (Philosophical Transactions of the Royal Society B, 2021, DOI: 10.1098/rstb.2019.0765) demonstrated that somatic pattern memories in regenerating organisms are realized through bistability and multistability, not through explicit representation but through the existence of multiple stable states in a constraint landscape.
Law and Levin (Theoretical Biology and Medical Modelling, 2015, DOI: 10.1186/s12976-015-0019-9) modeled bioelectric memory mathematically, showing how resting potential bistability in cells creates memory without any representational machinery. A cell “remembers” its state by persisting in one of two stable voltage configurations. The memory is not stored anywhere; it is the constraint closure of ion channel dynamics. Perturb the cell, and it may shift to the other attractor, a different “memory,” that then persists without further input.
The Durant Experiment: Memory as Constraint, Not Blueprint
The most striking experimental demonstration comes from Durant and colleagues in the Levin lab. Their 2017 paper (Biophysical Journal, DOI: 10.1016/j.bpj.2017.04.011) showed that brief bioelectric manipulation can permanently alter planarian body plans. Expose planarian fragments to gap junction blockers for 48 hours, and approximately 30% regenerate as two-headed worms. Here is the crucial finding: these two-headed planarians remain two-headed through subsequent rounds of regeneration, even though the original perturbation is long gone. The “memory” of being two-headed persists indefinitely.
This result directly contradicts Platonic interpretations of morphogenesis. If organisms were accessing an abstract morphospace or consulting an ideal template, transient perturbations should be corrected; the system should return to the canonical form. Instead, we see path dependence. The constraint landscape was reshaped by the intervention, and the system now relaxes to a different attractor. There is no reference back to an ideal; there is only forward propagation of whatever constraints currently obtain.
Durant and colleagues extended this in 2019 (Biophysical Journal, DOI: 10.1016/j.bpj.2019.01.029), showing how early bioelectric signals establish anterior-posterior polarity, and how altering those signals creates cryptic phenotypes, organisms that appear morphologically normal but harbor altered bioelectric states that manifest in subsequent regeneration. The memory is not in the genes, not in the morphology, but in the voltage pattern, a distributed constraint closure that persists because it is dynamically stable.
The implications are profound. Memory does not require representation. It requires constraint closure that makes certain future states more likely than others. A system “remembers” insofar as its past has narrowed its future possibilities. This is what I have called the Embedded Memory Framework (EMF): memory is the reduction of admissible future states due to past dynamics. No storage required. No retrieval mechanism. Just constraints that close on themselves and persist because they can.
Part V: Structure Without Substance (Why Relations Are Primary)
Ontic Structural Realism
If constraints are the mechanism and closure is the principle, what is the ontology? The answer emerging from philosophy of physics is radical: there are no substances, only structures. Relations are ontologically primary; objects are derivative.
James Ladyman and Don Ross’s Every Thing Must Go: Metaphysics Naturalized (Oxford, 2007, DOI: 10.1093/acprof:oso/9780199276196.001.0001) is the landmark statement of ontic structural realism (OSR). They argue that our best physical theories, quantum mechanics, quantum field theory, general relativity, systematically undermine the notion of objects with intrinsic properties. Quantum particles fail the principle of identity of indiscernibles; spacetime points lack individuality in general relativity; fields are more fundamental than particles in quantum field theory. What remains invariant across theory change is not the inventory of objects but the structure of relations.
The slogan “relations without relata” captures the radical thesis. This is not merely epistemic humility (we can only know structure) but ontological commitment (only structure exists). Ladyman’s original 1998 paper (Studies in History and Philosophy of Science, DOI: 10.1016/S0039-3681(98)80129-5) distinguished epistemic from ontic versions; subsequent work by French and Ladyman (Synthese, 2003, DOI: 10.1023/A:1024156116636) and French’s comprehensive The Structure of the World (Oxford, 2014, DOI: 10.1093/acprof:oso/9780199684847.001.0001) developed the framework in detail.
French and Krause’s Identity in Physics (Oxford, 2006, DOI: 10.1093/0199278245.001.0001) provides the technical foundations. They show that quantum particles “need not satisfy the law of identity,” a claim that would be incoherent for ordinary objects but is straightforwardly true for entities that can be permuted without physical consequence. Their earlier paper with Redhead (British Journal for the Philosophy of Science, 1988, DOI: 10.1093/bjps/39.2.233) demonstrated that quantum statistics violate the principle of identity of indiscernibles, suggesting that quantum particles lack individual identity altogether.
Rovelli’s Relational Quantum Mechanics
Carlo Rovelli’s relational interpretation of quantum mechanics (International Journal of Theoretical Physics, 1996, DOI: 10.1007/BF02302261) pushes this further. Rovelli argues that quantum mechanics describes not absolute states but relations between systems. “The physical content of the theory has not to do with objects themselves, but the relations between them.” There are no observer-independent values; there are only relative facts, facts that hold for one system relative to another.
This dissolves the measurement problem by denying its presupposition. The puzzle arose from assuming that quantum systems have definite states that measurements reveal. If there are no absolute states, only correlations between systems, then “measurement” is simply the establishment of correlation. No collapse, no mystery, just relational facts all the way down.
The connection to constraint-based explanation is direct. If only relations exist, then what we call “objects” are crystallizations of relations, patterns of constraint that maintain themselves under interaction. Identity is not intrinsic; it is relational. Persistence is not substance; it is constraint closure. What survives under perturbation is what remains relationally stable. Everything else was always already nothing.
Eastern Philosophy Already Knew This
The convergence with certain Eastern philosophical traditions is striking and non-accidental. Nāgārjuna’s Mūlamadhyamakakārikā (second century CE) developed the doctrine of emptiness (śūnyatā): phenomena are empty of intrinsic existence (svabhāva). This is not nihilism; it is the claim that nothing exists independently of the relational web that permits it. The two truths doctrine distinguishes conventional truth (how things appear) from ultimate truth (their dependent origination). Westerhoff’s Nāgārjuna’s Madhyamaka (Oxford, 2009) and Garfield’s translation provide accessible introductions.
The structural parallel with OSR is precise. Both deny intrinsic natures. Both affirm relations as primary. Both dissolve the question “what is X really?” by showing that “really” presupposes a substance metaphysics that closer examination undermines. The methods differ, Nāgārjuna used dialectical analysis; modern physicists use mathematical formalism, but the destination is recognizably the same.
This convergence is evidence of something real. When independent traditions, using different methods, across thousands of years, arrive at the same structural insight, coincidence becomes implausible. What persists under recursive examination across scales, substrates, and cultures is more likely to be tracking something invariant than something parochial.
Part VI: The Free Energy Principle (What Survives the Critique)
Friston’s Framework
Karl Friston’s free energy principle (FEP) has become enormously influential in neuroscience and theoretical biology. His 2010 paper “The free-energy principle: a unified brain theory?” (Nature Reviews Neuroscience, DOI: 10.1038/nrn2787) has over 6,500 citations and spawned an industry of active inference research. The core claim is that biological systems minimize variational free energy, a measure of the difference between their internal model and the environmental states they encounter. This provides a principled explanation for perception (minimizing prediction error by updating models), action (minimizing prediction error by changing the world), and learning (optimizing model structure).
The mathematical framework is sophisticated. Parr, Pezzulo, and Friston’s textbook Active Inference (MIT Press, 2022, DOI: 10.7551/mitpress/12441.001.0001) provides a comprehensive treatment. The free energy principle connects to predictive processing (Rao and Ballard, Nature Neuroscience, 1999, DOI: 10.1038/4580; Clark, Behavioral and Brain Sciences, 2013, DOI: 10.1017/S0140525X12000477; Hohwy’s The Predictive Mind, 2013, DOI: 10.1093/acprof:oso/9780199682737.001.0001) and can be derived from first principles as a consequence of maintaining existence under changing conditions.
The Unfalsifiability Problem
Yet the FEP faces a serious critique: Friston himself has acknowledged that the principle is not falsifiable in the usual sense. It is more like Hamilton’s Principle of Stationary Action, a variational principle from which dynamics can be derived, not a contingent empirical claim. Colombo and Wright’s philosophical analysis (Synthese, 2021, DOI: 10.1007/s11229-018-01932-w) argues that the FEP’s epistemic status is “muddled,” it functions sometimes as an empirical hypothesis, sometimes as a mathematical framework, and sometimes as a conceptual schema, and these roles are not clearly distinguished.
Friston’s response to the “dark room problem” (Frontiers in Psychology, 2012, DOI: 10.3389/fpsyg.2012.00130) is instructive. Critics asked why organisms don’t simply minimize uncertainty by shutting themselves in dark rooms and dying quietly. Friston’s answer involves priors: organisms come equipped with expectations about environmental fluctuation, including expectations that they will act and explore. This resolves the puzzle but at the cost of packing more and more content into the prior. If the prior can be adjusted to fit any behavior, the framework becomes unfalsifiable.
What We Can Keep
The solution is to strip the FEP of anthropomorphic inflation while retaining its useful structure. What Friston describes is constraint satisfaction under energetic bounds, systems maintaining themselves by minimizing the divergence between their internal states and environmental conditions, subject to the thermodynamic costs of doing so. This is not “inference” in the folk-psychological sense. It is dynamics under constraints. The brain is not literally a scientist testing hypotheses; it is a physical system whose states are coupled to environmental states through constraint propagation.
When so interpreted, the FEP becomes a specific instantiation of the general principle we are articulating. Systems that persist are systems that maintain constraint closure. One way to maintain closure is to minimize surprise, keeping internal states within viable bounds relative to external fluctuations. This is not cognition; it is thermodynamics with feedback. The mathematical formalism is valuable precisely because it makes predictions about neural dynamics, perceptual phenomena, and adaptive behavior. But the cognitive vocabulary is optional and potentially misleading.
Part VII: Causation as Constraint Propagation
From Hume to Pearl
The concept of causation has been contested since Hume declared it nothing but constant conjunction plus psychological habit. The interventionist revolution in causal reasoning, led by Judea Pearl and James Woodward, provides a rigorous alternative that aligns with constraint-based explanation.
Pearl’s work on causal diagrams (Biometrika, 1995, DOI: 10.1093/biomet/82.4.669) and his subsequent book Causality introduced the do-calculus, a formal framework for reasoning about interventions as distinct from observations. His 2010 introduction (International Journal of Biostatistics, DOI: 10.2202/1557-4679.1203) provides an accessible summary. The key insight is that causation is about what happens under intervention, not merely what correlates with what. A causes B if intervening on A changes B; correlation is evidence for but not proof of causation.
Woodward’s Making Things Happen (Oxford, 2003, DOI: 10.1093/0195155270.001.0001) develops an interventionist theory of causal explanation. Explanations are causal when they describe relationships that would be stable under at least some interventions, when changing the explanans would change the explanandum in systematic ways. This captures what distinguishes genuine explanations from merely predictive correlations.
Constitutive vs. Efficient Causation
Kaiser and Krickel (British Journal for the Philosophy of Science, 2017, DOI: 10.1093/bjps/axv058) distinguish constitutive from efficient causation. Efficient causes are what we typically mean by “causes,” events that bring about subsequent events. Constitutive causes are different: they explain what something is, not how it came to be. The arrangement of atoms constitutively explains the solidity of a table, though no atom “caused” the table to become solid.
This distinction matters for cognition debates. When someone claims that bioelectric patterns “cause” regeneration, we should ask: efficient or constitutive? If efficient, what is the mechanism? If constitutive, we are not invoking a new causal agent; we are describing what the regeneration process consists in. Confusing constitutive with efficient causation creates phantom agencies, treating the description of a process as though it were an additional cause of that process.
Constraint-based explanation dissolves this confusion. Constraints neither push nor pull; they eliminate. What remains after constraint application is not caused by the constraints in the efficient sense; it is what the constraints permit. This is why constraint-based explanations are so powerful: they explain without invoking additional causes. They show why certain outcomes were the only ones consistent with the operative constraints.
Part VIII: Fractals and Scale Invariance (The Geometry of What Survives)
Mandelbrot’s Insight
Benoit Mandelbrot famously observed: “Clouds are not spheres, mountains are not cones, coastlines are not circles, and bark is not smooth, nor does lightning travel in a straight line.” Euclidean geometry describes idealized forms that rarely appear in nature. Fractal geometry describes what actually persists.
This is not merely a mathematical curiosity. It is diagnostic of constraint satisfaction. Hutchinson’s contraction mapping theorem (Indiana University Mathematics Journal, 1981, DOI: 10.1512/iumj.1981.30.30055) proves that iterated function systems with contractivity s < 1 have unique fixed points, attractors. The fractal structures we observe in nature are these fixed points. They are what remains after repeated application of constraints.
Why don’t spheres and cones appear in nature? Because they cannot persist under iteration. Cones erode. Spheres collapse under gravity. Straight lines diffuse. What survives repeated interaction with physical processes is what we actually observe: branching structures, rough surfaces, self-similar patterns at multiple scales. Fractal geometry does not describe some hidden Platonic realm; it describes the statistical signature of constraint satisfaction.
Biological Scaling Laws
West, Brown, and Enquist’s model for allometric scaling (Science, 1997, DOI: 10.1126/science.276.5309.122) demonstrates this principle in biology. They derive the ubiquitous 3/4-power scaling of metabolic rate with body mass from constraints on resource distribution networks. The fractal branching of circulatory and respiratory systems is not arbitrary; it is what minimizes transport costs under space-filling constraints. The geometry is selected because it satisfies thermodynamic constraints better than alternatives.
Bak, Tang, and Wiesenfeld’s self-organized criticality (Physical Review Letters, 1987, DOI: 10.1103/PhysRevLett.59.381) shows how systems naturally evolve toward critical states characterized by power-law distributions. Stanley’s review (Reviews of Modern Physics, 1999, DOI: 10.1103/RevModPhys.71.S358) connects this to universality classes in statistical mechanics. The same scale-invariant structures appear in earthquakes, forest fires, neural avalanches, and financial markets, not because these systems share mechanisms, but because they share constraint geometry.
Cross and Hohenberg’s comprehensive review of pattern formation (Reviews of Modern Physics, 1993, DOI: 10.1103/RevModPhys.65.851) catalogs how ordered patterns emerge from constraints in systems far from equilibrium. Gierer and Meinhardt’s reaction-diffusion theory (Kybernetik, 1972, DOI: 10.1007/BF00289234) explains how activator-inhibitor dynamics generate the spots and stripes we see in animal coats. Diego and colleagues (Physical Review X, 2018, DOI: 10.1103/PhysRevX.8.021071) show that pattern-enabling features are determined purely by network topology, by constraint structure.
Part IX: Indigenous Epistemologies (65,000 Years of Empirical Validation)
Songlines as Constraint Paths
The analysis takes an unexpected turn when we recognize that the pattern which connects has been encoded, with remarkable precision, in knowledge systems designed to transmit constraint patterns across generations without writing.
Aboriginal Australian songlines are not merely mnemonic devices or cultural artifacts. They are constraint-propagation paths through physical landscape. Lynne Kelly’s The Memory Code (Allen & Unwin, 2016) documents how Aboriginal Australians encoded navigation, ecology, law, kinship, and cosmology in sung narratives that must be walked to be fully activated. The singing and the walking together regenerate the country. Knowledge that does not remain coupled to land, season, and practice does not survive transmission. Songlines are literally falsification-resistant: if following the song leads you off a cliff, the song gets corrected.
Reser and colleagues (PLOS ONE, 2021, DOI: 10.1371/journal.pone.0251710) experimentally tested Aboriginal memory techniques against Western methods. Medical students using songline-like spatial encoding achieved significantly stronger recall (Kendall’s W = 0.43) than those using memory palace techniques (W = 0.37). This is not cultural romanticism; it is empirical measurement. The knowledge system developed under survival pressure outperforms systems developed in libraries.
Tempone-Wiltshire and Yunkaporta’s recent analysis (Psychotherapy and Counselling Journal of Australia, 2025, DOI: 10.59158/001c.143975) articulates the epistemology underlying these practices. Knowledge exists within relationships, not in isolated propositions. Yunkaporta’s Sand Talk: How Indigenous Thinking Can Save the World (2019) makes this vivid: knowledge drawn in sand cannot be reified. It must be regenerated through practice. This is constraint-based epistemology: only what can be reproduced survives.
The Dreaming as Constraint Structure
The Dreaming (Dreamtime) is often misunderstood as a mythological past. It is more accurately understood as atemporal constraint structure, the generative network from which actual configurations emerge. What Ladyman and Ross call “real patterns,” Aboriginal cosmology calls the Dreaming. The Dreaming is not a “realm”; it is the ongoing constraint-satisfaction process through which Country maintains coherence.
This is not analogy; it is convergent discovery. Cultures that survived for long horizons under survival pressure did not center creation myths alone. They encoded elimination myths. What gets cut. What fails. What cannot be spoken without collapsing. The pattern which connects was already known to peoples who lived within it for sixty-five millennia.
Kampanelis, Elizalde, and Ioannides (SSRN, 2023, DOI: 10.2139/ssrn.4594688) found that Aboriginal trade routes correlate with current economic activity patterns, evidence that the constraint-satisfaction encoded in songlines tracked real geographical and resource structures with remarkable accuracy. This is not romantic primitivism; it is recognition that empirical validation comes in many forms, and 65,000 years of continuous practice constitutes a long-running experiment.
Part X: Why “Nothing” Cannot Persist (Dissolving the Primordial Question)
Grünbaum’s Critique
“Why is there something rather than nothing?” has been called the fundamental question of metaphysics. Adolf Grünbaum’s analysis (British Journal for the Philosophy of Science, 2004, DOI: 10.1093/bjps/55.4.561) reveals that it contains a hidden presupposition: the “Spontaneity of Nothingness” (SoN), the assumption that nothingness is the ontologically natural default state requiring no explanation while existence demands justification.
Grünbaum traces this assumption to second-century Christian creatio ex nihilo doctrine, imported into philosophy without rational justification. Once we recognize the presupposition, the question dissolves. We can equally ask: “Why should there be nothing rather than something?” Without independent justification for privileging nothingness, the original question becomes a non-starter.
Grünbaum’s earlier paper (British Journal for the Philosophy of Science, 2000, DOI: 10.1093/bjps/51.1.1) develops the critique in detail. The physical vacuum has positive zero-point energy, undergoes measurable fluctuations, and exhibits definite structural properties. True metaphysical “nothing,” no spacetime, no laws, no potentiality, cannot be addressed by physics because it has no properties to analyze.
Constraint-Based Dissolution
The constraint-based answer is more direct: “nothing” cannot persist because it cannot support constraint. The moment any difference persists, it does so by excluding alternatives, and exclusion under finite resources is constraint satisfaction. No extra metaphysical floor is needed.
This reframes the question from a metaphysical “why” to a dynamical “how come.” The universe is not determinate because it must be; it is determinate because determinacy is what survives recursive elimination. That is a Darwinian answer applied to ontology itself. What we observe is not explained by necessity or design; it is explained by persistence.
“Nothing” is mathematically and conceptually unstable. Pure unconstrained possibility is paradoxically self-constraining, the absence of all constraints is itself a maximal constraint (on what can be distinguished). Quantum vacuum fluctuations, Gödel-type self-reference, and topological considerations all suggest that coherent “nothing” cannot be maintained.
Part XI: Derivation of Intelligibility Conditions
From Thermodynamics to the Floor
The boldest claim of this essay is that the conditions for intelligibility itself, distinction, identity, relation, constraint, are not additional axioms that must be assumed. They are derivable consequences of constraint satisfaction under thermodynamic bounds.
Distinction emerges from boundary maintenance. Landauer’s principle shows that maintaining a bit, distinguishing 0 from 1, requires energy expenditure. Without energetic work, distinctions dissolve into thermal noise. Distinction is not a metaphysical primitive; it is the signature of boundaries maintained against entropy.
Identity emerges from constraint closure. What makes something “the same” over time is not a magical soul or haecceity; it is invariance under perturbation. An entity persists as itself insofar as its constraint network regenerates through change. The Montevil-Mossio framework makes this precise: identity is closure that maintains itself. The Durant experiments demonstrate it empirically: morphological identity is bistable attractor, not Platonic form.
Relation emerges from constraint propagation. If Rovelli and the structural realists are right, relations are not between pre-existing things; things are crystallizations of relations. Constraint propagation is inherently relational, constraints couple systems, create dependencies, enable and forbid. Relation is not an addition to an object-based ontology; it is what remains when substances are dissolved.
Constraint is the fundamental primitive. But unlike other proposed primitives, constraint comes with a derivation: thermodynamic state-space structure. The Second Law is not a stipulation; it is a statistical consequence of vastly more high-entropy than low-entropy microstates. Constraint emerges from combinatorics plus dynamics. We do not need to posit it; it falls out of the mathematics of possibility.
The Selection Principle That Precedes Categories
This means that what philosophers have called “categories of understanding,” the preconditions for any thought or experience, are not a priori impositions by mind onto world. They are the structure that any world capable of supporting thought must already have. Intelligibility is not a human projection; it is the shadow cast by constraint satisfaction.
The framework is not an ontology, epistemology, or cosmology. It is a selection principle on what can remain available to any of those enterprises. It operates beneath the level at which traditional philosophical questions are posed. It explains why there is anything to know, not merely what we know.
This is why the framework feels austere, even deflationary. It does not add furniture to reality. It removes it. Ontologies describe what passed the filter. Epistemologies describe how filtered things are tracked. Cosmologies describe how filters evolve. The work here is at the level of the filter itself.
Part XII: Falsification and the Discipline of Loss
Popper, Lakatos, and the Scientific Research Programme
A framework that cannot lose is not a framework but a faith. Karl Popper’s Logic of Scientific Discovery (Routledge, 1959, DOI: 10.4324/9780203994627) established falsifiability as the demarcation criterion between science and pseudoscience. A claim is scientific not because it is true but because it specifies what would prove it false.
Imre Lakatos refined this in his methodology of scientific research programmes (DOI: 10.1017/CBO9780511621123.003). Lakatos distinguished the “hard core” of a research programme, the central commitments that define it, from the “protective belt” of auxiliary hypotheses that can be modified. A programme is progressive if its modifications lead to novel predictions; it is degenerating if modifications serve only to shield the core from refutation.
Antony Flew’s “Theology and Falsification” (1950) applied this logic to religious claims with devastating effect. Flew observed that theistic claims often die “the death of a thousand qualifications,” each apparent counterexample is absorbed by reinterpretation until nothing empirical remains. The same diagnosis applies to “basal cognition”: if every observation is compatible, nothing is tested.
How Constraint-Based Explanation Can Fail
Constraint-based explanation is falsifiable. Here are the failure conditions:
- Violation of Landauer bound: If information processing occurred without thermodynamic cost in isolated systems, the foundation would collapse.
- Constraint closure without robustness: If organizationally closed systems showed no greater persistence than open ones, Montevil-Mossio would be wrong.
- Path independence in bioelectric memory: If the Durant results failed to replicate, if transient perturbations reliably corrected toward canonical form, the attractor interpretation would fail.
- RAF failure at realistic catalysis levels: If autocatalytic closure required unrealistic levels of catalytic promiscuity, Kauffman-Hordijk-Steel would be wrong.
- Scale non-invariance: If the fractal signatures and scaling laws broke down systematically, the constraint-survival interpretation would need revision.
- Superior predictions from cognitive frameworks: If “basal cognition” generated predictions that constraint-based models could not match, predictions about specific outcomes, intervention effects, or system behavior, the cognitive framework would earn its place.
None of these failures have occurred. The framework survives because it can fail and has not.
Part XIII: The Question That Remains
What Does “Basal Cognition” Add?
So the challenge stands, and it is simple. What concrete, falsifiable prediction does “basal cognition” make that is not already made, tested, and explained by constraint propagation under thermodynamic limits? How would we tell if it is wrong?
This is not a rhetorical question. It is the minimal bar for any scientific claim. If “basal cognition” predicts something specific, some pattern of behavior, some response to intervention, some measurable quantity, that constraint-based models do not predict, then it is doing real work. State the prediction. Design the experiment. Run the test.
If, on the other hand, “basal cognition” makes no predictions that differ from constraint satisfaction, then it is not an alternative explanation. It is a vocabulary preference. It may inspire research. It may motivate funding. But it does not add mechanism. And mechanisms, not metaphors, are what science ultimately trades in.
The phenomena are real. Biological systems exhibit stunning efficiency, robustness, and adaptability. The empirical work documenting these phenomena is invaluable. What is contested is whether labeling them “cognitive” does explanatory work, or whether it merely relabels what constraint-based frameworks already explain while immunizing itself from falsification.
What Persists Is What Constrains
Strip away the metaphysical excess, and a simpler principle remains. Differences that persist are differences that constrain what comes next. Persistence is evidence of constraint satisfaction. That rule applies at every scale. It explains why there is determinate existence rather than indeterminate nothing without appealing to necessity or intention. Nothing cannot persist. Anything that does persist has already passed a filter.
This is the pattern which connects. Not information, not mind, not structure, not process, but the invariant condition under which any of those could exist long enough to be observed, measured, or thought.
Gregory Bateson asked the question in 1979. The answer was always implicit in his formulation: a difference which makes a difference. But what makes a difference make a difference? Persistence. And what makes things persist? Constraint satisfaction under thermodynamic bounds.
Bateson circled the answer without landing. The Indigenous knowledge systems encoded it without formalizing. The Eastern philosophers described it without mechanizing. The physicists derived it without seeing its scope. What emerges from convergence across traditions, methods, and millennia is not coincidence. It is consilience, independent lines of evidence pointing to the same invariant.
What persists is what constrains. Everything else is commentary.
Appendix: Key Scholarly Citations by Domain
Thermodynamics and Information Theory
- Landauer, R. (1961). DOI: 10.1147/rd.53.0183
- Bennett, C.H. (1973). DOI: 10.1147/rd.176.0525
- Bennett, C.H. (1982). DOI: 10.1007/BF02084158
- Bennett, C.H. (2003). DOI: 10.1016/S1355-2198(03)00039-X
- Bérut, A. et al. (2012). DOI: 10.1038/nature10872
- Jarzynski, C. (1997). DOI: 10.1103/PhysRevLett.78.2690
- Crooks, G.E. (1999). DOI: 10.1103/PhysRevE.60.2721
- Seifert, U. (2012). DOI: 10.1088/0034-4885/75/12/126001
- England, J.L. (2013). DOI: 10.1063/1.4818538
- England, J.L. (2015). DOI: 10.1038/nnano.2015.250
Biological Organization and Constraint Closure
- Montévil, M. & Mossio, M. (2015). DOI: 10.1016/j.jtbi.2015.02.029
- Moreno, A. & Mossio, M. (2015). DOI: 10.1007/978-94-017-9837-2
- Mossio, M. & Moreno, A. (2010). PMID: 21162371
- Mossio, M. et al. (2009). DOI: 10.1093/bjps/axp036
- Kauffman, S.A. (1986). DOI: 10.1016/S0022-5193(86)80047-9
- Hordijk, W. & Steel, M. (2004). DOI: 10.1016/j.jtbi.2003.11.020
- Hordijk, W. et al. (2011). DOI: 10.3390/ijms12053085
- Xavier, J.C. et al. (2020). DOI: 10.1098/rspb.2019.2377
Bioelectric Research and Memory
- Levin, M. (2014). DOI: 10.1113/jphysiol.2014.271940
- Levin, M. & Martyniuk, C.J. (2018). DOI: 10.1016/j.biosystems.2017.08.009
- Durant, F. et al. (2017). DOI: 10.1016/j.bpj.2017.04.011
- Durant, F. et al. (2019). DOI: 10.1016/j.bpj.2019.01.029
- Pezzulo, G. et al. (2021). DOI: 10.1098/rstb.2019.0765
- Law, R. & Levin, M. (2015). DOI: 10.1186/s12976-015-0019-9
- Hopfield, J.J. (1982). DOI: 10.1073/pnas.79.8.2554
Structural Realism and Physics
- Ladyman, J. & Ross, D. (2007). DOI: 10.1093/acprof:oso/9780199276196.001.0001
- Ladyman, J. (1998). DOI: 10.1016/S0039-3681(98)80129-5
- French, S. & Ladyman, J. (2003). DOI: 10.1023/A:1024156116636
- French, S. (2014). DOI: 10.1093/acprof:oso/9780199684847.001.0001
- French, S. & Krause, D. (2006). DOI: 10.1093/0199278245.001.0001
- Rovelli, C. (1996). DOI: 10.1007/BF02302261
Causation and Methodology
- Pearl, J. (1995). DOI: 10.1093/biomet/82.4.669
- Pearl, J. (2010). DOI: 10.2202/1557-4679.1203
- Woodward, J. (2003). DOI: 10.1093/0195155270.001.0001
- Popper, K. (1959). DOI: 10.4324/9780203994627
- Lakatos, I. (1970). DOI: 10.1017/CBO9780511621123.003
- Grünbaum, A. (2004). DOI: 10.1093/bjps/55.4.561
Free Energy Principle
- Friston, K. (2010). DOI: 10.1038/nrn2787
- Parr, T. et al. (2022). DOI: 10.7551/mitpress/12441.001.0001
- Rao, R.P.N. & Ballard, D.H. (1999). DOI: 10.1038/4580
- Clark, A. (2013). DOI: 10.1017/S0140525X12000477
- Colombo, M. & Wright, C. (2021). DOI: 10.1007/s11229-018-01932-w
Indigenous Knowledge Systems
- Kelly, L. (2016). The Memory Code
- Yunkaporta, T. (2019). Sand Talk
- Reser, D.H. et al. (2021). DOI: 10.1371/journal.pone.0251710
- Tempone-Wiltshire, J. & Yunkaporta, T. (2025). DOI: 10.59158/001c.143975
Fractals and Scale Invariance
- Hutchinson, J.E. (1981). DOI: 10.1512/iumj.1981.30.30055
- West, G.B. et al. (1997). DOI: 10.1126/science.276.5309.122
- Bak, P. et al. (1987). DOI: 10.1103/PhysRevLett.59.381
- Cross, M.C. & Hohenberg, P.C. (1993). DOI: 10.1103/RevModPhys.65.851
What survives is what can. The framework survives because it can.
