Thermodynamic Realism: In-Principle Determinacy and Ethics as Necessary Modeling
Thermodynamic Realism: In-Principle Determinacy and Ethics as Necessary Modeling
Abstract
Hume's is-ought problem has resisted dissolution because the standard framings have either ignored the semantics of the ought-operator or underestimated what physicalism plus thermodynamics actually delivers. We argue that the problem dissolves in principle once two facts are recognized. First, "ought" is a domain-bound operator whose application requires an agent with sustained negentropy conditions; sentences deploying it outside this domain are malformed in the structural sense that "what was before time?" is malformed. Second, within the operator's domain, ought-statements describe physically determinate facts about coupled-system thermodynamics across the agent's actual embedding, evaluated across the open-ended horizon physics permits. The dissolution is conditional on three formal questions: the principled coarse-graining of relevant systems, the principled scalarization across Pareto frontiers, and the rigorous specification of the coupling-weighted contribution function we sketch in §IV. The time-horizon question, traditionally a fourth, is not an open question on this framework: it is exogenously fixed by physics and not subject to analyst choice. Ethics, on this account, is the modeling activity through which computationally bounded agents navigate a determinate but inaccessible territory; moral cognition, moral emotion, moral disagreement, and moral progress are all consequences of modeling being done by bounded agents tracking real structure. The contribution is to specify what the dissolution requires, what work remains, and what kind of object ethics turns out to be.
I. The Two Theses
This paper defends two claims that together dissolve the is-ought problem.
The two theses are independent in principle but mutually reinforcing in practice. Thesis 1 explains why the gap appears unbridgeable in standard formulations: the operator is being asked to apply outside its domain, and no inference can produce a referent where there is none. Thesis 2 explains why, within the domain, the gap was never there: the descriptive and normative are the same physical structure described at different levels of compression.
A note on Hume scholarship is in order. The argument here engages the standard reading of Hume that has shaped subsequent moral philosophy, in which the descriptive and normative are taken to be categorically separate domains. Contemporary Hume scholarship1 reads the original Treatise passage more narrowly, as a point about the inferential structure of certain moral arguments rather than a wholesale ontological separation. Our target is the structural argument as it has functioned in the philosophical literature, not Hume's specific historical position.
II. The Domain Argument
Some operators are domain-bound by their semantics. The temporal operator "before" presupposes time. The question "what was before time?" is not a hard question with an unknown answer; it is malformed, because the operator presupposes the very thing the question places it outside of. The Hawking-Hartle no-boundary proposal exploits this structural feature: certain cosmological questions dissolve once one recognizes that the operator they deploy has a presupposition the question violates.
We claim "ought" has the same structure. "Ought" presupposes an agent — a self-maintaining system that models its own persistence conditions. The question "why ought one persist at all?" places the operator outside the conditions that give it its domain, and is malformed for the structural reason given.
This is not a definitional move. It is a claim about what kind of operator "ought" is, supported by the observation that every coherent first-person use of "ought" can be paraphrased as a statement about constraints on the agent's sustained negentropy conditions. Uses that resist this paraphrase are uses where the operator has been detached from any agent's conditions, and these are precisely the uses that generate the appearance of an unbridgeable is-ought gap.
The North Pole analogy formalizes the point. Within the domain of a sphere's surface, "north" is well-defined and every position has a determinate northward direction. At the pole itself, "north" stops applying because the operator's preconditions are not met. The question "what is north of the North Pole?" is not asking a hard geographical question; it is misapplying an operator. The is-ought gap, on this analysis, is the same kind of misapplication.
Hume's original passage in the Treatise (III.i.1) observes that authors typically proceed from descriptive premises and then suddenly introduce normative conclusions "without observing any reason for this manner of expression." The framework here gives the reason: when the inference stays within the agent-and-embedding context that grounds the ought-operator, there is no leap because there is no gap; when the inference appears to leap, the operator has been detached from its grounding.
III. Physicalism and the Inductive Base
The physical thesis defended here rests on physicalism, adopted not as metaphysical preference but as the maximally inductively justified position given the unbroken pattern of physical explanation across all investigated phenomena. Every domain that has been seriously investigated has yielded a physical explanation. Heat became molecular kinetic energy. Life became chemistry under selection pressure. Mental illness became neurobiology. The pattern is exceptionless across the entire recorded history of science.
To demand more than this before treating physicalism as a working axiom is not rigor; it is epistemic paralysis. The framework therefore takes physicalism as foundational in the same spirit one takes the expectation that the sun will rise tomorrow: not as logical necessity but as the maximally justified inductive inference from a complete and unbroken record.
What physicalism commits us to, for present purposes, is that every phenomenon has a physical explanation, that mental states are physical states, and that moral facts, to the extent they exist, are facts about physical systems. It does not commit us to the claim that the lowest level of description is always the most useful. The arrangement of physical components is just as physical as the components themselves.
IV. The Determinacy Claim and the Coupling Function
The substantive claim of the framework is that within the operator's domain, ought-facts are physically determinate, even when the relevant computation is intractable.
To see why this matters, consider a multi-agent multi-level conflict. Two agents have conflicting persistence conditions. They are embedded in nested systems whose stability depends on various distributions of resources, behaviors, and arrangements among them. The traditional view is that no fact of the matter exists about which configuration is "correct" — this is the apparent indeterminacy that has driven moral relativism, error theory, and constructivist accounts.
We claim this appearance is wrong. If physicalism holds, and if entropy and information are well-defined at every level of the physical hierarchy, then every configuration of the multi-agent system has a determinate thermodynamic accounting across any specified embedding. The accounting is a physical quantity, not a stipulation. The fact that no embedded agent can compute it does not affect its existence.
This is the same structural point one makes about chess. A chess position has a determinate game-theoretic value — win, lose, or draw with optimal play — even though the game tree is too large for any agent to traverse. Computational intractability does not imply ontological indeterminacy. Confusing the two has been one of the persistent errors of philosophical analysis when it engages with mathematical structure.
The Time Horizon Is Not a Choice
A standard objection to thermodynamic accounts of value is that the relevant time horizon is analyst-dependent. Different choices of T yield different optimal configurations, so the framework appears to require a stipulation that physics does not provide.
This objection mistakes the framework's commitment. The relevant horizon is not chosen; it is the open-ended horizon physics permits. Maximum T is whatever the universe allows agents to exist; below that, an agent's effective horizon is whatever its actual embedding can sustain. The horizon is exogenous to analyst choice. It is fixed by the physical regime in which agents and negentropy maintenance are possible, and it stops applying outside that regime in the same way the ought-operator stops applying outside agents.
This commitment has a substantive consequence: any strategy that buys local persistence by drawing down embedding stability fails when evaluated across the embedding's actual remaining lifetime. Cancer cells, parasitic strategies that destroy hosts, totalitarian states that exhaust their populations, civilizations that destroy their resource base, all share the structure of optimization on a horizon shorter than the agent's own existence requires. Under open-ended horizon, such strategies fail by physical accounting, not by stipulation.
Coupling-Weighted Contribution and the Coarse-Graining Question
The coarse-graining problem is the question of which systems count, and how much, in the agent's negentropy accounting. We propose that the answer is not analyst choice but a physical fact: systems contribute to an agent's sustained negentropy capacity in proportion to their actual coupling to that capacity.
Let S be an agent and let {Ni} be the nested systems in which S is embedded. Let K(S, Ni) be the coupling coefficient between S and Ni: the degree to which loss or degradation of Ni would reduce S's sustained negentropy capacity over the open-ended horizon. Then the contribution of Ni to the relevant accounting is weighted by K(S, Ni), and the total is a sum over all coupled systems.
where C(Ni, t) is the negentropy capacity of Ni at time t.
The coupling coefficient K is itself a physical quantity. It can be operationalized as the partial derivative of S's sustained negentropy capacity with respect to Ni's state, integrated over the relevant horizon. Tightly coupled systems (cellular machinery, biosphere) have high K; loosely coupled systems (distant trade partners, abstract institutions) have low K; uncoupled systems have K = 0 and do not enter the accounting at all.
The coarse-graining question therefore reduces to the question of how to compute K rigorously. This is a tractable problem in the sense that it inherits the formal apparatus of coupled-systems analysis, sensitivity analysis, and information-theoretic measures of dependency between subsystems. It connects to existing work on objective coarse-graining in non-equilibrium thermodynamics2 and to the maximum entropy formalism developed by Jaynes. We do not solve the coarse-graining problem here; we claim it is the right problem to be solving, and that its solution is not a matter of analyst choice but of physical analysis.
The Pareto Frontier Question
Even with horizon and coarse-graining specified, multi-agent conflicts often present a Pareto frontier of non-dominated configurations rather than a unique optimum. Different configurations may sustain different agents at different capacities without one configuration strictly dominating another.
The replacement of "persistence" with "sustained negentropy capacity" as the maximand shrinks but does not eliminate this frontier. Configurations that destroy capacity in any coupled system are dominated by configurations that preserve it; this rules out a wide class of putative solutions. What remains is the genuine plurality of arrangements that distribute capacity differently across coupled agents and systems.
The framework does not currently specify a principled scalarization across this residual frontier. The scalarization may itself be physically determined by deeper analysis of coupled-systems dynamics — for instance, by selecting the Pareto-optimal configuration that is most robust to perturbation, or that maximizes total integrated capacity over the open-ended horizon — but we do not derive such a principle here. We mark this as the remaining open question of the three originally identified, and we treat it as an active research target rather than a permanent residue.
V. The Necessity of Ethical Modeling
The intractability of the full computation is not an accident or temporary limit. It is structural. The agent is part of the system whose accounting it would need to compute. Any computation the agent performs is itself a process within the system, and increases the complexity of what would need to be tracked. Embedded agents cannot, in principle, compute their own optimal configurations exactly.
What agents can do, and what they have always done, is run models of the underlying structure. Models compress the determinate but inaccessible territory into tractable approximations. A model is good when it tracks the territory well enough to keep the modeling agent persisting and capacitated; it is bad when it fails to. Agents whose models are bad enough do not persist, which is one reason agents tend to have models that are at least minimally adequate.
This claim is supported by substantial work across cognitive science. Moral emotions function as fast heuristics that compress information about action-pattern consequences for sustained cooperative arrangements. Predictive processing accounts of cognition treat the brain as a hierarchical modeling system minimizing prediction error, and moral cognition fits this account directly. The cognitive science of cooperation has documented the specific computational shortcuts that underlie human social judgment.
Three consequences follow.
First, moral cognition is real cognition responsive to a real territory. It is not a separate normative faculty operating outside physics. It is the cognitive form taken by coupled-system modeling in agents whose persistence depends substantially on coordination with other agents. The realism is naturalized but genuine.
Second, moral intuitions are heuristics with the strengths and predictable failures of heuristics. They evolved for ancestral embedding scales and predictably misfire in novel ones. Face-to-face harm is intuitively salient while statistical and diffuse harm is not. Kin and reciprocity partners are weighted while strangers are not. Immediate consequences are weighted while multi-generational consequences are not. The framework predicts the observed pattern of moral cognition and its failure modes.
Third, moral progress is real. Refinement of moral models toward better tracking of the underlying structure is genuine progress, not mere change. Models can be more or less accurate to a determinate territory, and history can be read as the long process of agents discovering that earlier models were inadequate to embeddings that later analysis revealed.
The framework does not predict universal moral convergence in finite time, because the territory is computationally inaccessible and models are bounded. It predicts asymptotic convergence in the limit of unbounded computation across stable embeddings, with the asymptote being the determinate structure itself.
VI. Affect as Telemetry
Emotions, on this account, are not separate from moral cognition but constitutive of it. They are the agent's internal telemetry on its own thermodynamic state and on the state of arrangements it depends on. Fear tracks predicted boundary violation. Anger tracks systemic friction with another agent's behavior pattern. Disgust tracks contamination of a system the agent is embedded in. Love tracks the integration of predictive models across cooperative agents. Joy tracks high-fidelity alignment with viable trajectories.
This is not metaphor. The affect systems of social mammals are physically implemented prediction-error and prediction-confirmation signals operating on the agent's models of self and embedding. Calling them "telemetry" is a description of what they do, not a literary device. They are how the model communicates with the agent that runs it.
The distinction between necessary and unnecessary cognitive cost matters here. An agent thinking carefully about a real problem in its environment is paying a metabolic cost for genuine modeling work; this is necessary cost. An agent maintaining a false self-model, or sustaining inconsistent representations of its embedding for purposes of deception, is paying a metabolic cost for inconsistencies the agent itself generates; this is unnecessary cost. The framework predicts that unnecessary cost is what moral cognition tracks when it tracks deception, and that the empirical signature is the cost of self-generated inconsistency, not total cognitive load. This handles the empirical messiness of the claim that lying is metabolically expensive, where skilled deception is often metabolically cheap. The framework's narrower prediction matches the evidence better than the broader claim.
VII. The Realism Question
The framework supports a strong moral realism with one careful qualification.
The realism is strong because ought-facts within the domain of an agent are physically determinate, not relative to the agent's preferences, and not constructed by any procedure. They are facts about the physical structure of the agent's embedding under open-ended horizon.
The qualification is that the realism is agent-and-embedding-relative in the technical sense developed in contemporary moral realism debates.3 The same physical configuration can ground different ought-facts for different agents because their persistence conditions and embeddings differ. This is not relativism in the usual sense, because each agent-and-embedding pair has determinate ought-facts grounded in physical structure rather than in convention or stance.
The position extends Railton's reductive moral realism by grounding moral facts in coupled-system thermodynamic structure rather than in objective interests. Where Railton's view requires a separate account of how facts about objective interests become motivating for agents, the present framework has the simpler structure that the agent is the persisting system whose negentropy conditions ground the relevant facts; motivation is built into the metaphysics rather than added on top.
The position extends Foot's neo-Aristotelian naturalism in a different direction. Where Foot grounds goodness in life-form flourishing, with life-forms providing built-in standards, the present framework extends the analysis below the level of life-form to any self-maintaining system, while distinguishing agents from non-modeling self-maintainers by their capacity to represent their own persistence conditions. The work that life-form does for Foot is done here by modeling capacity: an agent's standard of flourishing is constituted by what its modeling capacity allows it to be, not by membership in a biological category.
VIII. The Moorean Open Question
Moore's open question argument, in Principia Ethica, holds that for any natural property N identified with goodness, the question "but is N actually good?" remains meaningful, which shows N and goodness are not identical.
On the present framework, the apparent meaningfulness is a grammatical artifact. Once "good" is properly understood as a domain-bound operator referring to coupled-system thermodynamic structure across an agent's embedding, the question "is sustained negentropy capacity actually good?" has the same structure as "is what's north of a position actually north of it?" The question can be asked, but the appearance of substantive content depends on the operator being detached from its domain. Reattach the operator and the question closes.
This response is not a refusal to engage Moore. It is a diagnosis of the open question phenomenon as the same kind of artifact the is-ought gap turned out to be. Both are products of treating domain-bound operators as if they could float free of their domains. The persistence of the open question intuition is evidence of how naturally human language permits this kind of detachment, not evidence of an ontological gap.
IX. Predictions
The framework generates predictions at several levels.
At the level of moral models, it predicts that moral systems coexisting with stable embeddings should track those embeddings' actual constraints better than moral systems coinciding with embedding collapse. This is testable in historical and anthropological records, with the qualification that "stability" must be operationalized carefully and confounds are extensive.
At the level of cognitive substrate, it predicts that moral cognition shares processing characteristics with predictive modeling of nested systems generally, and that the affect systems implementing moral telemetry are continuous with the affect systems implementing physiological homeostasis. This connects to existing work in predictive processing, interoception, and the cognitive science of cooperation.
At the level of metabolic accounting, it predicts that sustaining inconsistent self-models within cooperative arrangements imposes costs distinct from accurate modeling of the world. The relevant cost is the cost of self-generated inconsistency, not total cognitive load. This is narrower than naive claims about the cost of deception and matches the existing mixed evidence better.
At the level of moral disagreement, it predicts that disagreement decomposes into two phenomena: genuine differences in correct ought-facts between agents in different embeddings, which are not resolvable without embedding convergence, and model-quality differences between agents in similar embeddings, which are in principle resolvable by attending to the actual structure. The framework predicts the observed pattern: persistent disagreement across cultures with substantially different embedding conditions, gradual convergence within cultures sharing embedding conditions over historical time.
At the level of formal research, the framework generates a research program: rigorously specify the coupling function K; derive a principled Pareto scalarization where one is physically determined; and develop tractable computational approximations to coupled-system thermodynamic accounting for ethically relevant cases.
X. What the Framework Does Not Claim
The framework does not claim that current moral intuitions are correct. It claims they are heuristic models, evaluable by how well they track the underlying thermodynamic structure. Some current intuitions will not survive scrutiny. Some current dissents from intuition will turn out to be model-improvements.
The framework does not claim that any embedded agent can compute the optimal action. It claims that agents run models, and that ethics is those models. The existence of a determinate optimum does not imply its computability. The framework is therefore not a recipe for action; it is an account of what action-guidance is doing when it works.
The framework does not claim that the remaining open formal questions have been solved. It claims the time-horizon question is dissolved by recognizing that the horizon is not analyst-chosen but physically fixed; it claims the coarse-graining question reduces to specification of the coupling function K; and it claims the Pareto scalarization question is a tractable formal problem rather than a metaphysical residue.
The framework does not extend to questions outside its domain. Questions about value or obligation phrased independently of any agent or embedding are not answered; they are diagnosed as malformed.
XI. Conclusion
The is-ought problem persisted partly because philosophy treated "ought" as a domain-free operator and partly because philosophy underestimated what physicalism plus thermodynamics actually delivers. Once "ought" is recognized as domain-bound, large regions of normative discourse stop generating an apparent gap. Once thermodynamics is taken seriously as the substrate at every level of the physical hierarchy, what remains of the apparent gap turns out to be either misapplication of the operator or a computationally hard but in-principle determinate question about coupled-system thermodynamic structure.
Ethics is what computationally bounded agents do when they need to act in a territory whose full structure they cannot compute. The activity is real. The territory is real. The models are partial. The work of refining the models is the work of moral inquiry, and it has direction because the territory it tracks has structure.
The contribution of this framework is to identify the dissolution as conditional on a small number of formal questions, to argue those questions are tractable rather than residual, to dissolve one of the questions (time horizon) by recognizing it is not a choice but a physical fact, and to relocate ethics from a domain where it appears to lack physical grounding to a domain where it appears as the necessary modeling activity of agents in physically determinate but computationally inaccessible conditions.
[1] See Cohon (2008), Hume's Morality: Feeling and Fabrication, and Garrett (2007) on the Treatise III.i.1 passage. ↩
[2] The objective coarse-graining literature in non-equilibrium statistical mechanics, including work building on Jaynes (1957) and developed in contemporary stochastic thermodynamics (Seifert 2012), provides the formal apparatus for principled coarse-graining choices grounded in system structure rather than analyst preference. ↩
[3] The technical sense of agent-relative moral realism is developed in Schroeder (2007), Slaves of the Passions, and in subsequent literature on the metaphysics of reasons. The position defended here differs in grounding the relativization in physical structure rather than in agents' desire-sets. ↩
References
Boyd, R. (1988). How to be a moral realist. In G. Sayre-McCord (ed.), Essays on Moral Realism. Cornell University Press.
Cohon, R. (2008). Hume's Morality: Feeling and Fabrication. Oxford University Press.
Foot, P. (2001). Natural Goodness. Oxford University Press.
Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience 11(2), 127–138.
Garrett, D. (2007). Hume. Routledge.
Greene, J. (2013). Moral Tribes: Emotion, Reason, and the Gap Between Us and Them. Penguin Press.
Haidt, J. (2012). The Righteous Mind. Pantheon.
Hawking, S. and Hartle, J. (1983). Wave function of the Universe. Physical Review D 28(12), 2960–2975.
Hume, D. (1739). A Treatise of Human Nature.
Jaynes, E. T. (1957). Information theory and statistical mechanics. Physical Review 106(4), 620–630.
Kolmogorov, A. N. (1965). Three approaches to the quantitative definition of information. Problems of Information Transmission 1(1), 1–7.
Landauer, R. (1961). Irreversibility and heat generation in the computing process. IBM Journal of Research and Development 5(3), 183–191.
Moore, G. E. (1903). Principia Ethica. Cambridge University Press.
Nowak, M. (2006). Five rules for the evolution of cooperation. Science 314(5805), 1560–1563.
Railton, P. (1986). Moral realism. Philosophical Review 95(2), 163–207.
Schrödinger, E. (1944). What Is Life? Cambridge University Press.
Schroeder, M. (2007). Slaves of the Passions. Oxford University Press.
Seifert, U. (2012). Stochastic thermodynamics, fluctuation theorems and molecular machines. Reports on Progress in Physics 75(12), 126001.
Shannon, C. (1948). A mathematical theory of communication. Bell System Technical Journal 27, 379–423, 623–656.
Tomasello, M. (2016). A Natural History of Human Morality. Harvard University Press.
Comments
Post a Comment