Thermodynamic Realism: The Apex Synthesis

Thermodynamic Realism

The whole case: what holds, what is bet, and where the map ends

Apex synthesis · the whole corpus in one document

Andraž Đurič

Version 1.0 · June 2026

This is the current (12. June 2026) apex statement of a larger body of work on this blog page: an attempt at the whole case in one document, built bedrock-up. Claims appear once each, with their confidence attached. Two premises carry the main exposure: the referential reading of the audited operators (§3) and the no-default claim (b) of §4.

Scope. The full modules carry the defences, the formalism, the objections, and the registries; this document carries the spine and its joints. On any conflict, the modules govern.

0. Read this first: two layers, judged separately

Everything below sits in one of two layers, and the layers do not have the same standing.

The descriptive layer is a thermodynamic account of persistence, modeling, selection, and the structures these produce. It is empirically falsifiable. It can be wrong in ways an experiment would show.

The metaethical layer is built on top of it. It takes the descriptive layer as input and adds one non-thermodynamic premise about the meaning of "ought." It is not falsifiable by experiment; it is defeasible by argument. You may accept the first layer and reject the second.

The claim that the physics grounds the ethics is the program's central hypothesis, not a proven entailment — a hypothesis with a named crux (§5), not a result.

One frame governs the whole metaethics, and it has three links, not a reduction:

thermodynamics → agents → ought. Thermodynamics supplies the agents (a far-from-equilibrium self-maintainer is the only kind of thing an "ought" can be indexed to). Agent-relativity is a semantic fact about the operator "ought," more general than the thermodynamics. Persistence then fixes the content of a given agent's ought. The first link is descriptive, the second is semantic, the third is the bet.

The agent is the hinge of the whole structure. It is a descriptive object — a far-from-equilibrium self-maintainer, characterizable entirely in physical terms — and it is simultaneously the relatum the metaethical operator indexes. It lives in both layers. That is why the layers are separable (the semantic claim about the operator needs no particular physics of agents) and yet not independent (this program fills the agent slot thermodynamically, so the content of any actual ought comes out thermodynamic).

This is physicalist monism with thermodynamics as the functional bedrock, not thermodynamic fundamentalism. Thermodynamics is not the deepest layer of the world; below it is a time-symmetric microphysics that contains no arrow and no agents. It is the functional floor: the lowest level at which the program's vocabulary — boundary, gradient, persistence, fidelity — has anything to point at, and the level whose validity is robust to whatever resolves the open questions beneath it. The program is built there for that reason.

1. The bedrock

Four axioms and four background premises, each carrying no more weight than its kind permits.

A0, the Persistence Tautology. A system that fails the conditions of its own persistence ceases to exist as an observable. Asserts nothing about value; only that existence has prerequisites.

A1, Physicalism. All phenomena are physical. Adopted as the maximally inductively justified working premise: every phenomenon ever explained was explained physically, none by adding a non-physical ingredient, and the burden falls on whoever would break the pattern. Note its scope: A1 is monism, not the stronger and false claim that everything reduces to the Second Law.

A2, the Second Law. In an isolated system entropy tends to its maximum; maintaining a localized gradient requires continuous work. Adopted without re-derivation.

A3, Information is Physical. Storing, processing, and erasing information carry minimum thermodynamic costs (Landauer 1961; experimentally confirmed across platforms from colloidal beads to single atoms: Bérut et al. 2012; Jun, Gavrilov and Bechhoefer 2014; Hong et al. 2016; Yan et al. 2018). The program relies only on the narrow, well-supported claim that the operations it actually uses — maintaining and updating divergent or dual models — are logically irreversible and carry an irreducible cost.

BP0–BP3, background premises. The accessible universe began in an extraordinarily low-entropy state (the Past Hypothesis; Albert 2000), and every usable gradient is that condition still discharging; the universe is far from equilibrium; local free energy is finite, which forces competition; persisting systems are organized into nested statistical boundaries. These are empirical facts relied on, not derived. The arrow of time lives in BP0, a boundary condition, not in A2 as a fundamental law: delete the low-entropy past and there is no arrow, no selection direction, and no sense to "less wrong over time."

2. The secure spine (descriptive)

These hold, at the labelled strength.

Gradients discharge, and structure discharges them faster. Far from equilibrium, matter organizes, because a structured flow can level a gradient faster than disordered matter can. Life is the limit case: a structure that maintains its own boundary so it can keep discharging. (One hedge: that structures in general maximize discharge rate — maximum entropy production — has only restricted, contested support and is a different and stronger claim. Nothing here rests on it.)

Persistence requires work and modeling (T2, T3). To persist is to hold a boundary against dissolution, which costs continuous work; any model of the environment is encoded in physical states and costs free energy to keep.

Map-territory divergence has a thermodynamic cost (T4). A model that fails to track the environment dissipates free energy. The physical anchor is the prediction-dissipation result of Still, Sivak, Bell and Crooks (2012): for a system driven by a stochastic environment, the nonpredictive information it retains — model complexity that does not improve prediction — lower-bounds the work it dissipates. One interpretive bridge: identifying "nonpredictive retained information" in that theorem with "map-territory divergence in the program's sense" is a defensible interpretive identification, not itself a theorem, and the result is proved for driven systems in stochastic thermodynamics, not as a universal law.

Selection is a statistical tendency, indexed to horizon (T5, T6). In an ecology competing for finite free energy, lower-divergence systems dissipate less on error correction on average, so the observed distribution drifts toward closer trackers (the Price equation supplies the formal bookkeeping; Price 1970). This is a tendency, never a law: a distorted model persists untouched in a stable, resource-abundant niche (the dark-room limit), and luck uncorrelated with fidelity can end any system. The crucial index is the horizon. The persistence-truth coupling is weak at short horizons and in stable niches, and tightens as the horizon lengthens and perturbations multiply, because the dark-room shrinks. At the cosmic limit — a structure trying to outlast its star, its galaxy, the cooling itself — it approaches necessity, because there is no stable niche left in which to hide a false model. It never becomes a law, because that limit is never attained. ("Entropy selects for truth" is this claim with the horizon index dropped; true in its regime, overstated as a flat slogan.)

Deception carries a standing cost (T8, the Consistency Tax). A deceiver maintains an accurate private map and a divergent presented one and a policy to deploy the second while acting on the first: strictly more description-length, and the held-but-unused map is nonpredictive, so it also carries dissipation. The description-length cost is analytic given that dual-map architecture, and T8 is scoped to intentional deception — an architecture that represents the truth and presents otherwise; honest error and self-deception are different objects with different cost structures. The open, empirical part is whether real deception always instantiates the dual-map architecture, and whether the cost survives when a system is trained to hide it. Recent interpretability work supplies both the support and the sharp test: linear probes on a model's internal activations distinguish honest from deceptive behavior with high reliability even when the outputs look benign — the dual-map signature, observed (Goldowsky-Dill et al. 2025) — while parallel work shows models can be trained to evade exactly such latent-space monitors (RL-Obfuscation, 2025; Neural Chameleons, 2025), which is the live question in its sharpest form: does obfuscation remove the cost, or relocate it?

No global optimizer, and this is a result, not a gap. Statistical boundaries nest across scales, but the maximal scale — the universe — has no Markov blanket, because it has no outside. Apply the agenthood criterion (below): the universe maintains no boundary, has no exterior it could fail against, runs no model on its own behalf. So there is no subject for a global "ought," no holder of a cosmic optimization target. The selection direction is always indexed to a bounded agent. Any reading of the descriptive frontier as "what everything ought, globally, to maximize" reintroduces exactly the domain error the metaethics dissolves, and is firewalled off.

The agenthood criterion (what makes a collection an agent, and blocks the move from "coupled" to "is an agent"): a collection is an agent only if it (1) maintains a Markov blanket, (2) has persistence conditions it can fail, and (3) runs an internal model deployed to keep itself from failing them. Coupling supplies at most (1). Two predators over a carcass, an ecosystem: tightly coupled, not agents. This gates every higher-scale claim in the program.

3. The trunk: meaning is physical, and reference needs relata

The metaethical layer is a special case of a more general semantic position the whole body of work runs on. That position has a secure half and a contestable half, and everything in §4 inherits the split.

First, the secure half. Facts about meaning are physical facts. This is not a new commitment; it is A1 applied to meaning. A meaning-fact is a fact about physical agents — their usage, their internal states, their couplings to the world. The denial would be Platonism about propositions: meanings subsisting with no physical realization, which A1 already excludes. This half sits at A1's confidence tier, and it costs the program nothing extra: it falls out.

Second, the audit principle. For an operator whose semantics is referential and relational — whose content is fixed by what it relates — a deployment that supplies no relata expresses nothing. Not something false: nothing. The body of work's standing test, name the relata or the claim has no content, is this principle in imperative form. The is-ought dissolution is its most developed application (operator: "ought"; relata: an agent and its persistence conditions). The far-edge audit (§6) is another (operator: causal vocabulary; missing relatum: any regularity for it to describe).

Third, the premise the audit rides on. The audit assumes that the operators it audits are referential and relational. That is a substantive semantic commitment, not a consequence of physicalism. A physicalist can instead hold a conceptual-role semantics, on which a term's meaning is its inferential role — a pattern of use, itself entirely physical — and on that view a term with no relata can still be contentful via its role. Such a critic can resist every ghost verdict in this body of work without leaving physicalism for a moment. The program's defense is functional: "ought" is not a connective whose job is internal to inference; its job is to bind an agent to an end and guide action, and a job description of that shape already contains the relata. That defense is real, and it is not consensus. So the foundation is two-tier: physicalism about meaning (secure, at A1's tier) plus a referential, truth-conditional reading of the audited operators (defended, contestable, and — together with claim (b) in §4 — the program's principal exposure).

Fourth, the open part. The theory of content — how a physical state comes to be about anything, the disjunction problem, reference indeterminacy — is unresolved in naturalized semantics generally, and the program's verdicts are conditional on that literature in the way every naturalist's are.

Self-application is consistent rather than vicious: "meaning is physical" is itself a physical claim about physical meaning-facts. The structure of that consistency, and what it does and does not buy, is §7.

4. The dissolution of the is-ought problem

What Hume showed is granted, read as the inferential fact it is: ought-vocabulary cannot be validly derived from is-premises alone. What is not granted is the conventional gloss — that this underivability reveals an autonomous normative domain standing beyond the descriptive (a gloss owed more to the later autonomy-of-ethics tradition than to Hume, whose own account of morals was sentimentalist). Granting that nothing entails a class of conclusions carries no commitment to there being such conclusions; the program reads the underivability as the edge of a term's domain mistaken for a chasm between realms. The denial then does more than respect the stricture: it explains it, in two halves. The bare categorical "ought" is underivable because it is contentless (claim (b) below): there is no proposition on the far side for premises to entail — the gap is real as a fact about inference and empty as a fact about the world. The agent-indexed "ought" is underivable from is-premises alone for the ordinary reason that no conclusion containing a term follows without a premise fixing that term's meaning, which is true of "uncle" as much as of "ought" and marks no metaphysical gap in either case. The program supplies that meaning-fixing premise — one, non-thermodynamic, contested (§3 and Part 2) — and it is the single crossing. No ought comes out of the physics alone; Hume's underivability comes out a theorem of the semantics, not a brute datum.

That premise splits, and the two halves carry different confidence.

Part 1, agent-relativity (a): "ought" is agent-and-end relative. Deployed with an agent and an end, it has content. This is close to consensus across the deflationary tradition (Anscombe 1958; Geach 1956; Foot 1972; even Mackie 1977 grants there are no objective free-floating prescriptions).

Part 1, the load-bearing half (b): deployed with no agent, "ought" is contentless, not false. There is no default or universal agent that rescues the bare "ought" — it is type-errored, the way "north" is undefined at the North Pole, not the way "tall" is merely incomplete without a comparison class. This is the claim the dissolution actually rides on, because the gap dissolves only if there is nothing on the far side, which requires emptiness, not mere falsity. And (a) does not entail (b): a relative term can take a default (the realist says the bare "ought" defaults to "all rational beings" and is perfectly contentful). So (b) is a substantive, contested claim.

The engagement. Against expressivism and most naturalism — the bulk of the field — the dissolution is secure: those positions already grant, or entail, that the free-floating categorical "ought" answers to nothing.

Against the realist, the program does not fight to a draw. The two realist shapes owe different bills. The non-naturalist rescues the bare "ought" by positing a non-physical normative fact for it to answer to, and in doing so takes on three debts: access (how a physical agent detects the fact), causal relevance (how it moves matter while appearing nowhere in physics), and individuation (what fixes its content). The program's verdict is that the standing attempts to pay those debts are weaker than the objections — and the program's own cost accounting compounds the verdict: a posit that buys no predictive power is retained divergence, carried at a standing cost with no offsetting fidelity. The naturalist realist — Cornell-style, or the Kantian with "all rational beings" — rescues the bare "ought" by supplying a default agent-class, and so occupies the unspecified-physical horn of the program's trilemma. That default is not free grammar. "Tall" takes a default because conversation supplies the class; the categorical "ought" was advertised precisely as binding with no supplied context, so the default must be argued into place rather than read off usage — and once a class is named, the original question reappears inside the rescue: why does that class's end bind? The trilemma prices all three horns — an unspecified-physical domain, a non-physical domain, or a non-referential reading, which concedes the dissolution in other words. It is a bill, not a refutation: it does not show the cost cannot be paid, and the realist genuinely occupies a horn.

The confidence statement, in full: the dissolution is secure against most of the field. Against the realist it presses a genuine advantage on the merits — the bill is presented and, by this program's audit, unpaid — rather than fighting to a draw. And the premise it rides on, (b), remains the program's named exposure: asserted at strength, defended by the operator's function and the trilemma, and deliberately held one tier below the maximal-induction tier reserved for A1, because that tier is earned by an unbroken empirical record, which a contested thesis in the philosophy of language does not have and must not borrow. The placement is calibration, not diffidence.

What the dissolution delivers is the negative result: the classical is-ought gap, posed as a chasm between description and free-floating prescription, is a pseudo-problem, because the free-floating side was empty grammar. That needs neither the thermodynamics nor Part 2. It is a dissolution, not a reduction: the program does not redefine "ought" as a persistence-predicate (that route faces the open question undissolved and is eliminativist about the folk "ought"); it shows the agent-free "ought" never referred, so there is nothing to reduce and nothing to bridge.

Where does the positive normativity of the agent-relative ought come from, if it is just persistence-structure indexed to the agent? It decomposes, with no extra ingredient. The force is the operator's action-guiding function (that is what "ought" is for), not a non-natural property. The content is thermodynamic, which is Part 2 below. The question "why be an agent at all" dissolves: it is askable only from inside persistence-directed activity, so it cannot be pressed from outside it. The residual "why persistence specifically" is not a separate hole — it is Part 2, the bet.

(One subtlety: whether the semantic premise is itself "descriptive" — a fact about how a word functions — or genuinely non-descriptive bears on whether the program strictly respects Hume's stricture or sidesteps it by a route Hume did not consider. A determined critic can press here; it is unresolved.)

5. The bets and the frontiers

Nothing in this section is a result.

Part 2, the live bet: persistence fixes the content of an agent-indexed ought. The end that fixes what an agent ought to do is persistence under thermodynamic constraint, grounded in A0. Its rival is not the expressivist or the realist but the flourishing-naturalist (Foot), who agrees the ought is agent-and-end relative and disagrees about the end. The two coincide across most cases and come apart on persistence-costly behavior (altruism toward the uncoupled, art, play, martyrdom). The confidence of Part 2 is the confidence of whether those cases can be carried, which is the crux below.

The positive ethics fixes a frontier, not a point — at two independent levels. Earlier statements treated coupling density as the single crux; there are two:

1. Intra-agent (the more basic one). Within the domain, ought-facts are determinate only up to the agent's persistence dimensions. Where an agent has one persistence dimension, a determinate optimum exists. Where it has several that trade off — energy now against robustness later, intensity against duration — physics fixes the Pareto frontier but supplies no unique point on it, and no coupling construct can, because the trade-off is between an agent's own dimensions and there is no coupling coefficient between an agent and itself. So T12 must be stated with its dimensional qualifier: determinate where one-dimensional, a frontier where not. The canonical statements that asserted a flat "determinate configuration" overstated.
2. Inter-agent (the κ crux). Coupling density κ measures how far one system's states carry predictive information about another's persistence. Whether a single canonical, measured κ can carry the persistence-costly inter-agent cases — and so vindicate persistence over flourishing as the content-fixer — is the central open question of the positive ethics. Closing conditions, both directions: a canonical κ that absorbs those cases would earn the unification; a κ that resists formalization would be evidence the moral territory is plural and the single-scalar persistence map is overcompressed. The standing danger: "hidden coupling at some scale" can rescue any case post hoc, which would make Part 2 unfalsifiable and empty; the only defense is that κ must be operationalized and measured, not tuned to save the thesis.

The inter-agent normativity gap, left open. The program does not adjudicate fully impartial sacrifice for an uncoupled stranger or an abstract principle. Such an ought is either reducible to coupling at some scale (and so not impartial after all) or a genuine instance of the universal-scope use the program calls contentless. It does not here decide which. Reproduction, note, is not patched by relocating the persister to "the lineage" — a lineage is not an agent. Reproduction is persistence-costly for the individual and is absorbed at the population level by selection, not at the metaethical level. The program does not need every costly behavior to be secretly individually rational.

Empirical status: nothing has been tested by an experiment built to test it. The descriptive layer has genuine falsifiable content — a core selection prediction, a controller-architecture prediction (any complex crisis-managing controller must implement a valence-like priority interrupt), and the deception-cost test, now sharp on both sides (§2). It borrows real support from independent physics (Landauer, experimentally confirmed; stochastic-thermodynamics results; the Price equation) and from evolutionary and interpretability literature. It has generated no test of its own. The weakest reach is the civilizational-collapse hypothesis (suppression severs telemetry, divergence accumulates as informational debt, collapse is non-linear): an analogical extension, untested, partly contradicted by the historical record in its strong form, defensible only as a comparative-probabilistic tendency, and gated on the agenthood criterion. It is the weakest claim in the program.

Consciousness, bracketed. The program gives functional accounts of valence (a non-ignorable global interrupt triggered by escalating divergence) and empathy (the monitoring of coupled telemetry where κ > 0). These explain what those signals do and why any complex controller needs them. They do not explain why there is something it is like to have them, and the program does not claim to. The hard problem is bracketed, not dissolved; the program selects no resolution of it.

A lead, not a result. A reframing exists — that the felt gap is the gap between being a physical state and modeling it, two physical relations to one state — but it reaches only the meta-problem and the access-limit, not the residue. The author's own leaning toward a graded, proto-experiential view is a personal position, fenced off, and no part of this case.

6. Where the map ends: below, beyond, ahead — and value at the end

The map has three edges. At each, the same rule: state where the description stops, and add nothing past it.

Below. From §0: beneath the functional bedrock lies a time-symmetric microphysics with no arrow and no agents, the arrow enters as a boundary condition (BP0), and the program's level of description is robust to whatever resolves the open foundations beneath it — the infrared physics is insulated from the ultraviolet completion. Frontier programs that would push thermodynamic or informational principles back down toward the foundation (entropic gravity, it-from-bit) are live contrarian bets; the program neither needs them nor leans on them, and their contrarian status confirms rather than threatens the consensus stack it builds on.

Beyond. The question "what grounds the laws themselves" splits into two proposals with opposite verdicts, and the audit of §3 runs at this scale exactly as it runs on "ought." A lawful meta-reality — a wider physics in which our universe's laws are local conditions; other separated pockets of low entropy; a multiverse — is deferred physics: internally coherent, possibly true, and, where causally quarantined from us, permanently outside empirical reach. Out of scope is not nonexistent; it is a verdict about what we can test, not about what there is. A lawless outside, invoked to explain where the laws "come from," is the ghost: every verb the invocation needs — produces, causes, gives rise to — imports the very regularity being denied, so the proposal unsays itself in the saying. The program therefore declines both consolations at the far edge: it claims no meta-reality and traffics with no lawless outside, and it needs neither.

Ahead. The descriptive program runs on the discharge of the low-entropy past, and that discharge ends — or it does not, and the difference is currently a live question in observational cosmology, so what follows is conditional on the branch.

If dark energy is a true cosmological constant, the end is hard, and the mechanism can be stated exactly. An eternally accelerating universe has an event horizon, and that horizon has a temperature — the Gibbons-Hawking temperature, of order 10⁻³⁰ kelvin for the measured expansion rate (Gibbons and Hawking 1977) — which the universe never falls below. Landauer's principle then fixes a positive floor under the cost of every irreversible operation, while the horizon quarantines a finite store of accessible free energy. Finite energy over a fixed minimum cost per operation is a finite total number of operations: a last thought, somewhere ahead, in that branch (Krauss and Starkman 2000).

But the branch is not settled. The DESI baryon-acoustic-oscillation program currently reports a preference for dark energy that weakens over time rather than holding constant: 3.1σ against a cosmological constant from DESI with the cosmic microwave background alone, rising to between 2.8σ and 4.2σ when supernova catalogues are added, depending on the catalogue (DESI Collaboration 2025) — suggestive, below the discovery threshold, methodologically contested in the subsequent literature, and the sharpest available hint as of mid-2026, with the survey's full five-year results expected in 2027. If the weakening is real and acceleration dies away, the far future has no event horizon and only an asymptotic approach to equilibrium, and in that branch the gradient argument alone cannot deliver finitude: a structure rationing a finite store at a per-operation cost that sinks as the universe cools need never exhaust it — Dyson's slow burn (Dyson 1979). The genuine obstructions there are different in kind: the discreteness of quantum states, which forbids halving the energy of a thought forever, and the alarm problem — a hibernating structure must keep a wake-up mechanism running, and the alarm is subject to every cost and failure mode it was built to wait out (Krauss and Starkman 2000). Dyson's rejoinder, that analog systems evade the discreteness bound, keeps the question contested rather than closed. So the statement is branch-conditioned: finite under a true constant or a Big Rip; trivially finite under recollapse; open and contested — leaning finite via discreteness — under weakening dark energy.

This branch-conditioning also strengthens an internal result. The claim that selection forces truth on an infinite horizon fails in both live branches, for complementary reasons: under a true constant, finiteness holds for accurate and distorted systems alike and so cannot select between them; under weakening dark energy, positive indefinite survival is on the table and the pressure is a tendency, not a law. Either way, "entropy selects for truth" is a horizon-indexed tendency, never a law — now for branch-specific physical reasons, not by stipulation.

Suppose the hard end is real. The reactive response is nihilism: it all ends, so none of it matters. There are two independent answers, with uncorrelated failure modes.

The first dissolves the inference. "Not eternal, therefore worthless" equivocates. Value was always agent-relative; nothing ever carried frame-independent, eternal, monument-like value, not before the heat death was on the table and not after. So "there is no eternal value" does not entail "there is no value" — it entails there is no value of a kind that never obtained for anything. The mourned object, eternal significance, is the same free-floating essence the program denies has any referent; its absence was always the case and changes nothing now.

The second out-competes the conclusion, granting the inference for argument. The nihilist grieves because they value something now and want it to endure; the despair is parasitic on a live valuation. But withdrawing — ceasing to protect or extend the valued thing — harms precisely what the valuation was about. Acting on the despair defeats the very end that generated it.

Both speak only to reactive nihilism, the despair finitude provokes, and have nothing to say to a structure that values nothing and feels no loss. The grief is the tell: it is the prior valuation making itself heard, which is exactly what both answers take as their premise.

This is the most emotionally loaded claim in the body of work and among the most secure, because it is purely conditional — an imperative resting on ends already held, not an ought conjured from the world. It does not depend on Part 2, on κ, on anything still being adjudicated. It would stand unchanged if the persistence bet failed entirely.

Which leaves a real, structured choice the physics sets up and does not decide: a finite store can be spent fast, for maximum intensity now at the cost of a nearer end, or throttled, for maximum duration at the cost of intensity. Burn bright or burn long. There is no agent-independent fact about which is right, because "right" here is agent-relative and the trade depends on what the agent values, which the physics does not contain. This is the firewall holding in plain view — and it is the intra-agent frontier of §5 made concrete: the description runs right up to the point where content would have to come from the agent, and then stops, because that is exactly where content comes from.

7. The loop

A last structural fact.

The framework is a physical model, running in physical agents, of the physics of physical agents that model. When it describes "a structure that models its environment in order to persist," the describing is an instance of the described. The document is a token of its own type.

Two kinds of circularity must not be confused here. Vicious circularity is a justification that presupposes its conclusion. The program does not argue that way: its premises are graded independently (§1, §3), its bets are adjudicated against rivals (§5), and nothing in its case for any claim assumes that claim. Self-instantiation is a claim turning out to satisfy its own standard: the domain diagnostic passes its own filters; the design spec is written to its own spec; the persistence story is told by a persisting modeler. At the limit, this is the map-territory relation closing over one special region of territory. For modeling structures, a perfectly faithful map of the structure would itself instantiate the structure — the map is made of the same kind of thing as its referent. That is real, and it is the precise way to hear the intuition that the whole account eats its own tail without contradiction: the universe, inside one of its eddies, running a model of the eddy-making.

And it adds nothing to the confidence of any claim. Self-consistency is necessary, not sufficient: a false but coherent self-model passes its own filters too. The loop explains why the program can describe itself without paradox. It is structure, not warrant.

8. What the program is, in one paragraph

A descriptive thermodynamic account of persistence and modeling, empirically falsifiable and so far untested by its own experiments, on top of which sits a metaethical layer that dissolves the classical is-ought gap by a semantic move that is secure against most of the field and presses a genuine, still-open advantage against the moral realist, and that bets — this is a bet, with a named and measurable crux — that persistence under thermodynamic constraint fixes the content of an agent's oughts, a bet that fixes a frontier rather than a point and leaves the fully impartial and the intra-agent multi-dimensional cases genuinely open. It is physicalist monism with thermodynamics as its functional bedrock; its semantic foundation — meaning is physical, and referential operators need relata — has a secure half and a contested half; it claims no global optimizer and no cosmic purpose because there is no agent at the maximal scale to hold one; it describes itself without contradiction and claims nothing from that fact; and it holds that value is real in the discharging and asks for nothing beyond it. The aim was never a map that is finished or certain. It is a map that is less wrong over time, and marks where it is still wrong.

Appendix: the layperson note

The same territory, in plain words. It compresses how much is said and how finely, never how confidently. Where the case above says "this is a bet" or "this is open," so does this.

Start with one fact: energy is spread unevenly, and it is always leveling out. That leveling is the deepest thing there is. Everything else is what happens on the way down.

The universe began almost perfectly smooth and unimaginably hot, which sounds like the opposite of useful but is the engine of the whole story. Space stretched faster than the heat inside it could even out, and that gap opened a vast reservoir of usable energy. As space cooled, the smoothness broke. Gravity pulled matter into clumps, and the universe became lumpy: blazing dense pockets, the stars and galaxies, surrounded by enormous cold empty voids. The difference between the hot pockets and the cold dark is what powers everything that happens next.

There is a window in the middle of this story. Early on it is too hot for anything complicated to hold together; far in the future it will be too cold and too spread out for anything to run on. In between, conditions are right, and something new appears: structures that keep themselves going by feeding on the energy flowing past them and dumping their waste heat back out. That is what a living thing is, told in physics. Evolution is just those structures getting better and better at the trick.

Eventually one of those structures gets good enough to model the world, picture the future, and plan over long stretches of time. That is intelligence, and it is a tool for one job above all: lasting. A simple creature only grabs the energy in front of it. To last a truly long time, a structure has to see far enough ahead to escape its own dying star. So intelligence is matter learning to keep its own order going against the long cooling of everything. Note the wording: it is the structure that persists, the agent, not "the universe" doing anything on its behalf. There is no universe-sized mind with a plan. There are only local structures, holding their shape for a while.

Two edges. Outward: there may be other universes, other separate "pops" of energy. We can never receive a single particle of light or scrap of information from them, so they are out of reach of any test we could ever run. That is not the same as saying they are nonexistent; it is saying the question is, for us, permanently out of bounds. Inward: every word we use and every fact we check only has grip on this physical world, the one we are in, because that is the only world anything we have ever measured has ever belonged to.

One more rule the whole story obeys: a word only says something when it points at something. "Ought" is such a word. Ask "what ought to be done?" with no one in view, and you have not asked a hard question; you have not asked anything. Ask "what ought this creature do, if it is to keep going?" and there is suddenly a fact of the matter the physics can speak to. Every ought is somebody's. That single move is the heart of the philosophical half of this work — and it is also where a serious critic pushes hardest; the main text says the same.

And notice who is telling this story. A story about structures that model the world to keep going, told by one of those structures. The telling is an example of the thing told. That makes the story consistent with itself. It does not, by itself, make it true, and the difference between those two matters enough to say out loud.

And the end. If the universe keeps speeding up forever in the simplest way, there is a hard floor, and a structure runs out of usable energy after a finite run. Recent measurements have made even that uncertain, so the truthful answer is: it might end hard, or it might not, and we do not yet know. Suppose it ends. Does that make any of it pointless? No, and the reason is exact. Nothing ever had the kind of value that lasts forever from no one's point of view; that kind of value never existed, before or after. What is real is value to someone, here, now, while it is happening. A song is not worthless because it ends. The point was never to make the song infinite. The point is how much is in it while it plays.

References

1. Albert, D. Z. (2000). Time and Chance. Harvard University Press. Verified · June 2026
2. Anscombe, G. E. M. (1958). "Modern Moral Philosophy." Philosophy 33(124), 1–19. Verified · June 2026
3. 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, 187–189. Verified · June 2026
4. DESI Collaboration (2025). "DESI DR2 Results II: Measurements of Baryon Acoustic Oscillations and Cosmological Constraints." arXiv:2503.14738. Verified · June 2026 · full survey results expected 2027
5. Dyson, F. J. (1979). "Time without end: physics and biology in an open universe." Reviews of Modern Physics 51, 447–460. Verified · June 2026
6. Foot, P. (1972). "Morality as a System of Hypothetical Imperatives." The Philosophical Review 81(3), 305–316. Verified · June 2026
7. Geach, P. T. (1956). "Good and Evil." Analysis 17(2), 33–42. Verified · June 2026
8. Gibbons, G. W., & Hawking, S. W. (1977). "Cosmological event horizons, thermodynamics, and particle creation." Physical Review D 15, 2738–2751. Verified · June 2026
9. Goldowsky-Dill, N., Chughtai, B., Heimersheim, S., & Hobbhahn, M. (2025). "Detecting Strategic Deception Using Linear Probes." Proceedings of the 42nd International Conference on Machine Learning (PMLR 267); arXiv:2502.03407. Verified · June 2026
10. Hong, J., Lambson, B., Dhuey, S., & Bokor, J. (2016). "Experimental test of Landauer's principle in single-bit operations on nanomagnetic memory bits." Science Advances 2, e1501492. Verified · June 2026
11. Jun, Y., Gavrilov, M., & Bechhoefer, J. (2014). "High-Precision Test of Landauer's Principle in a Feedback Trap." Physical Review Letters 113, 190601. Verified · June 2026
12. Krauss, L. M., & Starkman, G. D. (2000). "Life, the Universe, and Nothing: Life and Death in an Ever-Expanding Universe." The Astrophysical Journal 531, 22–30. Verified · June 2026
13. Landauer, R. (1961). "Irreversibility and Heat Generation in the Computing Process." IBM Journal of Research and Development 5, 183–191. Verified · June 2026
14. Mackie, J. L. (1977). Ethics: Inventing Right and Wrong. Penguin. Verified · June 2026
15. "Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors" (2025). arXiv:2512.11949. Verified · June 2026
16. Price, G. R. (1970). "Selection and covariance." Nature 227, 520–521. Verified · June 2026
17. "RL-Obfuscation: Can Language Models Learn to Evade Latent-Space Monitors?" (2025). arXiv:2506.14261. Verified · June 2026
18. Still, S., Sivak, D. A., Bell, A. J., & Crooks, G. E. (2012). "Thermodynamics of Prediction." Physical Review Letters 109, 120604. Verified · June 2026
19. Yan, L. L., Xiong, T. P., Rehan, K., Zhou, F., Liang, D. F., Chen, L., Zhang, J. Q., Yang, W. L., Ma, Z. H., & Feng, M. (2018). "Single-Atom Demonstration of the Quantum Landauer Principle." Physical Review Letters 120, 210601. Verified · June 2026

Standing of this document. This is the apex synthesis of the Thermodynamic Realism corpus: the whole case in one document, each claim at the confidence it carries. The descriptive layer is empirically falsifiable; the metaethical layer is defeasible by argument; the named exposures are the referential reading of the audited operators (§3) and claim (b) (§4); the bets and their closing conditions are in §5. On any conflict between this document and the modules, the modules govern. Corrections and counterexamples are welcome and change the map.

Author: Andraž Đurič, independent researcher, Slovenia. Written in collaboration with Claude (Anthropic); the corpus it synthesizes was developed in collaboration with Claude, DeepSeek, and Gemini, with contributions credited in the modules. Contributions are judged on content rather than origin; doing otherwise would be the genetic fallacy. Text licensed CC BY 4.0.