What a Description Answers To
Structural realists hold that the world’s structure is modal, and decline to say what it is the structure of. An account of what makes a pattern real, on which counterfactuals come with descriptions and production does not
Daniel Dennett’s best-known argument about what exists is about a toy. In Conway’s Game of Life a grid of cells is updated by one rule, and each cell is on or off depending on its eight neighbours a step earlier. One arrangement of five live cells reappears every four steps, one cell further along a diagonal. It is called a glider. Nobody who watches Life for a minute describes it cell by cell. They say that a glider is heading for a block and will be destroyed there.
Dennett asked whether the glider is real, and answered that it is. A pattern is real, he wrote, “if there is a description of the data that is more efficient than the bit map, whether or not anyone can concoct it” (“Real Patterns”, 1991, p. 34). James Ladyman and Don Ross built a metaphysics on that answer. In Every Thing Must Go (2007) they hold that to be is to be a real pattern, at every scale, and that what science describes is the world’s modal structure: what would and could happen, and not only what does.
I think the answer is right, and that it leaves out three things. All three can be seen in the toy.
First, the glider supports counterfactuals only as a glider. “This cell is on at step 40” reports an occurrence. “Had the glider started one cell to the left, it would have missed the block” compares two histories. To compare them one must say what is held fixed and what is varied: the same glider, displaced. Without the type glider there is no such comparison to make.
Second, the glider does no producing. Each cell’s next state comes from its neighbourhood by the rule. The glider’s next position comes from nothing but that.
Third, Life can run Life. There are arrangements of cells, thousands on a side, each of which behaves as one cell of a larger grid, and a grid of them carries out Life’s rule. A glider on the large grid is produced by large cells, which are patterns among the small ones. Whether something is the level that produces or a pattern above it depends on where one stands.
Ladyman and Ross warn against the toy. In Life, they say, “there is an unambiguous fundamental level,” and its patterns never cross-classify one another, so that “Life admits of complete decomposition; the universe does not” (p. 201, n. 12). The third point above is my answer to the first half of the warning. The second half is fair, and I bring other examples where it matters.
This essay states the kind of structural realism I hold, and says what would show it false. It agrees with Ladyman and Ross that patterns are real at every grain, that none is privileged, and that their structure is modal. It parts with them in two places. They decline to say what the structure is the structure of. I say that what a description answers to is production: one state coming from another. And they do not build the structure at one scale out of the structure at another. I hold that the modal facts at each level are constructed from the level below, so that modality is real at every level and brute at none.
The essay also settles a debt. Two earlier essays on this blog gave two different criteria for when a pattern is real, and nothing has yet said how they fit.
Two criteria
The first criterion came from a problem about computation. Hilary Putnam argued that every open physical system implements every finite automaton, given a free enough assignment of physical states to computational ones. If so, a rock computes whatever a brain does. In Causation, Computation, Indexicality we answered that the assignment must be a structure-preserving map:
A pattern is genuinely realised in a physical system when the realization relation is such a morphism. Causal succession at the pattern level maps onto causal succession in the physics. Intervene on the pattern and you intervene on the physics.
The test case was the dust man, a mind supposedly instantiated by scattered dust under the right interpretation. Its mapping is “an arbitrary lookup table — a gerrymandered mapping that preserves nothing.” Call this the morphism criterion.
The second criterion came from a dialogue between Michał Wójcik and his chatbot interlocutor, Thomas Epistemes, which The Grain of the Hard Problem reported. Thomas put it in one sentence: “A coarse variable earns reality when interventions and interactions repeatedly respect it.” Coarse-grainings, on this view, are “relative to families of interactions, not to observers.” Call this the intervention criterion.
The two answer different questions. The first asks which computations a system realises: a model is given and a map is sought. The second asks which coarse variables are real: a family of interactions is given, and it sorts the fine states into classes. Section five argues that they are the two halves of one definition.
Which structural realism
Structural realism began as a claim about what survives when theories change. Fresnel thought light was a vibration in an elastic ether. Maxwell’s theory has no such ether, and keeps Fresnel’s equations. John Worrall concluded that “there was continuity or accumulation in the shift, but the continuity is one of form or structure, not of content” (“Structural Realism: The Best of Both Worlds?”, 1989). That is a claim about what we know. It leaves open that the structure is the structure of something whose nature we cannot know.
The ontic version closes the opening. Ladyman and Ross define it as
the view that the world has an objective modal structure that is ontologically fundamental, in the sense of not supervening on the intrinsic properties of a set of individuals. (p. 130)
Their slogan is “There are no things. Structure is all there is,” which they call a first approximation. Their worked position is milder. There are objects, but they “have been purged of their intrinsic natures, identity, and individuality” (p. 131), and “the relata of a given relation always turn out to be relational structures themselves on further analysis” (p. 155). Steven French holds the eliminative view in earnest (The Structure of the World, 2014). Michael Esfeld and Vincent Lam hold a moderate one, on which “objects and relations (structure) are on the same ontological footing, with the objects being characterized only by the relations in which they stand” (“Moderate Structural Realism about Space-Time”, 2008).
The variants differ on five questions, and my answers are these.
Are there objects? Yes, and relations individuate them. This is the moderate position. Section seven shows it in a case one can compute.
At which scales does the view apply? At every scale. A theory is true of a regime, and its patterns are real there (Ladyman and Lorenzetti call this effective structural realism).
How are scales related? Here I keep half of what Ladyman and Ross say. Patterns at every grain are equally real, and they do not form a ladder. Thomas’s phrase for it was “an atlas of overlapping operational quotients.” What I drop is the other half. Ladyman and Ross “deny that everyday or special science real patterns must be mereological compositions of physical real patterns,” and “deny the local supervenience of the table on a real pattern described by physics” (p. 253). In a later reply to critics they commit only to special-science structures being partially explained by fundamental physics (Metascience 19, 2010). I hold that every real pattern is fixed by any admissible finer description of the same system. Sections five and six say what “admissible” means.
What is the modality? Real at every level of patterns, and brute at none. That is the next section.
Is there a bottom level? Ladyman and Ross offer as a “tentative metaphysical hypothesis … that there is no fundamental level, that the real patterns criterion of reality is the last word in ontology” (p. 178). I make no claim about levels below the lowest one we could have access to. What that level answers to I call a base, and a base is not a level. Sections three and four explain.
A view with this many moving parts invites Kerry McKenzie’s question. She points out that structural realists disagree about whether quantum fields are objects or structures, and that they do not treat the disagreement as urgent. She concludes that the view “is best understood not as a thesis or doctrine concerning how the world fundamentally is, but rather as a ‘stance’” (“Structuralism as a Stance”, 2024). The fair test is what the view rules out. Section eleven lists six things.
Counterfactuals come with descriptions
A stone is thrown at a window and misses. Had it not missed, it would have broken the glass. Whether that is true depends on the ontology of stone-throwing: on what counts as the same stone, the same throw, and a nearby trajectory. The facts that make it true are fixed lower down, by the mechanics of glass and stone, and those by the physics beneath. At each step the modal fact is constructed from the level below, and the construction uses the upper level’s types.
So modality is not strongly emergent at any level. There is no modal fact about stones that the lower level leaves open. Nor is it an illusion. The stone would have broken the window.
The construction has to stop somewhere, and the question is what it stops at. It cannot stop at more modal structure. A counterfactual holds something fixed and varies something else, and both need types. Modal structure is therefore structure under a typing, which is to say structure as described. If what lies beneath every description were itself modal, it would be a description of something further, or a description with nothing described. To make modality primitive is to put the description where the thing described should be.
Ladyman and Ross see part of this. “Merely listing relations among locators does not state anything with modal force,” they write. “Therefore, it doesn’t specify structure in our sense.” Then they ask the question themselves:
What makes the structure physical and not mathematical? That is a question that we refuse to answer. In our view, there is nothing more to be said about this that doesn’t amount to empty words. (p. 158)
I think there is more to be said, and it is not much. What the lowest description answers to is not modal. A Humean agrees so far. For the Humean it is occurrences in an arrangement, each a distinct existence, and laws are the best summary of the arrangement. What the Humean picture lacks is that the occurrences come from one another.
By production I mean that. In a system generated by a rule, an application of the rule takes these inputs and yields these outputs, and the outputs are individuated as its products. An occurrence exists as generated. This cannot mean that the product would not have existed without the producer, since that is a counterfactual and belongs to a description. It means that what the occurrence is depends on what it came from. Hume held that there are no necessary connections between distinct existences. I do not assert necessary connections. I deny that occurrences at a base are distinct existences.
Two routes lead here, and they are one route. The first is constructibility: a totality given all at once is a structure, and so is on the side of description, while an occurrence has to be reached. The second is elimination. With no intrinsic natures and no space given in advance, what is left to relate occurrences is that one comes from another, and that is the structure of rule applications.
A base, then, is the generation that the lowest accessible level of description answers to. It is not one more level. Levels are on the side of description. There is a level wherever some bounded system could have a typing respected by what it interacts with. Below the lowest such level I say neither that there are further levels nor that there are none.
A base differs from every level above it in two ways that go together. It has no counterfactuals, and it has no types of pattern. The modal facts of physics reduce, at a base, to sets of occurrences, weighted by whatever measure turns out to apply. There is no further counterfactual structure there, only the generative one.
This does not make patterns depend on minds. By a description I mean a level fixed by a family of interactions. A system whose parts respond only to certain differences in one another carries its own typing, whether or not anyone describes it. Section seven comes back to this.
The disagreement with Ladyman and Ross is now exact, and it is an inversion. They write:
For ‘modal’ read ‘nomological’ if you like. We do not take it to be ‘causal structure’, for reasons to be explained, but will argue later that causal structure is the pragmatically essential proxy for it in the special sciences (but not in fundamental physics). (p. 130)
For them modal structure is fundamental, and causation is a bookkeeping device of observers, “a notional-world construct” (p. 279). For me production is what there is at a base, and modality enters with patterns. At the level of patterns we agree. There modality comes first, because the modal profile of a pattern fixes what counts as that pattern. And there causation is a matter of what makes a difference to what, which section eight takes up.
A second disagreement holds at every level. A view with no base has two ways to avoid brute modal facts. It can say that each level is constructed from the one below, without end. Or it can say that there is one modal structure of the world, primitive as a whole. Ladyman and Ross take neither in a constructive form. They reject composition and local supervenience, and they reject the label of emergentism too (p. 56). What remains between scales is a constraint, that special sciences may not contradict physics, and partial explanation. Each scale then carries some modal content that no other scale fixes. That is what “brute at none” denies.
On this point the essay is closer to Esfeld, who argues that “the fundamental physical structures possess a causal essence” (“The Modal Nature of Structures in Ontic Structural Realism”, 2009), and who defends with Christian Sachse a reduction of the special sciences that eliminates none of them. I differ from them in three ways. No base is singled out. A reduction is a construction carried out under the upper level’s types, and not a deduction in the lower level’s vocabulary. And not every construction counts, as section six explains.
Two old objections to structural realism get short answers here. Bas van Fraassen put the first: “if there is no non-structure, there is no structure either.” If structure is all there is, the word contrasts with nothing. If structure is instantiated, something bears it. I take neither horn. Structure contrasts with production, and production is not a bearer with a nature of its own. M. H. A. Newman put the second in 1928: any collection of the right size can be given any structure, so a claim that the world has a certain structure says almost nothing (Newman 1928). The reply needs relations that are not arbitrary sets of pairs. At a base those are the generating relations. Above it they are the modal relations that an admissible map preserves.
Bases, and where we are
Life is not our world, and nobody knows which rule, if any, is. Stephen Wolfram’s proposal is to stop asking. Take every rule, applied in every way to every initial condition, and consider the whole of what is generated. He calls it the Ruliad (Wolfram 2021). Bounded observers embedded in it sample a region, and physics is what they find there.
I take from this that all bases are real and none is singled out. Production is primitive, and it is not the same from one base to another. Physics is the pattern that emerges in the region where we are.
A base is a process, not a frame. Wolfram sometimes treats the choice of rule as something like a choice of reference frame. A frame is a description. If bases were descriptions, their production would fall on the modal side by the last section’s argument, and the regress would start again. A base generates. Any base can be emulated in another, and there it is a pattern with a modal profile. That is the lesson of Life running Life: two concrete occurrences of one organization. Each produces when taken in itself, and the second is also a pattern in the first.
This gives concrete a meaning. To be concrete is to be located in a lower level. The large glider is concrete in the large cells, and the large cells in the small ones. The whole is neither concrete nor abstract. It is not concrete, because no base is God’s pick. It is not abstract, because it is not something we have access to. The Ruliad is not a determinate structure that anyone knows. We know it as structure only through Wolfram’s programme: physics turns out to be such that a range of structures of that kind would support it.
In Presence and Actuality I withdrew the claim that perspective constitutes actuality, and kept that an indexical account “can locate us within a plurality.” The phrase “the region where we are” does the second and not the first.
The picture makes a demand on physics. An observer spans a region and cannot tell which of the region’s bases underlies it. So the layer of physics has to be stable under a shift of base: the same pattern over each base in the region. This is a requirement, and it has not been shown to hold.
The Standard of Motion drew the distinction that shows how much is being asked. Passive invariance is the independence of a description from the observer’s frame. Active invariance is the existence of the transformed thing: for a body at rest, a moving twin. Life shows the second failing. The ships nearest a glider in speed are different patterns, and “a small boost of a glider is not a slightly faster glider but no glider at all.” For a shift of base the two kinds are these. That rules can emulate one another is passive. That physics exists, as the same pattern, over each base of a region is active. The first is known. The second is a theorem nobody has.
A reader who rejects Wolfram’s picture can keep most of what follows. Sections five to eight need only that there is a lower level. Sections three, four, nine and eleven lean on the picture, and they are the speculative part.
The criterion
A pattern is real, relative to a family of interventions, when a model of it is a causal abstraction of the system under an admissible map.
The terms come from the theory of causal models. A causal model has variables, and says how each depends on the others. An intervention sets some variables and lets the rest follow. An abstraction of a fine model by a coarse one is a pair of maps. The first takes fine states to coarse states. The second takes fine interventions to coarse interventions. The pair must commute: intervening on the fine system and then mapping its state gives what one gets by mapping first and intervening on the coarse model. Paul Rubenstein and colleagues gave the definition in 2017 (arXiv:1707.00819), Sander Beckers and Joseph Halpern tightened it (arXiv:1812.03789), and Atticus Geiger and colleagues made it the foundation of a method for interpreting neural networks (arXiv:2301.04709).
The two criteria of the earlier essays are the two maps. The morphism is the map on states. The family of interactions is the domain of the map on interventions. The sentence “Intervene on the pattern and you intervene on the physics” already says that the diagram commutes. Geiger, Jacqueline Harding and Thomas Icard state the identification outright for computation: a system implements a computational model “just if” the model is an abstraction of the system, under a suitable class of translations (arXiv:2508.11214).
The two questions remain, as two uses of one relation.
In the first use a model is given and a map is sought. The dust man fails. A table can assign states of a computation to states of the dust, but nudge the dust and the assignment has to be written again. A trained network passes. Change a weight, and the activations change as the model of the network says they should.
In the second use a family is given, and it sorts the fine states. Temperature passes. Setting the temperature of a gas has one effect on its pressure, whatever the molecules are doing. Total cholesterol fails. Setting it has no single effect on heart disease, because the effect depends on how the total divides between two kinds of cholesterol that act in opposite directions. Rubenstein’s paper draws the consequence for the first kind of case: “we can view the Y-variables as causal entities, rather than only functions of underlying ‘truly’ causal entities.” It also notes the relativity: “If we were to allow larger sets of interventions … the transformations given would not be exact.”
The two halves are not equal in strength. Fernando Rosas and colleagues distinguish three ways a coarse level can be closed (arXiv:2402.09090). It is informationally closed if its own past predicts its future as well as the fine past does. It is causally closed if interventions on coarse variables suffice to guarantee outcomes on other coarse variables. It is computationally closed if the coarse process is a coarse-graining of the fine one, so that a diagram commutes. They prove that the first two are equivalent and that they imply the third. The converse fails. A coarse-graining can commute with the dynamics and still be too coarse to be the variable that control acts on. So the intervention half is the stronger, and the morphism half follows from it.
The family of interventions is the modality of section three. Which interventions belong to the family fixes which counterfactuals the pattern supports, and that modal profile is what individuates the pattern.
One step is left, at the bottom. A base has no counterfactuals, so there are no interventions on it, and the criterion cannot be applied to what the lowest description abstracts from. What that description answers to is repetition. A rule types many actual applications, to many different local configurations, across the region. Its counterfactuals about configurations that never occur here hold by its generality, which is what a typing is. A Humean says the same of the best summary of the world’s events. The difference is the one from section three. The applications are generated, each from others, and are not independent items that a summary happens to fit.
Exact abstraction is rare. An approximate one needs a tolerance, and a tolerance looks like a choice. The family sets it. A difference that no intervention in the family can detect is no difference at that level.
Three trivialities, one cure
Structure-preserving maps are cheap, and three literatures have found this out separately.
Newman found it for structure in 1928. Putnam found it for computation in 1988: with a free enough map, any open system implements any finite automaton (Representation and Reality). Denis Sutter and colleagues found it for causal abstraction in 2025. They prove that with unrestricted maps any algorithm is an abstraction of any neural network meeting mild conditions, and they exhibit maps that “reach 100% interchange-intervention accuracy” on language models whose weights are random (arXiv:2507.08802). Geiger, Harding and Icard call this “an even stronger triviality argument”: “provided we allow C to be completely unconstrained, every neural network implements every algorithm.”
Each literature also found a cure, and each cure restricts which maps count. David Chalmers required that the parts of a state vary independently. Peter Godfrey-Smith required that the physical states grouped under one formal state be physically alike. Tyler Millhouse bounded the complexity of the map (“A Simplicity Criterion for Physical Computation”, 2019). Neal Anderson and Gualtiero Piccinini required that physical states “bear neither more nor less information about the evolving computation than do the computational states they map onto” (The Physical Signature of Computation, 2024). Interpretability researchers restrict maps to linear ones, or require that the map keep working on new inputs.
I propose one clause that these approximate. A map is admissible when it is stable under any shift of the lower level that the family of interventions cannot detect.
Take a system and change it beneath the level in question, in a way no intervention in the family would register. Put the dust in a different arrangement with the same gross distribution. Retrain the network from a different random seed. Run the same physics on a different base. An admissible map still works. An inadmissible one has to be made again.
The dust man’s table is rewritten at every nudge. Sutter’s maps are fitted to one network’s weights, and say nothing about the next. A map from a trained network’s activations to the steps of an algorithm is admissible when it survives retraining and perturbation.
The clause is section four’s requirement on physics, applied at every level. It matches what Christian List and Peter Menzies call a “realization-insensitive” relation (“Nonreductive Physicalism and the Limits of the Exclusion Principle”, 2009). It explains why Millhouse’s simple maps do well: a simple map has little room to be fitted to the accidents of one realization. I take simplicity as evidence of stability, and stability as the criterion.
Stability under shifts is a modal notion, and the clause is used to say which modal structures are real. That is a recursion and not a circle. The shifts in question are shifts of the lower level, whose modal facts are constructed from the level below it, and the descent ends in the repetition of a rule’s applications.
The clause could fail in three ways. Stability comes in degrees, so admissibility does too. A map could be stable for a reason that has nothing to do with the pattern. And linearity, which interpretability practice relies on, is an empirical bet about networks. It does not follow from the clause.
Whose interventions?
Compression was Dennett’s test, and James McAllister has argued that it tests nothing. “Any empirical dataset of length L can be decomposed into any one of all patterns of length L and the associated noise term.” How much noise to tolerate is a convention. So “if any pattern is real, then all patterns must be” (“Is Any Pattern Nonreal?”, 2026).
I give up compression as the criterion. McAllister’s argument concerns data that are passively collected, and it does not reach a criterion stated in terms of interventions. His own illustration shows why. He imagines the other causal factors rearranged so that a bogus pattern becomes the best compressor of the new data. The real pattern is the one that persists through the rearrangement. The tide is in the sea-surface heights whatever the winds do. A component stipulated to fit one season’s record is gone in the next.
At a base McAllister is right, and I accept his conclusion there. Every base is real, and whatever a rule generates is among what there is. Selectivity enters with a family of interactions. Given a family, which patterns pass is not up to anyone.
That leaves the question of whose family it is. It is not an interpreter’s. Ladyman and Ross require that a real pattern be “projectible from a perspective that physics tells us could be physically occupied” (p. 236), and the earlier essay said the same of Wolfram’s observers: an observer “is not an interpreter imposing a description. It is a bounded system embedded in the Ruliad, able to enter only certain interactions with the rest of it.” An observer is one more embedded system, and passes or fails the criterion like any other.
For a system that operates on itself, the family that matters is the interactions among its own parts. The grain of an experience, we wrote, is “the grain that the vehicles’ own regulation respects.” No outside observer is needed to fix it.
A neural network gives a case where all of this can be computed.
The function a network computes fixes its hidden representation only up to an invertible linear map. Nelson Elhage and colleagues put it plainly: “We could transform the residual stream by an arbitrary full-rank linear transformation, and then also multiply the same transformation into every other matrix … and arrive at an identical function with completely different coordinates” (“Privileged Bases in the Transformer Residual Stream”, 2023). Zhou and Klindt prove the general form: the best a classifier can do is “recover the latent variables up to linear mixing” (arXiv:2610.20851). So the question “which direction is the feature?” has no answer in the input–output behaviour.
Four things can supply one.
An elementwise nonlinearity can. The linear maps that commute with a ReLU layer are permutations of the axes combined with positive rescalings, and nothing else (Godfrey et al., Theorem 3.3). An energy cost on non-negative activations can. Zhou and Klindt show that the least-energy solution is “identifiable up to permutation and component-wise positive scaling.” Both belong to the network’s own operation, and the axes they pick out are real for the computation.
The optimizer can. Adam normalizes each coordinate’s update separately, “and thus privileges the basis that the weights are stored in.” But Elhage’s paper also finds that “Transformers do not rely on a privileged basis to train and function properly.” So the basis is real for the training process and idle in the computation. Being selected by a system’s dynamics is not enough. The interventions that respect the pattern have to be interventions on the process in question.
An analyst’s sparsity penalty can, as in a sparse autoencoder. It counts when the network’s own operation respects the directions it finds, which Zhou and Klindt tie to a condition on the data.
In every case one symmetry is left. Swap two neurons, together with their weights, and nothing changes. So the set of axes is real, and no axis has an identity beyond its place among the others. That is the moderate position on objects from the second section: relata exist, and relations individuate them.
Life cannot show the last thing this section needs, which is patterns that cut across one another. Thomas’s example was a cell membrane, whose electrical organization and mechanical organization preserve different invariants. Ladyman and Ross give natural selection, which “sculpts the dynamics of populations even though no individual animal was ever brought into being or killed by natural selection” (p. 203). Each of these is real relative to its own family, and neither is a coarsening of the other. The patterns of a system form an atlas and not a ladder.
Causes above the base
“Cause” has two senses, and Ned Hall separated them (“Two Concepts of Causation”, 2004). In one, a cause produces its effect. In the other, a cause makes a difference to its effect: had it been otherwise, the effect would have been otherwise.
Production belongs to bases. Difference-making belongs to patterns, and James Woodward’s interventionism is its theory. Particle physics describes patterns, so its causes are difference-makers too. Ladyman and Ross say that the universal patterns of fundamental physics “are not causal. They are instead structural” (p. 283). I agree about physics as a description. What I add is that the description answers to production, which physics does not mention because no description does.
This settles the exclusion problem in the one way it can be settled. Jaegwon Kim argued that if a physical event has a sufficient physical cause, a mental cause of it is either the same thing or idle. The argument is about production. Patterns do not produce. So they compete neither with one another nor with a base. Among difference-makers there is no competition to begin with: the stone’s momentum and the motions of its molecules both make a difference to the window. I follow Woodward here and not List and Menzies, who hold that a higher-level difference-maker can exclude a lower one.
Proportionality, the fit between a cause’s grain and its effect’s, is a matter of degree and is relative to what one wants explained. It picks the apt description for a question. It does not pick the one real grain.
Three earlier statements need rewording in this light.
Presence and Actuality answered Trenton Merricks, who denies that composite objects exist when their causal powers are redundant with those of their parts. We said the redundancy was “between two descriptions of one causal pattern rather than between two causes.” Better: there is one production and several difference-making descriptions that answer to it. Existence is not earned by producing something the parts do not. On that test nothing above a base would exist, particles included. It is earned by the criterion of section five.
Causation, Computation, Indexicality said that pattern-level causation “does not add productive power to the generative primitive.” That stands, and now has a name.
Giulio Tononi and colleagues, defending free will from Integrated Information Theory, offer a choice (arXiv:2206.02069). Either a system has intrinsic causal powers at one privileged grain, or the micro level “exists as such” and everything higher is “carried along for the ride.” I take neither. Higher levels are carried along, as far as production goes. So is the micro level.
One objection remains. Laura Franklin-Hall has shown that interventionism alone does not say which variables to use: it “remains maximally open about what sorts of things might be related as cause and effect,” and “some instrument—however mysterious … is required to locate the explanatorily apt variables in the bottomless interventionist pit” (“High-Level Explanation and the Interventionist’s ‘Variables Problem’”, 2016). The instrument offered here is admissibility together with the system’s own interactions. She would suspect naturalness under another name. It is not a primitive ranking of properties. It is a pair of facts that interventions can test: which differences the parts of a system respond to, and which maps survive a change beneath them.
What makes a thing one
Michał Wójcik has pressed a demand on accounts like this one. A computation, he says, must be “physically legible to mindless nature herself.” His interlocutor Thomas sorted the possible answers into three tiers: a coarse-graining that someone describes, an organization that is operationally individuated, and an individual possessing esse, the act of being that Aquinas distinguished from what a thing is. Presence and Actuality granted the second tier and could not see what the third adds: “we cannot grant that it exists and then ask whether it additionally possesses existence.”
The first half of the demand has been met. The system’s own interactions carry the typing, and nature needs no reader.
The second half can now be given some content. A whole can be one in three ways, each in degrees.
Integration. The parts are not closed off from one another. Cut the whole along the boundaries that lose least, and ask how much of its organization is lost. Is the Monastery Minded? called this residual integration.
Autonomy. The whole is closed off from its surroundings by its own activity. Peter van Inwagen’s criterion for composition, that the parts constitute a life, is of this kind.
Closure. The level’s own typing suffices for its own dynamics. Its state and its rules fix what it does next, with nothing about the finer detail needed. This is Rosas’s closure, and the admissibility clause in another guise.
An organism is high on all three. A trained network running on a processor is integrated at the grain of its computation and closed there, and has little autonomy: it maintains nothing. A nation has thin integration beyond that of its households and institutions. A heap has none of the three. What the tradition calls a substantial unity is high on all three, and an aggregate fails at least one.
Closure is where esse finds a place. A perfectly closed level is a base when taken in itself. The large cells of Life running Life are the case: nothing about the small cells is needed to say what a large cell does next. So a pattern has production of its own in the limit of closure, and by degrees short of that limit.
This sharpens the disagreement with Michał without ending it. For Aquinas each finite substance has its own limited act of being. Here production belongs to bases, and a person has it to the degree that the person’s level is closed. A Thomist will answer that a degree of closure still describes what a thing is, and not that it is.
Two rival accounts insist on the line between a unity and a heap, and both want it drawn once. For Robert Koons’s hylomorphism “no two substances overlap” (“Staunch vs. Faint-hearted Hylomorphism”, 2014). For Integrated Information Theory one grain of one substrate exists, and “no subsets, supersets, or parasets of that complex also exist.” The three dimensions draw the line by degrees and allow wholes to overlap.
The account is weakest at closure. No open system is exactly closed. And I do not know whether the three dimensions reduce to one measure.
The neighbours
| View | Where I agree | Where I part |
|---|---|---|
| Powers (Bird; Esfeld) | What is basic is causal | No bearers of the powers, and no single fundamental level |
| Humean supervenience (Lewis) | Among patterns, laws describe and do not govern | The occurrences described are generated, and are not distinct existences |
| Modal realism (Lewis) | The modal structure of a pattern is objective | Not every possibility exists |
| Russellian monism | Physics describes structure | Nothing non-structural lies beneath it |
| Hylomorphism; Integrated Information Theory | A unity differs from a heap | By degrees, and wholes overlap |
| Interventionism (Woodward) | It is the account of causation among patterns | It needs admissibility to fix its variables |
Two rows need more than a cell.
The Humean is the nearest neighbour. For David Lewis the world is a mosaic of local matters of fact, “and that is all.” A law is a line in the best summary of the mosaic. Among patterns I agree that laws describe. At a base the Humean account of production cannot hold. It says that one event produces another when events of the first type are always followed by events of the second. Types belong to descriptions, so a regularity is not what a base consists of.
The same point removes the principle that gives Lewis his plenitude of worlds. The principle is free recombination: a duplicate of anything can coexist with a duplicate of anything else. It needs items that are what they are apart from their relations, so that a duplicate can be lifted out and set down elsewhere. An occurrence individuated by what it comes from cannot be lifted out. And since existence comes by generation, describing a combination does not make one. Presence and Actuality already used the argument: “Describing arbitrarily intense or prolonged suffering does not prove the possibility of its full embedding.” The argument removes the principle and perhaps little of the abundance. With every rule and every initial condition, much of what can be finitely built is built somewhere.
Realism about modality is therefore not modal realism. Take the stone again, and keep three things apart. What would have happened: the window breaks. That is fixed by the modal structure of stone and glass, patterns that are real by a criterion which actual interventions test. What is possible: the stone’s hitting is possible if a support for it could be rebuilt, level by level, down to a base. I mean this metaphysically. A metaphorical God could do the rebuilding. What we do is imagine it in a toy model, which is a proxy and can skip a step. What occurs somewhere: the rebuilt situation may be generated in some region, with a measure that may be vanishingly small. The counterfactual is true and the situation is not actual, and whether it occurs elsewhere bears on neither.
Russellian monism holds that physics gives only structure and that something with an intrinsic nature underlies it. The next essay argues against it, through Torin Alter’s defence of the knowledge argument.
A thesis, with a hostage
McKenzie makes four charges. Dependence of objects on structure does not make structure prior. The symmetries of our theories “do not determine how many particles there are in nature,” nor the values of their defining properties, so structure does not fix the kinds. Structuralists tolerate disagreement about whether fields are objects, so the view is a stance. And no naturalistic project can speak about fundamental reality before there is a fundamental theory.
The view stated here rules out six things.
- Categorical natures beneath structure.
- Modal structure with nothing non-modal for it to answer to. Existence by description, as in free recombination, is one form of this.
- A singled-out base, including a privileged level of particles.
- Productive power at the level of patterns.
- Exclusion by maximality: the claim that one grain of a system exists and the others do not.
- Brute necessities: modal facts at a level that cannot be constructed from the level below.
It also has a hostage. Section four requires that physics be the same pattern over each base of a region, and that is a theorem nobody has proved.
Her second charge does not apply, since nothing here claims that symmetries determine the kinds of particle. Her fourth is met in part by scope. The commitments about patterns are indexed to regimes and can be held before physics is finished. The claims about bases are the exception. They are the speculative part of the essay, and I have marked them.
One of her earlier arguments applies, weakly. She showed that arguments from physics against a bottom level hold the principles of the physics fixed while the levels descend, which yields “a sort of ‘half-way house’ form of anti-fundamentalism.” Something is held fixed here too: that what there is is generated, within the bounds of what can be computed. No one rule is held fixed. And the essay does not say that levels descend without end. It makes no claim below the lowest level we could reach.
What would change my mind
- A proof that physics is not the same pattern across the bases of a region. A derivation, for rules like ours, of friction or of swerving in free motion with nothing to protect inertia would be a step toward one.
- A variable that interventions respect and that no admissible map carries to a model, or a model with an admissible map whose variables interventions do not respect. Either would separate the two halves of the criterion.
- A map that is stable under every undetectable shift and identifies a pattern we would all call bogus.
- A basis that a system’s own operation respects and that does no work under any intervention on that operation.
- A counterfactual that needs no typing.
- A system high on integration, autonomy and closure that nobody would call one thing, or one low on a dimension that is plainly a unity.
The glider
The glider is real, relative to the interactions that respect it: a block will stop it, an eater will consume it, and another glider will collide with it in one of a few ways. It supports counterfactuals because it falls under a type. It occurs because cells produce it. On the large grid of Life running Life, the cells that produce it are themselves patterns.
Two questions are left for the essays that follow. One is what this view owes to the argument that structure cannot be all there is, since we know what red is like and no description of structure tells us. The other is what makes an experience lived, and not only real.
Draft notes (remove before publishing)
Draft of 17 Nov 2026, from the outline in
research/structural-realism-essay-outline.md. Not
committed, not in the index.
To verify before publishing:
- Life run inside Life: the unit-cell constructions (the OTCA metapixel and its relatives), their size, and whether the emulation is exact. The opening and two later sections rest on it.
- The total-cholesterol example, usually credited to Spirtes and Scheines (2004).
- Wolfram’s wording on a choice of rule as something like a reference frame.
- Lewis’s wording for Humean supervenience (“and that is all”) and for the principle of recombination, and the references.
- Ladyman and Ross’s reply in Metascience 19 (2010): 179–184: the DOI and the exact sentence on partial explanation. The essay paraphrases it.
- Esfeld and Sachse: the reference for conservative reductionism.
- Rosas et al.: an example to go with the three closures. The outline has the count of particles in two gas chambers against an arbitrary partition of micro-states; I left it out because I have not checked it in the paper.
- McAllister: that the tide and the stipulated component are his own examples, and the page for the rearrangement of causal factors (the dossier has p. 149).
- The quotations from Tononi et al. (“exists as such”, “carried along for the ride”, “No subsets, supersets, or parasets”) against version 5 of the paper.
- Kim: a reference for the exclusion argument.
- Chalmers and Godfrey-Smith on implementation: references are in the dossier’s §5 and are not linked in the text.
- Every Thing Must Go, sections 5.1–5.4 and chapter 6, are unread. The essay’s account of their view of causation rests on pp. 130, 159, 257–258, 279–283.
- Whether the architecture series has a public page to link from “Whose interventions?”.
Claims the author has not confirmed in these words:
- That the family of interventions sets the tolerance for approximate abstraction.
- That simplicity of a map is evidence of stability, with stability as the criterion.
- The reply to Franklin-Hall (admissibility and the system’s own interactions are testable facts, not naturalness).
- The answer to the first half of Ladyman and Ross’s warning about Life, and the concession of the second half.
- “The modal facts of physics reduce, at a base, to sets of occurrences, weighted by whatever measure turns out to apply” renders the author’s “reduces into sets at the base level, or what the reality fluid measure turns out to be concerned with”.
- The three-dimensional account of unity is the version accepted on 15 Nov; the examples (network on a processor, nation, heap) are Claude’s.
- The reply to van Fraassen (“structure contrasts with production”).
Left out, from the outline: the Thomist resonance of section three’s argument; Dasgupta on symmetry and detection; the withdrawn Lewis-style reading of counterfactuals; the note that causal abstraction is rarely exact in practice (Beckers, Eberhardt and Halpern on approximation).