What we should believe about Unobservables

Introduction

Early modern science didn’t begin as a neat, prediction-first enterprise. It inherited the older ambition of natural philosophy: not only to describe what happens, but to say what’s really there and what causes it—what sits “behind” appearances. The ideal picture was: find the right basic furniture of the world, write down the laws it obeys, and then explain events by showing how the laws plus initial conditions yield them. That aspiration wasn’t a side-hobby; it was often treated as the proper aim of a unified science.

By the nineteenth century, that aim came under pressure from inside physics itself. Some scientists argued that the drive for deep causal stories had pushed theory into metaphysical overreach—strong claims about hidden mechanisms that outpaced what experiments could actually control. Instead of “peering beneath the veil,” they urged a more disciplined goal: build theories that organise phenomena and guide expectations without pretending to read off the world’s ultimate furniture. In that setting, Mach’s slogan—science as an economy of thought—lands as a methodological proposal: treat theories as tools that simplify and coordinate experience, economising memory and effort, rather than as literal pictures of what reality must contain. [1]

The realism/anti-realism debate is one way philosophers later sharpened this tension. Scientific realists tend to treat the success of mature theories as a reason to believe (at least approximately) what they say about unobservables; anti-realists deny that success automatically licenses that further step. In contemporary form, the contrast is often framed through constructive empiricism and structural realism.

Constructive empiricism (van Fraassen) claims that “Science aims to give us theories which are empirically adequate; and acceptance of a theory involves as belief only that it is empirically adequate.” [2] The point is not that theories are meaningless, but that success warrants belief about what can be saved at the level of observable phenomena, while remaining agnostic about whether the theory’s unobservables are real.

Structural realism (associated with Worrall’s 1989 “best of both worlds” proposal) accepts much of the realist motivation—scientific success is not a miracle—while granting the anti-realist’s historical warning that theories can be replaced. [3][10] The proposed compromise is that what tends to survive theory change is not a stable catalogue of entities “in themselves,” but stable structure: relations, constraints, and mathematically expressible dependencies that later theories preserve or refine. [3][10]

Modern science, in practice, is often more pragmatic than either camp. Scientists routinely rely on models that work—sometimes without a settled interpretation of what the model’s hidden posits “really are.” That is exactly the pressure point for this essay: what does success actually license us to believe about unobservables, especially once models are embedded in systems that act?

Thesis

Model success most securely licenses constraints and difference-makers—dependencies that survive tests: change X and Y shifts as expected—rather than ontological commitment to a full catalogue of unobservables, especially when the model is embedded in systems that act. In deployed feedback loops, deployment can change incentives, change behaviour, and change the data the model later sees, shifting the problem the model is “solving.”

By “difference-makers” I mean the parts of an account you can lean on counterfactually: if X varies (within a stated scope), Y should shift in a repeatable way, and we can say what would count as failure. By “full ontology” I mean stronger claims about what the model’s hidden posits are, over and above the role they play in tracking those dependencies.

This stance is adjacent to structural realism, but more cautious about what “structure” buys you once models are used inside systems that adapt. The emphasis is not only on what survives theory change, but on what remains stable under intervention, regime shifts, and feedback.

On this view, a model’s real explanatory value includes stating the feedback-sensitive limits: which interventions preserve the dependency, which perturbations flip it, and what would count as a breakdown once the system adapts.

Bridge: “If we often rely on models we don’t fully interpret, what happens when we embed them in systems that act?”

Evolutionary game theory gives a clean bridge from “success” to what we’re entitled to lean on. In EGT, what gets selected is not the truest internal picture, but the strategy that does best under a payoff structure. [9] That makes the epistemic point vivid: different internal “stories” can lead to the same outward success, so success most directly supports claims about which distinctions make a difference under selection pressure, not which inner representation is uniquely correct. The family resemblance to Bayesian learning and modern machine learning is straightforward: training optimises expected loss under the data and objective you give it.

The deployment twist is that once a model is used to make decisions, the system can adapt: incentives shift, behaviour shifts, and the data the model later sees can shift too. The point is not that the model stops working; it’s that we need to say where it should keep working, what would break it, and what changes when the world responds—i.e., when the model becomes part of the causal loop.

Criterion / Standard

You don’t certify an ontological claim just because a model works. You earn the right to lean on parts of a model when you can say what its success tracks, what would break it, and what remains warranted once the model starts acting in the world. Put bluntly: warrant attaches to articulated dependencies and failure modes, not to “whatever hidden furniture the model posits.”

A claim involving unobservables is “good enough to lean on” when:

• Scope is explicit: it states the conditions under which success is being assessed (data regime, interventions, environment, incentives).
• Difference-makers are identified: it says what variations would change the outcome (inputs, constraints of the system, payoff/loss objective), not just that the model fits.
• Testable “what-if” claims are available: you can say “if we change X, we should see Y,” and you can actually run or approximate that test (experiment, controlled change, natural comparison, stress-test).
• Underdetermination is acknowledged: it notes what success leaves open (multiple internal stories can produce the same observable performance). [5]
• Breakdown conditions are named: it specifies what would count as failure (new conditions the model hasn’t seen, the system responding to the model, the target being gamed, the measuring setup changing, missing interactions).
• Action embedding is handled: it says what changes when the model becomes part of the system (people adapt, incentives shift, the data-generating process moves) and what that does to the warrant.
• Commitment is proportional: it distinguishes what success licenses strong reliance on (stable constraints/relations and intervention-resilient difference-makers) from what it licenses only weakly (ontological commitment to a full catalogue of hidden entities/states).
• Unobservable handles are anchored: when the model relies on hidden variables, it states how they connect to interventions and measurements (what would move them, how they would show up, and what would count as disconfirmation).

A simple way to see the contrast is to treat snowflakes as a testbed for what success licenses. The more reliably the morphology shifts with temperature and supersaturation, the more we are tempted to treat the underlying unobservables as “what’s really there.” But the snowflake story also shows why that step needs discipline. The same success that lets us use temperature and supersaturation as dependable difference-makers also creates pressure to posit microphysical items we never directly observe—attachment behaviour, nucleation barriers, premelting effects, and the rest. The epistemic point is not that these unobservables are meaningless; it’s that model success most securely warrants reliance on the stable dependencies and intervention handles the model tracks, not automatic ontological commitment to a full catalogue of hidden entities. [12] 

Objection

Success isn’t a truth-meter for unobservables. History gives many cases where successful theories later turned out to be wrong in their deep posits (the pessimistic induction), [4] and even at a time the evidence can fit more than one incompatible story (underdetermination). [5] So the safest stance is constructive empiricism: treat success as warrant for empirical adequacy and practical use, not ontological commitment to the unobservables a model happens to posit. [2]

Response

I’ll avoid the word belief unless I mean ontological commitment. By warrant I mean something weaker: the right to rely on certain claims as stable difference-makers within scope, even if we remain agnostic about what, in the deepest sense, exists.

The objection is right about one thing: success isn’t a truth-meter for deep ontological posits. A model can be predictively strong while its internal story is wrong, and the same performance can often be generated by more than one incompatible internal picture. A standard historical worry—sometimes called the pessimistic induction—is that past scientific successes did not reliably preserve their deep posits: theories that worked extremely well were later replaced, and their hidden entities were abandoned or radically reinterpreted. [4] So success on its own cannot certify ontological commitment.

But constructive empiricism is too blunt, on its own, for the kinds of models we actually lean on—especially models we probe, stress, and embed. In practice, we don’t just ask whether a model matches observations. We ask whether its success is stable under intervention: do outputs shift in the expected way when we deliberately vary inputs, constraints, and environments?

Crucially, the variables that do the explanatory work are often not themselves observable. The “X” we test is frequently a latent parameter or mechanism-level quantity we access indirectly—via boundary-condition changes, instruments, and controlled perturbations. When a model repeatedly gets right not just the outputs, but which unobservable handles move them—and when that pattern survives robustness checks and regime changes—its success earns more than mere fit. It earns warrant to rely on stable constraints and difference-makers for counterfactual control within a stated scope, even if underdetermination remains. [5] What it still does not earn, by default, is a blank cheque to treat one particular inner story as the ontology.

Handoff

As I argued in the thesis, a model’s real explanatory value includes stating its feedback-sensitive limits: which interventions preserve a dependency, which perturbations flip it, and what would count as breakdown once the system adapts. In practice, the “difference-makers” we rely on are often not directly observable—they are latent handles or constraints we reach only indirectly through instruments and controlled changes. Treating those handles as indefinitely “agnostic” isn’t neutral; it leaves the model’s scope and sensitivities unspecified, which is exactly where bad surprises live.

Once a model is deployed inside a decision system, success has to be reinterpreted again. In feedback loops, the model is not just describing the world; it helps shape the data it later sees. Proxies become targets: people optimise what is rewarded, and measures used for decision-making can be distorted under pressure. [6][7][11] Here the older realism question (“what should we believe about unobservables?”) sharpens into a practical epistemic question: what exactly is the deployed system treating as evidence, and how does that change under optimisation pressure?

That sets up the next step: a model-belief problem. If the model is mis-specified, deployment doesn’t merely expose the error—it can amplify it. Optimisation pressure makes the proxy look increasingly “real” inside the system, while the neglected variables and side-effects grow in the background. Distinguishing Goodhart-style failure modes helps here, because it makes explicit how optimisation can break the link between a measure and what it was meant to stand in for. [12] The next essay asks how to recognise that dynamic early, and what kinds of tests and governance keep proxy-success from turning into epistemic failure. [11][12]

Notes

1. Stanford Encyclopedia of Philosophy. (n.d.). “Ernst Mach.” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/ernst-mach/ (Accessed: 28 Jan 2026).
2. Stanford Encyclopedia of Philosophy. (n.d.). “Constructive Empiricism.” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/constructive-empiricism/ (Accessed: 28 Jan 2026).
3. Stanford Encyclopedia of Philosophy. (n.d.). “Structural Realism.” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/structural-realism/ (Accessed: 28 Jan 2026).
4. Stanford Encyclopedia of Philosophy. (n.d.). “Realism and Theory Change in Science.” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/realism-theory-change/ (Accessed: 28 Jan 2026).
5. Stanford Encyclopedia of Philosophy. (n.d.). “Underdetermination of Scientific Theory.” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/scientific-underdetermination/ (Accessed: 28 Jan 2026).
6. CNA. (2022). “Goodhart’s Law: Recognizing and Mitigating the Manipulation of Measures in Analysis.” CNA. https://www.cna.org/reports/2022/09/Goodharts-Law-Recognizing-Mitigating-Manipulation-Measures-in-Analysis.pdf (Accessed: 28 Jan 2026).
7. Nichols, S. L., & Berliner, D. C. (2007). “Collateral Damage: How High-Stakes Testing Corrupts America’s Schools.” ERIC. https://files.eric.ed.gov/fulltext/ED508483.pdf (Accessed: 28 Jan 2026).
8. Alexander, J. M. (2002/updated in SEP). “Evolutionary Game Theory.” Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/game-evolutionary/ (Accessed: 28 Jan 2026).
9. Worrall, J. (1989). “Structural Realism: The Best of Both Worlds?” Dialectica, 43(1–2), 99–124. doi:10.1111/j.1746-8361.1989.tb00933.x.  
10. Krakovna, V., Uesato, J., Mikulik, V., Rahtz, M., Everitt, T., Kumar, R., Kenton, Z., et al. (2020). “Specification gaming: the flip side of AI ingenuity.” DeepMind Blog (21 Apr 2020). https://deepmind.google/blog/specification-gaming-the-flip-side-of-ai-ingenuity/  
11. Manheim, D., & Garrabrant, S. (2018). “Categorizing Variants of Goodhart’s Law.” arXiv:1803.04585. https://arxiv.org/abs/1803.04585  
12. https://aftercertainty.net/index.php/snowflakes-lawless-explanation/