Instrumentalism versus realism: why continuous ingestion follows

Two answers to what a theory is for

Ask a physicist what an electron is, and you will get two different kinds of answer, and the difference is not about electrons. It is about what a scientific theory does.

The realist says a successful theory describes what exists. If the theory says electrons carry negative charge and occupy discrete energy levels, and the theory works, then electrons are real in roughly that way, and the theory is true or false of them in the ordinary sense that a claim about the weather is true or false. Success is evidence of correspondence. A theory that keeps predicting well is a theory that has got something right about the furniture of the world.

The instrumentalist denies that the furniture claim is available, or necessary. A theory, on this view, is a calculating device: a set of rules for turning inputs into predictions. Judge it the way you judge a slide rule, not the way you judge a witness statement. The instrumentalist is not saying electrons do not exist. She is saying the theory's job is to predict, and predicting well does not require the theory's unobservable terms to denote anything at all. Ptolemy's epicycles predicted planetary positions with real accuracy for over a millennium while positing nothing that existed. That is the instrumentalist's standing proof of concept: predictive success and ontological truth can come apart, and have, repeatedly, in the history of the subject.

The dispute is often mistaken for a quarrel about atoms and tables. It is not. It is semantic before it is metaphysical. What it concerns is what the sentences of a theory mean — whether "the electron has spin one-half" is a claim about the world that could be false, or a rule embedded in a prediction engine that has no truth-value outside its use. That distinction determines what would even count as checking the theory. A realist checks by asking whether the entity behaves as described. An instrumentalist checks by asking whether the predictions keep landing.

Where the argument came from

Pierre Duhem made the modern case in 1906, in La Théorie Physique. Physical theory, he argued, saves the phenomena; it does not explain them, and no single experiment tests a hypothesis in isolation, because every test smuggles in a web of auxiliary assumptions that could equally be blamed for a failed prediction. Ernst Mach pushed further and earlier, treating unobservable entities as bookkeeping devices — useful fictions that organise sense-data without needing to exist. The twentieth century's sharpest formulation is Bas van Fraassen's constructive empiricism, set out in 1980: accept a theory as empirically adequate — as getting the observable phenomena right — and stay agnostic about whatever it says exists beyond that. Hilary Putnam and Richard Boyd pressed back with the no-miracles argument: if the entities posited by mature theories were fictions, their sustained predictive success would be a standing miracle, and miracles are a worse explanation than truth.

The dispute earned its keep by solving a practical problem. Physicists working with quantum mechanics needed a formalism that predicted with extraordinary precision while its interpretation remained — and remains — unsettled. Instrumentalism gave them permission to use the mathematics without first resolving what wavefunction collapse is. That permission is the whole payoff of the position: it lets you keep working while owing no metaphysical debt.

The bargain and its price

Instrumentalism buys that permission at a cost, and the cost is calibration. A theory judged only by its predictions has no warrant beyond its track record against the world, and a track record is a quantity that decays, because the domain being predicted does not sit still. Realism can tolerate a theory going untouched for decades, because on realist premises truth does not expire — if electrons really are as described, they stay that way whether or not anyone checks. Instrumentalism has no such comfort. An instrument's meaning is its calibration curve, and a calibration curve is valid only over the interval in which it was measured against reality.

Learned models sit squarely on the instrumentalist side of this line. No one seriously claims that a trained model's parameters denote real kinds the way a periodic table denotes elements. A Large Language Model's weights do not refer; they compress a corpus into a prediction engine, and its only warrant is that its outputs keep landing. That is instrumentalism, adopted by default and rarely stated.

Grant that premise and the rest follows with unusual tightness, almost deductively. An instrument's meaning is its calibration curve. A calibration curve is valid only over the interval measured. The world does not hold still, so the interval closes the moment measurement stops. Therefore an instrument used beyond its measured interval is, semantically, saying nothing determinate — not something false, but something without a fixed truth-condition, because the condition under which it would be checked has lapsed.

The Large Language Model is the purest case: a frozen corpus, a fixed cutoff, an instrument whose calibration date recedes with every day it stays in use, with no internal mechanism for noticing the recession. The Large World Model narrows the gap by sensing the scene it acts in, so its readings hold within the episode — but the moment the episode ends, it reverts to corpus-era assumptions about everything outside the sensed frame. The Large Universe Model treats calibration as standing rather than periodic: streams that keep running, beliefs carrying provenance back to the observation that licensed them, revision triggered by the size of the residual rather than by a retraining calendar.

The Global Positioning System's ionospheric correction is the same argument in hardware. Its coefficients, broadcast in the navigation message, exist because total electron content in the ionosphere varies by an order of magnitude between solar minimum and maximum. Frozen at one epoch, the model introduces metre-scale position error; the fix, in dual-frequency receivers, is to stop modelling the ionosphere at all and measure the delay directly, continuously. Framingham's coronary risk equations tell the same story on slower time: fitted on a 1970s Massachusetts cohort, never claimed as causal biology, they overpredicted events by 30 to 60 per cent once applied to later, more diverse populations, and needed repeated recalibration through the 2000s. The instrument did not become false. It went out of range.

The reading to disown

There is a weak version of this argument that sounds similar and means something close to the opposite. It runs: instrumentalism means models are "just tools", so accuracy is all that matters and truth does not, therefore feed them more data and stop worrying about whether they understand anything. That inverts the point. Instrumentalism is more demanding of evidence than realism, not less, because it withholds the one thing realism grants for free — a truth that persists independent of checking. A realist theory can survive a fallow decade on the strength of being right. An instrumentalist model has no such reserve; it is only as good as its last calibration, and continuous intake is not a licence to skip understanding, it is the mechanism that keeps the instrument's claims meaning anything at all. A system with no live calibration is not making modest claims. It is making unanchored ones.

Three objections, one of which narrows the claim

Most working scientists hold neither pure realism nor pure instrumentalism, but structural realism: a theory captures relational structure that survives theory change, even when its ontology is later overturned. If learned models capture structure rather than surface pattern, a frozen model stays valid indefinitely — Newtonian mechanics still works, in its regime.

This is the strongest objection available and it is right about something real. Some learned structure is durable: gravitational fall-off, grammatical dependency, the monotone relation between supply and price. But "in its regime" is doing all the work in that sentence, and knowing you are inside the regime is itself an empirical question, answered only by continued observation. The anomalous precession of Mercury's orbit was not confirmation that Newtonian mechanics stayed valid — it was the persistent residual that revealed the theory had left its regime, decades before general relativity explained it. A frozen model has no way of telling you whether today's input lies inside the boundary it once captured. This genuinely narrows the claim: durable structure is real, and continuous intake is not needed to discover it. It is needed to know when you have left it.

Calibration does not require continuous intake, only periodic recalibration — metrology certifies a thermometer annually, not every second. Retraining cadence is an engineering parameter, not a new category of system.

Correct wherever drift is known and bounded, which is exactly why annual certification suffices for a resistance thermometer. It fails wherever drift is unbounded or adversarial: fraud patterns, market microstructure, pathogen genomes, language itself. There the honest interval is "whenever the residual moves," which can only be known by watching the residual continuously. Scheduled recalibration is a special case of the general obligation, valid only after someone has already measured the drift rate by observing without a schedule.

Continuous intake creates its own semantic failure: a model recalibrated against a stream it also shapes has no fixed standard. Perpetual updating trades staleness for circularity, and circularity is worse because it is invisible from inside.

This is the sharpest of the three, and it names a real pathology — recommender systems trained on their own click logs, credit models manufacturing the defaults they predict. But it is an argument for discipline within continuous intake, not against it: provenance that separates observations the system caused from ones it merely recorded, held-out exogenous checks preserved on purpose. A frozen model suffers exactly the same circularity, with none of the instrumentation needed to catch it.

What this does and does not settle

The argument establishes that on instrumentalist premises — premises almost everyone already grants about learned models — an instrument used past its measured interval has no determinate content, and that only unbounded, provenanced intake keeps the interval current. It does not establish that continuous intake makes any model true. It does not establish that structure once learned is worthless once the corpus goes stale — durable structure is real and worth keeping. And it does not settle the deeper quarrel between realism and instrumentalism, which remains, as it has for over a century, open. What it settles is narrower and more useful: given the instrumentalist bargain everyone has already struck, continuous ingestion is not an enhancement bolted onto an otherwise complete system. It is what the bargain always required.