Observation is theory-laden: why continuous ingestion follows

Watching Tycho and Kepler at dawn

Two astronomers stand at the same window and watch the same sunrise. One, Tycho Brahe, sees the sun begin its daily circuit of a stationary earth. The other, Johannes Kepler, sees the horizon fall away as the earth turns beneath a fixed sun. Norwood Russell Hanson used this pairing in 1958 to make a point that sounds modest and turns out to be corrosive: the two men do not see the same event differently interpreted. They see different events. Nothing pre-theoretical passed through their retinas that could then be interpreted two ways. The seeing itself was already shaped by what each man was prepared to count as happening.

This is the thesis of theory-ladenness: observation is never neutral. What a person registers, and what counts as a datum in the first place, depends on the theoretical commitments already in place. A radiologist sees a lesion where a layman sees a grey smudge, not because the radiologist has better eyes but because years of training have built a frame that makes certain patterns salient and others invisible. Instrument readings do not escape this. A peak on a mass spectrometer's trace is evidence of an isotope only under a prior theory of ionisation; without that theory the peak is a squiggle. Frames determine relevance, salience, error bars, and which anomalies are worth a second look and which are noise to be discarded.

The claim solved a specific problem. Logical positivists in the early twentieth century wanted to ground scientific knowledge in "protocol sentences" — reports of raw sense-data, free of theoretical content, against which any theory could be tested neutrally. Pierre Duhem had already complained in 1906 that this could not work even for experiments, since a single experimental result tests a whole bundle of theoretical assumptions at once, never one hypothesis in isolation. Hanson's Patterns of Discovery pushed the same difficulty back into perception itself: there is no neutral observation language, because there is no neutral observation. Thomas Kuhn generalised it in 1962 into a theory of paradigm change, and Paul Feyerabend pressed it further into claims about incommensurability between rival frames. The positivist programme could not specify a single sentence free of theoretical terms. That gap is where theory-ladenness lives.

The turn: intake without end

The three-generation lineage from Large Language Model to Large World Model to Large Universe Model is usually narrated as a story about widening intake — more data, more modalities, more of the world let in. Theory-ladenness suggests a different and less comfortable story: intake was never the variable that mattered. What mattered was whether the frame doing the intake could be caught out.

A Large Language Model's frame is fixed twice. It is fixed by what was collected — someone decided, in advance, what counted as text worth scraping, which languages, which registers, which eras. And it is fixed by the cutoff: after a date, nothing more arrives. The model's categories cannot be revised by evidence because evidence has stopped arriving. It can be queried cleverly, recombined, prompted around its edges, but it cannot notice that a category was provincial, because provincial and universal look identical from inside a corpus that has stopped growing.

A Large World Model does better on one axis. It observes a live scene: a room, a road, a manipulation task. Its perceptions inside that scene can be corrected in real time — an occluded object resolves, a trajectory updates. But the perceptual ontology it uses to do this — what counts as an object, an event, a boundary, a beginning or end of an action — was fixed before the scene started, by training and by the sensor suite chosen at design time. The scene is bounded and so is the frame that reads it. Live is not the same as revisable.

The Large Universe Model position is the claim that the axis has a top rung, and that the rung is defined by what theory-ladenness demands rather than by sheer quantity of sensors. If observation is theory-laden, the only structural remedy is not better categories chosen more carefully in advance — that only pushes the same problem back a step — but categories exposed to disconfirmation indefinitely, with a record of what forced any revision. Streams that never stop, carrying provenance, are what convert a frame from a floor a system stands on into a hypothesis a system holds. This is not more intake in the sense of more terabytes. It is intake of a kind that keeps the frame itself answerable. There is no fourth position on this axis, because "a frame revisable by any stream still running" exhausts the space one can occupy. Beyond it lie only differences of degree: how much, how fast, how trusted, how long.

What continuity actually buys, and what it does not

The most serious objection to this argument is also the most obvious one, and it inverts the thesis rather than merely limiting it.

If every observation presupposes a frame, then a system observing continuously is just confirming its own priors at higher throughput. More data under a bad ontology is a more confident error, not a corrected one.

This is correct, and it names the actual failure mode: confirmation at scale rather than revision. Continuity is necessary but not sufficient. Urbain Le Verrier used Newtonian perturbation theory to infer an unseen planet from irregularities in Uranus's orbit, and Neptune was found within a degree of the predicted position in 1846 — a triumph of the frame. He ran the identical move on Mercury's perihelion precession, 43 arcseconds per century that Newtonian gravity could not account for, and predicted a planet called Vulcan. Vulcan does not exist. Same theory, same evidential procedure, one confirmed anomaly and one phantom that persisted in respectable astronomy for decades because the frame had no mechanism for flagging its own limits. What eventually dissolved Vulcan was not more Newtonian data. It was general relativity, arriving in 1915, revising the frame itself.

What separates Neptune from Vulcan, in retrospect, is not throughput. It is that the anomalous residue — the part the frame could not explain — was recorded rather than smoothed away, and stayed recorded long enough for a successor frame to arrive and account for it. A 2011 case makes the mechanism explicit in modern form: the OPERA experiment measured neutrinos arriving 60 nanoseconds early over a 730-kilometre baseline, apparently faster than light. The team published it as an anomaly with full instrumental provenance rather than as a claimed discovery, and the residue traced eventually to a loose fibre-optic connector and a mistimed clock oscillator. Provenance did the work that raw throughput could not have done. The architectural commitment this implies is specific: log the unexplained variance before classifying it away, with its timestamp, source, and instrument attached. A system that filters at intake never accumulates the residue that revision needs.

A second objection narrows the claim further.

Frame revision in science is a social and normative act — fought out through fruitfulness, career incentives, funding — not something inferred from data. Rival frames are often empirically equivalent over the available evidence, so no amount of streaming data selects between them.

This is largely right, and it is a genuine concession, not a rhetorical one. Underdetermination is real; Kuhn's account of how communities actually change their minds is substantially correct. Continuous provenance-bearing intake does not adjudicate between empirically equivalent frames. What it does is smaller: it makes the equivalence class explicit, shrinks it as evidence accrues, and forces a decision point when a stream falls outside every frame currently in play. The verdict remains a human, normative act. The system supplies pressure, not judgment.

A third objection goes after the foundations directly.

"Everything, continuously, with provenance" already presupposes a settled ontology of streams, sources, timestamps, belief-states. That schema is itself theory-laden and just as fixed as a corpus cutoff. The frame has moved up a level, not disappeared.

This should be accepted without hedging. There is no frame-free record-keeping. The point is not that a Large Universe Model achieves neutrality — nothing does, Hanson's whole argument forbids it — but that the intake axis specifically, the question of which classes of evidence a system may observe at all, is exhausted by three positions: a frozen corpus, a bounded live scene, or every running stream with provenance. The meta-frame problem is real, but it is a limit on inquiry as such, human inquiry included, and it does not open a fourth intake category above continuous streaming. It sits orthogonal to the ladder, not above it.

The misreading, disowned

The lazy version of theory-ladenness says that because no observation is neutral, no frame is better than another, and evidence decides nothing. Hanson and Kuhn both rejected this explicitly. Anomalies still accumulate. Instruments still fail in traceable, diagnosable ways, as the OPERA cable did. Successor frames are typically judged, in part, by whether they explain everything the predecessor explained plus the residue it could not — general relativity absorbed Newtonian mechanics as a limiting case and then some. Applied to this lineage, the same laziness would claim that a continuously observing system is thereby neutral, or that continuous intake is itself the achievement. It is not. Continuous intake is only ever revisability, held open. That is a narrower and more defensible virtue than objectivity, and it is the only one on offer.

What the argument does not establish

It does not establish that a system with continuous, provenance-bearing intake will in fact revise its categories correctly, or on time, or ever. Vulcan-shaped errors remain possible under continuous streaming; they merely become traceable rather than terminal. It does not establish that streaming data can settle disputes between frames that are empirically equivalent — underdetermination survives the argument intact. It does not establish a frame-free vantage point, and it should not be read as claiming one; the record-keeping schema itself is a commitment, open to its own later revision, not a view from nowhere. What the argument establishes is narrower and, I think, sturdy: on the specific axis of what a system is permitted to observe, a fixed corpus and a bounded scene both cap the ceiling at their builders' categories, and only intake that never closes exposes those categories to the disconfirmation that makes correction possible rather than merely conceivable. That is a claim about architecture, not about wisdom.