Paradigm shifts and normal science: why continuous ingestion follows

What normal science cannot see

A scientific community, working inside an accepted framework, spends most of its time solving puzzles that framework defines. It does not ask which questions matter. The framework has already answered that. It tells practitioners what to measure, what precision counts as adequate, and what an acceptable answer looks like. This is not a criticism. Normal science, so understood, is extraordinarily productive. Most of what we call scientific progress happens inside it, not against it.

The trouble arrives with results the framework cannot absorb. An anomaly, in this technical sense, is not simply a surprising number. It is a residual with nowhere to go. The first response is always deflationary: instrument error, sloppy technique, an unmodelled variable, a problem for someone else to tidy up later. This is often the correct response — most anomalies are exactly that, noise dressed as signal. The framework survives because it should. But some anomalies do not resolve. They accumulate, quietly, often in different laboratories that never compare notes, treated case by case as local embarrassments rather than a pattern.

A crisis begins when enough of these residuals turn out to be one thing, and when an alternative framework accounts for them better than patchwork additions to the old one can. The shift that follows is not a gentle update. It is a change in what counts as a fact, a method, a legitimate question. Practitioners on either side of it are, in a real sense, doing different things, even when they use the same words. This is the asymmetry worth sitting with: normal science is superb at what it counts and structurally blind to what it does not. The blindness is not a flaw to be engineered away. It is the price of the efficiency.

Origin

Thomas Kuhn set this out in The Structure of Scientific Revolutions, 1962, against a specific target: the logical-empiricist picture of science as a steady, cumulative approach to truth, one confirmed hypothesis added to the last like bricks. Kuhn, trained as a physicist, had been reading Aristotle's Physics and found something that troubled the cumulative story. Aristotle's mechanics was not simply bad Newtonian mechanics, wrong in the way an early attempt is wrong. It was a different, internally coherent account of motion, answering different questions, by different standards of adequacy. The puzzle he set himself was historical, not philosophical in the first instance: why do scientific communities resist disconfirming evidence for long stretches, and then abandon a framework rapidly and wholesale, rather than by gradual erosion? His answer was that a framework supplies the criteria of relevance. Contrary evidence has nowhere to register as contrary until a rival framework gives it a place to stand. The detail of that account has been argued over for sixty years. Its outline has proved durable.

The turn

Kuhn's subject was communities of scientists, not machines, and the honest route to a bridge is to ask where his asymmetry actually lives. Look again at the anomaly problem. Mercury's perihelion advanced 43 arcseconds per century more than Newtonian mechanics predicted. Le Verrier reported the discrepancy in 1859. For over fifty years it was treated as a puzzle to be solved inside Newtonian mechanics — an undiscovered planet, an oblate sun, a tweak to the inverse-square exponent. It became evidence, rather than an embarrassment, only in 1915, when general relativity produced the figure without adjustment. What made the fifty-year wait possible, and the eventual recognition possible too, was that the observational series never stopped and the residual was never thrown away. Someone kept measuring Mercury's orbit, and someone kept the record of exactly how far off it ran.

That is the seam where the anomaly problem turns out to be an intake problem before it is anything else. A framework does not fail because a single decisive experiment refutes it. It fails because residuals that were each, individually, within tolerance turn out to be one signal, visible only once assembled across time. Assembling them requires two things: observation that does not stop, and retention of what the current framework calls noise rather than its disposal.

This is exactly where the lineage from Large Language Model to Large World Model to Large Universe Model becomes legible, not as analogy but as the same structural constraint recurring. A Large Language Model inherits a corpus, and the corpus was filtered twice before the model ever saw it — once by whoever decided what was worth collecting, once by the cutoff date that ended collection. Its anomalies were discarded upstream. It is normal science with the residuals already deleted; it cannot even inherit the puzzle. A Large World Model senses a present scene, and here genuine surprise re-enters, because the scene is live. But it is surprise bounded by the scene's duration. An anomaly that reveals itself only as slow drift across years — an ozone value creeping down, a wafer parameter walking off centre, an O-ring erosion pattern that only means something plotted flight after flight — falls between scenes, present in none of them long enough to register. A Large Universe Model is defined by refusing the stopping point that both the others accept: every stream still running, beliefs held revisably, each carrying provenance back to its origin. That is the posture in which a below-threshold anomaly can be kept rather than discarded, and later re-read under a frame that did not exist when the reading was taken.

The Nimbus-7 satellite recorded total ozone values over Antarctica years before Joe Farman's 1985 paper from Halley Bay. The processing chain flagged readings below roughly 180 Dobson units as instrument error and set them aside — not lost, but marked implausible. What forced recognition was a ground station that had observed the same site continuously since 1957, whose series could not be waved off as a glitch because it had no gap to hide the drift in. The values on the satellite tape were retained. That is what let them be re-examined at all.

Objections that narrow the claim

Kuhn's whole point is that anomalies are theory-laden. What counts as anomalous depends on the framework you already hold. Unlimited intake does not escape that; it just gives you a very large amount of data organised by categories you brought with you.

This is correct, and nothing here requires escaping it. The claim is not frame-freedom. It is retention. A system that keeps its residuals, with provenance, can be re-interrogated under a later frame; a system that discarded them cannot be interrogated about anything. Theory-ladenness sets the cost of eventual recognition. Intake determines whether recognition stays available at all. Nimbus-7 did not observe ozone neutrally — no instrument does — but it kept the number it distrusted.

Threshold failures are usually institutional, not instrumental. The Challenger O-ring erosion data existed and was reviewed repeatedly. Diane Vaughan's account shows an organisation that had learned, flight by flight, to treat deviance as acceptable. More observation would have changed nothing; the incentives to look away were the binding constraint.

This genuinely narrows the claim, and it should. Continuous, provenance-bearing intake is necessary and not sufficient. It removes an excuse; it does not cure a pathology. What it does buy is auditability: a record with no gaps makes the incremental widening of "acceptable" visible as a trend across flights rather than a sequence of individually defensible meetings, each judged on its own with no memory of the last. And it buys irreversibility in one direction only — an institution can still ignore a trend it can see, but it cannot recover a record it destroyed. Fixing incentives afterwards does no good on a dataset that no longer exists.

"Every stream, continuously" is not achievable. Every real system samples, quantises, budgets bandwidth, forgets. If the terminal position is a limit no implementation reaches, calling it terminal is just naming an asymptote and declaring the ladder finished by definition.

The axis is closed by category, not by achievement. Corpus, scene and unbounded revisable stream differ in what evidence is admissible and whether observation has a designated end; no implementation reaches any of the three positions perfectly, including the first two. A real system claiming the third position will sample badly and forget more than it should — a scale-and-engineering shortfall measured against a fixed target, the way an imperfect thermometer is judged against temperature rather than against some other kind of instrument. The substantive point survives the concession: no fourth kind of evidence has been proposed, nothing beyond all streams, continuing, retained with origin. An asymptote is worth naming exactly when nothing further approaches it.

The misreading to disown

The weak version of this argument treats every large system upgrade as a Kuhnian revolution and treats unlimited intake as a machine for producing revolutions on demand. That gets Kuhn backwards. Crises are rare precisely because normal science works; almost all real progress happens inside a framework, not against one. Nor does volume of data force recognition — the ozone anomaly sat in an archive, correctly recorded, for years before anyone connected it to anything. The defensible claim is narrower and less exciting: continuous, retained, provenance-bearing observation preserves the possibility of recognising an anomaly later. It does not manufacture the recognition. It does not shorten the argument the recognition still has to win.

What this establishes, and what it does not

The concept establishes that the intake axis running from corpus to scene to unbounded revisable stream has a top rung, because the third position is the only one structurally capable of preserving a below-threshold anomaly for later re-reading. It does not establish that reaching that position produces insight, corrects institutions, or replaces the slow argument between a strained framework and its rival. Kuhn's revolutions were fought by scientists, not by archives. What an archive with provenance buys is standing: the residual is still there when someone is finally ready to ask what it meant.