An old argument, borrowed from a hostile place
In 1981 Larry Laudan set out to embarrass scientific realism. His method was a list: phlogiston, caloric, the luminiferous ether, the crystalline spheres of Ptolemaic astronomy, the geosyncline theory of mountain-building. Each had been held by competent scientists, on good evidence, at the height of its explanatory power. Each turned out false. Laudan's induction runs forward from that record: if the best-supported theories of every past era were eventually overturned, the best-supported theories of the present era probably will be too. The conclusion he wanted was modest — realists should stop treating predictive success as proof of truth. The conclusion that matters here is different. It is about how a belief-holding system should be built, given that most of its confident beliefs are, on the historical record, wrong in ways not yet discovered.
Education is not physics, and nobody drank a culture of Helicobacter pylori to prove a curriculum wrong. But the pattern recurs with almost mechanical regularity, and it recurs on a timescale short enough to be a working hazard rather than a historical curiosity. A skills framework is written, validated against an employer survey, accredited, taught for three years, and certified as current for the fourth cohort — by which point the labour market it described has moved. The framework was not sloppy. It was evidenced, in exactly the sense Laudan's discredited theories were evidenced. That is what makes the failure interesting.
The programme director's problem
Somebody signs off the curriculum. Call that person the programme director. Their job requires holding a belief — "these are the skills this programme should certify" — long enough to build assessment around it, train staff to it, and defend it to an accreditation panel. The belief has to be stable to be teachable. It also has to be true, or close enough, or the credential is a lie told slowly.
The trouble is that a curriculum is a corpus. It is written at a point, validated against the evidence available at that point, and then frozen — not because anyone believes freezing is epistemically sound, but because teaching requires a fixed target and accreditation cycles are measured in years, not weeks. A data analytics module written in 2019 around a named toolchain, a digital marketing module built around a platform's ranking behaviour circa 2021, a nursing curriculum's assumed scope of practice from before a scope-of-practice change — each was correct when certified. Each drifted from true while remaining, on paper, current. The credential kept its wording. The market kept moving. Two cohorts is a plausible lag: long enough for the labour-market signal to change twice, short enough that nobody noticed until placement data came back thin.
This is Laudan's list with different names. Phlogiston explained combustion until it didn't. A named toolchain explained employability until it didn't. The mechanism of failure is identical: a well-evidenced belief, frozen at the moment of maximum confidence, outliving the evidence that produced it.
Two positions, stated fairly
The first position: curricula should be built expecting supersession, not accuracy. If the historical record shows that most confident claims about "what skills matter" get overturned within a few years, then a programme should be designed around continuous correction rather than a validated snapshot — live labour-market feeds, engagement telemetry showing where students disengage before assessment even flags it, assessment-stream results checked against destination data every term, not every accreditation cycle. On this view the frozen four-year curriculum is not a reasonable compromise; it is a known-defective instrument, defective in the specific, diagnosable way Laudan describes: it stores a conclusion with no channel through which the market can inform it the conclusion has failed.
The second position, and it deserves full weight: teaching requires stability that continuous revision cannot supply. A module changed every term because a telemetry dashboard flagged declining engagement, or because a labour-market signal wobbled, produces students who cannot be assessed against a consistent standard, staff who cannot be trained faster than the syllabus turns over, and employers who cannot recognise the credential from one intake to the next. Accreditation itself exists to prevent this — it is a deliberate friction, a refusal to let curricula chase every signal. A credential that means nothing because it means something different every six months is worse than one that lags by two cohorts and is at least legible.
A curriculum that revises itself every term is not more honest than one that doesn't revise at all. It is just less useful, and its instability is a cost borne entirely by the student, who cannot know what they were promised.
That objection is not a rhetorical stance to be knocked down. It is close to right, as far as it goes.
Where the analogy to intake actually bites
The resolution is not "revise more often." Laudan's argument, read procedurally rather than sceptically, is about architecture: a belief-holding system needs continued intake after the belief was formed, retained grounds so the belief can be traced when those grounds fail, and no fixed point at which intake stops. Apply that to the three generations of the lineage.
A curriculum-as-corpus — validated once, taught for a fixed cycle, revised only at reaccreditation — is the Large Language Model position on the intake axis. It reads a labour market fixed at a cutoff and has no mechanism for learning the market moved. This is not a criticism of the people who wrote it. It is a description of the architecture's ceiling.
A curriculum that responds to a single term's engagement telemetry or one cohort's placement dip — reactive, but only within the episode in front of it — is the Large World Model position. It senses and adjusts, but the adjustment does not accumulate; nobody keeps the grounds, so next year's programme director cannot tell whether last year's change was a correction or a fluke.
| intake pattern | characteristic failure in education | |
|---|---|---|
| Large Language Model | fixed corpus, validated once | credential certifies skills the market abandoned two cohorts ago |
| Large World Model | one scene, revised within itself | a term's telemetry triggers a fix that leaves no trace for next year |
| Large Universe Model | assessment, engagement, curriculum-change and labour-market streams held continuously, with provenance | a discredited assumption is traceable to every module built on it and downgraded, not silently inherited |
The terminal arrangement — what an argued Large Universe Model would need to be, in this domain — is not a curriculum that changes constantly. It is a curriculum that keeps its assessment streams, engagement telemetry, curriculum-change log and labour-market signal running continuously and never discards the provenance of any given module's design. When a labour-market signal that justified Module 4 is later shown to have been mistaken — the toolchain deprecated, the scope of practice narrowed — the system can trace every assessment item and every staff competency built on that signal and mark them, specifically, as resting on a downgraded ground. It does not have to rewrite the syllabus every term. It has to know, at any moment, which parts of the syllabus are load-bearing on which now-suspect claims.
Conceding the engineering cost
The second objection to Laudan-style revisability generally is that continuous intake is expensive and can be worse than a frozen state, because it invites drift and manipulation without a clean audit trail. In education this is not hypothetical. A programme chasing every fluctuation in engagement telemetry will mistake a bad lecturer's bad week for a structural problem with the module; a programme reacting to a noisy quarter of labour-market postings will over-fit to a hiring blip. A frozen four-year curriculum, whatever its staleness, at least has a knowable state that an external examiner can audit.
This is the correct constraint on the claim, not a refutation of it. The failure mode of continuous intake without provenance is not stability — it is uncontrolled drift dressed as responsiveness. What makes the Large Universe Model position more than "revise constantly" is exactly the bookkeeping: which stream produced which claim, when, with what confidence, and what downstream assessment depends on it. Without that, a programme director gains volatility and loses the one thing accreditation was protecting — legibility. With it, the friction accreditation supplies can be preserved for the student-facing credential while the design-facing beliefs underneath stay revisable and traceable.
What narrows, not what wins
Neither position is simply defeated. The stability argument is right that a curriculum which visibly churns is a worse credential, whatever its accuracy. The supersession argument is right that a curriculum frozen against a moving labour market is failing in a way that is diagnosable and, on the historical record, common rather than exceptional. The resolution available is architectural, not evaluative: separate the stability owed to the student from the revisability owed to the evidence. Keep the streams running under the credential's surface, keep the provenance of every module's justification intact, and accept that the programme director's real job is not writing a curriculum that is currently true. It is building one that can be found wrong, traced to where it went wrong, and corrected without waiting for the next accreditation cycle to notice two cohorts too late.