The formula that fit the far infrared
In 1900, Rayleigh and Jeans wrote down a law for how much energy a hot cavity radiates at each wavelength. It came from equipartition: give every vibrational mode of the radiation field the same average energy, kT, and sum the modes. In the far infrared, long wavelengths, few modes, the law matched measurement to a few per cent. Pushed towards the ultraviolet, where the mode count grows without limit, the same law predicted infinite emitted energy from a warm wall. The equation did not know it had left the region that constrained it. It kept producing numbers in the same confident voice, and the numbers went to infinity.
A curriculum has the same structure. It is a formula fitted to a region: the cohort of employers surveyed, the skills audit conducted, the labour-market signals available at the moment of design. Inside that region it is often excellent. A data-analytics module written against 2019 job postings taught exactly the tools those postings asked for. Pushed two or three cohorts past its fitting window, it produces the curricular equivalent of infinite emission: confident certification of skills the market stopped paying for, delivered with the same authority as when the fit was good. Nothing on the transcript announces that the boundary has been crossed.
What actually diverges
The failure is not gradual. A programme director does not watch employability scores drift down by single percentage points each year until someone notices. More often the scores hold, sometimes rise, because the assessment instruments were also fixed at the same moment as the curriculum and are grading against the same stale target. Graduates pass. Portfolios look strong against the rubric. The rubric is the part that has gone quietly infinite: it is certifying fluency in a skill set the rubric itself has not updated to question. The divergence shows up somewhere else entirely — in a labour-market signal nobody was streaming into the review, in a hiring manager's private note that a "certified" graduate cannot do the one thing the job now requires.
This is the shape of a Large Language Model's failure, not by metaphor but by mechanism. A corpus frozen at a cutoff interpolates well inside its sampled region and extrapolates fluently, wrongly, outside it. A curriculum frozen at its last major review does the same. The output — a graduating cohort, a transcript, a set of badges — stays fluent while becoming detached from what it certifies, because fluency was assessed and market value was not.
Two positions, honestly held
The defensible response is to keep the streams open: assessment data, engagement telemetry, curriculum-change logs, and labour-market signals, all running continuously rather than sampled once every accreditation cycle. That is a move along the intake axis, from a Large Language Model's frozen corpus towards something closer to a Large World Model's sensed present — and arguably towards the terminal position, a Large Universe Model, where the streams never stop and every module's justification carries a date and a source.
But there is a second, equally serious position, and it deserves to be stated in its own voice before it is answered.
A programme director's job is to teach durable capability, not to chase a labour market that changes faster than any curriculum committee can meet. If you retool the syllabus every time a job-posting aggregator shifts its keyword weights, you produce graduates who are exquisitely tuned to a demand signal that itself may be noise — a temporary skills panic, a vendor's marketing cycle mistaken for a structural shift. Equipartition was elegant and wrong in one direction; chasing every telemetry spike is wrong in the other. Some curricula should be allowed to be unfashionable on purpose.
This is not a weak objection. Continuous intake has its own failure mode, which is thrashing: a programme that rewrites its assessment criteria every semester because engagement telemetry dipped, or because one influential recruiter tweeted about a new framework, teaches instability rather than competence. A student who begins a three-year programme and ends it having been reassessed against four different "current" skill sets has not been served by openness. They have been served by a director who mistook noise in the labour-market stream for signal.
Where the resolution actually sits
Planck's response to Rayleigh–Jeans is instructive here, and it cuts against the simplest version of "just gather more data." He did not fix the ultraviolet catastrophe by measuring harder in the infrared. He introduced a new theoretical constraint: energy exchanged between matter and radiation only in quanta of size hν. That is a bolder model, not a more measured one. The lesson generalises badly if taken as "theory beats observation" — because the theoretical move was itself forced by observation. Rubens and Kurlbaum's far-infrared measurements, at wavelengths of 51.2 and 152 micrometres, shown to Planck on 7 October 1900, ruled out the rival Wien law in exactly the region everyone had assumed was safe. Planck's formula followed within days. The new theory did not come from nowhere; it came from a stream that had just pushed past where the old formula had been trusted, and reported back that the old formula was wrong there.
For a programme director, the equivalent discipline is not to abolish stable curriculum design in favour of real-time syllabus chasing. It is to make the boundary of the evidence a tracked, dated fact rather than an assumed one. A module justified by a 2019 employer survey should say so, on the record, with an expiry implied by the date. When labour-market telemetry produces a signal strong enough to rule out the assumption behind that module — not a single tweet, but a sustained shift, the pedagogical equivalent of Rubens and Kurlbaum's measurement — that is the trigger for redesign, and it should be traceable to a specific, named piece of evidence rather than to institutional anxiety.
Provenance, not omniscience
Objection two, restated for this domain: continuous intake does not abolish extrapolation. A programme director still has to make three-year commitments about what will matter in a labour market nobody can observe yet. Adding more streams — engagement telemetry this term, employer signals this quarter — pushes the horizon of known evidence outward. It does not remove the horizon. The Large Universe Model, on this reading, is the same structural risk at a larger radius: still guessing about the unobserved, just with a longer run-up.
That is correct, and the claim made for the intake axis is narrower than it sounds. Open intake does not eliminate extrapolation. It eliminates unmarked extrapolation. A curriculum built on continuous streams, with provenance attached — this competency was last validated against labour data dated such-and-such, this module has had no confirming signal for six terms, this assessment rubric predates the last major market shift it should have responded to — lets a programme director see the boundary rather than discover it retrospectively, in an alumni survey, three cohorts too late. Rayleigh–Jeans was dangerous because its extrapolated region looked exactly like its measured region. A curriculum with dated, sourced justifications cannot make that mistake invisibly, even though it can still, honestly, be wrong about the future.
The rung this ladder actually has
None of this makes labour-market forecasting solved. Markets shift for reasons no telemetry captures in advance — regulatory change, a technology nobody was streaming data about because it did not yet exist. A Large Universe Model of education would not predict those. What it removes is the specific, avoidable failure: certifying a skill as current when every available stream has already reported, with a date and a source, that it stopped being current two cohorts ago. That is a narrower victory than "solving" curriculum design, and it is the only one the intake axis was ever going to deliver.