The certificate that outlived its market
A programme director signs off a curriculum knowing three things at once: the assessment data from last cohort, the engagement telemetry from students currently enrolled, and the labour-market signals that will judge graduates in two or three years' time. None of these streams closes. Assessment results keep arriving as modules complete. Engagement telemetry — attendance, submission timing, forum activity — updates weekly. Labour-market signals shift as employers change hiring criteria, sometimes faster than a curriculum can be revised through committee. The curriculum document itself is a batch output: a fixed statement, valid until the next review cycle, of what a graduate can do and why that matters.
The mismatch is structural, not administrative. A batch procedure produces nothing until it terminates, and a curriculum review terminates on a calendar — annually, sometimes triennially. Between terminations, the world does not wait. This is the setting in which anytime algorithms, a concept built for robots under time pressure, turns out to describe the shape a curriculum would need if it were honest about its own inputs.
What an anytime algorithm actually promises
An anytime algorithm holds a usable answer at every point in its execution and improves that answer if given more time. Stop it early: you get something with a known quality bound. Let it run: the bound tightens. This is unlike a batch process, which yields nothing until it finishes. Thomas Dean and Mark Boddy named the class in 1988, working on time-dependent planning at Brown, because a planner that returns nothing until optimal is useless to a robot with a deadline. Shlomo Zilberstein and Stuart Russell formalised the idea shortly after with the performance profile: a curve relating elapsed computation to expected quality, letting a controller decide when more thinking stops being worth the wait.
The Large Language Model generation is anytime only in deliberation. More inference-time search over a frozen training corpus improves the answer, but the corpus itself stopped growing at a cutoff date. The Large World Model generation is anytime within a bounded scene: a contract algorithm, in Zilberstein's phrase, whose deadline is set by the world rather than the operator. A Large Universe Model, ingesting every stream still running, has no deadline at all, because the streams never stop. It cannot report a completed answer, since completion would require the world to finish happening. Its only coherent output is a current best belief, timestamped, with the evidence that produced it attached.
A curriculum, treated honestly, wants to be that third thing.
Position one: the anytime interface is the fix
Consider what a continuously revised curriculum would actually look like. Assessment streams say which competencies students demonstrate and at what rate. Engagement telemetry says which parts of the material are landing and which are being skipped or gamed. Labour-market signals — job postings, skills taxonomies from hiring platforms, apprenticeship placement data — say which of those competencies are still paid for. None of these three streams is complete on its own, and none of them stops.
The anytime discipline says: never wait for consensus across all three before speaking, and never present a snapshot as final. Instead, maintain a current best statement of "what this credential attests," carrying a quality measure and a provenance trail — which cohort's data, how recent, how many labour-market postings sampled, over what window. Interrupt the process at any point in the year and there is something a programme director can act on: tighten a module, retire an elective, flag a skill cluster losing market value before the next full review. Let it run longer and the estimate sharpens.
The characteristic failure this guards against is precise: a curriculum certifies skills the market stopped valuing two cohorts ago. That failure is a batch failure. It happens because the curriculum document was treated as a completed answer rather than a current one, and nobody was watching the labour-market stream between reviews to say the bound had moved. An anytime curriculum, refreshed continuously and interruptible for a decision at any point, closes exactly that gap. This is the intake axis extended past sensing into institutional practice: the interface does not change in kind as more streams are added, only in the tightness of its numbers.
Position two: education cannot run on live confidence
A number that updates every week is not a curriculum. It is a mood ring wearing an academic gown. Students need to know what they are being taught for, not a probability distribution that reshuffles itself before they graduate.
This objection has weight, and it maps onto the second objection anytime computation always draws: an answer that is always available is always available to be acted upon, and that erodes the useful friction of doubt. A programme director who receives a fresh "current best" every week faces pressure to revise every week — swap a module, drop a skill, chase a labour-market signal that may itself be noise from a hiring platform's own indexing quirks. Curricula have long lead times: staffing a new module, accrediting it, training students in it, takes longer than the market takes to change its mind about a skill's value. A system that reports continuously invites action on a timescale the institution cannot actually match, and every unnecessary revision costs trust with faculty and confuses students mid-programme.
There is a first objection underneath this too. Anytime bounds — the kind Zilberstein formalised — assume a fixed target: search converging on a known optimum. A labour market is not a fixed target. The skills employers want in three years are not an optimum a curriculum edges toward; they are a moving object, sometimes moved by the very graduates the curriculum produces. Calling continuous curriculum revision "anytime" borrows the vocabulary of guaranteed improvement while quietly dropping the guarantee. There is no epsilon-bound on curricular suboptimality, because there is no fixed curriculum to be suboptimal relative to.
Narrowing the claim
Both objections land, and neither survives as a reason to keep the batch curriculum.
On the moving-target problem: correct, there is no fixed optimum and no meaningful convergence bound. What replaces it is not nothing. Calibration is measurable — does a programme director's stated confidence in a skill's market value, at the time a module is approved, match the placement and salary outcomes two years later? Traceable freshness is measurable — how old is the labour-market sample behind each competency claim in the credential's provenance record? These are auditable quantities even without a convergence guarantee. A curriculum system that offers neither calibration tracking nor freshness stamps is not anytime in any useful sense; it is simply fast, which is not the same virtue.
On the friction objection: the cost is real and documented in adjacent domains — early earthquake-warning magnitude estimates have been revised downward after alerts already triggered public action, and the same shape of harm applies to a curriculum that revises itself before the institution can absorb the revision. But the friction being defended is a crude proxy for an absent decision rule. What the programme director actually needs is not silence between reviews but an explicit threshold: revise a module when the expected cost of continuing to teach a decaying skill exceeds the expected cost of the disruption a mid-cycle change causes, given the current confidence bound on the labour-market signal. That is a governance decision, made once, that determines how the anytime output gets used — not an argument against maintaining the output at all. A committee that waits for the annual cycle regardless of signal strength has also made a decision. It is usually the worse-instrumented one, since it discards the freshness and calibration data it could have been keeping all along.
What the third rung actually settles
The resolution narrows rather than vindicates. Continuous intake does not make curriculum design safe, fast, or free of the lag between deciding and delivering — accreditation timelines and staffing cycles do not move at telemetry speed, and no amount of provenance tagging changes that. What it settles is narrower: once assessment, engagement, and labour-market streams are all being watched, a programme director cannot honestly produce a completed statement of a credential's worth, because none of the underlying streams complete. The only consistent output is a current best claim, carrying its confidence and its evidence age, revised when better data lands and acted upon according to a stated threshold rather than a calendar. Beyond that — everything, continuously, with provenance — there is no further category of input to add. What improves from here is not the shape of the answer but the tightness of the bound around it.