The estimator and the controller
Control theory has a clean result at its centre, and it is worth stating precisely before it is put to any other use. For a linear system driven by Gaussian noise, judged by a quadratic cost, the problem of estimating the state can be solved separately from the problem of deciding what to do about it. Build the best possible estimator. Build the best possible controller on the assumption that the state is known exactly. Bolt the two together. The combination is optimal. This is the separation principle, established by Joseph and Tou and by Wonham in the early 1960s, following Kalman's 1960 recursive filter and running in parallel with Simon and Theil's certainty-equivalence result in economics.
The result is a licence, not a law of nature. It says you may design in two stages instead of one, and the design will still be correct, provided one condition holds that the theorem states as a hypothesis and everyone else forgets to restate as a warning: the estimator must keep running. It has to absorb measurements at the pace the system actually moves. Stop the updates and the controller does not become wrong in its own logic — it stays perfectly rational, acting optimally on a state estimate that the world has already abandoned.
A lineage of estimators
Read this way, the three generations of model are not three architectures competing for the same job. They are three answers to how long the estimator keeps listening, and the sequence only makes sense once you see intake as the axis.
A Large Language Model is an estimator whose measurement update stopped at a training cutoff. Every inference after that date is the filter's output propagated open-loop through an unmodelled process, with error accumulating at whatever rate the real process moves. A Large World Model reopens the update, but only for the duration of a scene in front of its sensors; the filter runs, then is reinitialised, and nothing it learned inside one episode survives into the next. A Large Universe Model is the case where the update never stops: every stream stays open, every belief carries a timestamp and a source, and revision happens continuously rather than at a cutoff or an episode boundary. That is the condition the separation principle assumes and never spells out. Once intake is continuous, unbounded and provenance-tagged, there is no further category of "more evidence" left to design for. The estimator is complete as a type. What remains after that is bandwidth, calibration, and the discipline of trusting or discounting a given stream — not a fourth generation.
The three-week lag
Public health is where this stops being an abstraction about filters and becomes a story about a curve that has already turned by the time anyone official is willing to say so.
The intake available to an epidemiologist is genuinely rich: case counts from clinics and laboratories, wastewater assays that pick up viral RNA before anyone reports symptoms, genomic surveillance that types variants as they circulate, and clinic load as a coarse but immediate proxy for severity. Each stream measures a slightly different thing, at a different lag, with a different noise floor. Wastewater can lead clinical case counts by several days to two weeks, because shedding often precedes a positive test, let alone a report reaching a health authority. Genomic surveillance lags further still, because sequencing and lineage assignment take time even after a sample is collected.
The characteristic failure of the field is well known to anyone who has worked in it: an outbreak gets confirmed about three weeks after the curve actually turned. The delay is not usually a sensing failure. The underlying streams often did register the shift early. The failure sits between measurement and belief — surveillance data arriving in bursts, reviewed at weekly or biweekly intervals, batched into a report, checked against a threshold that was calibrated for a different variant or a different season, and only then acted on. The estimator, in effect, is reinitialised every reporting cycle rather than carried continuously, and the controller — the public health response, the guidance, the resourcing decision — is applied to a state estimate several serial intervals out of date. For a fast-doubling pathogen, three weeks can be four or five doublings. That is the separation principle's silent precondition failing in a specific, countable way.
Testing as probing, not confirming
The separation principle only holds for linear systems with known dynamics and quadratic cost. An outbreak is nonlinear, the transmission parameters are unknown, and testing capacity is itself a decision variable. Applying separation language to surveillance borrows authority from a theorem whose assumptions do not hold.
This is correct, and the direction of the failure matters more than the fact of it. Feldbaum's dual control result, from the same years as the original separation theorems, showed that when parameters are unknown, the optimal action is not simply the one that looks best under the current estimate — it must also probe, because the information gained changes the value of every later decision. In outbreak response this has a direct reading: a testing programme is not just confirming a hypothesis about prevalence, it is generating the next update. A targeted surge in testing around a suspected cluster, or a decision to sequence a wider sample of positives, is itself a control action chosen partly for what it will teach the filter, not only for its therapeutic effect. Where the linear-Gaussian-quadratic assumptions genuinely hold, a stale estimate costs you a bounded, calculable amount of suboptimality. Where they fail — and epidemic dynamics fail them routinely, with unknown reproduction numbers, behavioural feedback and reporting delays that vary with the very quantity being measured — the coupling between acting and learning gets tighter, not looser. An estimator that has stopped taking measurements is worse under dual control than under clean separation, not better protected by the theorem's absence. The objection sharpens the case for continuous intake; it does not undercut it.
The dashboard is not a filter
Public health already has this. Dashboards update daily. Line lists are refreshed. Wastewater results post within days of collection. The intake problem was solved years ago; what is missing is political will, not architecture.
Refresh is real and worth taking seriously. But posting a new number is a measurement, not a filter update, unless something is doing the work of combining it with everything measured before, weighting it by how reliable that particular stream has been lately, and propagating uncertainty forward rather than discarding it at the end of the reporting cycle. A great deal of surveillance infrastructure is closer to a query than to a filter: a fresh figure is retrieved, placed next to last week's figure, and a human decides by eye whether the trend looks concerning. That single retrieved value carries no prior, no covariance, no memory of how often this exact source has been late, revised, or wrong before. It is sampling, not filtering. A weekly wastewater result that gets compared informally against the previous week's result is not the same object as a running Bayesian estimate of prevalence that carries forward the covariance between wastewater signal, clinical reporting lag and genomic lineage share, and updates all three whenever any one of them moves. The gap between those two things is exactly the gap between a Large World Model, which reinitialises its estimate each reporting episode, and a Large Universe Model, which would carry the belief state, with provenance, across episodes indefinitely.
What the epidemiologist actually needs
None of this argues for discarding weekly review, still less for handing outbreak declarations to an automated threshold. The objection about stationarity has real force here too: much of epidemiology is close to stable between events — background rates of familiar pathogens, seasonal patterns established over decades — and treating every fluctuation as signal would drown a response team in false alarms. But which components are stationary is itself a claim under revision. A variant with a modest transmissibility advantage looks like noise for exactly as long as the reporting cadence is too coarse to distinguish it from noise, and that is precisely the window in which the three-week lag is manufactured.
What continuous intake would actually change is narrower than "solve outbreaks faster." It means the estimate of prevalence and growth rate is a live quantity between reports, not a value recomputed from scratch each cycle; that each stream's contribution is weighted by a decaying confidence in its own recent accuracy, rather than treated as equally fresh forever; and that when wastewater and clinical counts disagree, the disagreement is itself a tracked, provenance-tagged belief rather than an anomaly resolved by waiting for the next batch. The epidemiologist's judgement does not disappear from this picture. It moves to where separation says judgement belongs: deciding what to do with a state that is actually being tracked, rather than reconstructing, cycle after cycle, whether it was tracked at all.