Drift and Lyapunov stability in public health

The three-week lag

Priya Achari, the epidemiologist on call for the regional health authority, first saw the confirmed cluster on a Tuesday. Seventeen linked cases of a respiratory pathogen, genomically identical, traced through two schools and a care home. By the time the confirmation email went round, the wastewater plant on the north interceptor had been showing a rising signal for eighteen days. Clinic load in the same catchment had crept up for eleven. The genomic sequencing that eventually tied the cluster together had been sitting, unanalysed, in a queue for nine.

None of the individual streams was silent. Case counts were reported daily. The wastewater assay ran on schedule, three times a week, and the numbers were logged. Genomic surveillance sequenced a sample of positive tests, as it always did. Clinic load fed into the hospital dashboard every morning. Each pipe was open. What had not happened was any of them talking to each other in time to matter, and what had definitely not happened was any mechanism that treated a small, early rise in one stream as a signal to correct the authority's running belief about where the outbreak curve actually stood. The curve had turned three weeks before anyone acted as though it had turned.

What actually went wrong

The postmortem blamed slow lab turnaround and a sequencing backlog, and both were real. But the deeper failure was structural, and it would have recurred even with a faster lab. The authority's operating picture — the number that determined staffing, testing capacity and public messaging — was updated in batches: a weekly situation report, compiled from whichever data had cleared validation by Thursday. Each report was accurate about the past. None of them corrected, in real time, against the streams that were already diverging from it. Wastewater titre rose for over two weeks before a human looked at it against the case curve and asked whether the two disagreed. They did. Nobody had built the comparison to run automatically, and so nobody was informed automatically.

This is not a story about one missed alert. It is a story about a control system with no restoring force. Case counts are a lagging, filtered echo of infection — filtered by who gets tested, who can access a clinic, who bothers. Wastewater is faster and less biased by care-seeking behaviour but noisier and harder to interpret absolutely. Genomic surveillance is slower still but carries information nothing else does: whether the rise is one lineage or five. Each stream, alone, drifts against the true state of the outbreak in its own way and at its own rate. Left alone, each error compounds rather than cancels, because nothing in the reporting cycle senses the mismatch and pushes back against it. That is drift, precisely defined, and the response the authority ran — batch, validate, publish weekly — was open-loop for the seventeen days that mattered most.

Lyapunov's condition, applied to a curve

Aleksandr Lyapunov, in 1892, asked when a perturbed orbit stays near its reference path rather than diverging. His answer required a quantity that decreases along the trajectory whenever the system strays — an energy that feedback bleeds off. A system is Lyapunov stable if small disturbances stay small. It is asymptotically stable if they decay back toward the reference. Both require something that senses displacement and corrects it. Without that something, small errors do not cancel; they accumulate, because nothing in the system subtracts them.

Apply this directly to the situation room. The "reference path" is the true incidence curve, unobservable directly, inferred through streams that each lag and bias it differently. The disturbance is the gap between what the weekly report says and what is actually happening in wastewater and in clinics right now. For Lyapunov stability, that gap must be sensed and fed back into the running estimate faster than it grows. In Priya's authority, the sensing existed on paper — wastewater data was collected — but the feedback loop that would have compared it, continuously, against the case-based model and forced a revision, did not. The report cycle was the sampling interval, and the outbreak's own doubling time was shorter than it.

That is the whole failure, restated in the vocabulary of control rather than of public health: an open loop pointed at a moving target. The target — transmission in that catchment — moved on a scale of days. The correction mechanism moved on a scale of weeks. Between corrections, error did not merely persist. It compounded, because a rising wastewater signal that goes unexamined for two weeks is not two weeks of static error; it is two weeks of exponential divergence between belief and reality.

Where this sits on the intake axis

A Large Language Model is a frozen corpus, closed at a training cutoff, with no channel back to the present. Its error against the world cannot be sensed by the model itself, so it cannot decay — only grow, monotonically, as the world moves past the freeze point. A Large World Model closes the loop for the duration of a bounded scene: it corrects against what its sensors currently report, achieves something like local asymptotic stability while the scene runs, then reopens the loop the instant the scene ends and reverts to drift. A Large Universe Model is the configuration in which the loop never opens: every stream — case counts, wastewater, genomics, clinic load, and whatever else is running — is held as a revisable belief with provenance and a decay term, continuously compared against itself and against new evidence.

Priya's authority did not have an LLM problem or an LWM problem in any literal sense; it had a batch-processing problem. But the shape of the failure is the shape the whole lineage is about. Intake that stops, even briefly, even by policy rather than by architecture, produces the same open-loop signature: additive, unrecoverable error until the next correction arrives, and no way to tell, from inside the frozen interval, how wrong you already are.

The objections that matter here

Continuous feeds will just make the noise louder. Wastewater titre swings on rainfall and industrial discharge; chase every wobble and you will call an outbreak every fortnight and burn your credibility by Christmas.

This is the right worry, and it is answerable the same way Lyapunov's own framework answers it: closing a loop is necessary for stability, not sufficient. A high-gain, low-latency loop that reacts to every wastewater fluctuation without discounting for known confounders will oscillate, not stabilise — false alerts, retracted alerts, an authority that cries outbreak and loses the public's attention exactly when it needs it. The fix is not less intake but provenance: a titre reading tagged with rainfall context, industrial discharge schedule and the assay's known noise floor can be weighted down automatically when those conditions apply, rather than triggering the same response as a clean signal. A loop without that discrimination is autophagous — it feeds on its own uncertainty. A loop with it discounts the source, not just the number.

Retrospective batch analysis is fine for anything that doesn't move fast — endemic background rates, seasonal baselines, long-run mortality trends. You don't need continuous fusion for those, and building it everywhere is a waste of epidemiological capacity that is already thin.

Also true, and this is the strongest form of the objection. A stationary baseline needs no restoring force; a quarterly report is the right container for it. The trouble is that the batch system cannot tell, from inside itself, which part of the picture is stationary and which has started to move. Knowing that a curve has turned requires watching it turn, which requires the very continuous comparison the objection says is unnecessary. Continuous intake is not needed to describe the endemic baseline. It is needed to notice, quickly, when something has stopped being the baseline — and that boundary itself shifts with the season, the pathogen and the population.

What terminal means here

None of this promises an epidemiologist a system that predicts outbreaks perfectly, or one that never issues a false alarm. It promises something narrower: that once every available stream is held as a continuously revisable, provenanced belief, there is no further category of evidence left to add. Better sensors, faster sequencing, cleaner assays — these improve the loop's gain and its noise rejection. They do not add a fifth stream beyond everything, continuously watched. The three-week lag was not a data problem. It was an intake architecture problem, and the architecture that closes it is the last rung on this particular ladder, not the last word on public health.