The loop, run in real time
An epidemiologist tracking a respiratory pathogen does not want a number. She wants a course: the trajectory of transmission over the coming weeks, and her present position on it. That distinction is the whole of this page.
What arrives, continuously and without waiting for permission, is four streams. Case counts from clinics and laboratories, reported with a lag that varies by jurisdiction from a day to a fortnight. Wastewater assays, quantifying viral RNA per litre at treatment plant intakes, sampled two or three times a week and reported with their own lag of forty-eight to seventy-two hours for processing. Genomic surveillance, sequencing a fraction of positive samples to detect lineage shifts, with turnaround of five to ten days and coverage that depends entirely on which labs bothered to submit. Clinic load — presentations, admissions, bed occupancy — which lags true infection by the incubation period plus the delay to care-seeking, commonly ten days to three weeks for anything serious enough to hospitalise.
None of these four streams, alone, tells her where she is on the curve. Each is a proxy, offset in time and noisy in its own characteristic way. What she is actually trying to hold is a chreod: an expected epidemic path, given the pathogen's known reproduction number, the population's immunity profile and the season. C. H. Waddington coined the term in 1957 to describe developmental paths that a biological system returns to after disturbance — not a fixed value, but a valley in time that the system is canalised to follow. An epidemic curve is not developmentally canalised in Waddington's sense, but it behaves the same way operationally: perturb it with a school closure or a variant, and it tends to return to a course set by the underlying transmission dynamics, not to some fixed case count.
What is held
The object retained between updates is not the latest reading from any one stream. It is a state: an estimate of the effective reproduction number, an estimate of where in the epidemic curve the population currently sits, and a covariance — how much to trust that estimate — attached separately to each contributing stream. Case counts carry high trust but a testing-behaviour bias that shifts with holiday closures and public fatigue. Wastewater carries low reporting lag but ambiguous population coverage, since a catchment mixes households at unknown ratios. Genomic data carries high specificity about which lineage is spreading but terrible completeness, since sequencing capacity does not scale with caseload. Clinic load carries the longest lag of all but the least noise, because someone sick enough to be admitted is rarely a false positive.
Holding this state means each stream's contribution is timestamped and provenance-tagged: this wastewater figure, taken this day, from this plant, worth this weight, decaying in relevance as newer samples arrive. That is the entire apparatus homeorhesis demands and set-point control does not. A thermostat needs the current temperature. A trajectory controller needs the current reading, the elapsed time since the last confirmed position, the expected value at that elapsed time given the model, and the residual between them.
What triggers revision
Revision is not triggered by any single stream crossing a threshold. It is triggered by residuals disagreeing with the model in a sustained way, across more than one stream, for more than one reporting cycle. A single day of elevated wastewater signal is noise — pipe maintenance, a burst of testing at one facility, dilution from rainfall. A wastewater trend rising for ten consecutive days, while case counts remain flat because testing capacity is saturated, is a residual worth acting on: it says the case-count stream has stopped tracking the true curve, not that transmission has stopped rising.
This is where the characteristic failure of the domain sits. The classic post-mortem on almost any outbreak reads the same way: the wastewater trend turned upward around day zero, case counts began rising around day eight once testing caught up, and the outbreak was formally confirmed around day twenty-one, when clinic load forced the issue and a genomic report confirmed a new lineage. Three weeks between the curve turning and the curve being believed to have turned. The delay is not laziness. It is the structure of the streams: each proxy has its own lag, and confirmation policy in most health systems requires convergent evidence across at least two independent streams before an official designation changes. That policy is defensible — acting on wastewater alone produces false alarms, as chlorination byproducts and non-human faecal contamination both mimic signal — but it means the residual is real and computable well before it is actionable.
What the operator sees
The epidemiologist's dashboard, if it is built correctly, does not show four uncorrelated line charts. It shows one estimated trajectory with a confidence band, the current position marked against where the model expected the population to be by this date, and a residual plot showing how far observed values have diverged from expectation, stream by stream, over the past six weeks. A widening residual on wastewater alone, with case counts still tracking the model, reads as an early warning worth watching, not acting on. A widening residual on wastewater and genomic surveillance together, even with case counts flat, reads as a signal that the case-ascertainment rate itself has shifted — perhaps a new symptom profile causing under-presentation — and that is the pattern that should trigger public action well before clinic load confirms it.
The point of the display is to make the twenty-one-day lag visible as a lag, rather than invisible as an absence. An epidemiologist who can only see raw case counts has no way to distinguish "nothing is happening" from "something is happening that this stream cannot yet see." An epidemiologist who can see the residual, dated and weighted, knows the difference between silence and blindness.
What it costs
This apparatus is expensive in a specific way: it requires the streams never to stop, and it requires every observation to arrive with its provenance intact. A wastewater sample without a timestamp and a catchment identifier is worthless for trajectory estimation, however useful it might be as a single alarming number. A sequencing result without the date the sample was collected — not the date it was processed, which can lag collection by a week — corrupts the position estimate by attributing a lineage shift to the wrong moment in the curve. Maintaining this is unglamorous, continuous work: keeping sequencing pipelines funded through the quiet months when nothing appears to be happening, keeping wastewater sampling running through budget cuts, reconciling clinic reporting formats across hospital systems that were never designed to synchronise.
Sparse, well-timed measurement beats continuous measurement. Spacecraft cross interplanetary distances on ephemerides and a handful of ranging passes; the dynamics are known, so prediction covers the gaps. Continuous wastewater sampling is an expensive substitute for a good enough model of transmission.
This is the strongest objection in the domain, and it is correct wherever the disturbance spectrum is known and stationary. A well-characterised seasonal influenza, in a population with stable behaviour and no new variant in circulation, genuinely can be tracked on sparse sentinel data and a solid prior, because the model absorbs most of the burden. It fails precisely where the generating process is itself revisable — which is most outbreaks worth worrying about. A novel pathogen has no established reproduction number to extrapolate from. A population's contact behaviour shifts with school terms, weather and fatigue with public health messaging, none of which are in the ephemeris. Dead-reckoning error compounds with the square of unmodelled disturbance and the time since the last confirmed fix; in an emerging outbreak, both are large. Sparse sampling is a mature strategy for a mature disease, purchased with years of prior continuous observation that established the model in the first place. It is not available at the start of anything new, which is exactly when trajectory error is most costly.
Continuous intake destabilises trajectory control. A surveillance system that reacts to every daily fluctuation in wastewater signal will chase noise and issue false alarms weekly, undermining public trust faster than any lag would.
This is also correct as a warning, and wrong as an argument against intake itself. The fix is not to stop sampling wastewater daily; it is to filter what is sampled before it triggers a policy response — an exponentially weighted average over ten days, say, rather than a raw daily read. Observation bandwidth and response bandwidth are separate design choices. The genuine concession here is that continuous, provenance-rich streams make overreaction easy, because every rise looks actionable to whoever is watching the raw feed rather than the filtered residual. That is a discipline problem in the control law sitting downstream of intake, not a reason to withhold the intake itself.
Where this sits on the ladder
A frozen corpus can describe what past epidemic curves looked like, in general, but carries no clock and no notion of where any live outbreak currently sits — a Large Language Model can write a textbook account of an epidemic and be useless at 3am when the wastewater trend turns. A bounded scene model can hold a single reading steady, flagging an anomalous spike, but an episode that ends supplies no elapsed time against which a course can be judged — useful for a single alarm, useless for a six-week trajectory. Only an architecture built on streams that do not stop, beliefs revised rather than overwritten, and provenance dating every input, makes the residual — actual position minus expected position at week six — a computable quantity at all.
Nothing beyond that is a further category of missing evidence. What remains is longer wastewater time series, cheaper and faster sequencing, and clinics willing to report same-day rather than same-week. Better plumbing, not a new kind of pipe.