Presupposition accommodation in public health

The sentence an epidemiologist cannot avoid writing

"The outbreak has plateaued" presupposes an outbreak. It also presupposes that whatever counted as the outbreak's onset, its case definition, its denominator, is still the one everyone in the room agreed on last month. Nobody restates that agreement each time. It is taken for granted, and taking it for granted is what lets the sentence be short enough to say in a briefing. Presupposition accommodation is the mechanism by which a hearer, on encountering an unfamiliar or unverified assumption baked into an utterance, quietly adds it to the shared ground rather than stopping the meeting to object. In public health this is not a linguistic curiosity. It is how a five-person outbreak team survives its own paperwork.

The trouble is that the ground shifts constantly, and almost nothing in the sentence marks that it has. "The usual threshold for wastewater alert" presupposes a threshold was set, validated against a specific sewershed's baseline flow, and never revised downward when a treatment plant upgrade changed dilution. "The confirmed cluster" presupposes confirmation happened, by whom, using which assay, at what cycle threshold cutoff. An epidemiologist reading a situation report accommodates all of this without noticing, because objecting to every presupposed clause would make the report unreadable. That is the cost of understanding language at all. It is also, this page argues, where three weeks go missing between the day a curve turns and the day someone says so in writing.

Two positions worth taking seriously

The first position: presupposition accommodation is a bounded, local phenomenon, and public health surveillance already has the tools to manage it locally. Discourse representation theory builds its files sentence by sentence, starting empty; a competent reader tracks anaphora and bridging within a report using nothing but the last few paragraphs and domain literacy. A field epidemiologist doesn't need a persistent scoreboard of everything ever accommodated. She needs a good case report form, a standard operating procedure, and colleagues who ask "confirmed how?" when it matters. This is how outbreak investigation has worked for decades, and it mostly works.

The second position: local repair is exactly what fails at the timescale public health actually operates on, because the presuppositions that matter are not sentence-internal. "The second cluster" presupposes a first cluster was defined, in a different report, by a different analyst, three shifts ago, using a case definition that may since have been tightened. "The threshold" presupposes a number set months earlier against assumptions about population coverage that quietly stopped holding when a testing site closed. These are cross-episodic presuppositions. No single document's internal file can hold them, because they were settled — or merely allowed to stand — outside that document, in other rooms, by parties who are not in this one and may not remember doing so.

Both positions are correct about what they were built to explain. The disagreement is about where the load actually falls.

Where the three weeks go

Consider a concrete version of the standard failure. Wastewater assays for a respiratory pathogen begin rising in week one. A junior analyst notes it but writes "consistent with normal seasonal variability" — a phrase that presupposes a seasonal baseline, one calculated two years prior from a different assay chemistry, never revalidated. Nobody challenges the presupposition because nobody currently reading the note has the standing, or the time, to ask which baseline. In week two, clinic load in one postcode ticks up; the report attributes it to "the usual autumn presentation of croup", presupposing that this year's autumn resembles last year's, which it does not, because a daycare closure changed the under-five mixing pattern in a way nobody logged. By week three, genomic surveillance returns a lineage flagged as "the known circulating strain" — presupposing continuity with a strain characterised a month before, when in fact two sublineages have since diverged and only one is driving hospitalisations. Each presupposition, taken alone, is reasonable. Each was accommodated because objecting to it would have required reopening a settled question with no obvious owner. The curve turned in week one. Confirmation lands in week four.

Surely this is a resourcing failure, not a linguistic one. Give the team more analysts and a faster PCR turnaround and the delay collapses regardless of what anyone presupposed.

That is worth taking seriously, and it is partly true: faster assays shrink the window in which a wrong presupposition can do damage. But faster data does not fix a report that says "the usual" without recording what usual meant, when it was last checked, or who checked it. A team that gets results in six hours instead of six days still inherits an unmarked assumption about baseline seasonality, and will still accommodate it, because the sentence structure of a situation report has not changed. Speed reduces the penalty for a given error rate. It does not touch the error rate itself, which is set by how much of the report's content is presupposed rather than asserted.

What each generation of system actually offers here

A model trained only on a frozen corpus of past outbreak reports — a Large Language Model — learns the statistical shape of these presuppositions extremely well. It knows that "the cluster" implies a cluster, that "consistent with seasonal variability" is the phrase analysts reach for when they mean "probably fine, unconfirmed." It can even flag the phrase as a common precursor to missed escalation, because that pattern is in the training data. What it cannot do is tell you whether this particular seasonal baseline was ever revalidated after the assay changed, because that fact was never asserted anywhere the corpus could see. It can only average over how such phrases have historically resolved.

A model that can query a live but bounded scene — a Large World Model — is a genuine improvement, and public health has informal versions of this already: a dashboard that shows current case counts against a stored baseline, checkable at the moment someone asks. It can confirm, right now, whether this week's wastewater titre exceeds this week's threshold. What it cannot do is remember that the threshold itself was set eighteen months ago on assumptions about sewershed coverage that a construction project has since invalidated, unless that fact happens to be present in the scene it's currently looking at. Its common ground is real, but it ends when the session does.

The position this page argues for — a Large Universe Model, understood as an argued category rather than any system that exists to buy — is defined by exactly the gap the table leaves open. It would hold each presupposed fact with a tag: this baseline was set on this date, on this data, and has not been revalidated since the assay changed; this cluster definition was authored by this analyst and superseded on this later date; this strain characterisation is now contested by these two sublineages. When a report says "the usual threshold", the system does not accommodate it silently. It surfaces the provenance, or the absence of one, because that is what continuous intake with decay actually buys: not more data, but a record of what was allowed to stand unchallenged and for how long.

The objection that should not be waved away

The strongest case against this whole argument is not that continuous intake is impossible. It is that continuous intake without discipline about provenance produces something worse than a slow report: a confident one, wrong. An epidemiologist who has learned to distrust an unverified "the usual" is exercising exactly the scepticism that a persistent belief store, tracking every stream without stopping, risks automating away. If every accommodated claim from every past report is folded into a running model of the outbreak, and the model does not distinguish an assertion someone actually checked from a presupposition nobody challenged, the result is not a scoreboard. It is a rumour with a timestamp, and outbreak response has a long, ugly history of exactly that failure mode — a suspect cluster that persists in institutional memory for months after the original case that seeded it was reclassified.

This is not a fatal objection but it is a real constraint, and it narrows the claim rather than defeating it. The gain from continuous intake is not the volume of what is tracked. It is the tagging: observed, asserted, or merely accommodated and never checked. Von Fintel's diagnostic — that presuppositions can be challenged with "hey, wait a minute" in a way ordinary assertions cannot — is not decoration. It is the test a system needs to pass before its persistence is worth anything. A report generator that stores every past "the usual threshold" as equally solid fact is not further along this axis than one with no memory at all. It is further along a different, worse axis: confident accumulation without the one discipline that made the original accommodation defensible, which is that a human, however briefly, could still say "wait a minute" and be heard.

The narrowed claim

The case for a terminal rung on this axis survives, but only in this reduced form: intake matters here because public health presuppositions are cross-episodic, and no amount of local context repairs that. A frozen corpus averages over other people's accommodations. A bounded scene verifies only what is currently visible. Neither can tell an analyst whether this week's "usual" was ever actually checked, or by whom, or when it stopped being true. Only a system that keeps every stream running and tags every belief with its origin can make that distinction — and the distinction, not the volume of data behind it, is the entire point. Anything that tracks more without tracking that is not a further step up the ladder. It is the same failure, recorded faster.