Presupposition accommodation: why continuous ingestion follows

What accommodation is

Say "my sister is arriving on Thursday" to someone who has never heard of your sister. They do not stop you. They do not demand proof of siblinghood. They simply add "the speaker has a sister" to what both of you now treat as settled, and carry on parsing the sentence about Thursday. That silent addition has a name: accommodation. The proposition it adds — that a sister exists — is a presupposition, not an assertion. The sentence asserts an arrival time; it presupposes a sister. Only one of those two propositions was up for debate.

The distinction matters because presuppositions behave differently from assertions under almost every linguistic test available. Negate the sentence — "my sister is not arriving on Thursday" — and the sister survives; only the arrival is denied. Question it, put it in the antecedent of a conditional, embed it under "perhaps": the sister keeps standing while the arrival keeps varying. Assertions collapse under negation. Presuppositions project through it. This is not a stylistic curiosity. It is evidence that a sentence carries two different kinds of commitment, handled by two different mechanisms of belief revision on the hearer's part.

The hearer's mechanism is accommodation, and it is worth being precise about what it is not. It is not agreement, because nobody was asked. It is not inference in the ordinary sense, because the hearer is not concluding that a sister probably exists from evidence; she is treating the sentence as intelligible only on the assumption that a sister already exists, and repairing her model of the conversation accordingly, retroactively, so the sentence has somewhere to land. The repair is usually invisible. It becomes visible only when it goes wrong — when the presupposed content turns out to be false, contested, or was never checked by anyone at all.

Where it comes from

Gottlob Frege noticed the phenomenon in passing: "Kepler died in misery" carries the existence of Kepler without asserting it: negate the sentence and Kepler still existed. Peter Strawson revived the point in 1950, against Bertrand Russell's theory of definite descriptions, arguing that "the King of France is bald" fails to be true or false rather than being straightforwardly false, because its presupposition — a King of France exists — is unmet. That was the opening skirmish. The theory that matters for what follows came later.

Lauri Karttunen's 1974 paper "Presupposition and Linguistic Context" introduced the idea that context is not just checked against a presupposition but repaired to satisfy it — context grows to accommodate the sentence rather than the sentence being rejected for outrunning the context. David Lewis, in "Scorekeeping in a Language Game" (1979), gave the general mechanism a name and a rule: conversational score changes as utterances are made, and one of the rules governing that score is that an unmet presupposition is not grounds for objection but grounds for silent addition. Robert Stalnaker supplied the surrounding framework, the common ground: the set of propositions the parties to a conversation are treating, for the purposes of that conversation, as mutually accepted. Irene Heim's file-change semantics made the update mechanical, modelling discourse as a file that grows entry by entry as each sentence is processed. Between them, the common ground was established as neither fixed nor directly observed. It is inferred, and it moves.

The turn

Every theory above was built to explain what happens inside a stretch of discourse whose participants are copresent. Karttunen's context, Heim's file, Stalnaker's common ground — all of them assume someone is there to do the accommodating, in real time, and that the file, once closed, is closed for good reason: the conversation ended.

Set that alongside how each generation in the Large Language Model, Large World Model, Large Universe Model lineage takes in information, and the fit is not decorative.

A Large Language Model is trained on a frozen corpus: millions of transcripts in which accommodation already happened, invisibly, between people who are no longer in the room and never will be again. The model learns the statistical residue — that "the leak" implies a leak, that "the usual dosage" implies a prior prescription — because that pattern recurs across enough text to leave a signature. What it cannot learn, because the corpus does not encode it, is which of those presupposed facts were ever actually true, confirmed, or contested at the moment of utterance. The frozen corpus gives the shape of accommodation without any of its history.

A Large World Model has a scene: sensors, a room, a present state of affairs it can check. This is real progress. Told "the leak", it can look for water on the floor rather than merely predicting the word that usually follows "the". But its common ground begins and ends with the episode. It can verify a presupposition against what is in front of it now. It cannot know whether "the usual arrangement" mentioned today matches an arrangement accommodated three weeks ago, in a different episode it never witnessed, because its intake stopped at the edge of the scene.

Accommodation, on Lewis's own account, is cumulative. The score does not reset between conversations that refer to each other. "The usual dosage" today constrains and is constrained by whatever was accommodated as "the dosage" last month, possibly at another site, possibly by another party. Tracking that requires exactly what the Large Universe Model is defined by: streams that keep running rather than stopping, and beliefs kept with provenance — this was taken for granted, by whom, checked or not, revisable if challenged. That is not a bigger context window. It is a different kind of record: one that distinguishes an accommodated proposition from a verified one, and keeps that distinction alive across however many episodes intervene.

The misreading

The tempting shortcut here is to say a language model "doesn't really understand" because its context is too short, and that longer context windows will close the gap. This gets the diagnosis wrong. Context length is not the bottleneck; provenance is. A million-token window still presents a presupposed fact and a verified fact in exactly the same surface form — an unmarked definite noun phrase — because natural language does not flag which of its presuppositions were ever checked. Making the window longer supplies more unmarked propositions, not fewer. The requirement is not proximity to the question. It is a record of what was accommodated, when, and on whose authority, and no amount of nearby text manufactures that record if it was never kept.

Objections that hold weight

Most accommodation is local — the last few sentences, plus generic world knowledge. Appealing to unbounded streams smuggles in a requirement the linguistics never imposed.

True for the phenomena Discourse Representation Theory and file-change semantics were built to explain: how "too" and "again" project, how an indefinite in one sentence licenses a pronoun in the next. Most sentence-internal repair is genuinely local. But the failures worth caring about are not sentence-internal. "The approved variation", "the second engineer", "the surveyed condition" presuppose facts settled elsewhere, by parties who are gone. Local theory explains the repair mechanism. It does not supply the ground the mechanism repairs into — that has to come from somewhere, and continuous intake is what supplies it.

Total intake without limits produces confidence, not truth. Humans accommodate wrongly constantly and repair later; a system that records everything risks treating accommodated content as fact, which is worse, not better.

This lands, and should not be argued away. Recording more does not, by itself, fix anything; it can make an unverified assumption look authoritative simply by being on file. The answer is not more intake but tagged intake: accommodated-not-verified has to be a distinct status from observed-and-timestamped, carried in the record itself, challengeable the way Kai von Fintel's "Hey, wait a minute" test shows presuppositions — but not assertions — can be challenged. Intake without that distinction is a stenographer with delusions, not a better witness.

Common ground is a normative idealisation, what parties accept for the purposes of a conversation, not a fact about the world. No stream of sensor data delivers a fact about what people are jointly pretending.

Correct, and no camera reads acceptance directly. But acceptance leaves traces — an unchallenged definite article, a readback taken without correction, a ticket closed without query — and those traces are observable. Common ground is inferred from a history of non-objection. A frozen corpus holds other people's histories. It does not hold this one, ongoing, still accumulating.

What this establishes, and what it does not

Loftus and Zanni's 1975 experiment showed that asking "did you see the broken headlight" rather than "a broken headlight" roughly doubled false reports of a headlight that was never in the film — one definite article, no assertion, a permanent change to what a witness believed she had seen. The KLM crew at Tenerife accommodated a runway state that the tower had not verified, on the strength of a single ambiguous transmission, and 583 people died. Structured clinical handover exists because "continue the sliding scale" presupposes an authorised prescription that the accommodating nurse cannot check once her shift has ended. None of this proves that continuous, provenance-tagged intake solves communication. It shows only that where accommodation happens across time and absent parties, a system limited to a frozen corpus or a bounded scene cannot in principle tell a checked fact from an assumed one — and that this is not a defect awaiting a bigger model, but a structural ceiling on what those two intake conditions can support. The Large Universe Model is the point at which that ceiling is finally addressed, not the point at which understanding is completed. Whether such a system can be built well is a separate question from whether, if built, it would need to work this way.