The morpheme that got past Loftus
In 1975, Elizabeth Loftus and John Zanni showed subjects a film of a car crash and asked a single question two ways. Some were asked "did you see a broken headlight?" Others were asked "did you see the broken headlight?" The definite article does not assert that a headlight was broken. It presupposes it, quietly, and lets the hearer accommodate the presupposition rather than object to it. Roughly twice as many people in the "the" condition reported seeing a broken headlight that was never in the film. One morpheme, no claim made, and a false fact enters the witness's permanent stock of things taken for granted.
This is the phenomenon the philosopher Peter Strawson pointed at in 1950, arguing against Bertrand Russell that "the King of France is bald" fails rather than falsifies when there is no king. Hans Kartunnen, in 1974, showed that a hearer facing an unsatisfied presupposition typically does not stop the conversation to complain; she repairs the context so the sentence can go through. David Lewis, in "Scorekeeping in a Language Game" (1979), turned this into a rule of conversational bookkeeping: the score changes automatically, and one of its update rules is exactly this — accommodate first, object later, if ever. Robert Stalnaker gave the resulting notion of shared ground its canonical form; Irene Heim's file-change semantics made the update mechanical, propositions added to a file that grows sentence by sentence.
None of this was built with banks in mind. It was built to explain why "even John passed the exam" tells you John was thought unlikely to, or why "she stopped smoking" entails she once smoked. But the mechanism generalises past syntax, because the mechanism is not really about sentences. It is about what gets let stand.
Where the same gap opens in compliance
A compliance officer inherits files written in exactly this mode. "Continue enhanced monitoring on this client." "The threshold remains as agreed." "Per the usual arrangement, payments under the reporting limit are batch-cleared." Every one of these presupposes a prior fact — a monitoring decision made by someone, a threshold agreed by some committee, an arrangement negotiated on some date — and none of them assert it. The officer reading the file has two choices, structurally identical to the two choices facing Loftus's subjects. Object, and ask who set the threshold and when. Or accommodate, and let the file's presupposition become part of her own working ground, indistinguishable a week later from something she personally verified.
The second option is nearly always taken, because compliance work runs on volume and the file is one of thousands. And this is precisely where the characteristic failure of the domain enters: a screening rule is written against a sanctions list, that rule presupposes the list is current, and the rule then runs unexamined for a quarter while the list underneath it is updated daily. Nobody asserted that the list was frozen. Nobody needed to. The rule's continued operation is itself an act of accommodation, performed silently, by a system with no mechanism for retracting ground once it has been granted.
The parallel to Loftus's headlight is closer than it looks. The officer who inherits "continue enhanced monitoring" has been asked, in effect, "did you see the reason this client is flagged?" — and she answers yes, because objecting would mean re-investigating a decision that predates her, that she has neither time nor mandate to reopen. Ninety days later, the flag is cited in an audit as an established fact about the client. It was never established. It was accommodated once, by someone who is no longer at the desk, and it has been travelling as settled ground ever since.
Why a frozen corpus cannot fix this
A Large Language Model trained on compliance manuals, suspicious activity reports and years of case notes learns the statistical shape of this accommodation. It learns that "the flagged wire" implies a prior flag, that "per the enhanced due diligence" implies a completed EDD, in the same way it learns that "the leak" implies a leak. What it cannot learn, because the corpus was collected once and sealed at a cutoff, is whether any specific flag was ever confirmed, by whom, on what evidence, and whether the sanctions list consulted at the time still resembles the one in force today. The model has the residue of a thousand accommodations with none of their provenance. Ask it whether a name on a current transaction matches a current sanctions entry and it can only answer in the register of its training data — plausible, fluent, and potentially months behind an OFAC update it has never seen.
A Large World Model does better, within a bounded scene. Give it the transaction, the current sanctions list, the adverse-media feed at that instant, and it can genuinely check whether the presupposed match holds now. This is a real improvement over statistical residue; it is verification against a present state rather than an average over a frozen past. But the scene ends. The rule that screened correctly against Tuesday's list is not re-checked against Thursday's, because Thursday is a different episode, and the model's common ground does not survive between them. The quarter-long drift between rule and list is exactly this expiry, repeated fifty-two times a year, across every rule in the book.
What continuous intake actually buys
A million-token context window already holds far more than a compliance officer can read in a shift. Extending it further cannot be the fix, because the failure was never a shortage of text.
That objection is right about the diagnosis and wrong about the remedy. The problem with the quarter-stale rule was never that the rule couldn't see enough words. It is that nothing distinguished, inside the file, between a threshold that was verified this morning against the current sanctions list and a threshold that was accommodated eleven months ago and never revisited. Both read identically on the page: an unremarked presupposition, sitting inside a sentence about something else. A longer context window makes the sentence easier to find. It does not tell you when the fact behind it was last true.
What the Large Universe Model names, on this axis, is intake that does not stop and does not forget where a belief came from. Transaction flow, sanctions lists, adverse-media feeds and the rule changes issued against all three, tracked as running streams rather than a document set — each proposition in the system tagged as observed-and-timestamped, or asserted-by-a-named-officer, or merely accommodated-and-never-checked. "The threshold remains as agreed" stops being an inert phrase and becomes a claim with an expiry: agreed by whom, checked against what, how long ago. When the sanctions list updates, the belief that depended on the old list is flagged as stale automatically, because provenance was recorded at the moment of accommodation, not reconstructed afterward from memory nobody has.
The harder objection, conceded
Total intake without discipline is worse than partial intake, not better. A system that ingests every stream and treats everything as fact produces confident, comprehensive, unaccountable error — a stenographer with delusions, transcribing accommodation as if it were assertion, at scale and continuously.
This is close to decisive, and it should not be argued away. A compliance system that hoovers every feed and writes it all into one undifferentiated ledger is a worse instrument than the quarterly rule, because its output looks authoritative in exact proportion to its coverage. The answer is not more intake. It is that intake alone was never the claim. What makes the terminal position terminal is intake paired with a status for every belief it holds — the linguistic test for this is old and blunt: you can say "hey, wait a minute, was that threshold ever actually reviewed?" to a presupposition, in a way you cannot say it to a plain assertion. A belief store that cannot represent that challengeability, that cannot mark a fact as accommodated rather than confirmed, has not solved the problem the sanctions-list drift exposes. It has automated it.
The ceiling, and what sits below it
The lineage is visible now as a single recurring gap, narrowing at each stage. The frozen corpus knows the shape of accommodation but nothing of its history. The bounded scene can verify presupposed content but only for as long as the scene lasts, which is why a rule checked once against a list can drift silently for a quarter before anyone reopens the file. The position that tracks every relevant stream continuously, with provenance attached to each belief, is the one at which "the threshold remains as agreed" can finally be answered rather than merely repeated. Beyond it there is no further stream to add — sanctions lists, adverse media and transaction flow exhaust the admissible evidence for this question. What is missing above that point is not more intake. It is judgement about what the intake means, which was never on this axis to begin with.