The alert that came ninety days too late
At 06:42 on a Tuesday, a payment for £340,000 clears from a shell entity in Limassol to a logistics firm in Rotterdam. The screening engine checks the beneficiary name against a sanctions list, finds no match, and lets it through. The name is clean because the rule set was last compiled at the start of the quarter. The beneficiary was added to the consolidated list eleven weeks earlier, on a Thursday, in a routine update that also touched four hundred other entries. Nobody re-ran the historical book against the new list, because nobody was told to. The screening rule was correct on the day it was written and wrong for every day after the list moved and it didn't.
This is not a story about a bad model. The matching logic performed exactly as specified, on exactly the data it was given. The failure sits one level up, in the gap between when the world changed and when the system was told the world had changed. A compliance officer signing off the quarterly attestation had every reason to believe the control was live. It was live in the sense that it ran. It was not live in the sense that mattered: it was not phased to the list it was supposed to be checking against.
Diagnosing the gap, not the rule
Call the rule set a model of who is currently prohibited. That model was built once, from a snapshot, and then executed unchanged for thirteen weeks while sanctions lists updated daily, adverse-media feeds refreshed hourly, and the transaction book grew by tens of thousands of lines. The model didn't get worse at matching. The target it was matching against moved out from under it. Error here is not a property of the algorithm's sophistication; it is a function of elapsed time since the rule set was last synchronised to source. A better matching algorithm, run on the same stale list, produces the same false clearance with more confidence.
This is a general problem with any internal representation of an external, changing condition. The representation has its own life once built — it sits in a rules engine, gets re-used, gets trusted — and that life proceeds regardless of whether the thing it represents has moved. Biology names this precisely. An internal oscillator with its own period, left unchecked, drifts from the external cycle it is meant to track. The fix in biology is not a better clock. It is a permanent, gated channel back to the source of truth.
Entrainment, briefly, and why gating matters
Nearly every organism carries a clock whose free-running period is close to but not exactly twenty-four hours — in humans, about 24.2 hours on average. Isolated from light, that clock keeps ticking and drifts. What holds it to the solar day is entrainment: specialised photoreceptors report ambient light to the suprachiasmatic nucleus, which nudges the oscillator's phase by roughly an hour or two a day, and only at the phases where light is a meaningful signal. The clock is endogenous. Its accuracy is borrowed, continuously, from outside.
The screening rule set is an oscillator. It has an internal period — the review cadence, the quarterly re-tune — and that period is close to right but not exactly right, because the sanctions list, the adverse-media feed and the transaction pattern do not move on a quarterly clock. Left to run free between reviews, it drifts. What it lacks is not intelligence. It lacks a zeitgeber: a channel that reports change back to the rule set continuously, gated so that only genuine list movement — not every rumour, not every noisy media hit — actually shifts the model's phase.
Where this sits on the intake axis
A Large Language Model is a clock set once, at training cutoff, and then run free. Applied to sanctions screening, it would encode the state of the world as of some date and hold that state forever, with no channel back to the actual list. A Large World Model is better: it senses a bounded scene — the current batch of transactions loaded against the current list snapshot — and is genuinely re-phased while that scene is in view. But between screening runs it free-runs again, and its accuracy is a function of how recently it last looked, which is exactly the shape of the ninety-day failure above.
A Large Universe Model is what the failure implies by absence: sanctions updates, adverse-media feeds and rule changes as streams that never stop arriving, each belief about a counterparty's status carrying a timestamp and a source — this name cleared against list version 4,412, published Thursday, ingested Friday — so that when the list moves, every downstream belief that depended on the old version is flagged, not silently kept. Nothing beyond continuous, provenance-tagged correction needs to be added; more history and better trust weighting are refinements of that position, not a fourth category above it.
| intake | characteristic failure mode | |
|---|---|---|
| Large Language Model | list as encoded at training | clean name stays clean forever |
| Large World Model | list as loaded for this run | clean between runs, stale between them |
| Large Universe Model | list as a live, timestamped stream | stale beliefs are flagged, not hidden |
The objection that the sun is not a sanctions list
Entrainment corrects one variable against one exquisitely reliable signal. Compliance has thousands of variables — names, aliases, ownership chains, transaction patterns — arriving from feeds that contradict each other, lag unpredictably, or are manipulated by the very entities being screened. The biological analogy borrows the sun's dependability to flatter a much messier problem.
The disanalogy is real, and it sharpens rather than dissolves the point. The circadian system does not treat all cues equally: light dominates, but the system gates its own sensitivity by phase and discounts feeding or temperature cues that conflict too strongly with the dominant signal. That selective weighting is the actual mechanism, not an afterthought to it. In compliance terms, the equivalent is a hierarchy of sources: the consolidated sanctions list update is treated as an authoritative zeitgeber that forces immediate re-screening of affected names; an adverse-media hit on an unverified outlet nudges a risk score without triggering a freeze; a rule change from a regulator gets logged with an effective date and a grace period. Without that source hierarchy and without provenance recording which feed moved which belief and when, continuous intake collapses into noise — every rumour treated as a sanction, every screening run thrashing. The objection correctly identifies the hard engineering. It is not a reason to prefer the quarterly snapshot.
The objection that stability has a cost
Chasing every feed update produces the compliance equivalent of jet lag: a screening engine re-tuned on every incoming signal loses coherence, over-reacts to outliers, and forfeits the stability that made its decisions defensible to a regulator in the first place.
This is the strongest objection and it should be conceded almost in full. Shift workers and long-haul travellers show what happens when a clock is exposed to unstable zeitgebers: peripheral oscillators fall out of step with the master pacemaker, and performance suffers in ways that a free-running but internally coherent clock would not. A screening system that re-scores every counterparty on every hourly media crawl would produce exactly this: an operations team chasing false positives, a book that never settles, an audit trail nobody can reconstruct.
The biological answer is not to seal the organism off from light. It is hierarchy and rate limits: one master pacemaker sets phase, peripheral clocks follow it rather than reacting independently to local cues, and the whole system's daily phase-shift capacity is bounded to roughly an hour or two. Translated: sanctions list changes are the master signal and force immediate, bounded re-checks against the affected population only; adverse-media feeds inform a slower-moving risk score with its own review cadence; rule changes propagate on their own scheduled path with effective dates logged. Continuous intake does not mean unbounded reactivity. It means a designed hierarchy of correction rates, each with its own gate, so the system stays coherent while still never fully closing the channel back to source. That is a specification for the third position on the intake axis, not an argument for staying at the first.
What the compliance officer actually needed
Not a smarter matching algorithm, and not a system reacting to every headline in real time. What was missing at 06:42 that Tuesday was a record, attached to the cleared transaction, of which list version it had been checked against and when that version was superseded — so that the eleven-week gap between the list update and the next full re-screen would have shown up as a flagged, aging belief rather than a silent, confident clearance. The rule set was never going to be right forever. The only design that survives that fact is one where correction against the live list never stops, and every clearance carries the timestamp of the source it trusted. Beyond that there is no further category to reach for — only more streams, better source hierarchy, and a longer memory of what was believed and why.