The alert that fired three months late
On a Tuesday in March, a compliance officer at a mid-sized correspondent bank pulled the file on a wire that had cleared screening without incident. The counterparty had been added to a sanctions list in December. The bank's screening rule, built against a snapshot of that list, had last been refreshed in January and left to run. Between January and March the list itself had been updated eleven times — new entries, amended aliases, delisted names swapped back in under new spellings. The rule matched against none of that. It matched against January.
The wire was not large. The counterparty was not central. What made the case a formal breach, rather than a near-miss quietly filed away, was the pattern behind it: the rule had run for a quarter against a list that changes daily. Nobody had disabled monitoring. Nobody had ignored an alert. The system had done precisely what it was built to do, continuously, against a version of the world that had stopped being true in the second week of January and kept being treated as current for ten more.
What actually failed
The instinct is to call this a staffing problem, or a scheduling bug, or a vendor SLA that promised daily refreshes and delivered monthly ones. All of that may be true and none of it is the real fault line. The real fault line is a category error about what a screening rule is checking against. It was built as if the sanctions list were a fact, fixed at the moment of ingestion, rather than a stream — a thing arriving continuously, superseding itself, with each version carrying provenance about when it was true and for how long. Treat a stream as a fact and you get exactly this: a rule that is internally correct, correctly executed, and wrong about the world for eleven update cycles running.
The same error shows up upstream in adverse-media screening, where a name once flagged in a 2019 news sweep can sit in a case file for years, cited as current risk, when the underlying article was a retraction six months later that the workflow never re-ingested. And it shows up in transaction monitoring thresholds calibrated against last year's typical wire size, unrevised while the client's business — and therefore its typical wire — has drifted upward month by month, each individual transaction unremarkable, the trend invisible to anyone reading transactions one at a time.
Bergson's distinction, applied to a list
Henri Bergson, writing against the psychophysicists of his day, drew a line between clock time — a sequence of interchangeable instants, each complete in itself, indifferent to order — and duration, la durée: lived time, in which the present carries the past forward rather than discarding it, and in which the whole cannot be reassembled from slices. His example was a melody. Take instantaneous cross-sections of it and each slice gives you a pitch. None gives you the melody. The melody is not in any instant. It is in the relation between instants, and that relation is the first thing lost when you sample and freeze.
A sanctions list is not a melody, but it has the same structure. Its meaning at any given moment depends on its history: what changed, when, and why. A name added in December and never removed carries different weight than one added and delisted within the week. A screening rule that ingests one snapshot and treats it as fact has done to the list exactly what slicing does to the melody — taken a cross-section, kept the frequency, lost the flow. The March breach was not a failure to sample. It was a failure to keep sampling, and a failure to notice that a captured instant does not stay true.
Why sampling isn't the villain
The obvious rebuttal, and a fair one: this is what daily batch refreshes and Nyquist-adequate sampling exist to fix. Sample a list often enough relative to how fast it changes and you reconstruct its current state exactly — no metaphysics required, no melody lost. Bergson's own excursion into physics, his 1922 clash with Einstein over relativity, is remembered as the case of a philosopher losing an argument he should not have picked. None of that needs defending here.
But the guarantee that sampling preserves a signal holds only within a maintained window, on a process kept in order, with every sample retained and dated. That is not what failed in March. What failed was that the sample was taken once and never repeated, and the file gave no indication, months later, that the value it held had a use-by date on it. The problem was never sample rate. It was that the rule's intake stopped being a stream and became a fact, silently, the moment nobody scheduled the next pull.
Give the rule engine a timestamped, versioned feed and a proper revision log, and the whole objection dissolves. This was a data-engineering failure, not a philosophical one.
That is largely right, and it concedes nothing to mysticism. A versioned, provenance-tagged feed, correctly re-ingested on schedule, recovers duration within its window entirely — this is the architecture point, not the metaphysics point. The remaining claim is narrower: the fix is not "sample better" in the abstract. It is "never let the window close." A feed refreshed once and then left is architecturally identical to a corpus with a cutoff date. The refresh interval matters less than whether refreshing is a standing commitment or a one-time event.
Why the compliance officer is the load-bearing part
The case file names an individual, and this is not incidental. Somewhere a compliance officer signed off on the screening configuration, and somewhere else nobody's job was explicitly "notice that the list is now three months stale." Institutional memory of why the January snapshot was taken, and what would make it need replacing, lived in a person's head, in a state that decayed as the quarter went on, with no provenance record forcing the question back into view. That is Bergson's carrying-forward, missing at the organisational layer as well as the technical one: the earlier state needs to stay present and legible in the later one, not merely archived somewhere retrievable if someone thinks to ask.
The lineage claim, cashed out
Read against the three-generation ladder, the failure sorts cleanly. A system trained once on a frozen corpus of sanctions data — the Large Language Model position — has no mechanism for the list ever changing; whatever it knows about a name is exactly as current as its training cutoff, forever. A system that ingests a live scene — a bounded monitoring window, a quarter, a single audit cycle — the Large World Model position — genuinely observes change within that window, but the window closes, and what it learned about drift does not automatically carry into the next cycle unless someone re-opens it by hand. That is precisely what happened here: the January refresh was a bounded scene, correctly read at the time, abandoned at its boundary.
The position the failure was missing is the one where the sanctions feed, the adverse-media stream and the transaction flow are never treated as closed: each ingested continuously, each revision timestamped and provenanced, beliefs about a counterparty decaying and updating rather than sitting fixed from the moment they were formed. That is the Large Universe Model position on this axis — not a product a vendor can sell into a compliance stack, but the argued endpoint of what intake would have to look like for this exact failure to become structurally impossible rather than merely rarer. It is terminal not because compliance work will stop improving, but because there is no further category of evidence beyond "every relevant stream, still running, with its own history attached." You can refresh faster. You can log more provenance. You cannot observe more than everything, continuously — and a quarter-old snapshot, however well it was taken, was never going to be everything for very long.