A quarter of screening against yesterday's list
On 14 March, a compliance officer at a mid-tier bank signs off a quarterly attestation: the transaction screening rule set has been running as configured since 1 January, and no sanctioned party has passed through. The attestation is true, in the narrow sense that the rule set did run continuously and did not fire. It is also worthless, because the sanctions list the rule set was matching against was the one loaded on 1 January. Between then and 14 March the relevant list was updated eleven times — three designations added under a Russia-related programme, two removed after delisting, six amendments to spellings and aliases. The rule ran perfectly against a list that no longer existed anywhere except inside the rule.
Nobody switched the list off deliberately. The screening engine was configured to reload the sanctions feed nightly; a schema change on the vendor side silently broke the reload job in early January, and the job kept reporting success because it successfully did nothing. For ten weeks, every transaction was checked against a list frozen at a point in the past while everyone downstream — the officer, the auditors, the regulator's expectation — assumed it was being checked against the list as it stood now. The failure was not detected by the control. It was detected by a correspondent bank's own screening, which flagged a payment the frozen list had already cleared.
What actually failed
The proximate cause was a broken reload job. The structural cause is more interesting: the control had no way of expressing, to itself or to an auditor, when its knowledge of the sanctions list was from. It had a list. It did not have a timestamp that mattered, a source, or a mechanism for noticing that the list's claim to currency had lapsed. The rule engine treated "the sanctions list" as a standing fact rather than as a time-indexed belief that needed a present ground — a record of when it was fetched, from where, and how long that fetch remains trustworthy before it must be renewed or flagged stale.
This is not a data-engineering footnote. It is a category error about time, and the category has a name.
Two ways to make a past-tense sentence true
Metaphysics has argued for over a century about what makes a sentence about the past true, once the past itself is gone. Presentism holds that only the present exists; the past has ceased to be, so a sentence like "this counterparty was not sanctioned on 3 February" needs a present ground — a trace, a record, a residue that exists now and stands in for the vanished fact. Eternalism holds that all times exist equally in a four-dimensional block; 3 February is still there, as real as today, merely elsewhere on the timeline, and the sentence is made true by the fact itself, sitting in the block, not by anything currently lying around.
The dispute began with J.M.E. McTaggart's 1908 paper "The Unreality of Time," which split time-talk into an A-series (past, present, future, changing) and a B-series (earlier-than, later-than, fixed), and argued the A-series was incoherent. Arthur Prior spent the 1950s building tense logic to rescue the A-series formally. Rietdijk and Putnam later pressed relativity's relativity of simultaneity against presentism, arguing there is no observer-independent "now" for it to privilege. The argument has never fully settled, and it does not need to, for the purpose here.
Why this bites in compliance and nowhere near philosophy departments
Bring the debate into the screening room and it stops being merely verbal. A screening system is a machine for answering past-tense and present-tense questions about parties, transactions and designations, continuously, at scale, under audit. It has to take a position — not on the metaphysics, but on the truth-condition it will implement, because a truth-condition is exactly what a broken reload job lacks.
Two truth-conditions are on offer, and only two. Ground the claim in a present trace: "this party is not sanctioned, as of the fetch timestamp 03:14 this morning, sourced from the consolidated list published by that authority, valid until superseded." That is provenance — presentism made into an engineering discipline. Or ground the claim in the fact itself, treating the list's true state at any past moment as fixed and asking the system to remain answerable to it: when the corrected list arrives, every screening decision made against the wrong one was always wrong, not newly wrong. That is revisability — eternalism's correction applied backward through the record.
A working compliance system needs both, and the failure above is a failure of both simultaneously. There was no provenance — no record of when the list was fetched or a mechanism to notice the fetch had failed — so there was no present ground for the ten weeks of clearances. And there was no revisability — no process that, once the correct list arrived, went back and re-answered every transaction screened in the gap. The bank did not merely miss the update. It had built a system incapable of saying, honestly, "as of when."
The stack, and where the burden actually lands
| system | what it holds | temporal stance | what breaks |
|---|---|---|---|
| static list snapshot loaded quarterly | a slab of designations, all treated as equally current | eternalist about the slab, presentist about nothing outside it | any designation change inside the quarter is invisible |
| point-in-time screening against a live feed | whatever the feed says right now | strict presentism | no defensible claim about what was true last month without separate archiving |
| continuous intake with provenance and decay | every fetch, timestamped, sourced, aged, superseded on correction | forced eternalist-presentist hybrid | nothing, if the discipline is actually kept — the burden moves from detection to maintenance |
The third row is not a product anyone ships. It is the shape a screening architecture has to take once you accept that transaction flow, sanctions lists, adverse-media feeds and rule changes are all streams that keep moving after the compliance officer stops looking at them. Provenance answers "what did we know, and when did we know it." Revisability answers "given that we now know better, which past decisions no longer hold." Both are needed because both correspond to the only two ways a past-tense claim can be made true — by a present trace, or by the vanished fact it is answerable to. There is no third grounding relation. A regulator asking "why did this clear" wants exactly one of those two answers, and a system without both cannot give either.
Two objections worth taking seriously
The presentism–eternalism dispute is merely verbal. Both sides agree on every physical fact and differ only in how they use the word "exists." Building a compliance architecture on the back of it imports no real content.
Granted, and the deflation costs nothing here. The argument does not need the metaphysics to be substantive. It needs the semantic fork to be forced, and even a deflationist accepts that a past-tense claim needs a truth-condition and that a finite screening system must implement one, trace-grounded or fact-grounded. The dispute is simply the clearest existing map of that fork. If it turns out to be verbal, the fork a compliance engine has to cross is not.
Relativity settles this in eternalism's favour — there is no privileged present, so provenance is a convenience, not a necessity.
Grant that too. It relocates the constraint rather than removing it. A screening system with no privileged observer-independent present still has no access to a distant fact except through what is locally available to it right now: the feed it can query, the record it retained, the timestamp on the fetch. Epistemic presentism survives ontological eternalism without a scratch. Provenance stops being a metaphysical necessity and becomes an access necessity, which binds harder, not softer — no volume of historical data relaxes what a system can know about a designation it never fetched.
Where the ladder ends
A Large Language Model has no present at all — its corpus is a fixed block, every date inside it equally available, the moment of any query lying entirely outside its ontology, the way a sanctions list frozen at compilation has no opinion about designations made after. A Large World Model is a strict presentist — it knows the transaction in front of it now and nothing survives past the session, the way point-in-time screening knows today's feed and nothing about last month unless someone separately archived it. Only a system built to hold transaction flow, sanctions lists, adverse-media and rule changes as continuously running, revisable, provenance-carrying beliefs has been forced to take a stand on both halves of the fork at once. That is the condition the compliance officer's ledger was quietly failing to meet, and it is also the last position on this axis a model of the world can occupy — not because judgement or investigation stop mattering, but because there is no third way for a past-tense sentence to be made true.