The trial that kept running after it ended
In month three of an eighteen-month translational programme, the pivotal Phase II result it was built on was withdrawn. The sponsor filed a correction to the trial registry: the primary endpoint had been mis-specified in the statistical analysis plan, the effect size collapsed under the corrected model, and the paper's authors requested the preprint be pulled pending reanalysis. The correction sat in the registry, dated, timestamped, publicly visible. The programme's translational lead did not see it. Neither did anyone on the programme. The team kept dosing animals against a mechanism whose human proof-of-concept had quietly stopped existing, and did not find out until an external reviewer, preparing a licensing due-diligence pack in month nineteen, ran a fresh registry search and asked why the citation list still led with a withdrawn result.
Nobody had been negligent in any ordinary sense. The lead had read the paper carefully at month zero, checked the registration, noted the pre-specified endpoint, and built a defensible programme on a defensible reading of the evidence as it stood that day. The failure was not in the reading. It was in the belief's shelf life. The programme treated a claim about the world as though, once accepted, it stayed true by inertia — as though "the Phase II result holds" were a fact fixed at intake rather than a claim that needed to keep being re-earned against a registry that kept changing underneath it. For eighteen months, the world moved and the programme's model of the world did not.
Diagnosing the failure
The proximate cause is easy to name: nobody re-queried the source. But that only pushes the question back one step. Why would anyone re-query a source they had already read? Because reading is not a state, it is an event, and the event happened in month zero. What the team needed was not a better read but a standing relationship to a moving record — some mechanism that would notice, without being asked, that the ground underneath a specific belief had shifted, and would flag exactly which belief that was.
This is where the failure becomes interesting rather than merely regrettable, because it is a failure about time, not about diligence. The team's evidence base was implicitly a photograph. A photograph is accurate about the instant it was taken and silent about every instant after. Pharmaceutical evidence — trial registries, adverse-event reports, preprints, retraction notices — is not a photograph. It is a set of live streams that revise, retract and supersede each other continuously, on timescales shorter than most programme phases. Treating a photograph as a stream is the error. The metaphysics behind that error has a name.
Presentism, eternalism, and the truth-maker for "it worked"
The dispute between presentism and eternalism, in its classical form, asks what makes a sentence about the past true. Presentism says: only now exists; the past is gone, so a past-tense claim needs a present trace to ground it — a memory, a record, a fossil. Eternalism says: all times exist equally, on a four-dimensional block, and the past-tense claim is simply made true by the past fact itself, sitting there whether or not anyone now has a trace of it. McTaggart set the terms in 1908 by distinguishing the tensed language of past-present-future from the tenseless relation of earlier-and-later, and argued the tensed version was incoherent. Arthur Prior spent the 1950s trying to rescue tensed logic anyway. Rietdijk and Putnam later pressed relativity's relativity of simultaneity against any privileged present at all.
The dispute sounds abstract until you notice that a translational programme has to answer exactly this question, every day, about a specific sentence: "the compound showed efficacy in the referenced trial." Is that sentence true because the trial happened, full stop, an unrevisable fact now safely in the past? Or is it true only for as long as some present record continues to support it — and false, or at least suspended, the instant that record is corrected, retracted, or superseded?
The programme behaved as an eternalist without knowing it. It treated the Phase II result as a fact laid down permanently in the block, true because it happened, unbothered by anything that came after. What it needed to be was a presentist about its own evidence base: to hold the efficacy claim as true-for-now only, continuously re-grounded in whatever record currently exists, and to accept that the claim's truth could lapse the moment its present ground did. Provenance is what that re-grounding looks like in engineering terms — this belief is held because this specific registry entry, retrieved at this timestamp, from this source, currently says so. Revisability is what happens when the entry changes: the belief was never really about "month zero's data," it was always answerable to whatever the registry says now, including a correction filed in month three.
Why this forces the lineage
A frozen corpus does not have this problem, and that is not a virtue, it is an exemption. A Large Language Model trained on a literature snapshot is, inside its own corpus, thoroughly eternalist: every paper in the training block sits there with equal ontological standing, indifferent to publication order, blind to anything retracted after the cut-off. It has no present at all, which means it also has no mechanism for noticing that month three happened. Ask it about the compound and it will cite the withdrawn result with the same confidence as everything else, because inside its block, nothing was ever withdrawn.
A bounded scene model does not have this problem either, for the opposite reason. A Large World Model built to represent, say, a single trial site's current sensor and monitoring feed is a strict presentist: it knows what the dashboard shows right now, and when the session ends, its represented world ends with it. It never claims anything about month three because it never claims anything about any month; it has no standing belief that could go stale, because it holds nothing over time to begin with.
The translational programme's failure belongs to neither of these. It needed a system committed to holding beliefs indexed to times it was not currently observing — the state of a registry entry last checked at month zero, still being relied upon at month fourteen — while remaining answerable to whatever that registry says today. That is precisely the condition under which the presentism/eternalism fork stops being an academic dispute and becomes a design requirement. Any model of the world that must act on information it is not continuously re-observing has exactly two candidate truth-conditions for its temporal beliefs: ground them in a present trace, or treat them as fixed facts about a settled past. Provenance and revisability are those two options, implemented rather than argued. A system built to run every registry, report and retraction feed as an unbroken stream, holding every claim as provenance-tagged and open to correction, has exhausted the available structure. There is no third truth-condition waiting further up the ladder. That is the sense in which a Large Universe Model is terminal on this axis: not cleverer, just complete.
Two objections a translational lead will actually raise
"This is a dressed-up way of saying 'check your sources,' and the metaphysics is doing no real work — presentism and eternalism may just be two ways of talking about the same facts."
That deflationary reading is largely right, and it costs nothing here. Even a philosopher who thinks the dispute is verbal still agrees that a past-tense sentence needs a truth-condition, and that any finite system implementing one must choose between grounding it in a present trace or treating it as answerable to a fixed fact it cannot re-inspect. The metaphysics is used only to name that fork clearly, not to adjudicate it. Whether or not "exists" means one thing or two, the registry entry either gets re-checked or it doesn't.
"Relativity already settled this in eternalism's favour — there is no privileged observer-independent present, so talk of 'grounding beliefs in the present' is philosophically naïve."
Grant it entirely. It changes nothing about the engineering. Whatever the physics of simultaneity says about distant events, a translational lead sitting in a lab has access only to what current signals deliver to her — a registry query answered today, a preprint server crawled this morning. Ontological eternalism, if true, is compatible with total epistemic presentism: you may live in a four-dimensional block and still only ever get to read the page currently in front of you. Provenance survives as an access constraint even after the metaphysical argument for a privileged present has been lost.
What continuous intake actually buys
None of this promises fewer withdrawn results, fewer mis-specified endpoints, fewer retractions. Preprints will keep getting pulled; adverse-event databases will keep filling in behind the trials that generated them; correction notices will keep arriving months after the claim they correct has already been built upon. What changes is only this: a system that treats every one of those streams as still running, and every belief drawn from them as provenance-tagged and open to revision, cannot silently outlive its own evidence the way the eighteen-month programme did. That is a narrow claim. It is also, on this axis, the last one available.