The check nobody scheduled a system for
A phase III trial has an enrolment window measured in months, a safety database updated in near real time, and a monitor whose job is to hold, somewhere in working memory, every criterion that governs who may still be recruited. That job description is a prospective memory task dressed up as a role. The monitor is not being asked to remember data. The data sits in the electronic data capture system, complete and retrievable. The monitor is being asked to remember to look — at the right moment, against the right signal, before the next site randomises the next patient.
This is worth stating plainly because trial oversight is usually described in the language of process: standard operating procedures, data safety monitoring boards, interim analyses on a locked schedule. All of that is real. None of it removes the underlying cognitive structure. A quarterly DSMB meeting is a time-based intention with an unusually good prosthesis. The gap it does not close is the interval between meetings, during which a signal can arrive that invalidates part of the protocol and nobody with authority to act is looking.
What arrives
Four streams run concurrently and none of them pause for the others. Enrolment feeds report new randomisations, screen failures and the running tally against inclusion and exclusion criteria, usually site by site with lag measured in days. Safety signals arrive as adverse event reports, serious adverse event narratives requiring expedited review, and increasingly as passive signals from wearables or lab panels that a site has not yet formally coded. Protocol amendments arrive as documents: a change to a dosing threshold, a narrowed inclusion criterion, a new contraindication flagged by a sister trial or a regulator's safety communication. Site telemetry arrives as operational noise — consent form versions in use, deviation logs, query resolution times — most of it irrelevant to safety on any given day and some of it the first trace of a systemic problem.
None of these streams is difficult to monitor individually. The failure mode lives in the gaps between them: a safety signal in stream two logically bears on an inclusion criterion in stream one, but nothing forces that connection to be made before the next patient is screened.
What is held
What a monitor is actually asked to carry is not the data, which is stored perfectly by the trial's own systems, but a set of live conditional intentions: if renal function markers in the safety database cross a threshold, then re-examine the eligibility of every patient enrolled under the current creatinine clearance cutoff. If a competitor trial reports a signal on the same drug class, then check whether it changes anything here. If an amendment tightens an exclusion criterion, then verify no one enrolled since the amendment breaches it, and flag anyone already enrolled who now would not qualify.
Each of these is a time-based prospective memory task with an event-based trigger buried inside a stream the monitor is not continuously watching. The content of the intention is never the problem. Ask the monitor a week later what the rule was and they will state it correctly. What fails is noticing that the triggering condition has occurred, amid the unrelated activity of running an actual trial — recruitment calls, query resolution, site visits, the ordinary churn that occupies the hours between the moment a signal appears and the moment someone happens to reopen the eligibility criteria and cross-reference it.
What triggers revision, and what doesn't
The characteristic failure is not a missed adverse event report. Those get read. It is a cohort that keeps enrolling for months against criteria a safety signal has already quietly invalidated, because the signal lived in one stream, the criteria lived in a protocol document, and the connection between them was never anybody's standing task — it was supposed to be somebody's memory.
This is where the distinction between retrospective and prospective failure matters clinically as well as cognitively. The safety database is not wrong. The protocol is not ambiguous. Every fact needed to stop enrolling a specific subgroup exists, correctly recorded, somewhere in the trial's own infrastructure. The failure is that no mechanism held the belief "this inclusion criterion is currently valid" as a thing with a shelf life, watched the evidence that supported it, and forced a re-evaluation the moment that evidence moved. The criterion was treated as settled at protocol approval and revisited only at the next scheduled interim analysis — a time-based check, run against a self-initiated cue, exactly the condition under which human prospective memory is worst.
What a Large Universe Model position changes here
A Large Universe Model, in the sense argued for on this axis, is not a smarter monitor. It is a system whose intake never closes across enrolment feeds, safety signals, amendments and telemetry simultaneously, and which holds each governing belief — this criterion is valid, this threshold is current, this cohort is compliant — with provenance: which data supported it, when, and what would count as a revision. The mechanism that matters is not alerting. It is that the check on eligibility criterion is no longer a standing task assigned to a person's memory. It is a consequence that fires automatically the moment the safety database updates in a way that touches the criterion's own evidentiary basis. The time-based task — remember to reconsider this in light of anything relevant — becomes an event-based one, because the relevant stream is already being watched and the criterion is already linked to it.
What the monitor sees
The honest version of this is not a dashboard that clears itself. It is a smaller number of flags, each carrying the belief it revises, the evidence that moved, and the history of prior revisions to that same criterion. A monitor looking at a flag on inclusion criterion 4.2 sees that it was set with 95% confidence on the original renal function data, that a new safety signal from an independent source shifted the supporting evidence three days ago, and that fourteen currently enrolled patients would fail the criterion under the revised threshold. That is a fundamentally different cognitive task from noticing, unprompted, that criterion 4.2 might now be stale.
The cost, honestly stated
Trial monitoring already has automated alerting. Electronic data capture systems flag out-of-range values; safety databases run coded MedDRA queries continuously. This is not new.
That is correct about execution, and it should be conceded fully. A coded query that fires when a lab value crosses a pre-declared threshold is a cron job, and cron jobs have existed in trial safety systems for years. What they cannot do is recognise that a condition nobody coded in advance now bears on eligibility — because the query was written before the evidence that would justify it existed. The distinctive claim for continuous, provenance-carrying intake is narrower: the trigger can be a change in the evidentiary basis of a belief, not only a threshold someone anticipated writing into the query specification months earlier.
Alarm fatigue in clinical monitoring is already severe. Sites and sponsors both routinely override or silence automated queries. Adding more continuous cross-stream checking will produce more flags, not better judgement.
This is the objection that should worry a reader most, and it is largely right. A system that streams everything and alerts on everything manufactures noise faster than it removes the underlying forgetting; monitors will do to a wider flood exactly what ICU staff do to non-actionable alarms. The only honest answer is that the position being argued for is not "more alerts." It requires that every flag carry its own provenance and confidence, so a monitor can allocate attention by evidentiary weight rather than by volume. A system that only fires alarms without that structure has not reached the position described here; it has reproduced the failure in a faster medium.
What continuous intake does not supply, and should not be credited with, is the decision to stop enrolling, the authority to amend the protocol, or the liability for having missed the signal for three months before the mechanism existed to catch it. Those remain exogenous, contested, and unimproved by wider intake. The claim on this page is narrower than that: after every relevant stream is held continuously, with provenance, there is no further category of evidence left to admit. There is only the harder, separate question of who acts on what it shows.