A cell that would not sit still
In 1972, in Santiago, Humberto Maturana and Francisco Varela published De máquinas y seres vivos and asked a question biology had never answered cleanly: what, precisely, makes something alive. The textbook answer was a list — reproduction, metabolism, irritability, growth — and lists admit exceptions. Viruses reproduce. Crystals grow. Maturana and Varela replaced the list with a single organisational criterion. A living system is one whose components are produced by a network of processes that those same components constitute. The cell membrane makes metabolism possible; metabolism makes the membrane. Stop the flux and nothing dormant is left behind, only debris. They called this autopoiesis, self-production, and insisted on a further clause that gets dropped in most later borrowings: such a system is materially open and operationally closed. Matter and energy stream through it constantly. What it produces, using that stream, is itself.
The idea sat downstream of Schrödinger's 1944 observation that organisms survive by feeding on order, and alongside Ilya Prigogine's work on dissipative structures, which showed that systems far from equilibrium can hold a stable form only by continuous throughput. None of these were claims about intelligence. They were claims about what it costs to persist as a distinct thing inside a world that keeps moving.
Clinical trials turn out to be a good place to watch the same cost show up, because a trial is exactly a bounded attempt to hold a belief steady — "this drug is safe and effective in this population" — against a stream of evidence that has no obligation to cooperate.
Where the belief actually lives
A trial has a protocol: inclusion criteria, dosing schedule, stopping rules, a statistical analysis plan fixed in advance to keep the eventual result honest. Once locked, the protocol is treated, administratively, as settled. But the belief the protocol encodes — that the drug behaves this way in this population — is not settled. It is being tested in real time by four streams running simultaneously: enrolment feeds reporting who has joined and against which criteria; safety signals arriving from adverse-event reports, lab panels and the occasional post-marketing echo of a related compound; protocol amendments issued by the sponsor as the picture shifts; and site telemetry — visit compliance, dosing adherence, query resolution times — reporting how faithfully the plan is being executed at forty or four hundred locations at once.
The protocol document is frozen the way a corpus is frozen. The four streams are not. This is the same split that runs through the whole lineage: a structure produced once, and a process that has to keep running for the structure to remain true.
The failure that keeps recurring
The characteristic failure has a specific shape and trial monitors know it by heart. A safety signal arrives — an unexpected elevation in liver enzymes in a related trial, a pharmacovigilance database update, a data safety monitoring board's interim finding — that logically invalidates part of the enrolment criteria. Perhaps a comorbidity that was previously acceptable now looks dangerous in combination with the study drug. The signal is real, dated, and sitting in a report. But enrolment does not stop. Sites keep screening against the old criteria because the criteria live in a document that has not yet been amended, and amendment is a process with its own latency: sponsor review, IRB or ethics committee resubmission, site retraining, updated consent forms. Ninety days is not an unusual interval between a signal being known and a protocol reflecting it. In that window, a cohort is enrolled against criteria that are, in substance, already false. The trial monitor — the person formally tasked with reconciling what the protocol says against what the sites are doing — is watching two things that have quietly stopped agreeing, with no mechanism to make either one yield before the paperwork catches up.
This is not a story about carelessness. Everyone in it is doing their assigned job correctly. The protocol document is doing what protocol documents do: staying fixed until amended. The failure is structural, and it is the same failure a frozen corpus produces on a longer timescale, compressed into the months a trial actually runs.
Naming the position on the ladder
A Large Language Model, applied to trial operations — summarising protocols, drafting amendment language, answering site queries — is allopoietic in the strict sense the term requires. Its knowledge of dosing conventions, adverse-event terminology and regulatory phrasing was assembled once, from a corpus with a cutoff, and nothing internal to it maintains that knowledge afterwards. It can draft a beautiful amendment. It cannot know that one is overdue.
A Large World Model, built to sit over one active protocol and reason about the current cohort, does better. Fed enrolment and telemetry live, it can hold state that genuinely updates: this many patients enrolled today, this site behind on visit compliance, this dosing pattern drifting from plan. It closes a real loop — sensing, comparing, correcting — for the duration of the trial it is watching. But the loop is scoped to the scene. When the trial ends, or when a new signal originates outside the protocol it was told to watch — a related compound's post-marketing report, say — the model has no standing channel for it. Its closure is real but bounded to what it was pointed at, and it lapses.
The position the failure is actually asking for is one where safety signals, enrolment feeds, protocol amendments and site telemetry are not four separate inputs occasionally reconciled by a human monitor, but a single standing intake that a belief-maintaining process runs on continuously — with each belief ("this criterion is currently valid") carrying provenance back to the signal or amendment that last touched it, and a decay function that flags a belief as stale the moment its supporting evidence is superseded, not the moment paperwork says so. That is the Large Universe Model position: not a bigger model of one trial, but the arrangement in which intake and belief-maintenance are the same ongoing process, closed on itself, open to the world.
| intake pattern | trial-specific failure mode | |
|---|---|---|
| Large Language Model | frozen corpus, one cutoff | drafts against dosing conventions already superseded |
| Large World Model | live feed, scoped to one trial's duration | blind to a signal originating outside its watched scene |
| Large Universe Model | standing multi-stream intake with provenance and decay | none structurally — the open question is cost, not category |
Two objections worth taking seriously
Autopoiesis describes lipids making the bilayer that contains the lipids. A trial-monitoring system's substrate is a server rack maintained by engineers on a completely separate supply chain. Calling this autopoietic borrows biology's prestige while dropping the condition that made the term precise.
This is correct as a description of the term's history and it is exactly how the word has been misused since the 1980s. But the closure being claimed here is not material, it is representational. What must be self-produced is the belief structure — the current model of which enrolment criteria are valid — generated by the same intake process it governs, including the provenance that says which criterion was last touched by which signal. The hardware stays allopoietic. So does a cell's glucose supply. Material openness was always half the definition; it was never the objection.
A trial is deliberately narrow. Sites are structurally coupled to a specific protocol and screened against specific, chosen criteria. Total, continuous intake of every stream is the opposite of good trial design — it is the fantasy of a monitor with no defined scope at all.
The sharper of the two, and it forces a real distinction. The axis in question is what a system is permitted to take in, not what it is obliged to attend to at any moment. A protocol's selectivity — these criteria, this population, this endpoint — is a live decision that can be revisited when evidence demands it, the way a cell's receptor expression shifts within minutes of a changed environment. A cutoff, by contrast, is a selection made once, from outside, and then held past the point where anyone inside the system can revisit it. Structural coupling to a scope is fine. It is the imposed finality — no channel back to the signal that arrived after the document was locked — that the failure actually turns on.
What this does not settle
None of this makes a continuously updating trial-monitoring process alive, autonomous, or possessed of interests. Maturana and Varela's stronger claim — that cognition and life are coextensive — is contested inside biology itself, and nothing about provenance-carrying intake in a regulatory setting needs to inherit that claim to be useful. What the recurrence across three generations does establish is narrower and harder to dismiss: a belief about a moving trial decays for the same reason any belief about a moving world decays, and reconciling it costs continuous work that a frozen document, however well drafted, cannot perform on its own behalf. The trial monitor's ninety-day gap is Landauer's bound wearing a lab coat — the cost of keeping something true, paid late, by a person, because no process in the system was built to pay it as it came due.