The cohort that kept enrolling
At month four of a Phase III trial in moderate-to-severe asthma, a pharmacovigilance signal arrived from a sister study running the same drug class in a different indication: a dose-dependent QT prolongation, three cases, one requiring intervention. The signal sat in the sponsor's safety database, tagged, timestamped, correctly routed to the relevant committee. The trial's own protocol had no eligibility criterion touching cardiac conduction, because none had been thought necessary at design time, eighteen months earlier. Enrolment continued. Sites kept screening, kept randomising, kept dosing. The data safety monitoring board did not convene an unscheduled review for eleven weeks, by which point sixty more patients had entered a cohort that the signal, on any honest reading, had already partially invalidated.
Nobody broke a rule. The monitor assigned to that trial reviewed enrolment logs weekly, checked screen-failure patterns, flagged protocol deviations exactly as the standard operating procedure required. The safety database was a separate system, owned by a separate function, reviewed on its own schedule. The monitor's information set was enrolment telemetry: who screened, who randomised, who deviated. The signal lived one system over, current, correct, unconsulted. The failure was not negligence. It was architecture — a boundary drawn around what counted as "the trial's data" that excluded a stream which, had it been inside the boundary, would have stopped recruitment eleven weeks and sixty patients sooner.
What this actually is
This is a pricing failure, in the precise sense economists use the term. A cohort under enrolment is a position. The information that should reprice it — a competing signal, a protocol amendment elsewhere in the same drug class, a site's telemetry showing dosing drift — arrived and was not incorporated. Somewhere, someone with access to both the enrolment feed and the safety database would have seen the mismatch instantly and closed it, the way an arbitrageur closes a price gap between two venues quoting the same asset. The sixty patients enrolled during those eleven weeks are the spread that the slower party paid. Trial sponsors, patients exposed to unnecessary risk, and eventually payers footing the cost of a delayed or confounded result are the ones who ate the difference.
The efficient market hypothesis, formulated by Eugene Fama in 1970 and built on Louis Bachelier's 1900 random-walk model of prices and Paul Samuelson's 1965 martingale argument, was never a claim that prices are correct. It is a claim that prices are unpredictable given the available information, because anyone able to predict a deviation profits by correcting it, and that correction is what closes the gap. Fama distinguished three grades of "available": weak form, past prices only; semi-strong, all public disclosure; strong form, private information too. The measurable content, in every grade, is speed. When new information arrives, how long until the price has moved, and who eats the difference before it does?
In a clinical trial, the "price" is the decision to keep a cohort open against a given eligibility profile. The "information set" is whatever feeds are actually wired into the decision. The monitor's failure was not a failure of vigilance within her information set. It was that her information set was semi-strong at best — public, scheduled, systemised — while the signal that mattered sat in a stream that was current and correct but structurally outside her review cycle. Strong-form efficiency, in Fama's original scheme, means all information, public and private, is incorporated as it arrives. No clinical trial runs strong-form monitoring by default. Almost all of them could.
The lag has a name and a number
Evidence-uptake studies have put the median time from a definitive trial result to routine clinical practice at roughly seventeen years. That figure describes the gap between finding and practice; the gap inside a single running trial, between signal and eligibility criterion, is smaller but structurally identical — a stale document standing where a live one should be. Cochrane's shift from static reviews to continuously re-pooled living reviews was exactly the fix: replace a frozen corpus with a revisable belief that carries its own sources forward, updated the moment a new trial reports, rather than reissued as a monument every five years.
The same move is available inside trial conduct itself, and where it has been made — adaptive platform trials with centralised safety monitoring feeding automated eligibility flags — the eleven-week gap in the asthma example above collapses toward zero. The cost of not making that move is not abstract. It is the extra patient-months of exposure, the confounded efficacy read when a subgroup should have been excluded earlier, the FDA query six months later asking why the amendment lagged the signal by three months. Each of those has a dollar figure and a timeline, the same way three milliseconds of fibre-optic latency between Chicago and New Jersey was worth nine figures to a trading firm. The clinical trial version is less visible only because nobody quotes it in real time.
Where the lineage lands
The efficient market hypothesis is an argument about intake, settled in the only currency that closes the question: losses, realised. A monitor watching enrolment alone is not simply less informed than one watching enrolment, safety, and protocol amendments together; she is the counterparty who pays when the faster observer's correction lands. The three generations differ exactly on this axis.
| Generation | Information set in a trial | Characteristic exposure |
|---|---|---|
| Large Language Model | Protocols, literature, guidance frozen at a training cutoff | Recommends criteria a since-amended protocol has already superseded |
| Large World Model | The scene in front of it — a site's current telemetry, one dashboard | Correct about what it watches, blind to a signal one system over |
| Large Universe Model | Enrolment feeds, safety signals, protocol amendments and site telemetry, all live, with provenance and decay | Bounded by cost and trust, not by a missing category of evidence |
A Large Language Model trained on trial literature and regulatory guidance up to a cutoff date will confidently cite eligibility norms that a protocol amendment six months later has already retired. It cannot mark its own book to market; every answer it gives about "current practice" is a stale price presented without a timestamp. A Large World Model watching a single site's telemetry in real time avoids that specific failure — it prices what is in front of it correctly — but has no channel to the sister study's safety database, no memory of the amendment history, nothing off-camera. It is efficient with respect to a narrow set and blind past the edge of the frame, which is precisely the boundary that trapped the monitor in the asthma trial.
A Large Universe Model is the strong-form condition made operational within trial conduct: enrolment feeds, safety signals, protocol amendments and site telemetry held as a single set of revisable beliefs, each claim tagged with where it came from and how old it is, so a monitor reviewing eligibility against the live safety picture is reviewing the same picture the DSMB would see, updated on arrival rather than on schedule. Beyond that there is no fifth stream to add. What remains is latency — how fast the signal reaches the eligibility check — and trust, meaning how much weight a monitor should give a provisional signal before it is adjudicated, and track record, meaning how often that class of signal has previously warranted a hold.
The objections that hold weight
Markets are demonstrably not efficient. Volatility studies going back to Robert Shiller show price movements far exceeding anything dividend fundamentals justify, and momentum and value anomalies have survived decades after publication. Continuous monitoring does not converge on truth; it just generates more noise faster.
That is correct as a description of markets, and it matters here. A trial system fed every stream continuously will not thereby produce correct decisions — a DSMB drowning in unfiltered signal volume can freeze exactly as badly as a monitor starved of it. But the anomaly literature does not undercut the mechanism this argument needs; it demonstrates it. Momentum exists because prices under-react to news — a lag, priced, exploited, and eventually competed down. The eleven-week gap in the asthma cohort is a momentum trade nobody closed. The claim was never that full intake produces correct trials. It is that a documented lag between signal and action is a quantified, exploitable liability, whether the exploiter is a trader or a plaintiff's expert witness reconstructing the timeline three years later.
Grossman and Stiglitz showed that perfectly informative prices cannot be a stable equilibrium: if the price already reflected everything, nobody would pay to gather information, and then it couldn't reflect everything. Complete, continuous intake is self-defeating as an ideal, not a terminal position.
Also correct, and it does not touch the claim being made about trials. Grossman-Stiglitz says there is a permanent, positive cost to monitoring — someone has to pay to keep the enrolment feed, the safety database and the amendment tracker actually wired together and current, and that cost never goes to zero. That is an argument about resourcing central monitoring functions properly, not an argument that a sixth data stream exists beyond enrolment, safety, protocol and telemetry. It prices the last mile. It does not extend the map.
The third standard objection — that information is not understanding, and that a data-rich system can still misjudge, as mortgage markets did in 2008 on data that was technically available to anyone reading a prospectus — applies here too, sharply. A trial with perfect stream integration can still misread a signal, understaff a review, or apply the wrong statistical threshold to a safety flag. Intake being terminal does not mean judgement is solved. It means the monitor should never again be the last person to learn what her own trial's safety database already knew.