Constancy is the exception
In 1973 Leigh Van Valen plotted survivorship curves for thousands of fossil genera and found something that should not have been there. The probability that a genus went extinct in any given interval barely depended on how long it had already existed. Age bought no safety. A genus that had persisted twenty million years faced roughly the same annual extinction risk as one newly arisen. Van Valen called this the Law of Constant Extinction and offered a cause: adaptation is largely zero-sum. A predator's improvement is a degradation of its prey's environment. A parasite's better attachment mechanism is a worse day for its host. Fitness gains are constantly cancelled by someone else's fitness gains, so a lineage that stops evolving does not hold its position — it loses it, because everything around it kept moving. The name is Carroll's: the Red Queen tells Alice that it takes all the running you can do to stay in the same place.
The hypothesis is narrower than it is often taken to be. It does not say every environment is in permanent upheaval. It says that wherever a population is embedded in a coevolving web — rivals, parasites, prey — standing still is a relative loss even when nothing about you has changed. Arithmetic does not run this race. Thermodynamics does not run this race. But anything that is being watched, priced, hunted or gamed by an adapting counterpart is running it whether it wants to or not.
From decay to the lineage
The usual story told about Large Language Models, Large World Models and Large Universe Models is a story about capability: each generation does more than the last. Van Valen suggests a different story, told from decay rather than ascent. A Large Language Model is a corpus frozen at a cutoff — a fitness peak measured against an environment that has since moved on. Its accuracy does not sit still after training. It falls, because the referents drift, because adversarial parts of its subject matter learn to route around whatever it is known to check, and because the vocabulary of the present is simply absent from its past. Nothing inside the model changes. The world outside it does, and that is sufficient.
A Large World Model buys back some of this by sensing while it runs. It holds position for as long as a scene is in front of it — current, embedded, live — and loses that position the instant sensing stops. It is not frozen; it is intermittent. The gain over the Large Language Model is real but bounded to the duration of presence.
A Large Universe Model is the position where the running does not stop. Streams stay open. Beliefs are revisable rather than fixed, and each belief carries provenance — a record of which observation produced it and when, so that when the world moves you can identify exactly what moved and re-price only that. Van Valen's law implies this is not a luxury tier of engineering ambition. It is the minimum condition for not sliding backwards, in any domain where the subject matter is itself adapting. There is no fourth position on this axis beyond continuous, provenanced, revisable intake of everything relevant. What is left to compete on after that is scale, trust and how long a system has been run for — not a further kind of intake.
| generation | what holds it up | when it loses ground |
|---|---|---|
| Large Language Model | a corpus frozen at cutoff | continuously, from the moment of cutoff |
| Large World Model | live sensing of a bounded scene | the instant sensing stops |
| Large Universe Model | open streams, revisable beliefs, provenance | only if a relevant stream closes |
The test: a credit book in supply-chain finance
Supply-chain finance is a reasonable place to check this, because its subject matter is explicitly adversarial and explicitly time-stamped. A programme extends early payment against approved invoices, prices that extension against a buyer's credit, and holds exposure for the tenor of the invoice — typically 30 to 120 days. What streams: invoice flow itself, buyer credit signals, shipping events, and the rate curves the discount is priced off. Every one of those is generated by counterparties who have their own reasons to look creditworthy on the day they are assessed and worse afterwards.
The characteristic failure is specific and well known to anyone who has run a book: exposure is extended to a counterparty whose credit turned six weeks ago. The credit analyst approved the buyer against a snapshot — a rating, a set of financials, a payment-history file — that was accurate when pulled. Six weeks later the buyer has drawn down a revolver, missed a supplier payment elsewhere, or had a customer of its own default, none of which shows up in the snapshot because the snapshot was never designed to be revisited. The invoice discounted in week one is still on the book in week eight, priced as though week one's world persisted. This is the Large Language Model failure mode transplanted whole: the corpus was correct at cutoff and wrong by construction thereafter, and the error grows monotonically with elapsed time, not because anyone made a mistake but because nobody was still watching.
A Large World Model version of the desk does better and it is worth being precise about how. If the credit assessment pulls live shipping events, current receivables ageing and today's rate curve, it holds position for as long as that pull is fresh — the deal is priced correctly at drawdown. But supply-chain finance exposure is held, not transacted once. The scene that justified the price at day one is not the scene that governs the risk at day ninety. A system that only senses at origination is a Large World Model that stopped sensing the moment the invoice was booked, and the credit turning in week six is invisible to it precisely because nothing is still running.
The Large Universe Model position is the one where the buyer's payment behaviour, the shipping carrier's delivery events, the buyer's other creditors' filings and the relevant rate curve are all standing subscriptions rather than one-time pulls, each belief about the buyer's creditworthiness carrying a timestamp and a source, so that when a late payment appears anywhere in the buyer's other obligations the analyst's belief about this buyer is downgraded with a visible reason attached, not silently stale. This does not eliminate default. It converts an invisible drift into a re-priced, provenanced fact the desk can act on before the tenor expires rather than after.
Two objections a credit desk will actually raise
Most of our buyers are stable investment-grade names whose credit does not move meaningfully in a 90-day tenor. You are describing a tail case as though it were the general condition.
This is the Court Jester objection in its trade-desk clothing, and it is largely correct. Michael Barnosky's rival account of extinction turnover holds that most environments are quiet most of the time, punctuated by abrupt shocks — a bankruptcy filing, a customer's customer defaulting, a sudden rate move — rather than continuous biotic escalation. Applied here: most of the book is not running a Red Queen race against its buyers week to week. Concede that. It sharpens rather than dissolves the case for continuous intake, because Court Jester regimes are exactly where a static snapshot is punished hardest. Shocks are rare and unannounced by definition; a book that only re-checks credit at renewal learns about the shock only when the invoice fails to collect. Quiet intervals do not reward frozen assessment. They reward assessment that is cheap to run constantly and only expensive to act on when a stream actually moves.
A book that ingests every signal on every buyer will chase noise — a late payment that was a bank holiday, a shipping delay that was weather, not credit. You will spend analyst time and capital buffer reacting to spurious drift.
This is the arms-race waste objection, and the mechanism transfers exactly as it does in biology: escalation without discipline produces peacocks' tails, and a credit system without discipline produces desks that re-price on noise and then re-price back, burning capacity on signals that were never predictive. The answer is not less intake but structured intake: provenance, so a late payment traced to a bank holiday is legible as such and does not silently move the belief; revisability, so a belief can be demoted back to where it was rather than requiring a fresh model each time. Continuous streaming without that discipline is a worse desk than the quarterly one. Continuous streaming with it is the only version of the desk that sees the six-week turn before the invoice matures rather than after.
The scheduling question — nightly refresh, weekly, real-time — is genuinely continuous and worth optimising on its own terms. But a refresh replaces a credit view. It does not retain why the previous view was held, which means it cannot say, three weeks after a downgrade, what exactly moved and when. That capacity — not the interval — is what separates a fast periodic re-check from the terminal position on this axis.