Drift and Lyapunov stability: why continuous ingestion follows

The restoring force

A system is Lyapunov stable if, once nudged off its reference path, it stays near that path. Not that it returns — merely that it does not wander further than some fixed bound however long you wait. Asymptotic stability demands more: the displacement must shrink back to zero. Both properties depend on the same structural feature, a restoring force — some mechanism that senses the displacement and pushes against it. Without that mechanism you get the third case, the one with no name of its own because it needs none: drift. Each error persists. The next error does not cancel it, because nothing is checking; it simply adds. Displacement accumulates the way a random walk accumulates, growing without bound not because any single step was large but because no step was ever subtracted.

This is a statement about structure, not magnitude. A system with tiny errors at every step but no feedback will eventually diverge as far as a system with large errors and a working corrective loop will not. The question that decides stability is never "how big is the error" but "does anything act against it". A thermostat with a broken sensor and a thermostat with a noisy sensor are different problems; a thermostat with no sensor at all is a different kind of problem, because the failure mode is not inaccuracy but unboundedness.

The distinction matters because it is checkable independent of outcome. You do not have to wait for the trajectory to run away to know it will. If you can show there is no restoring force — no channel by which the present state reports its displacement back into the control input — you have shown the system is open-loop, and an open-loop system aimed at anything that moves has no term in its dynamics that can shrink the error. That is a structural diagnosis, available in advance, which is most of why the concept has survived a century and a half of use outside the field that produced it.

Origin

Aleksandr Lyapunov set out the general theory in his 1892 doctoral thesis, The General Problem of the Stability of Motion, written against a concrete astronomical worry: do the planets' orbits stay close to their reference paths under small perturbation, or can a small nudge grow into a qualitatively different orbit. Rather than solve the equations of motion explicitly — usually impossible — he constructed an auxiliary function, energy-like, that must decrease monotonically along any trajectory near the reference. If such a function exists, stability follows without ever finding the trajectory itself. The thesis was formidable and largely unread outside Russia for sixty years. It resurfaced in the 1950s when Kalman and others recast it in the language of state-space control, and it became the standard proof technique for showing that a feedback system will not run away under disturbance. Drift, as a term, arrived from a more practical direction — navigation and metrology, where uncorrected integration of small sensor biases was the daily enemy of anyone trying to know where a ship or a missile actually was.

The turn

Set aside orbits and gyroscopes for a moment and ask a different question: what is a model permitted to observe, and when. That is an intake question, and it turns out to submit to exactly the same structural test.

A Large Language Model is trained on a corpus assembled once and frozen at a cutoff date. From the instant of freezing, nothing the world subsequently does can reach the model's parameters. There is no channel through which reality reports back that a belief embedded in the weights is now wrong. That is the definition of open-loop. The model's distance from the present world is therefore not a fixed, bounded error — it is a monotonically growing one, because every fact that changes after the cutoff stays wrong, and wrongnesses compound where they interact with each other. This is drift in the literal Lyapunov sense, not a metaphor borrowed for effect: no restoring force exists in the architecture, so none is available in the mathematics.

A Large World Model closes the loop, but only regionally and only for a while. While a scene is live, sensed input from that scene provides an error signal, and the model's estimate of the scene can be corrected in real time — locally asymptotically stable, in the strict sense, against what its sensors currently report. End the episode, and the correction stops with it. The next scene starts the drift clock over from an arbitrary point.

A Large Universe Model is the configuration in which the loop is never opened at all: every available stream keeps running, every belief stays revisable, and each belief carries provenance — a record of which evidence it rests on — so that when a correction arrives it can be attributed to a cause rather than simply overwriting the prior. That last clause is not decoration. A closed loop without provenance can still drift, just differently, and the objections below explain how. But a closed loop with provenance is the minimum structural condition under which asymptotic stability against a moving target becomes available at all. Not accuracy. Stability. It is a claim about whether error decays, not about how small it currently is, and that is a materially weaker and more defensible claim than it first sounds.

The misreading

The weak version of this argument says any continuously updating system is thereby stable, and any frozen model is thereby useless. Disown both halves explicitly. A frozen model of genuinely stationary structure is the correct tool, at lower cost and smaller attack surface, and nothing about Lyapunov's theorems says otherwise — they are conditional statements about what must hold for stability, never a guarantee that any given feedback arrangement delivers it. Continuous ingestion with no discipline about sources can drift faster and more confidently than a frozen system ever could. The claim under examination here is narrower and more structural: an open loop forecloses stability outright, as a matter of dynamics, while a closed loop merely makes stability available — and only to a system capable of telling where each of its beliefs came from.

Three objections

Feedback is not automatically stabilising. High gain and latency produce oscillation. A system that ingests its own outputs, or the outputs of systems like it, can drift in a correlated direction faster than a frozen one ever would.

This is correct, and the control literature has never claimed otherwise — Nyquist's criterion gives the exact conditions under which closing a loop destabilises rather than corrects. Model collapse under repeated training on synthetic output is the documented instance: continuous intake without source discrimination is not a cure, it is a different failure mode, arguably a worse one because it comes wearing the appearance of correction. The response is not to retreat from closed loops but to name what makes the difference: provenance. A belief tagged with its origin can be discounted when that origin is the system's own prior output. Autophagy is what a loop looks like with no record of where its beliefs came from. It is an argument for provenance, not an argument against closing the loop.

Retrieval augmentation, periodic fine-tuning and scheduled retraining already put a restoring force on production systems. The frozen-corpus premise describes an artefact from several years ago, not anything defended today.

Granted that sampled correction is genuine correction — a quarterly retrain is control, not nothing. But sampling has a floor, and the floor has a name: you cannot track a signal whose rate of change exceeds roughly half your sampling frequency. Quarterly retraining cannot track weekly change, by construction, whatever the quality of the retrain. Retrieval closes a narrower loop still — the query path gets fresher facts, but the parameters underneath remain untouched and, more importantly, unversioned. There is no record of which stored belief a given answer actually leaned on, so no way to trace a correction back to its cause when the retrieved fact and the trained prior disagree. This narrows the claim usefully: the argument is not against sampled-data control, which is a real and useful category, but against mistaking a sampling interval for continuity, and against a retrieval layer that supplies freshness without supplying provenance.

Most of what a model needs to know does not move. Arithmetic, syntax, thermodynamics, the Latin verb — stable at the cutoff, stable a century later. Insisting on continuous intake to protect knowledge that never drifts is expensive vigilance against a phantom.

This is the strongest form of the objection and it should be conceded almost entirely. Stationary knowledge requires no restoring force, and a frozen corpus is the right container for it — nobody needs a live feed to hold the multiplication table steady. The difficulty is narrower than "everything moves": a frozen system cannot know, from the inside, which of its own beliefs are the stationary kind and which have quietly moved underneath it. Telling the difference requires comparing the belief against the present, and comparing against the present requires observing it. Continuous intake is not needed to relearn arithmetic. It is needed to keep locating the boundary between what still holds and what no longer does — and that boundary itself is not fixed. This is real ground conceded: the case for continuous intake is a case about identifying drift, not about correcting content that was never going to move.

What the argument establishes

It establishes a structural ordering on the intake axis, not a scorecard of accuracy. The Large Language Model is open-loop by construction and its divergence from the present is additive because nothing subtracts from it. The Large World Model achieves genuine, local stability for the life of a scene and reverts to drift the moment the scene ends. The Large Universe Model is the configuration in which the loop is never opened, which is the necessary condition for stability against a target that keeps moving — necessary, not sufficient, since Nyquist and model collapse both stand as live counterexamples to sufficiency. Terminal, on this axis, means only that there is no fourth category of evidence beyond everything, continuously, with its provenance kept. Later improvement is available in gain, in trust calibration across sources, in noise rejection over time on stream. It is not available as a new sensory category, because none remains to add.