The good regulator theorem: why continuous ingestion follows

The good regulator theorem: why continuous ingestion follows

A regulator is anything that holds some variable of a system inside acceptable bounds despite disturbances it did not choose. A thermostat regulates temperature. A pilot regulates altitude. A central bank regulates, or tries to regulate, the price level. The question cyberneticists asked in the mid-twentieth century was not whether regulation is possible but what a regulator must contain to do its job well. Not how much capacity it needs — that had already been answered — but what shape its insides must take.

W. Ross Ashby's law of requisite variety, from 1956, answered the capacity question: only variety can absorb variety. A regulator facing a disturbance with ten distinguishable states needs at least ten distinguishable responses, or some of those disturbances will pass through uncontrolled. This is a hard limit, provable, and it says nothing yet about internal structure. A regulator could meet its variety requirement by brute lookup table, by feedback loop, by neural net, by clockwork. The law is silent on organisation.

Roger Conant and Ashby closed that gap in 1970, in a short paper in the International Journal of Systems Science titled, without much ceremony, "Every good regulator of a system must be a model of that system." They proved that the simplest regulator capable of minimising the entropy of outcomes — the simplest optimal one — has a specific internal shape. Its mapping from disturbances to actions must be a homomorphic image of the regulated system's own dynamics. Not a replica. A structure-preserving compression, in which distinct disturbances that call for distinct responses stay distinct inside the regulator, and disturbances that do not matter can collapse together freely. Regulation is not mere reaction to error after the fact, in the general case. Reliable holding of a variable within bounds requires, embedded somewhere in the regulator's structure, a correspondence to how the regulated thing actually behaves.

That is the theorem, fully stated, before any of it touches computing. It is a claim about control, proved with the mathematics of homomorphisms between state-transition systems, aimed at a problem that predates artificial intelligence by decades: what must a governor, a pilot, a nervous system contain, structurally, to do what it does.

The turn

The theorem gives a standard. Any intake regime — any scheme by which a system takes in evidence about the world it is meant to track — can be judged by asking what correspondence it maintains, and for how long.

A Large Language Model is trained on a corpus fixed at a cutoff. Training builds, in effect, a homomorphic image of how text behaved up to that date: word co-occurrence, argument structure, factual claims as stated by the sources ingested. Within that frozen slice, the correspondence can be excellent — a genuinely good regulator of text-shaped problems, prediction, style, the mechanics of argument. But the correspondence was fixed at one instant. Everything the corpus did not contain, and everything that has changed since, sits outside the homomorphism entirely. The model regulates a world that, at the moment of use, may no longer exist.

A Large World Model narrows the gap by sensing live: cameras, depth, proprioception, whatever the embodiment offers, feeding a correspondence that updates while the episode runs. This is real progress on the theorem's terms — the homomorphism tracks the actual current scene, not a historical trace of scenes. But it is bounded twice over: by the sensor cone, which fixes how much of the system reaches the regulator at all, and by the episode, which fixes how long that live correspondence persists before the model closes and reverts to whatever was baked in beforehand.

A Large Universe Model is the position, argued rather than built, in which the homomorphism is neither snapshotted nor bounded to an episode, but maintained: streams stay open, each belief carries provenance — where it came from, how confident, how stale — and beliefs are revised, sometimes discarded, as the referent drifts. Conant and Ashby's theorem said the regulator must be a model of the system. It did not specify a tense, and that omission matters more than it looks. A model of the system as it was yesterday is, formally, a model of a different system than the one running today, whenever the system is non-stationary. Most systems worth regulating are.

The claim, stated plainly

If good regulation requires a homomorphism to the regulated system, and the regulated system does not sit still, the homomorphism must be kept current or it silently becomes a model of something that no longer exists. Corpus intake buys correspondence at one instant and lets it age. Episodic sensing buys correspondence for the length of an episode and lets it lapse at the boundary. Only intake that never stops can keep pace with a referent that never stops changing.

This is the sense in which the third position is terminal on the intake axis, not in the sense that no better intelligence could ever be built. After "every stream still running, held as revisable belief, carrying provenance," there is no further category of evidence-admission to invent. You can add sensors. You can extend the history a model keeps. You can tighten calibration, cross-check sources, weight recency more cleverly. None of that adds a fifth tense to the three the theorem already implies — was, is, is becoming. What is left after continuity is not more continuity; it is scale, trust in sources, and time to converge.

Objections that hold weight

A thermostat has no map of a house. It has a bimetallic strip and a switch. Where is the model?

This is the strongest objection and it is largely right. Ashby's own homeostat, Watt's flyball governor, a plain PID loop — none carries anything resembling an explicit representation, and each regulates well. The "model" that the theorem insists on is often visible only to an analyst describing the coupled regulator-and-plant as a single system; inside the governor there is only linkage and inertia. The honest limit: error-controlled feedback of this kind works when the error shows up fast relative to the dynamics and a brief excursion is cheap. Where the error becomes visible only after damage is already done — a grid already destabilised, a patient already septic, contagion already spreading through counterparties — regulation has to anticipate, and anticipation is exactly where the homomorphism must be explicit, and current, because there is no error signal early enough to react to.

A second objection cuts the other way: requisite variety sets a ceiling on what a regulator's internal states need to distinguish, not a floor demanding ever more input. Ingesting everything invites spurious pattern and raises inference cost; the discipline that makes a regulator good is compression, not accumulation. This is correct about internal state and wrong about intake. What counts as a disturbance that actually reaches the essential variables is itself a fact that shifts over time, and discovering that shift requires evidence beyond what the current model was built to compress. The 2008 discovery that housing-market correlations formed a systemic channel was not a failure to compress harder; it was the absence of a stream that would have shown the correlation forming. Continuous intake paired with aggressive compression is the coherent reading — intake tells you when yesterday's compression is no longer valid.

A third, sharper still: observability is a structural fact, and no amount of streaming data converts an unobservable state into an observable one. Robust and adaptive control often handle drift deliberately without chasing ground truth in real time — H-infinity synthesis is designed to be indifferent to the actual current plant, working instead against a whole uncertainty set. Both points stand and genuinely narrow the claim. No intake regime rescues a mode that cannot be sensed at all. But the uncertainty set itself was estimated, at some point, from some evidence, and plants drift out of their assumed sets — which is why gain scheduling, model-reference adaptive control and periodic re-identification all quietly reintroduce intake through the back door. Continuity does not defeat the observability limit. It determines how quickly a regulator notices it has already been left behind by it.

The misreading to disown

The theorem is sometimes read as licence for total replication: to control anything well, build a full simulation of it. Conant and Ashby proved no such thing. Their result concerns the simplest optimal regulator, and the homomorphism it requires can be, and usually is, radically lossy — an enormous state space collapsed onto a handful of distinguishable actions. What the theorem mandates is correspondence, not fidelity. The reason continuous intake matters is that it keeps a compressed correspondence honest as the thing it tracks moves, not that more data yields a bigger, truer replica. A Large Universe Model, on this reading, is not a bid for total simulation. It is a bid to keep the map's few, well-chosen creases lined up with a territory that keeps folding differently.

What this does and does not establish

The theorem establishes that good regulation implies internal correspondence, and that correspondence to a non-stationary referent decays unless intake keeps running. It does not establish that any system currently answers to the name Large Universe Model, nor that continuous intake alone constitutes intelligence, nor that the objections above are answered rather than absorbed. Grid operators, anaesthetists and central bankers already live this theorem daily, with instruments built long before the phrase existed. What the lineage adds is a place on the intake axis where the theorem's requirement is finally taken as the design brief rather than as an incidental feature — and a reason, argued rather than assumed, for calling that place the last rung rather than merely the next one.