communicoupling · concept 26 of 26

To control it, you
must become it.

Why does anyone who successfully manages a complex thing end up carrying a model of it inside them? Conant and Ashby proved it as a theorem: every good regulator of a system must be a model of that system. To control something well is, unavoidably, to become a simulation of it — which is why deep stewardship reshapes the steward.

Think of the manager who has internalised the rhythms of the shop floor; the doctor who "thinks like" the disease; the diplomat who can run the other side's reasoning in their own head; the parent who anticipates the child a beat before the child moves. We call it experience, or intuition, or having a feel for it. It sounds like a soft advantage — helpful, but not required.

Conant and Ashby's 1970 theorem says something stronger and stranger: successful regulation is not merely helped by a model of the system — it requires one, mathematically, and the best regulator's internal structure is isomorphic to the system it controls. You cannot steer what you have not, in some form, become. Below, a regulator learns to control a plant from scratch. Watch its internal table quietly turn into a copy of the system.

the system · Khow many disturbance states the plant has
6
model capacity · Cdistinct states the regulator may hold
6
sensor noise · νchance it misreads the disturbance
0%
control error
bad outcomes, running
model accuracy
regulator table vs system
control ↔ model
how tightly they track
Peek inside · the regulator becoming the system
each cell is the action taken for that disturbance, coloured by action · the regulator's row converges onto the system's as control improves
disturbanced = 0 … K−1
the systemaction that fixes each d
the regulatoraction its model now picks
isomorphism cells where the regulator mirrors the system
Control quality and model accuracy, over the learning run
two quantities, one curve — they rise together because, by the theorem, they are almost the same thing
0 rounds
model accuracy control quality capacity ceiling · min(C,K)/K
Watch what happens
A regulator with a full-capacity model starts blind and learns the plant from reward alone. Its internal table fills in one disturbance at a time, and control climbs in exact step with it.
control ≈ accuracy(model, system) ≤ min(C,K)/K Conant & Ashby: the optimal regulator R is a mapping R = h(system) — an image of the system inside the controller. Control can be no better than how faithfully that image is built, and the image is capped by the regulator's capacity.

To control is to model

The theorem has teeth, not just a moral.

Conant and Ashby made the setup exact. A system is buffeted by disturbances; a regulator observes each disturbance and picks a response; together, disturbance and response fix an outcome, and the regulator's job is to keep that outcome inside a good set. In the instrument, each of the K disturbances has exactly one action that fixes it — a permutation that is the system's own structure. The regulator does not know this map. It learns, round by round, from nothing but whether the outcome came out good.

The theorem's claim is that the only way to drive the outcome to a good value reliably is for the regulator to implement a mapping that is a function of the system's state — and a function of the system just is a model of it. Optimal regulation makes the regulator's internal disturbance→action table an image of the plant's disturbance→fix table. Not similar in spirit: isomorphic. This is why, in the peek-inside view, the regulator's coloured row does not merely improve — it converges onto the system's row cell for cell. The isomorphism readout is the fraction of cells that already match, and it climbs toward 1 in lockstep with control. The model is not a study aid the regulator consults on the side; the model is the good regulator, wearing the outcome's colours.

What to try

Three moves that show the theorem working.

01

Watch them rise together

Press Restart learning with capacity full (C = K). Model accuracy and control quality climb as a single braid — the control ↔ model correlation sits near 1.0. Better control is not accompanied by a better model; it is the same event seen twice.

02

Cap the model, cap control

Drag C below K. The regulator can no longer tell some disturbances apart, so its model saturates at C/K — and control saturates at exactly the same ceiling. Understanding withheld is control forbidden; you cannot buy back the gap with effort.

03

Reach the mirror

Run to convergence, or press The mirror. The regulator's row is now a recoloured copy of the system's, accuracy near 100%, error near 0. The best achievable controller's insides are, structurally, the plant itself.

The steward reshaped by the stewarded

Expertise is internalised isomorphism.

Strip away the mathematics and the theorem becomes a fact about anyone who has ever managed something hard for long enough. To keep a shop floor steady through every kind of shock, a manager must eventually carry, inside, a working replica of the shop floor — its bottlenecks, its moods, the way one delay propagates into three. The good clinician does not memorise a bigger rulebook than the bad one; they grow an internal model of the disease that runs a step ahead of it. The diplomat who consistently gets what their side needs can simulate the other side well enough to feel their objections before they are voiced. In every case the competence looks, from outside, like a suspiciously rich intuition. The theorem says it is something more literal: a model of the system, built precise enough to regulate it, lodged in a person.

This is the intimacy of stewardship, and also its cost. What you tend at depth, you come to resemble. The parent ends up carrying a high-resolution simulation of a particular child; the maintainer of an old codebase dreams in its quirks; the nurse anticipates the ward. The reshaping is not sentimental — it is structural, the same convergence the instrument draws in coloured cells. And because a good model is expensive to build and expensive to hold, it does not transfer cheaply: the reason deep expertise cannot simply be written down and handed over is that the model lives in the regulator's own structure, not in a document. To pass it on, someone else has to become it too. The steward is changed by the stewarded because control had no other door.

The limits of control without understanding

You cannot durably regulate what you refuse to model.

The theorem cuts the other way too, and this is where it bites hardest. A regulator that refuses to build an accurate internal model — that follows rules it does not comprehend, or insists on treating a varied system as if it were simple — has a hard ceiling on how well it can ever control. In the instrument, that is the capped-capacity run: the rule-follower who cannot distinguish the cases blurs disturbances together and is punished with a fixed, irreducible error rate, no matter how diligently it plays. This is the deep tie to requisite variety: there, control needs enough responses to match the world's disturbances; here, control needs enough internal model to know which response to fire. Variety in the repertoire and fidelity in the model are two faces of the same bound. Neither can be willed into existence.

But the model is itself a distinction, and every distinction has a blind spot. The regulator's internal copy is drawn in its categories, and whatever those categories cannot represent is, for the regulator, invisible — folded silently into some bucket where it will be mishandled. This is where the theorem opens onto second-order observation: to steer well you must model the system, but to keep steering well as the system changes, you must occasionally step outside your model and ask what it is failing to see. The good regulator is a model of the system; the wise one also knows it is only a model — and keeps a door open onto everything the model, by being a model, leaves out.

The mapping

Mechanism ↔ social life.

In the modelIn the world
the system / plantThe thing to be managed — the shop floor, the disease, the counterpart, the child.
the regulatorThe steward or controller trying to keep outcomes inside a good set.
the internal modelWhat the steward carries inside: the working replica built from experience.
model accuracyHow well the steward has actually grasped the system, not how hard they are trying.
the isomorphismThe regulator's structure coming to mirror the system's, cell for cell.
capped capacity CControl limited by understanding — a shallow model can only steer shallowly.

Where it tears

Three honest limits.

A "model" need not be conscious, or understood.

The theorem's model is a mathematical mapping, nothing more. A thermostat is a perfect little model of the relation between temperature and its switch, and it knows nothing whatsoever. So "to control it you must become it" is evocative but must not be over-read: the regulator must be isomorphic to the system, not enlightened about it. Much real expertise is exactly this — a reliable mapping in the hands or the gut, with no articulable theory attached.

It assumes optimal regulation and a tidy world.

Conant and Ashby prove their result for the best regulator under specific conditions — a well-defined system, a fixed good set, regulation pushed to the optimum. Real control is usually good-enough, not optimal, and at the margins a great deal of it is cheerfully model-free: heuristics, buffers, and reflexes that work without any internal replica. The theorem describes the ceiling that perfect control would demand, not the scrappy floor most controllers actually live on.

Becoming the system can cost you the outside view.

"You must become what you steward" can quietly romanticise over-identification. The steward who has fully become the system inherits its blind spots along with its structure — and loses the outside vantage needed to change it rather than merely hold it steady. A model good enough to regulate can be exactly the wrong instrument for reform, because it is built to keep things as they are. This is the pull back toward second-order observation: keep one foot outside the copy.