This chapter is about the piece of our inheritance we are most confident in. Ashby’s law of requisite variety is the one result the tradition can point to when asked whether any of this is proved, and the confidence is largely warranted. But the sentence we quote is not the sentence Ashby wrote, the conditions that made his result exact did not travel with it, and the second result we lean on almost as hard — the good regulator theorem — does not prove what we cite it for. Its own authors said so, in the paper, in print, in 1970.
We want to separate three things that we have allowed to blur together. The first is a theorem with a clean scope and a documented lineage, which holds. The second is a design maxim built out of that theorem by substituting one verb for another, which is useful and which we should stop presenting as though it were the theorem. The third is a claim about models and regulators that we have treated as settled mathematics and that is not settled mathematics, though something close to it has been proved elsewhere and can be cited instead.
None of this is an argument against Ashby, against Beer, or against the Viable System Model. It is an argument about our reading habits. Where the tradition’s mathematics is secure, we should be able to say exactly why it is secure and exactly how far it reaches. Where it is not, we should say that too, and cite what is. The correction proposed at the end of this chapter costs nothing: no experiment, no fieldwork, no revision of the model. It is a change in what we put in our footnotes.
Two sentences and one verb
Put the two sentences beside each other. Ashby wrote, variety can destroy variety. Beer taught, only variety can absorb variety. The substitution is not cosmetic. Destroy describes what a regulator does to the variety reaching it; absorb describes what it must be able to hold. Beer’s verb makes the law a design instruction, and Ashby’s does not. It is the better sentence, which is why it traveled; it also traveled without the conditions that made Ashby’s result exact.
That is the whole shape of the problem in miniature. A compression is not a falsification. Beer’s version is faithful to what Ashby proved in the sense that anyone who accepts the theorem should accept the design advice. What the compression drops is the apparatus: the stipulation the proof rests on, the constant that appears when the stipulation is relaxed, and the explicit statement that the result is a priori and therefore says nothing about any particular organization. When we quote the compression and attribute it to the 1956 book, we are quoting Beer under Ashby’s name, and we are quoting a design instruction as though it were a bound.
What the law actually says
An Introduction to Cybernetics, 1956, chapter eleven. Beer’s phrase for the result — only variety can absorb variety — is not Ashby’s. Ashby wrote: “only variety in R can force down the variety due to D; variety can destroy variety.”1 The absorb version is Beer’s idiom and it is the better sentence, which is part of the trouble: it travels well and it travels without its conditions.2
The conditions are worth stating. In logarithmic form the law says the variety of outcomes cannot be pushed below the variety of disturbances less the variety of the regulator. The proof rests on a stipulation about the outcome table which, if dropped, weakens the bound; Ashby’s own generalization patches this by introducing a constant that in practice absorbs the entire empirical content.3 More important, he derives the whole thing in Shannon’s notation and states the homology explicitly: a regulator’s capacity as a regulator cannot exceed its capacity as a channel of communication, and “his ‘noise’ corresponds to our ‘disturbance’, his ‘correction channel’ to our ‘regulator R’.”4
So requisite variety is not an independent discovery about the management of complexity. It is channel capacity, applied to management. That is not a demotion. It is a lineage, and it is the reason this law is secure in a way the VSM’s looser mathematical claims are not. A result that reduces to information theory inherits information theory’s clarity about what it assumes and what it delivers. We should say this more often than we do, because it is the strongest thing we can say about any formal claim in our possession, and because saying it makes visible how little of the rest of the model has anything comparable behind it.
The lineage also fixes the law’s reach. Ashby is explicit that the law is a priori: “The law states that certain events are impossible… It has nothing to do with the properties of matter.”5 For a designer that is a strength — you can rule out an arrangement before building it. For anyone who wants the law to say something about a particular firm on a particular Tuesday it is fatal, and it explains why almost nobody has tried to measure variety in a real organization. The few attempts that exist have not converged on what would be measured.
We should be careful about the status of that last observation. It is a claim about our own literature rather than a proof of impossibility, and it is the kind of claim that a counterexample would settle immediately: a documented case in which two teams applied a stated counting procedure to the same organization and arrived at comparable numbers. We are not aware of one. Until there is one, the practical consequence stands. The law licenses an argument about whether a proposed arrangement could work. It does not license a number, and the diagnostics that report variety figures as though it did are reporting the output of a counting convention rather than a measurement. Chapter 12 takes up the one lineage in organization theory that did make a comparable argument operational, and the contrast is instructive.
Who counts the states
There is also a hole in the law that nobody has closed. Variety is a count of distinguishable states, and nothing in Ashby fixes who does the distinguishing. Change the partition and you change the count; change the count and the law’s verdict changes with it. The observer therefore enters before the arithmetic begins.
Ashby saw this and said so; what he did not do was close it. Introducing variety, he warns that “a set’s variety is not an intrinsic property of the set: the observer and his powers of discrimination may have to be specified if the variety is to be well defined.”6 Two years later he put it more plainly: “If two observers differ in the distinctions they can make, then they will differ in their estimates of the variety.”7 The observer is therefore not a critic’s addition to the law. It is part of the concept as Ashby stated it, and it was lost in transmission along with the stipulation and the constant.
This is not a defect peculiar to Ashby, and it is not resolved by care. It is a structural feature of any formalism that begins from a count of states. What matters for us is that the same choice reappears at two other places in the model where we have also treated it as a technical detail. Someone draws a system boundary and decides what is inside the system and what is environment; Chapter 5 shows how much of the recursion argument depends on that decision. Someone decides what a transducer will preserve as a message crosses a boundary, which is the subject of Chapter 19. Those are not three problems. They are one political choice appearing in three places, and we have inherited it three times without noticing that we were inheriting the same thing.
Naming it does not close it. What naming it does is prevent us from presenting a variety calculation, a boundary, or a transducer specification as a finding when it is a decision, and it tells us where to look when two competent analysts using the same method produce incompatible diagnoses of the same organization.
The theorem that is not a theorem
The law survives inspection. The result we have built System Four on does not, and the evidence for that is inside the paper we cite.
Roger Conant and Ross Ashby, 1970: every good regulator of a system must be a model of that system.8 In the cybernetics community this is what System Four stands on. We invoke it constantly and read it rarely, and reading it is instructive.
The optimality criterion is minimization of the entropy of the outcome set. Nothing else. On that criterion a regulator which pins its system to a single constant catastrophic outcome scores perfectly: entropy zero, no surprise, optimal. Conant and Ashby’s good regulator is a reliable regulator, not a desirable one. For an organization this is the sharpest available point: the theorem underwrites System Four’s need for a model only after someone has separately established that the set-point being defended is worth defending — a question the mathematics cannot see.
Then there are the authors’ own concessions, all in the text. “Although not all optimal regulators are models of their regulands, the ones which are not are all unnecessarily complex” — the universal quantifier in the title is false on their own account. The proof, they note, “by avoiding all mention of the inputs to the regulator R and its opponent S, leaves open the question of how R, S, and Z are interrelated”: no sensing, no channel, no noise, no dynamics. And the assumption that the disturbance distribution is constant can be relaxed only to systems where it changes “slowly with time” — which excludes precisely the non-stationary environments that are System Four’s entire reason for existing.9
The second of those concessions deserves particular attention from a community whose model is a diagram of channels. A proof that mentions no inputs, no channel and no noise cannot be the warrant for a claim about how an organization senses its environment, and Chapter 4 sets out how thoroughly the channels were left unspecified elsewhere in the model as well.
Others have noticed. John Baez, in 2016, observed that the load-bearing lemma is close to trivial and that the leap from the regulator’s state is a function of the system’s state to the regulator contains a model is unargued.10 John Wentworth, in 2021, wrote that it “may be the most misleading title and summary I have ever seen on a math paper” and showed that on the definition given, the identity function qualifies as a model; his repair adds expected-utility optimality, information arriving over time, and a storage cost, and recovers a real result — the minimal optimal regulator’s internal state must be isomorphic to the Bayesian posterior.11 In 2025 Virgo, Biehl, Baltieri and Capucci published the scholarly version, in a preprint on good regulators for embodied agents: the paper “doesn’t strictly succeed, even in its own terms, in showing what its title claims.”12 Their reconstruction replaces being a model with having one, by way of an interpretation map under which an observer may attribute belief states to a regulator that update consistently with the system’s dynamics — at the price of conceding that the attributed model may be trivial.
From a complaint into a claim
That price is the useful part, and it is where the reconstruction earns its place in our literature rather than only in theirs.
A System Four whose environmental scanning nobody’s decisions depend on is the tradition’s most common practical failure. Chapter 2 treats it at length. We have always described it in the vocabulary of complaint: the scanning function exists, it produces material, and nothing downstream moves. Complaints of that kind are hard to adjudicate, because the people running the scanning function can point to its outputs and the people ignoring them can point to their workload, and there is no shared standard by which either side is wrong.
Under the reconstruction the complaint becomes a formal claim: the model attributable to this regulator is trivial. That is a statement someone can dispute with evidence, because it says something specific — that no assignment of belief states to this regulator does any work in predicting what it does. We are not claiming that the reconstruction has been applied to an organization; to our knowledge it has not. We are claiming that it converts a familiar grievance into the kind of proposition that can be argued about, which is more than the 1970 theorem ever gave us.
Meanwhile the result that does prove what Conant and Ashby are believed to have proved exists, and it is not theirs. Francis and Wonham’s internal model principle, 1976, states that a controller achieving asymptotic regulation robust to parameter perturbation must contain a copy of the dynamics generating the exogenous signals. It has a clean scope, it is taught in graduate control, and it has half a century of engineering behind it. The load under System Four is real. It is carried by Francis and Wonham.1314
The scope condition is worth keeping in view rather than quietly dropping, since dropping scope conditions is the habit this chapter is about. Francis and Wonham prove their result for a class of control problems with stated dynamics, and an organization is not obviously such a system. What their theorem gives us is a rigorous statement of the form we thought we already had, and a standard against which to judge whether an organizational reading of it is being argued for or merely assumed.
A derivation that went almost uncited
There is a further pattern here, and it is not confined to the good regulator theorem. Work that would have supported our claims formally has repeatedly been done outside the tradition and left unread inside it. Chapter 12 follows one instance through Galbraith, who arrived at a variety argument for organization design without citing Ashby at all.
Arvid Aulin’s law of requisite hierarchy, published in Kybernetes in 1979 — the same year as The Heart of Enterprise — argued formally that hierarchy is what compensates when regulatory ability is limited. It is Beer’s recursion given the derivation Beer never supplied.15 On 7 August 2026 OpenAlex recorded 43 citations for Aulin, against 1,040 for Conant and Ashby and 2,838 for Francis and Wonham.16
Those numbers should be read only as an order of magnitude. Citation counts drawn from a single database are a crude instrument: coverage differs by field and by decade, book citations are badly captured, and no count of this kind distinguishes a paper that was read from a paper that was listed. What survives the crudeness is the ratio. A formal treatment of the structural claim at the center of the VSM, published in a cybernetics journal in the same year as the book that argues for that structure, has been cited on the order of tens of times, while the two control-theoretic results discussed in this chapter have been cited on the order of thousands. We do not need a better instrument to conclude that we did not take up an argument that was available to us. What would settle the sharper question — whether Aulin’s derivation actually does establish recursion in the form the VSM needs — is a close reading of the paper against Beer’s claim, which is work this book does not do and which we would like to see done.
What the law becomes when someone has to measure it
The chapter has argued that we quote a law and use a vocabulary. There is now a place to check that, because one researcher had to turn requisite variety into a question an informed outsider could answer.
In Pfiffner’s 2017 instrument the requisite-variety items read: the organization’s “ability to react (i.e. the capability to react in a timely, appropriate and formally correct manner), scope of action (the set of suitable behaviours) and the range of offerings (available solutions to a problem)” are “always sufficient to master expected” internal or external challenges — and then, as a second item, unexpected ones.17
Nothing there is counted. There is no enumeration of the regulator’s states, none of the regulated system’s, and no inequality between them. What is asked is whether the people concerned can cope, and the answer is a judgment of adequacy on a scale. Ashby’s law has become a question about whether an organization seems equal to what it faces.
We should be fair about the alternative. Counting states is not available to a survey of bankers and consultants, and it is barely available to anyone; this chapter’s account of who counts the states is why. The instrument did the only thing that could be done. But that is precisely the finding. When the law had to be made operational by a careful researcher, what came out the other side was the vocabulary rather than the law — and the vocabulary is what this tradition has been quoting all along. Chapter 52 examines the same study’s further use of that item, which is less benign: its near-negation appears on the other side of the same regression.
What this chapter hands forward
Stop citing the good regulator theorem as though it were proved. Cite Francis and Wonham for the rigorous claim and Virgo and colleagues for the modern reconstruction. We have gone on invoking a 1970 result whose authors qualified its title inside the paper itself, and we take that as evidence that we have not read our own foundations closely enough. It is also the cheapest correction available to us.
Own the partition problem instead of inheriting it three times. Who counts the states, who draws the blanket, and who defines the transducer are one question wearing three coats. It is not a technical gap to be closed by better work. It is the point at which an observer, and therefore a politics, enters the model.
Receipt — one corrected citation practice; cost: free. Stop presenting Conant and Ashby’s title as a proved universal. Cite Francis and Wonham for robust regulation through an internal model, and Virgo and colleagues for the modern reconstruction. This is a practice change, not an experiment: it is retired only by a proof that restores the universal claim under conditions relevant to organizational regulation.
The receipt is deliberately the smallest one in the book. Chapter 9 asks why so little in this tradition has ever cost us anything, and part of the answer is that we have not often been in a position where being wrong was expensive. A citation practice is expensive to no one. If we cannot make this correction, the difficulty is not resources.