The Variety Logic: Galbraith, Boisot, Aulin, Friston

In 1968 Jay Galbraith spends a little over two months in residence at the Commercial Airplane Division of the Boeing Company in Seattle, with two graduate students. They interview, read internal documents and sit in meetings. Boeing management reads the text before it goes out as a Sloan School working paper.

He takes away two things. The structure of an organization is set by how much information it has to process under uncertainty. And a designer has two families of move and no others: reduce the need for information, or increase the capacity to handle it.

Those are variety↗^[1] attenuation↗ and variety amplification. The paper does not contain the word variety. It does not contain Ashby, Wiener, Shannon or cybernetic either, and neither does the article that followed six years later. We searched both in full.

He got there from a factory in Seattle, and he got there without us.^[2]

The oldest borrowing

The variety logic is the oldest thing in this book and the least Beer’s own, and he never pretended otherwise. Ashby is in every one of his major works, and requisite variety↗ is the load the whole structure rests on. Chapter 10 found him among the Explicit References of Brain of the Firm and moved on.

Chapter 3 took the narrower question: what Ashby proved, what Beer’s absorb compresses, and how little use we made of the one adjacent law that would have licensed the recursion we teach. None of that is restated here. The fact at issue now is harder to hold in one piece. The most striking evidence that the variety logic is real is that a management theorist rederived it in the early seventies from field data, having apparently never read a word of cybernetics. That same evidence shows that Beer’s vocabulary was sufficient to reach the logic and never necessary to it.

Our teaching tends to split that sentence in two. We quote the rediscovery when we want to show that the principle is no private notation, and we set it aside when we want to claim the principle as ours. This part of the book asks us to hold both halves at once. A literature that arrived at the same design space without our vocabulary is a corroboration and a limit in the same document, and the limit is the part we have never priced.

Each case below goes on the same ledger: what was found, when, who cited whom, what the evidence base actually was, and what the finding costs the cybernetic claim. The last column is the one we have been slowest to fill in.

Galbraith: the same design space, in another vocabulary

Galbraith published Designing Complex Organizations in 1973, and its compressed statement, “Organization Design: An Information Processing View”, in Interfaces the following year.^[3] The premise is a proportionality: “the greater the uncertainty of the task, the greater the amount of information that has to be processed between decision makers during the execution of the task.”^[4] The design consequence follows directly. An organization has two ways to reduce the information it must process (slack resources, self-contained tasks) and two ways to raise the capacity with which it handles what remains (vertical information systems, lateral relations).

That is attenuation and amplification with the labels changed, and the correspondence is close enough to state carefully. Galbraith’s variable is uncertainty. Beer did treat variety, in bits, as uncertainty; what neither he nor we have supplied is the argument that Galbraith’s task uncertainty is that quantity.^[5] The shape of the design space he describes is unambiguous: two directions of move, claimed as exhaustive, distinguished by which side of a mismatch they act on. A designer who has never encountered the word variety is working the same two levers, and knows there are only two. His two capacity moves, vertical information systems and lateral relations, are closer to Beer’s Second Principle, that channels must carry more variety than their source generates, than to amplification.^[6]

On one point his formulation is sharper than the one we usually quote. “The task information requirements and the capacity of the organization to process information are always matched. If the organization does not consciously match them, reduced performance through budget overruns, schedule overruns will occur in order to bring about equality.”^[7] Ashby’s law says that regulation fails when variety is short. Galbraith says that the books always balance, and names the currency in which the balancing is paid: schedule, budget, performance. Beer had said as much about the balance. His First Principle of Organization holds that managerial, operational and environmental varieties “tend to equate; they should be designed to do so with minimal damage to people and to cost.”^[8] Galbraith’s advance is the currency. Beer names damage to people and to cost in general terms; Galbraith names overruns an observer can check, and with them the claim takes an empirical shape. It tells an observer where to look when a design is under-specified, and what to expect there. Whether it is true can be put to data. Our own statements of the same idea have rarely been cast in a form that would let anyone ask.

This settles less than it might seem to. The paper asserts the matching claim without demonstrating it, and the currencies it names are those of an aerospace program in the late sixties, which is where Galbraith had been standing. The formulation does not prove the variety logic. Its value lies in stating the logic in terms a manager can act on and a researcher can measure, which is the step we did not take.

The lineage we can verify, and the part we cannot

The claim that Galbraith arrived without cybernetics is negative, and a negative claim is only as good as the scope attached to it. The Interfaces paper carries nine references: Cyert and March, Lawrence and Lorsch, Lorsch and Lawrence, March and Simon, Pounds, Simon, and three items of his own. The words Ashby, Wiener, Shannon, cybernetics and variety do not occur in it at all; his word throughout is uncertainty. The 1968 working paper behind the book is the same. We searched both in full.^[9] He states his lineage explicitly, and it points somewhere else entirely: “the cognitive limits theory of Herbert Simon was the guiding influence.”^[10] That is the Carnegie behavioral tradition rather than the Macy one, a lineage with its own account of why organizations take the shapes they do.

The scope has a limit, and we would rather state it than have it found. We have not been able to read the reference list of Designing Complex Organizations. A search of the full-text index of the Internet Archive’s scan of the book finds neither Ashby nor Beer, Wiener, Shannon or cybernetic, but a search of an index is weaker than a reading. The negative finding therefore holds for the 1974 article and the 1968 working paper, and holds provisionally for the 1973 book. Anyone repeating the claim should repeat it at that scope, and the sentence in this chapter is written so that it can be.

A second limit is more general, and it survives even if the book turns out to be clean. Absence of citation does not establish absence of influence. By the late sixties the vocabulary of information, channel and capacity was in wide circulation, and a researcher could take up the frame without ever meeting its sources. The negative search establishes that Galbraith did not derive his argument from ours and did not need ours to make it stand. The stronger question would be settled by documents rather than by texts: correspondence, teaching materials, the reference list we could not obtain. Until someone does that work, the defensible formulation is independence of derivation, not innocence of the surrounding air.

The evidence base needs the same care. It was a longitudinal case study inside a single division: Boeing’s commercial airplane division in Seattle, two months in residence with two graduate students, interviews, internal documents, observation of meetings, and the text reviewed by management before release.^[11] One site, one program, one period. The paper’s own contrast is with the cross-sectional data of Lawrence and Lorsch, and the single case bought depth at the cost of range. Stating this keeps us from overclaiming on behalf of a literature that is not ours to overclaim on. It also makes the comparison with our own record exact. A two-month case study inside one division is a modest evidentiary base, and it is still more field data than most of our own claims have ever been asked to rest on.

What the rediscovery costs, and what it does not

Two conclusions follow, and only one of them is comfortable.

The comfortable one is that the variety logic is not arbitrary. It was assembled from parts that other literatures found independently, so nobody invented it as a notation and then persuaded people to adopt it. When the model is charged with arbitrariness, with components chosen because Beer chose them, the independent arrival at the same two-move design space is the strongest answer available to us. It is a good answer, and we should keep giving it.

The uncomfortable one is the bill attached to the first. The fact that clears us of arbitrariness also denies us proprietorship. If the logic can be reached from a factory in Seattle by way of Herbert Simon, then the problem imposes it on anyone who looks hard enough at how organizations cope under uncertainty. Our claim on it is one of priority and of completeness: that we arrived first, and that we hold the logic in a more general form. Neither amounts to ownership. Priority is a historical question, and it settles largely in Ashby’s favor rather than ours. Completeness is a question about what the model adds once the variety logic has been granted, and it belongs to Part Four, where it has to be answered narrowly or not at all.

Two neighboring derivations

Two further arrivals belong on the same ledger, for different reasons.

In 1999 Boisot and Child distinguished complexity reduction from complexity absorption as alternative adaptive strategies: a system can cut the complexity it confronts, or it can take that complexity in and hold it in a repertoire of responses.^[12] Theirs is a neighboring literature with its own lineage, the Information-Space framework, and nothing in the paper suggests a dependency on Beer. Fairness to them as much as to us requires an accurate account of the evidence base. It is a conceptual paper illustrated at length with China rather than an analysis of a Chinese institutional dataset, and it is stronger on drawing distinctions than on measurement. The cost to us is modest but real. The attenuation–amplification distinction travels perfectly well without the VSM attached to it, in a vocabulary organization theorists already had to hand.

Aulin-Ahmavaara’s law of requisite hierarchy appeared in Kybernetes in 1979, the year of The Heart of Enterprise.^[12:1] It holds that hierarchy compensates when regulatory ability is limited, which reads in outline like a derivation of what Beer asserted. Chapter 3 records why that reading may be overstated: the hierarchy Aulin-Ahmavaara derives is one of control, and the paper mentions neither Beer nor recursion. It sits inside cybernetics, so it is no rediscovery in the Galbraith sense, and this chapter establishes no derivation in either direction between it and Beer’s recursion. Chapter 3 has already recorded how little the law was taken up. The figure appears here only to fix an order of magnitude: 44 citations recorded by OpenAlex on 29 September 2026, from a single database, an instrument crude enough that we use it for nothing finer.^[13] For this chapter the lesson is narrow. A neighboring formal result existed for a claim related to the recursion we teach and not identical with it, since Beer’s recursion rests on viable systems inside viable systems rather than on limited regulatory ability.^[14] It was published in a cybernetics journal in the same year as The Heart of Enterprise, and first stated four years earlier, and we did not build on it.

Friston: the mathematics Beer never had

The free energy principle is the first attempt since Ashby to derive regulatory structure from first principles about boundary maintenance with mathematics that runs. Beer’s derivation in The Heart of Enterprise had the same aim and ended in balance equations and a theorem. Take any system that persists and whose states divide by a Markov blanket into internal, sensory, active and external. It will appear on average to minimize variational free energy, an upper bound on the surprisal of its sensory states, so that its internal dynamics can be read as approximate Bayesian inference about its environment.^[15] Active inference extends this to action: the system either changes its beliefs to match the world or changes the world to match its beliefs.^[16]

It is tempting to call this variety attenuation and amplification restated with better mathematics. That is not quite right. The belief/action duality differs from the attenuate/amplify one. The two can be brought into correspondence, but doing so is an interpretive move, and it should be labeled as one rather than presented as a derivation. What they genuinely share is the ambition: a claim about any boundary-maintaining system, at any scale, in any substrate. That is Beer’s ambition, arriving a generation later with an apparatus he did not have.

Two things follow from that apparatus, and they point in opposite directions.

The first is that it runs. There is working code, there are published simulations, and there is a pipeline that leads from a generative model through simulation to model inversion against behavioral data.^[17] We have not run any of it ourselves and make no claim about the quality of any particular fit; we claim only that the pipeline exists. A VSM↗ diagram cannot be executed. Nobody has ever shown that a VSM claim is mathematically wrong, and survival under scrutiny of that kind has nothing to do with it. Beer wrote his central claims as equations and a theorem, and of the theorem he wrote: “The proof offered later will be a topological proof.” He also wrote that the equalities “require some extra variety, which is called redundancy”, and he did not measure the terms.^[18] The claims look precise and cannot be checked, which is a different condition from being wrong. That asymmetry matters more than any particular result in the free energy literature.

Explicit enough to be wrong

The second is that being explicit has a price, which is being corrected in public. The free energy literature has paid it, and ours has not.

In 2021 Biehl, Pollock and Kanai showed that the definitions of the Markov blanket are not equivalent across the corpus, and proved by counterexample that the original free energy lemma, taken at face value, is incorrect. They were explicit that their target was the formal argument rather than the underlying ideas.^[19] The following year Aguilera, Millidge, Tschantz and Buckley showed that even for weakly coupled linear systems the conditions the argument requires hold only within a very narrow band of parameters.^[20] Bruineberg and colleagues, in Behavioral and Brain Sciences, separated “Pearl blankets” from “Friston blankets”. The first is the Markov blanket as Judea Pearl defined it in 1988, an epistemic tool for inference in a network; the second is a metaphysical construct used to demarcate where an agent ends and the world begins. They argued that the literature persistently trades on the conflation.^[21] And Friston himself, asked directly, has said of the principle, as quoted in Chapter 9: “Like Hamilton’s Principle of Stationary Action, it cannot be falsified. It cannot be disproven.”^[22]

We enter none of this as a defeat for anyone. It describes a working relationship between a general claim and the criticism such a claim can attract, and the central concession is available in the author’s own words. The framework carries Beer’s ambition with mathematics Beer never had, and its architect declines to call it falsifiable. What survives the criticism is a generative machinery for building testable process theories. The claim to be a law does not survive. We recognize that position because it is our own, except that our version was written as equations and never in terms anyone could check, so it did not attract the criticism in the first place. Being told that your lemma is wrong is a cost. It is also a form of participation, and by staying imprecise we have avoided paying for it.

The blanket is not ours

One genealogical point needs recording, because claiming neighboring work as cybernetic inheritance is a habit we should not indulge. Markov blankets are Judea Pearl’s, from 1988, and their route into active inference runs through Bayesian networks and variational inference.^[23] Cybernetics did not bequeath them. The free energy literature does cite Conant and Ashby, but for the model claim rather than for the blanket, and running the two together produces a genealogy that will not hold.

The instrument we did not build

Galbraith’s descendants spent four decades operationalizing information-processing requirements, information-processing capacity and the fit between the two, with validated constructs, in a line of work running from Tushman and Nadler in the late seventies to supply-chain studies four decades later.^[24] Those are three points on a line rather than a survey of it, and the line is not uniformly strong. The direction of travel is unambiguous all the same, and we did not travel it. Nothing in principle prevents the variety claims from being given the same treatment. Disturbance variety, attenuators, amplifiers and response capacity are candidate constructs, and construct validation is a well-understood craft with a standard apparatus. After fifty years, the subject matter cannot be blamed for the absence of that work. We chose it by default, and the remedy is a study rather than an argument.

Beer anticipated part of the difficulty. Of his own variety calculations he asked: “Could we ever say: this is the variety of the chemical industry, say? Of course we could not.” The use of the measure, he went on, “is not the naive use of counting states, but of matching state generators.”^[25] A negative result from the study priced below would refute a stronger claim than the one Beer made, that variety is one measurable organizational quantity. The balance claim would then have to be stated in terms of gaps an observer can see, and nobody has yet done that.

The priced receipt

Receipt — construct validation of the variety quantities. Adapt established organizational-information-processing scales^[26] into a construct-validation study of disturbance variety, attenuators, amplifiers and response capacity. Design: a pilot in three organizations, comprising instrument design, thirty cognitive interviews, and two survey waves. Price: roughly forty researcher-days plus participant time. Pre-register convergence and discriminant-validity thresholds before the first wave. Negative result: if the proposed variety measures do not converge, stop treating “variety” as one measurable organizational quantity and specify the narrower constructs that do.