The Category Error as Shelter

The claim

Ask for a test of the Viable System Model, and the strongest reply this tradition has is not that the model has passed one. It is that the request mistakes what kind of thing the model is.

That defense is good. It has a serious literature behind it, people with no stake in the VSM hold it, and this book has not answered it. It deserves to be stated at full strength, priced, and set beside a neighboring tradition that made the same argument in public and paid what it costs.

Pickering’s version

Andrew Pickering’s history of cybernetics makes the argument for the whole field. On his account, cybernetics is about a shift from epistemology to ontology, from representation to performativity, and Beer’s work in particular displays a cybernetic ontology of unknowability and becoming.^[1]

Read that carefully, because it is no soft position. It does not say that the model is hard to test, or that testing it is premature. It says that the model was never a representation of an organization in the first place: it is a way of engaging with a system whose behavior cannot be known in advance, a set of moves rather than a set of assertions. On that reading, asking whether there really are five subsystems is like asking whether a hammer is true.

Pickering has the historical evidence for it. Beer’s practice was performative in exactly this sense: the biological computers, the Chilean operations room, the insistence that the model is built in the act of using it, the refusal to treat organizations as objects sitting still to be described. A reader who knows the archive will recognize Beer’s practice in Pickering’s account. Beer’s books say something more definite about what the model is, in his own words. Even in his most practice-minded pages he asks that models be open to checking: the model of the whole enterprise that senior managers carry in their heads fails, he writes, because “it is not made explicit, and cannot be interrogated or tested.”^[2]

If Pickering is right, better measurement does not help. It refines an instrument aimed at nothing, the structure Chapter 16 identified in the objections that land.

Beer’s own declaration

Beer did not leave the question to his readers. In Decision and Control (1966) he asked what kind of thing a model is and answered with two refusals. Can it be equated with a scientific hypothesis? “Clearly not; the model does not postulate the causal mechanisms that underlie events, it simply represents the pattern of the events themselves in advance”. Is it a scientific theory, then? “Again, it is not; it has no explanatory content. A model is simply a reflection of whatever is the case, which is explicitly made available for experimentation.”^[3]

Twelve pages later he said what that status buys. A scientific model “is a homomorphism on to which two different situations are mapped, and which actually defines the extent to which they are structurally identical.” If the mapping leaves out something that matters, “the model will lose in utility, but it cannot lose in validity. This is its peculiar strength.” Utility, in turn, is measured “by straightforward testing of the predictive value of the scientific model. If this predictive value is low, we had better start again.”^[4] Nineteen years later the same status reached practitioners as a maxim: a model “is neither true nor false: it is more or less useful.”^[5]

This is a declaration in two layers, made by the founder, in print. It is neither Pickering’s reading nor a hypothesis. Validity is secured by construction; utility is exposed to prediction. And the representational claim stays in. A model that “defines the extent to which [two situations] are structurally identical” says something about the organization, and the declaration protects that something from being wrong. The performative reading would give the claim up. Beer’s declaration keeps it and sets a single test beside it. Beer held the other reading as well: closing the chapter of the same book on controlling enterprises, whose model is the precursor of the Viable System Model, he called that model “no more and no less than a theory” and “a descriptive account of how enterprises actually work”, so the founder stood on both readings, one that can be wrong and one that “cannot lose in validity”.^[6]

What it would cost

The book’s answer is to price the performative reading, and the price is high enough that nobody in this tradition has been willing to pay it in public.

If the VSM is performative rather than representational, its representational claims must leave the teaching. All of them.

The neuroanatomical mapping goes. If the model is a way of acting, System Five corresponds to nothing in a cortex, and Chapter 1’s charge about what Beer actually wrote becomes moot, because the mapping was never making a claim. The necessary-and-sufficient claim goes: a set of moves is not the kind of thing that could be necessary and sufficient for the survival of an organization. Beer’s own presentation of The Heart of Enterprise as written in the belief that the necessary and sufficient conditions of viability had been established goes with it, though he had warned at the outset that the rigour would “fall short of the target.”^[7] The derived pathology catalogue survives only as a repertoire of things to try, no longer as predictions. Beer’s comparison of Ashby’s Law to the law of gravity goes,^[8] and so does every diagram taught as a picture of how organizations are.

The subtraction is large: most of what is currently taught, and all of what is currently claimed for the model when it is sold.

A community is entitled to take the performative reading. It is not entitled to take the performative reading when asked for evidence and the representational reading when teaching, selling or citing. That alternation is no philosophical position, and it is what makes the model unchargeable.

The price named above is also a revision rule, of the form Chapter 28 asks for: if the performative reading is adopted, these specific claims are withdrawn from the teaching material by a stated date. It costs nothing to state, and nobody has ever stated it.

The tradition that did it in public

One tradition in the cybernetic lineage has made exactly this argument, in print, about its own central construct, and accepted what followed.

Asked directly whether the free energy principle can be falsified, Karl Friston answers that the principle is what it is: a principle; that like Hamilton’s Principle of Stationary Action it cannot be falsified and cannot be disproven; and that hypotheses about the brain performing some form of Bayesian inference or predictive coding are what they are: hypotheses.^[9] Chapter 9 quotes the same passage for a different purpose. What matters here is its shape: an explicit two-layer declaration, made by the author, unprompted, in print.

The philosophical response divides three ways, and the division is instructive. Mel Andrews argues that demanding falsifiability of the principle is a category error, which is Pickering’s move made analytically.^[10] Matteo Colombo and Cory Wright hold that the epistemic status of the principle is unclear.^[11] Jelle Bruineberg and colleagues argue that its metaphysical claims cannot be justified by appeal to the success of the mathematical framework alone.^[12] Outside that debate, Anil Seth, placing predictive processing in the cybernetic line, calls the supporting evidence only circumstantial.^[13]

Look at what the declaration bought and what it cost. It bought the principle immunity from a demand that did not apply to it. It cost the community the right to claim empirical support for the principle from the success of the process theories, and it exposed the process theories to ordinary testing, where their evidence base is still described as limited.^[14] Immunity and exposure came as a pair. That is what an honest category-error defense looks like.

Set the two founders side by side. Friston put the principle beyond testing and left the process theories exposed. Beer put the model’s validity beyond testing and left its utility exposed to prediction. Both declarations pair an immunity with an exposure, and both were made in print. The difference lies in what the communities did next. Friston’s took the exposure with the immunity, and its process theories are being tested. Ours kept the immunity, left the prediction test unbuilt, and invokes the immunity only when asked.

The shelter has a version one floor down, in the formal layer, and Chapter 3 has already dealt with its substance. The model borrows authority from the Conant–Ashby theorem,^[15] and since a theorem makes no empirical hypothesis, that layer too is immune by construction. Beer came close to saying so for Ashby’s law, in the reply on tautology that Chapter 9 quotes. What has changed since Chapter 3 was drafted is that the theorem’s own standing is now under formal attack, and the attack comes from outside our field. Chapter 22 takes it up, because it is the one place where the adversaries are doing us a favor.

What this chapter does not establish

It does not establish that Pickering is wrong. This book has no argument that he is, and the honest statement of our position is that the performative reading has been neither refuted nor accepted.

Nor does it establish that the free energy principle community is a model in every respect. Its declaration has been criticized from three directions, and the criticisms are live.

What it establishes is that this tradition holds a defense it has never declared, the performative reading, beside one its founder did declare and it has not honored, the prediction test attached to Beer’s own status for the model. Declaring the first carries a price nobody has quoted; honoring the second carries one nobody has paid. A community that wants either defense without its price is doing something other than philosophy.

Receipt — the two-layer declaration, with the subtraction listed. Publish a two-page statement, signed by the certifying bodies, dividing the Viable System Model into a principle layer and a hypothesis layer on the template Friston set for the free energy principle, or on the one Beer set in 1966, in which validity is secured by construction and utility is tested by prediction: what the model asserts that no observation could contradict, and what it asserts that observation could. Attach to the principle layer the list of claims that must consequently be withdrawn from teaching and from client-facing material — the neuroanatomical mapping, the necessary-and-sufficient formulation, the Ashby–Newton comparison, and any prediction currently made from the pathology catalogue — with a date by which the withdrawal takes effect. Price: one workshop day to draft, one round of comment, one signature; the expensive part is not the writing but the consent, and the consent is the deliverable. The exercise is decisive in a way the book’s other receipts are not: it cannot be completed without someone giving something up, and if it cannot be completed at all, that is the strongest available evidence for Chapter 9’s central charge — not that the tradition has failed a test, but that it will not enter a position from which anything could be lost.

Handover

Five chapters have dealt with friends. The last deals with the people who are supposed to be the danger, and asks whether any of them is.