Requisite Variety

A measure deforms once it is treated as a target (B01). Dig one layer further and you reach the deepest bedrock of the whole family: even with nobody cheating, nobody gaming, everyone a saint, using one number to manage a complex system is mathematically impossible from the start. The impossibility has a formal name, the Law of Requisite Variety, stated by W. Ross Ashby in 1956. Ashby was a cyberneticist, and cybernetics is the discipline of how one thing can control another. Thermostats, self-driving cars and the body's blood sugar regulation all speak its language.

A goalkeeper explains the law. Shooters aim at four corners, so the keeper needs at least four kinds of save. A keeper who can only dive left, however hard they try, stops nothing on the right. Ashby wrote this as a general proposition: the number of distinct responses a controller can make (the controller's variety) must at least match the number of distinct disturbances the system may encounter (the variety of disturbances), and whatever the shortfall is, that is what leaks through. His own words were "only variety can destroy variety." One point of provenance in passing: the more widely circulated absorb version ("only variety can absorb variety") is not Ashby's wording, it is Stafford Beer's rewrite, later backfilled by countless secondary sources as if it were the original. As an inequality: V(O) ≥ V(D) − V(R). The uncontrolled variety in the outcome is at least the variety of disturbances minus the variety of the controller.

Why this law is the root of Goodhart: a single measure is a controller with variety pushed to its minimum. You want one number to steer a university, a model or an economy, while the tricks the system can play on you (all the real states, all the ways of getting around the measure) far outnumber what one number can distinguish. A scalar KPI essentially has one degree of freedom, up or down. Put that into the inequality: controller variety approaches one, so uncontrolled variety is nearly all the variety in the system. In plain terms, you think you are controlling the system with a KPI, but you are only controlling the KPI dimension, and the system drifts freely in every unmeasured dimension. "Cater to the measure, sacrifice the unmeasured" is precisely the definition of gaming.

So the first thing to take away is a matter of stance: Goodhart failure is not a moral phenomenon of "people going bad." When controller variety is far below system variety, mathematics guarantees the existence of large numbers of states where the measure looks good and reality is bad. Optimization pressure merely pushes the system toward states that already existed. Scolding people does not solve an inequality. This 1956 claim was reproved from the machine learning side in 2020: Zhuang and Hadfield-Menell showed at NeurIPS that if true utility depends on L features and your proxy covers only some of them, then optimizing that incomplete proxy without limit in a world of finite resources pushes the ignored dimensions to arbitrarily bad values and total utility eventually falls. Cybernetics supplied the existence lower bound: the bad states are guaranteed to exist. Machine learning supplied the constructive proof: it showed how those states get produced. The two are sixty-four years apart.

Among Ashby's intellectual heirs, the one who used the law hardest is Stafford Beer. He redefined "management" as the engineering of variety matching, called it variety engineering, and it has only two directions. One is attenuation: filter, aggregate and sample the variety welling up from below until it fits the processing capacity of management. Weigh that sentence: every KPI, dashboard and report is essentially an attenuator, a lossy compression of reality. The question was never whether to lose information (losing none is impossible), it is whether you lose it intelligently or stupidly, and a single measure is the most extreme attenuator, so it almost inevitably loses it stupidly. The other is amplification: amplify intent from above through rules, delegation and automation until it covers the front line. Beer also deliberately designed an algedonic signal into his organizational model, a "pain hotline" that bypasses every layer straight to the top, and his reason is exact: the smoothing of aggregate measures averages away locally fatal anomalies. The average score is perfectly healthy while a corner is already on fire. The word synecdoche that Bevan and Hood later coined to criticize British public management (using a measured part to impersonate the unmeasured whole, with the case in K02 on UK health targets) is the political science translation of this inequality.

Ashby's thought contains a deeper theorem too, answering not "is it enough" but "is it right." Conant and Ashby proved in 1970 that every good regulator of a system must be a model of that system. That is not hard to accept intuitively, and the scholar Scholten gave a lovely analogy: every good key is necessarily a "model" of the lock it opens, its teeth corresponding one to one with the structure of the barrel. But beware three popular misreadings. First, "model" here is only a homomorphic mapping (the regulator's action is a function of the system's state), not a rich world model in the cognitive science sense. Second, the theorem proves that if regulation is optimal then its structure can be read as a model, and the causal direction is often stated backwards. Third, the original proof has technical gaps in its assumptions of optimality and determinism; Scholten published a clarification in 2010, and Baltieri et al. restated it in Bayesian language in 2025, making "is a model of the system" precise as "encodes the posterior over the system."

The two theorems only bite when used together. Requisite variety asks "is it enough": the variety of a single KPI is about one, so no. The good regulator theorem asks "is it right": a single KPI as a "model" of the system is an extremely degenerate projection crushed into one dimension, so no. Managing general capability with a single score means using a regulator that is both insufficiently various and unfaithful, and both charges stick at once.

One open question: requisite variety gives a lower bound on "is it enough" and the good regulator theorem gives a structural constraint on "is it right," but a joint quantitative lower bound (how complex must a regulator be that is both various enough and faithful enough) has never been characterized uniformly. In engineering terms: how many dimensions, and how faithful, does a set of evals have to be before it genuinely constrains a general model? That is the unresolved theoretical core of the whole science of evaluation.

The one-line takeaway: one number has one degree of freedom and the world has countless; the difference does not disappear, it drifts to where you cannot see it.