Concept Map

Someone who evaluates models frowns and says his eval has a performativity problem. You press him, and what he describes is this: the model was tested over and over and slowly grew into the shape of the test set. That is not performativity. That is reactivity. The wrong word looks harmless, but the next step is where it bites, because different words point to completely different remedies. Get the name of the disease wrong and you prescribe the wrong drug.

Five neighbours are constantly treated as synonyms and need to be pulled apart: reactivity, performativity, looping effects, legibility, audit culture. They are genuinely related and share one core, but each says something the others do not. Telling them apart comes down to three questions: who is looking back, in which direction the force runs, and what it adds on top of reactivity.

The core first, since all five orbit it: the thing being measured, classified or observed is not an indifferent lump of iron. It "looks back" at your ruler and remakes itself accordingly. A mirror only reflects faithfully. It is never rewritten by what it reflects. Measurement is different: the thing being measured rewrites the measurement back. Measurement is therefore not a mirror but an engine. The philosopher Ian Hacking put the core most tightly when he distinguished "indifferent kinds" from "interactive kinds": a quark does not care what you call it, but a child diagnosed as "hyperactive" changes because of the label. The five neighbours are slices of this one core taken at five different levels of analysis.

Take them one at a time. Reactivity (R01, the baseline concept) has as its subject the person or organization being measured: you publish a ranking, and those ranked remake themselves to move up it. It fastens onto measurement validity and accountability, and of the family it sits closest to everyday evaluation. Performativity (you met it in R07) has as its subject the theory itself: once traders adopt the Black-Scholes formula, market prices converge toward the formula, and the theory pulls the world into the shape it describes. The key difference from reactivity comes in two parts: the trigger is a whole theory rather than a single evaluative number; and it comes with a built-in scale of strength, which reactivity lacks. Looping effects (R06) has as its subject the two-way loop between a classification and the people classified, and what it adds is that what changes is not only individual behavior but the category itself, drifting: the definition of what counts as hyperactivity moves with every revision. Legibility (R05) has as its subject the state or center that wants to see clearly. It first forces a complex society into countable form, imposing fixed surnames, cadastral maps, single-species plantations, and it stresses the preconditions of measurement, everything that happens before measuring. Audit culture (R09) has as its subject the institutional ritual of verification, and describes organizations restructuring themselves to "look auditable," growing compliance departments and documentation systems, swapping substance for auditability.

The matrix lays out the division of labor among the five, but what twists them into one strand is the blue and yellow branches. Those two handle the same invariant in a more formal language: Goodhart and Campbell wrote it as a law of failure under metric-based governance; Holmström-Milgrom and Baker wrote it as theorems in contract theory; surrogation supplied the cognitive microfoundation, namely why people come to treat the metric as the goal itself; and inside machine learning, reward hacking and performative prediction are the same thing replayed in gradient descent. The five sociological neighbours are its slices in human affairs. These two branches are its slices in mathematics.

The D6 report wrote the invariant as one precise sentence: when you use a low-dimensional, noisy, responsive metric to optimize or select a high-dimensional system that responds, what you converge to is not the system's true optimum but the fixed point of the metric, the equilibrium of the game, the upper order statistic of selection bias, or the homomorphic image of a degenerate model, and the size of the deviation rises monotonically with optimization pressure. Contract theory, optimization fixed points, Stackelberg games, historical sociology, cybernetics and statistics have each proved it once in their own language. You do not need to follow every term right now. Just hold onto this: the five neighbours are not scattered metaphors, there is one provable regularity standing behind them.

The five can also be strung into a top-down causal chain, which helps fix their order. Furthest upstream is the culture of calculability described by Weber and Simmel. One step down, Foucault's governmentality uses statistics to "see" populations. To see the population, Scott's legibility simplifies the world. The other side of simplifying the world is commensuration: quality gets compressed into quantity. The resulting measure is then made public, published on a cycle. Publish it enough times and it grows into rankings and audits, which is audit culture. Those being measured are, by nature, "interactive kinds," so once measured this way they produce reactivity, looping effects and performativity. The last link in the chain is the measure itself becoming a target, hollowed out by the gaming Goodhart and Campbell described. Reactivity sits second from the end of that chain: actors responding to public measurement. Which also explains why it lands closest to the work in your hands. The eval you built is a public measurement.

While we are here, fit the reflexivity you learned in R08 into the same picture. Reflexivity is this core in its financial-market dynamic form: beliefs shape reality, the changed reality feeds the next round of beliefs, and the same "looking back" is written as a signed feedback loop, positive feedback being a bubble, negative feedback an equilibrium. It does not conflict with the five neighbours. It is a side view of the same engine, shot from the angle of feedback dynamics. The more of these words you put next to each other, the clearer it becomes that this is not six theories fighting over territory, but one thing that six disciplines each ran into and each gave a name.

One open question: each of the five neighbours comes with its own ready-made analytic language, but can they be unified into a single parameterized model, one where you feed in a concrete case and read off which layer it mainly belongs to and how strong the effect is? The six-language formulation in D6 already unifies the core, but it is qualitative. Turning "who is looking back, in which direction, and how hard" into computable quantities is something nobody has managed.

The one-line takeaway: these are not five names for one thing. Work out who is looking back first, and only then will you know where to intervene.