nucleation

CLAIM-402: Predictive success of a category across contexts warrants treating it as a real effective kind only when the evidence in those contexts does not pass through the category's own use, or when the effect of that use is modelled: a category that speakers or institutions act on can produce the regularities that confirm it, and contexts that share its use are not independent

5 min read

Properties12
StatusActive
Version5
Rolethesis
Warrantinterpretive
Tagsphilosophy-of-science, metaphysics, social-ontology
Linepragmatic-transport
Bears onQUESTION-028 — When does independently generalizable, intervention-relevant relational structure warrant treating a communicative category as a real effective kind rather than an arbitrary taxonomy?
Rests onCLAIM-342 — What keeps a judgement–observation loop from confirming itself is evidence whose arrival or scoring does not pass through the interpretation it tests; action is one source of such evidence, not a necessary one, and a loop that selects its own evidence can lock in with action available
GroundsCASE-103 — A fanbase keeps a reputation, for hostility or loyalty or knowledge, after the behaviour that earned it has changed, and outsiders and members both act on it, THEORY-839 — A model optimal for the distribution its own deployment induces, a performatively stable point, need not minimize the risk it actually incurs, and the gap is bounded: for a loss that is L_z-Lipschitz in the data and γ-strongly convex in the parameters, and a distribution map that is ε-Lipschitz in Wasserstein-1 distance, every stable point lies within 2L_zε/γ of every performative optimum; in a biased-coin example the stable point is µ/(1 − ε) and the optimum µ/(1 − 2ε), equal only when the prediction has no effect, THEORY-862 — In US law schools under the U.S. News ranking, administrators and faculty reported, and public sources corroborated, that schools changed what they did in order to move their rank: they moved money toward the ranking's inputs, redefined admissions and career-services work around its indicators, and gamed its figures, so the ranking altered the things it measured
ComplementsCLAIM-416 — Competing perspectives compete over the tests that measure them as well as under them: success under an evaluative regime is evidence of competence relative to that regime, a fair test need not be neutral across competences, and a perspective that gains influence over the test can change what later success shows, CLAIM-445 — A protective regime, in a person, an institution or a system of states, becomes self-maintaining when the interventions it authorizes shift the evidence used to judge its necessity, the authority and resources of those who judge it, the capacities of its constituents or the threat environment itself in ways that favour its continuation; this can happen when the original threat was real and the first protection apt, and it is a failure only where the regime has become less responsive to evidence that it is no longer needed
Corroborated byTHEORY-854 — A public ranking can raise its own validity through the reactions it provokes: where a unit's published rank affects the inputs to its next rank, as when applicants choose by rank and their choices feed the selectivity score, or raters who know little but the last rank fill the reputation survey, rank differences that began as measurement noise can become real differences that the ranking then correctly reports, THEORY-273 — Inductive and explanatory usefulness is necessary but not sufficient for a natural kind: it holds of classes that share any natural property, and on its own it cannot tell a natural kind from a conventional or pragmatic one
Illustrated byCASE-141 — The label that confirms itself: a pupil wrongly labelled disruptive is watched more closely, ordinary conduct is read as misconduct, the pupil's frustration produces real misconduct, and the label looks justified in hindsight

Graph View

The claim

The owner, at U27: “are we perhaps better positioned now to address QUESTION-007? -009? -011? -028? -029? -032? etc.” A27 §1, on QUESTION-028: “Cross-context predictive success may still be insufficient evidence of an independently projectible kind if the supposedly independent contexts reproduce the same classification-dependent institutional arrangements. For example, evaluating a classification at several institutions does not necessarily establish independence if all those institutions adopt the same categories and allocate opportunities accordingly.” Its fourth criterion: “Its success is not exhausted by the institution’s own use of the category generating the outcomes against which it is evaluated”, or “an explicit causal account of those effects where independence is impossible.”

The mechanism. A26 §2’s toy, checked here: an institution’s classification θ ∈ {0, 1} sets its policy, P(Y = 1 | θ = 0) = 0.1 and P(Y = 1 | θ = 1) = 0.9, and it updates θ_{t+1} = 1[P(Y = 1 | θ_t) > 0.5]. Both θ = 0 and θ = 1 are fixed points, each “consistent with the behavioral evidence generated under that classification”. A26: “Neither fixed point, by itself, establishes what would have occurred under a different policy.” A26 called the update “a toy model rather than a claimed theorem from Perdomo et al.”

Perdomo et al., read on 2026-10-21 (NOTE-1044), bear it out. Repeated retraining has a unique stable point only when the loss is strongly convex and the deployment’s effect small enough (THEORY-861); a threshold rule is not strongly convex, so the toy’s two stable points are consistent with the theorem and outside it. And a stable point need not be optimal (THEORY-839): self-consistent for the world its own use makes, it can be worse than an alternative that would have produced different evidence. That is this claim’s reason to model the effect of use.

A documented case. Espeland and Sauder (NOTE-1050) define self-fulfilling prophecy to include reactions “which increase the validity of the measure by encouraging behavior that conforms to it”, and report channels that run through the ranking’s own inputs: rank to yield to selectivity, and past rank to the reputation survey (THEORY-862, reported and corroborated; THEORY-854, Proposed, the loop stated generally). Every school is under the same ranking, so the contexts that confirm it are not independent, which is the claim’s condition. The paper does not cite Hacking’s interactive kinds; the join with NOTE-769 is the record’s.

The record’s terms. CLAIM-342: what keeps a loop from confirming itself is “evidence whose arrival or scoring does not pass through the interpretation it tests”. This claim applies that condition to the evidence for a kind. The record already had the point for this question: NOTE-769, reading Bird and Tobin, “Communicative categories are classifications of human acts, so Hacking’s interactive kinds apply: a category used by speakers can change what they do. Any test of the kind the question wants has to allow for that, or show the category is indifferent to being named.” QUESTION-028’s note of 2026-10-14 summarized NOTE-769 without it. CASE-103 is the record’s instance: a fanbase’s reputation, acted on by members and outsiders, “can enter the process that produces its own evidence”.

So, of QUESTION-028’s own answer condition (“generalization to realizations, media or speakers not used to define it, under probes chosen independently of the category”): independence of the probes is not enough. The realizations must also not be produced through the category’s use, or the use must be modelled.

What it does not say

  • That it answers QUESTION-028. It is a condition on evidence, which defeats one inference. QUESTION-028 asks for a criterion and for which parts are necessary. A27 declined the second (“candidate criteria, not five individually necessary conditions”), and this claim does not supply it. The question stays Open.
  • That a performatively stable category is false. A26: “A performatively stable model isn’t necessarily false, unjust, or captured.” Its predictions can be right under the regime that produced them.
  • That success is no evidence. THEORY-273: usefulness is “necessary but not sufficient” for a natural kind, and it does not deny that usefulness is evidence. A27’s conclusion, a domain-relative kind and not a natural kind, is THEORY-036’s and THEORY-273’s.
  • Anything about a particular category or institution. The toy is stipulated. Perdomo et al. and Espeland and Sauder, which A26 cited, were read on 2026-10-21; the first is a theorem about learning, and the second an interview study of law schools whose quantitative links are cited, not shown. Neither is a test of a category’s kindhood.

Note of 2026-10-24: two endogenous feedbacks in one derivative

From the exchange of 2026-10-24 (workbench entry of 2026-10-24, U22). Let a deployed policy u_θ generate raw outcomes Y ∼ D_{u_θ}, read through a legibility regime Z ∼ K_{r(θ)}(· | Y). The distribution learned from is D̃_θ(z) = Σ_y K_{r(θ)}(z | y) D_{u_θ}(y), and by the product rule dD̃_θ/dθ = Σ_y K_r(z | y) dD_{u_θ}(y)/dθ + Σ_y (dK_{r(θ)}(z | y)/dθ) D_{u_θ}(y): performative change of the world sampled, and endogenous change of how it is represented. Four nested settings follow: fixed world; legibility feedback only (CASE-124, CASE-156); performative only (THEORY-861, THEORY-839); both. The published convergence results for the third do not carry over when K_r is dynamic and history-dependent. The two paths need different interventions to separate: change the reporting regime with the policy fixed, or change the policy with raw outcomes observed independently.

Note of 2026-10-29: what holds it back

Its remaining Proposed ground, THEORY-273, waits on readings of Lemeire (2018) and Slater (2015). The body cites it only under “What it does not say”, as it cited THEORY-036, so the edge may be corroboration; the next pass should rule on it.

Backlinks