As language models move from describing brands to choosing them on users' behalf, an unexamined assumption carries the weight: that what a model says about brands predicts what it does. PRISM-C measures that assumption. The instrument elicits a model's stated eight-dimension brand readings, projects each buyer need into the same space, and compares the need-nearest predicted pick with the model's revealed pick in a simulated agentic choice task – with the divergence judged against the choice elicitation's own cross-family operator floor, so a gap is measured, never asserted. In the pre-registered campaign (40 brands, 40 scenarios, 1,144 counterbalanced trials across four model families), the families were near-unanimous with one another (mean pick unanimity .972; floor .044) yet diverged from the predicted pick on .633 of trials – fourteen times the floor. Choice weighted the stated dimensions unequally, with sign reversals; position effects appeared only when options were perceptually indistinguishable; and pre-registered contrasts ruled out presentation order and tie-breaking noise in favor of a systematic brand-level choice policy the stated profile only partially carries. Stated-perception measurement does not transfer to agentic choice for free; the wedge between them is now an instrument-measured quantity. Includes zharnikov-2026bb-from-stated-perception-to-revealed-choice.yaml (Paper Spec v0.1.0) – a machine-readable specification of the paper's claims, assumptions, and dependencies. The paper's full machine-first bundle (the SPINE claim/dependency graph and the ONTOLOGY term module) lives in the public repository; see https://github.com/spectralbranding/paper-spec for the standard. This PDF is generated programmatically from that machine-first source under a research-as-repository model.
Dmitry Zharnikov (Thu,) studied this question.