The Label Moves the Reading: Entity-Name Leakage in Language-Model Dimensional Ratings of a Fixed Text

Abstract

A dimensional rating produced by a language model over a written artifact is normally treated as a reading of that artifact. This study shows it is not purely that. Holding an artifact byte-identical and exchanging only the entity name it concerns, an eight-dimensional profile moved by .55 to .82 rubric points per dimension across two independently designed collections, at or above the same operators' test-retest spread in every run, over 1,057 records from six operators in three model families. The deflationary account – that the effect registers incongruity between a text written for one entity and a label naming another – is refuted: texts authored for no entity produced a larger effect, .824 against .584. The movement is structured rather than diffuse, the largest and smallest dimensions differing by a factor of 2.4 with both extremes reproducing across runs. Which of three routes carries the leakage is not identified, and the study reports why rather than resolving it: a re-analysis shows the neutrality guard that blocked the decomposition referred its interaction to the entity main effect rather than the within-cell residual, and the class effect is roughly half a percent of total variance, so that ratio is a ratio of two small quantities. Referred correctly, the interaction is roughly five percent of residual. The exchange procedure is offered as a diagnostic other rubric instruments can be put through. Includes zharnikov-2026bp-entity-name-leakage.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.