Separating Instrument Drift from Brand Signal: A Pre-Registered Model-Version Tracking Instrument for AI Brand Perception (PRISM-T)

Every longitudinal claim made with a large-language-model observer is confounded by the observer: when a brand’s measured perception moves between epochs, either the brand’s public signal changed or the vendor shipped a model version that reads the same artifacts differently. PRISM-T resolves the confound with a two-panel identification strategy: a pinned panel of public artifacts captured once and hash-sealed – byte-identical inputs, so all movement is apparatus drift – and a live panel of fresh artifacts whose movement adds brand signal. The pinned panel yields a version floor, the across-time counterpart of the operator noise floor, grounded in classical test-retest reliability with the locus of error inverted from respondent to apparatus. A pre-registered campaign read a sealed 160-artifact panel under eleven real shipped versions in three model families, including an eighteen-month cross-generation pair, with controls, a mechanical operator-exclusion rule, and a simulation power analysis fixing the certifiable drift magnitude ex ante. No version pair moved the reading beyond the contemporaneous operator floor (largest signal-to-noise 1.19, CI .89, 1.55), a bounded, reassuring null: version-robustness is now a measured property with an interval, not an assumption. Includes zharnikov-2026ba-separating-instrument-drift-from-brand-signal.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.