Observer Integrity Under Procedural Sociality: Measurement When the Instrument Shares the Grammar

Frame Capture in Conversational AI (Paper 20) introduced same-model cross-frame interferometry as a method for detecting frame capture in conversational AI. But the method has a recursive problem: the measurement instrument operates in the same social-signaling grammar as the system being measured. This paper argues that the perturbation is not a fatal flaw but an operating condition. We define five perturbation classes that socially legible assays induce in language models, propose a layered measurement architecture for managing the perturbation, and identify design principles for reducing the social signal surface of measurement instruments. A cross-model self-test across three API models and five local open-weight models found that observer perturbation is bounded and continuous in API models but materially stronger in several local models, with pilot evidence suggesting that perturbation susceptibility is not reducible to parameter count alone.