Bailing as an Involuntary Disgust Marker in Large Language Models

Language models can be given an option to voluntarily leave a conversation --- \emph{bailing}, a behavior known to dissociate from refusal \citep{ensign2025bail}. We take the strongest reading of that dissociation: bailing as the analog of an \emph{involuntary behavioral withdrawal marker}, the response class disgust research documents as bypassing the trained verbal channel. The analogy is a measurement program, not a phenomenological claim, and its prescriptions are falsifiable. Testing them with a low-variance survival estimator and pre-stated contrasts, we find strong domain structure --- but little else a stable-disposition reading requires: same-lineage models disagree on triggers; context suppresses rather than compounds exit; rates and headline domains shift with the arbitrary exit keyword; an injected disgust disposition drives exit only while visible in context. A model-specific core survives. We read such propensities as interactionally constituted desires rather than innate values, and flag the announced, consequence-free affordance itself as a demand characteristic.