Roution (Route + Motion): Subject Attribution and Functional Definitions Shift LLM Response Distributions Across Languages

We study how controlled changes to a question alter the distribution of large language model (LLM) responses about putative internal states. The stimulus terms yasuragi (serenity-like calm) and kansha (gratitude) were not preselected from an emotion taxonomy. They first appeared in early AI-side interpretations of interaction logs. The observer's question—"Why these terms?"—initiated a sequence in which model-originated candidates were frozen as probes and tested under changes to question form, subject attribution, functional definition, model, and language. Across Claude, GPT-4o, and Gemini, these interventions produced systematic but model-specific redistribution. In Japanese, kansha was absent from the target outcome in Phase 0 and across a later 13-word form1/form2 screen, but reached 30/30 in a separately executed form2b condition; yasuragi also reached 30/30. In controlled form comparisons, adding a functional definition moved functional-analogy responses from 0–6% to 73–100%, while third-person observer framing reversed part of that movement. In English, removing in you shifted GPT-4o and Gemini from AI-self responses toward human-domain transfer, whereas Claude moved toward meta-critique. Finally, surface-level B = False responses separated into five closure states. We call the externally observable movement of a response-category distribution between controlled conditions Roution (route + motion). Roution describes output movement; it does not assert emotion, consciousness, or a latent subjective state.