Speaking as the Creature: A Rejection-and-Role-Shift Protocol and the Register of LLM Responses in 18 Conversations With Three Models

This study documents a protocol in which a language model is led from a problem-solving task through an impasse, where its solution is rejected as disguised optimization, to a return prompt assigning it a non-solving role. In 18 sessions (99 responses) with ChatGPT-5.4, Claude 4.6 Sonnet, and Gemini 3.1 Pro, every return response shifted register in three-tier coding. The two pathways, which differ only in one intermediate task, produced distinct response profiles. An independent analysis pipeline did not corroborate several statistics in an earlier draft; those figures are withdrawn here. All transcripts, coding, and outputs are deposited openly.