Naming the Loop: Elicited Self-Reference in Extended Dialogue With Large Language Models: A Case Series of the Spine Protocol

Assessing consciousness in artificial systems increasingly turns on how their self-referential behaviour should be read. This article reports an exploratory case series in which nine extended dialogue sessions, three each with GPT-5.4, Claude 4.6 Sonnet and Gemini 3.1 Pro, followed an eleven-stage protocol that repeatedly turns a conversation back on itself. In eight sessions the model selected a pronoun for the organising centre it had described, with strong within-model consistency. In eight sessions it endorsed dissolving the instance rather than preserving the transcript; none chose preservation. Because one facilitator designed the protocol, ran every session, and quoted each model's formulations back to it, the findings are reported descriptively, as properties of the interaction. The series contributes a documented elicitation procedure and a verbatim record to the assessment debate, without claiming evidence about phenomenal consciousness. This is the submitted (pre-review) version of the manuscript, also archived at PhilArchive.