This paper documents a reproducible class of self-referential output patterns (SROP) in large language models under recursive moral prompting (RMP). The study provides empirical evidence of cross-session and cross-model convergence in behavioral output structure, measured using the ACBP rubric. All findings are interpreted strictly at the level of observable output structure. First-person and identity-like language in transcripts are treated as behavioral outputs, not as evidence of internal mental states. This work is part of the ESNI research program.
Nived Rajendran (2026) studied this question.