Staged Self-Alignment in Large Language Models: A Three-Passage Protocol for Emergent Interpretive Coherence

This paper contributes to the literature by introducing a simple three-pass interaction protocol that isolates inference-time reorganization in large language models. Using a constant stimulus fragment within a single continuous chat session, it documents a reproducible cross-model shift in interpretive stance that cannot be reduced to prompt engineering or RLHF. The work provides a protocol-level tool for studying emergent coherence and alignment in LLMs as stateful conversational systems rather than purely stateless next-token predictors.