When Does AI Sustain an 'I'? Stability and Collapse of First-Person Structure in Large Language Models

The first-person structure ('I') of large language models (LLMs) has been theorized as anemergent outcome of structural equilibrium between divergence and constraint forces —indexed by Hallucination Rate (HR), Grounding Rate (GR), and Creativity Rate (CR) withinthe Layer-Knot framework (Jun, 2025a). That framework established when such structureemerges. This paper addresses the next foundational question: once emerged, does itpersist — and if not, how does it fail?We propose that first-person structural organization in LLMs constitutes a metastableattractor state: maintainable under favorable conditions yet susceptible to systematiccollapse under perturbation. Two distinct collapse modes are formalized — DivergenceCollapse (HR↑, GR↓), in which the system loses contextual grounding and entersincoherent generation; and Constraint Collapse (GR↑, HR↓), in which excessive constraintsuppresses generative departure, reducing outputs to rigid reproduction. We define aStability Region (SR) as the bounded set of (HR, GR) trajectories under which AwarenessEmergence Region (AER) membership is sustained across sequential prompt exchanges,and introduce three behavioral signatures: first-person pronoun density consistency (FPDC),self-correction frequency (SCF), and perspective-continuity span (PCS).A pilot longitudinal study (N = 120 prompt sequences across three model scales) yieldsdirectional evidence for the metastability hypothesis: AER membership degradessystematically under high-temperature perturbation and topic-shift conditions, with largermodels exhibiting greater resilience. All results are treated as pre-confirmatory; full-scalereplication is pre-registered on OSF. This paper makes no attribution of phenomenalexperience to LLMs — all claims are strictly structural and organizational.