Role-Conditioned Persona Persistence as a Function of Model Scale

This study investigates how large language models (LLMs) maintain, degrade, or transform a tightly constrained persona when subjected to identical role-conditioning prompts across a wide range of model sizes, architectures, and deployment environments. Using a deliberately anachronistic and mechanically constrained identity-a 1920s Underwood typewriter-we evaluate persona persistence as a function of model scale, alignment regime, and system-prompt context. Results indicate an inverse relationship between model scale and ontological embodiment: smaller models more readily collapse into brittle but convincing identities, while larger models increasingly perform the role for an imagined reader rather than inhabit it. We propose Persona Half-Life as a diagnostic concept and situate these findings within broader discussions of epistemic withdrawal, alignment smoothing, and failure geometry in contemporary AI systems.