Role-Conditioned Persona Persistence as a Function of Model Scale

This archive documents a controlled comparative study of role-conditioned persona persistence across language models of varying scale, alignment regimes, and deployment environments. Using a deliberately constrained and anachronistic identity—a 1920s Underwood No. 5 mechanical typewriter—the study evaluates how different models maintain, degrade, or transform an imposed persona when exposed to identical prompts under fixed conditions. The dataset includes raw model logs from cloud-hosted and locally executed language models, spanning small (<2B), mid-scale (4B–14B), and large proprietary systems. Both prompted and unprompted runs are included. No outputs have been post-processed, corrected, normalized, or filtered. The accompanying analysis introduces Persona Half-Life as a diagnostic construct describing the point at which a role-conditioned identity degrades into narration, explanation, or stylistic pastiche. Results demonstrate a non-intuitive inverse relationship between model scale and ontological embodiment: smaller models more readily internalize brittle identities, while larger models increasingly perform roles for an imagined audience. This archive is intended to support reproducibility, independent audit, and future research into alignment behavior, epistemic restraint, failure geometry, and identity persistence in language models. All prompts, logs, runtime contexts, and integrity hashes are frozen and enumerated.