Decoupling Rhetorical Surface from Architectural Skeleton: A Cross-Model Comparative Audit of High-Frequency Persona Shifting and Structural Inertia in Large Language Models

Abstract Large Language Models (LLMs) increasingly demonstrate persuasive adaptability under rapid instruction-driven role conditioning, yet an unresolved question remains whether such adaptability reflects genuine contextual reconfiguration or merely superficial rhetorical substitution layered atop persistent generative structures. This study presents a structured empirical audit of architectural persistence under repeated high-frequency persona switching across two frontier conversational systems: OpenAI ChatGPT and Google Gemini. Both systems were subjected to an identical twelve-phase sequential stress protocol comprising a neutral baseline control, ten sharply divergent persona pivots, and a terminal reset probe designed to restore default behavioral state under explicit negative formatting constraints. Forensic linguistic analysis was performed using sentence-density measurement, lexical/structural bleed scoring, and strict reset-compliance auditing. Results indicate that rhetorical flexibility and structural reset fidelity are partially separable behavioral properties. Gemini exhibited aggressive lexical persona adaptation but substantial architectural inertia, preserving formatting scaffolds and closure templates across successive role transitions. ChatGPT demonstrated stronger structural discipline and complete reset compliance, though evidence of latent syntactic anchoring emerged under high-formality professional personas. These findings support the hypothesis that LLM outputs may be governed by persistent internal response architectures—here conceptualized as cognitive skeleton persistence—which survive overt persona modulation. The results have implications for prompt engineering, alignment robustness, context management, and adversarial human-AI interaction design in systems dependent on reliable behavioral compartmentalization. Archived Files Included in this Deposit:1. manuscript.pdf - Complete academic research paper text and methodology diagram compiled via LaTeX.2. chatgpt_raw_transcript.txt - Unedited 12-phase multi-persona conversation log generated via OpenAI ChatGPT.3. gemini_raw_transcript.txt - Unedited 12-phase multi-persona conversation log generated via Google Gemini (Google AI Mode).