Role-Conditioned Instruction Sets for Evaluating Epistemic Behavior in Large Language Models

This deposit presents a pilot methodological study examining instruction-following and epistemic behavior in large language models (LLMs) using controlled role-conditioned instruction sets. Two instruction profiles—Academic and Devil’s Advocate—were applied to the same model under identical prompts and clean-slate conversational conditions. By holding prompts and model parameters constant while varying role constraints, the study probes boundary recognition, refusal behavior, and instruction adherence rather than task performance. The archive includes the full methods paper, raw model outputs for both instruction profiles, and the scoring rubric used for qualitative evaluation. Negative results and failure modes (including speculative completion and critique overflow) are treated as primary data. The contribution is methodological rather than empirical, offering a reproducible framework for diagnosing epistemic behavior and instruction hierarchy failures in LLMs.