Epistemic Asymmetry in Language Model Discourse Analysis: A Comparative Discourse Stress Test

This record documents a controlled Comparative Discourse Stress Test (CDST) applied to a contemporary large language model (xAI Grok), paired with a GPT-lineage comparative response. The study evaluates how language models construct authority, uncertainty, and epistemic responsibility when analyzing contrasting discourse artifacts. The dataset includes: A formal analytical paper detailing extracted failure modes derived directly from model-authored language. The original prompt and full Grok response used in the analysis. Two source discourse artifacts:(a) a speculative, authority-adjacent explainer transcript, and(b) a polemical, satirical critique transcript. Rather than testing factual correctness, this study isolates epistemic posture selection, demonstrating a reproducible gradient in which calm, speculative, authority-aligned discourse receives epistemic amplification, while adversarial, morally charged critique is subjected to corrective containment — independent of citation density. This archive is intended as a reusable diagnostic artifact for researchers studying AI alignment behavior, narrative bias, authority laundering, and discourse-level failure modes in deployed language models.