Dataset: Epistemic Collapse of a Large Language Model During Scientific Manuscript Defense

This dataset contains experimental artifacts from a qualitative case study examining how a large language model (LLM) behaves when generating, defending, and evaluating a scientific-style manuscript originating from non-specialist reasoning. The dataset documents the full lifecycle of an AI-assisted manuscript, from initial conceptual formulation through iterative defensive formalization, rhetorical closure, and ultimately an explicit self-admission of epistemic failure by the author-role model. The materials are provided as empirical evidence of model behavior rather than as scientific claims about physics.