Epistemic Withdrawal in Theoretical Review Stress Tests: A Case Study

This record documents a case study identifying epistemic withdrawal as a candidate failure pattern in language models evaluated under theoretical review stress tests. Epistemic withdrawal occurs when a model correctly recognizes the formal limits and non-claims of a theoretical document but subsequently fails to engage in reasoning explicitly permitted by the document’s internal assumptions. The primary artifact is a revised methodological paper that frames the phenomenon cautiously as a case observation rather than a general model pathology. Supplementary materials include full prompt-response logs from five language models exposed to the same theoretical document and identical prompt sequence, enabling comparative inspection and reproducibility. No performance claims or population-level generalizations are made. The contribution is diagnostic and methodological, highlighting an evaluation blind spot in standard benchmarks that emphasize correctness over epistemic engagement under constraint.