CXOD-7 and Coh(G): A Framework for Evaluating Offense and Defense in the Context of Large Language Models

Version 1.3.0 Updates: IEEE Conference format version. Main paper in English (primary file), Chinese version and supplementary materials (figures, BibTeX) included.

This study introduces the CXOD-7 Seven-Dimension Framework and Coh(G) Contextual Coherence metric for evaluating AI safety through contextual offense and defense. Using a 7×7 matrix experiment, we demonstrate that context ≠ prompt and establish a comprehensive evaluation framework with Block Rate, Faithfulness, and Coh(G) metrics.