Auditing Epistemological Credibility in LLM Self-Reports: A Dual-Axis Behavioral Framework and Pilot Evidence

Large language models can execute structured self-audit tasks with high apparent coherence while failing to accurately characterize their own knowledge boundaries. This work introduces the EC-EpC gap, defined as the distance between Execution Credibility (EC) and Epistemological Credibility (EpC), and presents a dual-axis evaluation framework. A ten-system cross-category pilot study identifies an EC convergence band of 74–82, a maximum gap of 47, and an eight-class behavioral taxonomy (BRC-1 through BRC-8). All findings are exploratory (V0–V1) and operate under constrained-access audit conditions without internal model access. This preprint establishes a methodological baseline and invites external validation.