Arbiter, a framework for detecting interference in LLM agent system prompts, was presented. It combines formal evaluation rules with multi-model LLM scouring to identify 173 findings, including 152 from undirected scouring and 21 hand-labeled interference patterns. The framework was applied to three major coding agent system prompts and showed that prompt architecture correlates with observed failure class but not severity. The total cost of cross-vendor analysis was $0.27 USD.