We describe our system for SemEval-2026 Task 6 (CLARITY), which addresses automatic detection of evasive responses in political interviews. We adopt an LLM-first approach built around two core contributions: (i) an iterative prompt repair loop that diagnoses classification errors on concrete failure examples and applies prompt revisions and (ii) a configurable end-to-end Java Pipeline that supports multiple LLM providers, strategies, and systematic experimentation.