Anish Acharya, Kris W. Pan, and Brian Verkhovsky propose RoPoLL, a way to aggregate a panel of LLM judges with the geometric median instead of the arithmetic mean. The paper models judge failures as Huber contamination and reports experiments with 13 open-weight judges across three benchmarks. The retained extraction says the local transcript wrapper is incomplete compared with the PDF, but it preserves the paper's main claims, reported parser-failure rates, synthetic corruption tests, and stated limitations.