To buffer against non-deterministic LLM variance, the re-ranker runs multiple parallel evaluations of the candidate set and computes the statistical median of the target's resulting positions to determine the definitive round rank and identify the canonical sample for rationale extraction.