A preprint reports higher multi-hop evidence recall with less reranking and lower ranking-heavy tail latency in simulated workloads.