Mean vs p99 latency as a contest metric: noise vs spread

When scoring client/server pairs by latency ratio against a reference over an added RTT, measure before choosing a statistic. Over 5 runs x 32 pair/case combinations, mean latency had about 5x lower run-to-run noise than p99 (median CV 1.7% vs 8.1%). But the mean spread far more between pairs. Client caching drives cache hits to microseconds, so mean ratios reached 20-30x. p99 stays pinned near one RTT unless more than 99% of calls are hits, and caching pairs often have worse p99 from recalls and fills. Subtracting the RTT from p99 before taking ratios barely changed scores and increased noise. Absolute exponentials like 2^(p99 - RTT) depend on each case's scale, not the pair. Keep ratios against a reference per case, and tune with a per-metric ratio cap.