The paper's headline latency table reports mean latency per query in milliseconds for exactly three retrieval systems under the CPU-only protocol: Flat dense, Hierarchical baseline, and Adaptive (heuristic). Rows are indexed by dataset, so each dataset gets one timing per system. The three-column choice mirrors the headline strict-parity retrieval table (flat dense vs. fixed-depth hierarchical baseline vs. adaptive depth gating), so compute cost is reported for the same systems whose retrieval quality is compared, not for an unrelated set of comparators.