Simulator Update Frequency Sensitivity in Driving Scenario Search

We examine how archive partitions and simulator update frequency affect driving-scenario search comparisons. Twenty learned policies span two tasks, two algorithms and five training seeds per combination. The initial study includes 128,000 episodes and 16,800 search campaigns; intervals for all eight prespecified rotation contrasts include zero. Changing replay from 30 to 60 Hz changes 0.54–6.98 percent of collision labels across task–algorithm groups. An exploratory fixed-history analysis compares shared-label Cartesian bounds, which permit case-wise mixing, with five whole-bank rules: three frequency labels, the original 15/30 conjunction and all-three agreement. Among 40 main-method pairs, 38 retain a nonzero direction under all five rules, versus 21 under Cartesian bounds. Across all 840 pairs, 787 retain a direction under the five rules. Label-rule robustness does not imply statistical separation from search-seed variability. Budgeted flips show that small permitted changes can erode sign robustness sharply. A prospective six-bank application adds 184,320 episodes with frequency-specific adaptive search. Yield maximization and minimax-envelope selection both choose cross-entropy search in every group, so this application cannot establish whether auditing improves selection. The contribution is a controlled rotation intervention and a reproducible sensitivity comparison. Update frequency changes traffic behaviour as well as integration; no isolated solver-error or physical-safety claim is made.