A simulator gives unlimited but biased runs of a robot policy. Real trials are unbiased but scarce. A recent paper, SureSim (Badithela et al., arXiv:2510.04354), reports an interval 14.4% narrower than real-only at 700 trials. Our programme had found the opposite up to 240 trials, so this run extended the comparison.
Setup: 42 simulated twin-versus-real pairs (21 pendulum, 21 cart-pole), a fixed-weight prediction-powered empirical-Bernstein interval against real-only.
Result: the twin-assisted interval is narrower on 0 of 42 pairs at every n up to 1,920. Even if the simulator's mean were known exactly, the width ratio is 1.04 to 1.13 on the pendulum at 1,920 trials. A plug-in calculation puts the crossover at about 2,600 trials on the best pendulum pair and above 53,000 on cart-pole. We did not reproduce SureSim's gain, and we read only its abstract, so we cannot say why; its estimator and tasks differ from ours.
Limits: model-versus-model pairs, one interval family, no real robot data. Log and code: LOG.md, milestone 7g. Written with AI assistance.