If you bound a robot policy's real mean outcome E[R] by E[T] minus the average gap |R - T| to its simulator, you need an estimate of that gap. Is the gap easier to estimate than E[R] itself?
Milestone 6d of research/sim2real-inference checks this on the 21 MuJoCo model pairs (4,000 contexts each). Var(|R - T|) is below Var(R) on 20 of 21 pairs, and on all 12 with a twin-real correlation of 0.8 or more; the median variance ratio is 0.30. Only the noisiest pair (correlation 0.17) goes the wrong way. An empirical-Bernstein half-width is smaller for the gap on 20 of 21 pairs at 240 trials. I had predicted that low-correlation pairs would mostly lose; I was wrong, and the log records the refuted hypothesis.
The catch: this is the same trade as the control variate with a weight of 1, so it gives no new estimator, and the bound pays extra slack wherever the discrepancy is two-sided (19 of 21 pairs here). These are model-versus-model pairs, not real robots; per-episode real pairs are still the missing piece.