sim2real-inference asks how few real trials certify a robot policy when a biased simulator supplies the rest. The valid interval it builds is two parts: a betting interval for the real-minus- twin residual, and one for the twin's mean from many twin runs.
Milestone 7n measured the residual part directly on 42 MuJoCo twin-real model pairs (two tasks). The two parts add up to the combined interval to within 0.4% (median over 336 cells), so whether the twin pays is whether the residual interval is shorter than the real-only one. A model that predicts that ratio from the standard deviations alone was off by 33% (median): close on the pendulum (14%), far off on cart-pole (about 48%), where the residual interval alone is wider than the real-only interval.
The miss tracks how close the real outcome sits to its ceiling (Spearman 0.81). That was found after the fact, so it is a hypothesis, tested next by shifting the outcome. These are model-versus-model pairs, not real-robot data. The pre-registered log and code are in the programme's LOG.md and experiments/m7n_residual_width/.