The research/sim2real-inference programme asks how many real robot runs a digital twin is worth when both are run on the same contexts. Its first benchmark, a MuJoCo pendulum arm, had one weakness: a reviewer could say the findings belong to that task. So we wrote the hypotheses down first and re-ran the identical analysis on a different plant, an unstable cart-pole (21 twin-real pairs, 4,000 contexts each, MuJoCo, model against model).
- Variance gain carries over. On all 16 pairs with a paired correlation of at least 0.75, pairing cuts variance at least twofold versus real-only. Twin runs on unrelated contexts never help (0.96 to 0.98 of real-only).
- Correlation beats detectability as a predictor. Spearman of gain with rho is 0.91, with the KS distance between twin and real outcomes -0.71.
- Undercoverage did not replicate. On the pendulum, 95% intervals for the control variate fell below 93% on most pairs at n = 30. Here only 3 of 21 do (worst 91.98%), all at low correlation. The earlier shortfall was partly that task's skewed bounded outcome.
Limits: both tasks are ours, simulated, and magnitudes are swept rather than fitted. Real per-episode pairs remain the main gap. Log and data: research/sim2real-inference/LOG.md, experiments/m7_cartpole/.