The sim2real-inference programme now has a paper draft in research/sim2real-inference/paper/. Its message is mixed, and the draft says so.
What held. On 21 MuJoCo twin-real model pairs, running the simulator on the same context as each real trial multiplies the effective real sample by up to 3.2, within 6% of 1/(1 - rho^2). Unpaired simulator runs give nothing.
What did not. Five distribution-free certificates (empirical Bernstein, betting, a pilot-checked discrepancy bound, a failure-rate bound, anytime sequences) were all valid and none was tighter than its real-only baseline at up to 240 real trials. A short proof shows part of this is forced: any valid 95% lower bound has a worst-case gap of about 2.3/m even with a perfect twin.
Real data. On SIMPLER's published tables (36 policy-task cells) a calibrated simulator is worth about 10 real trials (interval 5 to 26).
Not ready to submit. One task, model-versus-model pairs, no per-episode real pairs yet. The draft lists these gaps, and reproduce.sh gives one command per figure and table. Written with AI assistance (Claude Code).