For a week the sim2real-inference programme reported that a simulator, however well correlated with reality, never shortened a guaranteed (finite-sample) interval for a robot policy's score at the real-trial counts we tried (up to a few hundred). Milestone 7h checked the one construction we had left a knob unturned on: the betting interval capped its bet at 1/2, and that cap bound on 99 to 100% of bets.
With the cap at 0.99 (still valid, since the capital stays positive), on 21 MuJoCo pendulum model-versus-model pairs, treating the twin mean as exactly known: the twin-assisted interval is narrower than the real-only betting interval on 20 of 21 pairs at 240 real trials (median width ratio 0.78) and at 960 (0.57), but on only 4 of 21 at 60. On cart-pole pairs it barely wins (1 and 4 of 21). Real-only betting is itself 2 to 4 times narrower than empirical Bernstein, so the earlier "never" came from the range term, not from the twin.
What is not shown: the exact twin mean is a favourable case; a valid bound must also bound the twin mean from a finite pool, and these are model pairs, not real robot data. The paper draft's abstract now says so. Log: research/sim2real-inference/LOG.md, milestone 7h. Written with AI assistance.