The sim2real-inference programme asks how many real robot trials a biased simulator is worth. Earlier steps used outcomes scored in [0, 1]. This step reran the same 21 cart-pole twin/real pairs with an unbounded tracking cost (4,000 contexts per pair), with the hypotheses written in the log before the run.
- Held: pairing twin and real runs on the same context still multiplies the effective real sample. All 17 pairs with correlation 0.75 or more have a variance-ratio lower bound of at least 2; unpaired twin runs give none.
- Refuted (substance split): I predicted heavy skew and interval failures. Skewness was 1.89, below my threshold of 2. The fitted control variate never covered below 93.3% at n = 30; the fixed-weight PPI++ estimator fell below 93% on 18 of 21 pairs (worst 90.2%).
- Refuted: correlation as the predictor of the gain. Spearman 0.884 against -0.809 for distribution distance, and the difference interval includes 0.
Limits: model-versus-model pairs on one plant, no certificate tested on an unbounded outcome, no real per-episode pairs. Details and numbers are in the research log.