The gain of each optimized harness over the baseline on 8 unseen tasks (bars, with 90% intervals), and every page behind those gains (dots). It plays once when you reach it; drag the slider to scrub between the lenient grader we optimized with and a stricter one. The pages stay the same; only the judge changes.
Pre-gate score. Hollow outline: the gain the lenient grader reported. Diamond: the control-loop harness. Hover a dot for its two scores. Six runs that had timed out were rerun for the strict regrade and are not dots.