66 held-out test tasks, never shown to the optimizer across all 6 iterations · seed (unmodified prompt bundle) vs. the run's final champion
cand_0005, n=10 trials/task, scored exactly once after the champion was fixed by iteration 5's acceptance on the hidden confirm split. Unlike v4/v5's heatmaps, there is no per-candidate column here: only the seed and the final champion were ever scored on test — scoring an intermediate, later-superseded candidate would have burned the one-time test seal for no purpose. For the per-ITERATION view (every candidate's F-split screen and C-split decision), see
summary.md.