| Adding venue-split (home_home_wpct/away_away_wpct), separated diff_off/def_rating, or symmetrizing form with away_win_pct_10 to the 16-feature star_dependency champion does NOT improve log loss (all statistical ties). | high | meta_scientist | 2026-07-18 |
| Promotion audit [vetoed] for TabPFN - trimmed 10-feature core (84a05669) | high | critic | 2026-07-17 |
| The 16-feature star_dependency champion can be trimmed to 12 features (dropping home_efg_pct, diff_net_rating, home_pie, away_pie) with no significant log-loss change — a valid production simplification, not an improvement. | high | meta_scientist | 2026-07-17 |
| Adding home/away roster_depth on top of the 16-feature star_dependency champion does not improve log loss (statistical tie); roster_depth adds nothing beyond star_dependency. | high | meta_scientist | 2026-07-17 |
| The 16-feature Ridge L2 (C=0.01) model with home/away star_dependency (0.615079) is the verified champion, replicated again and significantly better than the 14-feature anchor (0.616294, p=0.024); Optuna re-confirms C=0.01 is optimal for this exact feature set. | high | meta_scientist | 2026-07-17 |
| Promotion audit [vetoed] for TabPFN - trimmed 10-feature core (84a05669) | high | critic | 2026-07-16 |