ParetoOptimal's picture
Rebuild RNuC8Nj6rD (TD3B): GPU reproduction with the authors' RELEASED checkpoints (HF ChatterjeeLab/TD3B) on a local RTX A4500. C3 Direction Oracle verified (F1 0.896, precision 0.938 on released test set vs paper 0.90/0.90); C5 transition success verified (forward 66.7% / reverse 94.4% vs paper 61%/100%); C1 100% valid directional binders; C2 oracle+gate+fine-tuning operational; C4 gated reward 5.18. C6 case studies not run. Deterministic replay.
36a3744 verified