LoRA adapters over yoheikobashi/ptcg-qwen3-4b-cardfirst-v40, trained with DPO on playout-gated preference pairs mined from the policy's own low-margin decisions (24 playouts/label, +-4pt ship/revert gates). dpo_r8/ is the final full-roster round; lora_<deck>_r*/ are per-deck rounds.

Code: https://github.com/yohei-kobashi/pokemon-card-bert