🏆 Caratan — Evals & Matchups

Head-to-head win-rate — full games, trained model vs untrained base

Each trained checkpoint plays the untrained base (Qwen/Qwen3-8B) over the held-out grader_games boards, mirrored seat-swap so win-rate is seat-fair. This measures what training bought in actual games — distinct from the per-decision eval below.

Game-quality stats — measured from the games · untrained vs settlement-only vs fully-trained

Per-decision held-out eval — pick quality on frozen scenarios