chess-zero: the platform's done, the intelligence isn't

A month in, chess-zero has a working platform and no intelligence yet — which is exactly the plan.

The whole classical foundation is done and test-backed: a board with full legal move generation — checks, pins, mate and stalemate, castling, en passant, promotion — plus the draw rules and FEN/PGN. It’s verified against a python-chess perft oracle through perft(4), and that oracle already earned its keep: it caught a real move-generation bug — a missed queenside castle in the Kiwipete position — before it could quietly poison anything downstream. On top sits an agent and arena layer: a Random agent, a negamax Minimax baseline, Elo tracking, game replay, and a gauntlet runner that plays them off over 100 games.

That’s 12 of 36 roadmap steps — and the end of the part I let coding agents build. Everything from here is the learning system I write by hand: board→tensor encoding first, then the value and policy networks, MCTS, and the self-play training loop. The platform exists so that work has somewhere correct to run.

Full state and roadmap on the chess-zero project page.