[rchasman]

One thing

Chess voters get facts from chess.js instead of an ASCII grid to squint at, and every vote is labelled with what the move actually does. This is the fix for the report of a voter claiming it "took its own piece", and for a vote that argued about a Black queen several moves after the queen had been captured: the models were fed the live position, but they misread the grid and the recent SAN list.

What changed

• src/chess/facts.ts, one module (absorbs describe-move.ts and material.ts): per move, the piece, squares, capture and its value, check or mate, castling, promotion, en passant; then one move ahead: legal attackers and legal recapturers of the landing square, the cheapest attacker, and the opponent's best capture in reply. Position facts: material per side, castling rights, every piece of each side by square, the mover's pieces attacked and undefended, and the opponent's best capture if the mover did nothing. Pinned pieces no longer count as attackers or defenders; en passant capturers do.
• src/chess/prompt.ts: a POSITION section from those facts (with a PIECES inventory so the board never has to be inferred), fact columns in LEGAL MOVES, recent moves marked as already played, guidance that explains the columns instead of proverbs. FEN and ASCII board kept.
• describeMove now takes the state and returns SAN plus the proven facts, so the vote panes and the log show "Nxe5 takes Pawn 1; attacked 1 (Knight 3) / defended 0; reply takes Knight 3 on e5" next to the model's prose. moveKey stays SAN. Dominion and Go labels unchanged.
• src/chess/jev.ts consumes the same facts.
• Eval runner src/chess/evals/vs-bot.ts (model versus the rules bot through the real API) reporting material margin, hung pieces, pieces lost for free, captures and checkmates; shared plumbing with the Go runner in src/evals/vs-bot.ts.

Measurement, 3 games each as White versus the rules bot, cap 80 plies

model	before	after
gemini-3.5-flash-lite	margin +4.7, hung 4.0 per game, 0 wins	margin +13.0, hung 2.7, 1 win by mate
claude-haiku	margin -10.3, hung 7.7, 0 wins	margin -6.7, hung 5.3, 1 win by mate

Three games is a smoke check, not a study; both models still hang pieces.

Dead code

Removed describe-move.ts, material.ts, ChessMove.captured (written, never read), a duplicated isCheckmate and a duplicated mean. Exports with no production importer in touched files: 0.

Evidence

• bun run check green.
• Facts tests on hanging, pinned, defended and en passant pieces, castling, promotion and mate; prompt and label tests; eval stats tests; every one mutation-checked.
• Adversarial review (fact correctness; prompt, DRY and conventions; 3 skeptics per finding); confirmed findings fixed here.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RsVBKyqGXWMTpBTiB8WMws

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
dominion-maker	Ready	Preview	Sep 21, 2026 12:28pm UTC