An honest reconstruction.

What chooses the move?

The site’s own JavaScript chooses the move. The locally bundled chess.js library handles legal moves and game state; it supplies no chess AI. No Stockfish, neural network, opening book or endgame database is involved.

Material and position are separate

Material uses pawn 1, knight 3, bishop 3½, rook 5 and queen 10. The king is handled through checkmate rather than a tradable piece value.

Each root candidate receives a searched material balance from the calculating side’s perspective. The best material score wins; tied candidates are ranked by their immediate position-play score. Exact ties use a deterministic alphabetical ordering of move notation. This separation is an implementation choice, documented here rather than hidden inside a blended number.

The position-play calculation

Mobility rewards the square root of available moves, counting a capture twice. Minor pieces and rooks receive defence credit. Pawns receive advancement and non-pawn-defender credit. The king receives mobility and castling terms, offset by exposure estimated through a hypothetical friendly queen on its square. Checks and mate threats add small bonuses.

The detail panel reports these terms for the engine’s side immediately after its candidate move. It does not subtract the opponent’s complete positional score. It reports the total, not the change from the starting position.

Search choices in this version

• Examine every legal root move and every legal reply: two plies.
• Beyond that, extend recaptures on the previous capture square, favourable-value captures, captures of undefended pieces (no opposing geometric attack on the capture square after the capture), promotions and immediate mating moves.
• When in check, examine legal evasions. At quiet extension nodes, allow the current material value to stand rather than forcing a capture.
• Use alpha-beta bounds to avoid material-search work that cannot change the result.
• Limit extensions to six additional plies, evaluated at the boundary, and approximately 6,000 visited nodes per root candidate. The limit is checked at node entry; remaining sibling visits may take the total slightly above it.

Limits apply independently to each candidate. A marked score means a boundary was reached, not that a quiet position was proved. A forced sequence beyond that boundary can be missed. This bounded search is a deliberate browser adaptation of the historical idea of following significant continuations.

Conventions that affect the answer

• Mobility uses legal moves. Defence uses geometric attacks, including pinned defenders.
• King exposure uses open queen rays and doubles captures. It does not replace the king in the rules library and re-test king legality.
• A future castling bonus requires at least one retained castling right; an immediate bonus requires a legal castle. The performed-castling bonus applies only to the root castling move.
• A mate threat means the calculating side would have a legal mate in one if allowed another move. This hypothetical test is skipped when the opponent is already in check.
• Promotions and check evasions are followed as practical tactical extensions; these are explicit reconstruction conventions.
• Threefold repetition and the 50-move rule are automatically treated as draws in this interface rather than requiring a claim. Insufficient material uses chess.js detection, not an exhaustive solver for every possible dead position.

What the scores do not mean

A positional score of +15 is not a fifteen-pawn advantage. It is a sum of this procedure’s mobility, defence and other terms. Material and positional scores are separate columns, both oriented towards the calculating side. They are not win probabilities.

What remains to validate historically

A future fidelity pass should compare each convention against the full primary text, record alternative readings and test positions from the published analysis. Replaying the Glennie score is already possible, but the engine’s disagreements are not automatically historical errors or evidence of mistakes by Turing.