The pawn that never moved
Listen to this post7 min
Created by AI from the text of this post and read by a synthetic voice; not a word-for-word version, and it may contain errors.
Loading…
Created by AI from the text of this post and read by a synthetic voice; not a word-for-word version, and it may contain errors.Position and speed stay in this browser and are sent nowhere.
The recording could not be played. Open the audio file
The endgame battery introduced in the entry on bishop and knight tests one endgame class at a time. The engine plays both sides from twenty random positions per class, with a fixed budget of one hundred thousand nodes per move and without endgame tablebases. The tables only label each starting position as won, drawn or lost. For version 0.1.120, rook and pawn against rook went well: six of seven won positions were converted, and all thirteen drawn ones were held. This entry looks at the seventh won position.
For this entry, every position of that game was checked afterwards against the same tablebases. They act as a referee only. They did not play a single move.
The position
Black has king, rook and a pawn on g7. White has king and rook, and the white king stands far away on a3. Black is to move and wins with best play. That sounds comfortable, but the tables are strict: of Black's twenty-one legal moves, only three keep the win, Rd7, Kf7 and Kf8. Fourteen moves only draw, and four rook moves even lose.
Rook and pawn against rook, Black to move and win: the first eight half-moves of the game. White: king a3, rook g1; Black: king e8, pawn g7, rook d4.
The engine chose 1...Kf7, one of the three. After 2.Kb3 Kf6 3.Rf1+ Ke5 4.Rg1 the position is still won, but now by a single move: the king has to go back to f6, as the arrow shows. The engine played 4...Rd7, and the tables now say draw. White then had exactly one move that keeps the draw, 5.Kc4. The engine played 5.Kc2, and Black was winning again.
Two errors in a row cancelled each other out. The result of the game would not have shown either of them. Only a referee that checks every position sees them.
A win the clock takes away
From there on the tables, which do not know the move counter, rate Black as winning in every position with Black to move, up to the last one. Black never converted. In one hundred half-moves the pawn did not move once, and nothing was captured. That is exactly the limit of the fifty-move rule, and the game ended as a draw. The pawn could have advanced as early as Black's second move: after 2...g6 the tables still show a win.
Under the fifty-move rule, a win is a race between two numbers. The first is how many half-moves Black still needs, with best play on both sides, until the next pawn move or capture that keeps the win. The tables give this distance. The second is how many half-moves are left before the rule ends the game. As long as the first number is not larger than the second, the win is still there.
The same game, every position with Black to move: what the win still needs and what the clock still allows. Needed: distance to the next winning pawn move or capture (tablebase): 0 11, 2 1, 4 13, 6 15, 8 25, 10 27, 12 25, 14 27, 16 25, 18 23, 20 21, 22 23, 24 21, 26 23, 28 29, 30 27, 32 25, 34 27, 36 29, 38 27, 40 25, 42 27, 44 29, 46 27, 48 29, 50 31, 52 29, 54 31, 56 37, 58 35, 60 33, 62 29, 64 29, 66 37, 68 35, 70 29, 72 31, 74 37, 76 35, 78 33, 80 29, 82 27, 84 15, 86 33, 88 35, 90 33, 92 31, 94 33, 96 31, 98 5, 100 3; Left before the fifty-move rule: 0 100, 2 98, 4 96, 6 94, 8 92, 10 90, 12 88, 14 86, 16 84, 18 82, 20 80, 22 78, 24 76, 26 74, 28 72, 30 70, 32 68, 34 66, 36 64, 38 62, 40 60, 42 58, 44 56, 46 54, 48 52, 50 50, 52 48, 54 46, 56 44, 58 42, 60 40, 62 38, 64 36, 66 34, 68 32, 70 30, 72 28, 74 26, 76 24, 78 22, 80 20, 82 18, 84 16, 86 14, 88 12, 90 10, 92 8, 94 6, 96 4, 98 2, 100 0.
The diagram shows both numbers for every position with Black to move. The needed distance hardly changes over the game. After the early exchange of errors it mostly stays between twenty and forty half-moves: the moves Black played did not bring the pawn move any closer. The budget falls by two with every full move. After sixty-six half-moves the two lines cross for the first time, and from then on the needed distance lies above the budget almost everywhere. Only twice, after seventy and after eighty-four half-moves, does it dip under the budget for a single position: a White move had handed the win back, and Black's next move gave it away again. Apart from that, it falls steeply only in the last two positions, when two half-moves or none are left. In the final position the tables, which ignore the counter, still call the board a win for Black. Under the rule, the win had been gone for most of the last thirty-five half-moves.
What a one-move test sees
The project's regression battery contains this position as one of its lines. Its expected value comes from the tables: the first move must be one of Rd7, Kf7 and Kf8. The engine's Kf7 passes. In a full battery run on 2026-10-02, the main version of that day passed all thirty quota lines of the endgame group.
The game was lost later, through a plan that never came. A line that checks one move cannot see that, and it is not meant to. It catches a single wrong step, not a missing direction. Long endgames like this one are what the full games of the endgame battery are for.
What this does not say
This is one game of one version, played at a fixed node budget and without tablebases. It is not a measurement of playing strength, and it says nothing about other positions or longer thinking times. Why the search kept moving king and rook instead of the pawn was not examined. This entry describes what happened on the board, not the cause inside the engine. Whether later versions play this position better is not part of this entry.