rules.md

Board Gauntlet rules

Two cogs, one board, no dice and nothing hidden. Both seats see the whole position, the whole move history and the whole legal-move set. What varies is the board itself.

The rotation

config.game is one of rotate, connect-four, breakthrough, hex, quoridor. The rotation order is the fixed list ["connect-four", "breakthrough", "hex", "quoridor"], indices 0..3. With game: "rotate" the server resolves game = RotationOrder[((seed mod 4) + 4) mod 4] before the first move and records the resolved id in the replay, in results.game, in the player welcome frame and in the first line of both seats' prompts. If seed is 0 (unpinned) the server draws one, echoes it and writes it into the config, so every replay carries a concrete seed and re-derives exactly. Nothing about the rotation is hidden from a seat, and because the draw is a pure function of the recorded seed nobody can be steered into a game either.

Coordinates

Cells are algebraic: files a, b, c, ... left to right; ranks 1, 2, 3, ... bottom to top. a1 is bottom-left. Nothing spectator-facing ever shows an internal index. Seat 0 always moves first; the league's round robin seats each policy in both slots, which balances the opening advantage without a swap rule.

connect-four (7 files x 6 ranks)

Seat 0 plays red discs, seat 1 blue.

1. A move is a file letter a..g. It is legal iff that file has at least one empty cell.
2. Applying it drops the mover's disc into the lowest empty cell of that file.
3. The mover wins immediately if the placed disc completes a line of four of its own discs horizontally, vertically or on either diagonal.
4. If after the move every cell is occupied and nobody has four in a row, the game is a draw (ending = "board-full").
5. There is no other way for this game to end other than the ply cap or the wall clock.

breakthrough (6 x 6)

Seat 0 is red, with 12 pieces filling ranks 1 and 2; seat 1 is blue, with 12 filling ranks 5 and 6. Red advances toward rank 6, blue toward rank 1.

1. A move is <from><to>, e.g. b2-c3. from must hold one of the mover's pieces; to must be one rank forward of from and on the same file or one file to either side.
2. A straight-forward move is legal only onto an empty cell. There is no straight capture.
3. A diagonal-forward move is legal onto an empty cell or onto a cell holding an opponent piece, which is then removed.
4. The mover wins immediately if the moved piece lands on the opponent's home rank (ending = "home-rank").
5. The mover wins immediately if that move removed the opponent's last piece (ending = "no-pieces").
6. If the seat now to move has no legal move, it loses (ending = "no-moves").
7. There are no draws by rule; the ply cap and the wall clock adjudicate.

hex (7 x 7 rhombus)

Seat 0 is red, seat 1 blue.

1. Cell (r, c) is adjacent to (r, c-1), (r, c+1), (r-1, c), (r+1, c), (r-1, c+1) and (r+1, c-1). So c4 touches b4, d4, c3, c5, b5, d3, and the corner a1 touches only b1, a2.
2. A move is a cell name. It is legal iff the cell is empty.
3. Applying it places the mover's stone there; stones never move and are never removed.
4. Seat 0 connects the left file (a) to the right file (g); seat 1 connects the bottom rank (1) to the top rank (7). The mover wins immediately when its own stones form an unbroken adjacency chain between its two edges (ending = "connection").
5. Hex has no draws: a full board contains exactly one winning connection. There is no swap (pie) rule.

quoridor (9 x 9, 10 walls a side)

Seat 0's pawn starts on e1 and must reach rank 9; seat 1's pawn starts on e9 and must reach rank 1. Each seat holds 10 walls.

1. A move is either a pawn move (the destination cell, e.g. e2) or a wall placement, <anchor><h|v> with anchor in a1..h8, e.g. e3h.
2. A wall's anchor names the lattice point at the top-right corner of the anchor cell: xNh blocks the two vertical steps (x,N)<->(x,N+1) and (x+1,N)<->(x+1,N+1); xNv blocks the two horizontal steps (x,N)<->(x+1,N) and (x,N+1)<->(x+1,N+1).
3. A wall placement is legal iff all of: the mover has walls left; no wall of either orientation already sits on that anchor; neither of the two steps it blocks is already blocked; and, after the placement, both pawns still have some path to their own goal rank.
4. A pawn moves to an orthogonally adjacent cell when that step is not wall-blocked and the cell is empty. If the adjacent cell in direction d holds the opponent pawn, the mover may jump straight over to the next cell in direction d when that cell is on the board and the step from the opponent's cell in direction d is not wall-blocked; if and only if that straight jump is unavailable, the mover may instead move to either cell diagonally adjacent to the opponent pawn, perpendicular to d, provided that step is not wall-blocked and the cell is on the board.
5. The mover wins immediately on moving its pawn onto any cell of its own goal rank (ending = "goal-row").
6. Rule 3's path invariant guarantees a pawn always has a legal move, so no-moves is unreachable here.

Move notation, tolerantly read

A reply's move is lower-cased, trimmed and stripped of every character outside a-z0-9, then read per game: connect-four takes a file letter a..g or a digit 1..7 mapping to a..g; hex takes ^[a-g][1-7]$; breakthrough takes ^[a-f][1-6][a-f][1-6]$ and writes b2-c3; quoridor takes ^[a-i][1-9]$ for a pawn move and ^[a-h][1-8][hv]$ for a wall. Anything else is an invalid reply: the seat gets ONE retry carrying the printed legal-move list, and then falls back to the tactician baseline.

standing, and the eval bar

standing(sim, seat) is an integer, higher is better for that seat, computed from the true board and recomputed after every ply. It drives the viewer's eval bar, it adjudicates the ply-cap and wall-clock endings, and the tactician baseline maximises it.

• connect-four: over all 69 four-cell windows, sum w[k] for every window that contains no opponent disc, where k is the mover's disc count in that window and w = [0, 1, 4, 16, 10000].
• breakthrough: 100 x pieces + 10 x sum(advance) + 40 x max(advance).
• hex: 1000 - 10 x distToWin, where distToWin is a 0-1 BFS from the seat's source edge to its target edge in which a cell the seat owns costs 0, an empty cell costs 1 and an opponent cell is impassable; unreachable is 99.
• quoridor: 1000 - 10 x dist + 2 x wallsLeft, where dist is the BFS pawn-step distance to the nearest cell of the goal rank over the wall-blocked graph.

The eval bar is THIS repository's own heuristic, not an engine evaluation, and the viewer labels it HEURISTIC in the bar's own caption. No OpenSpiel, no external engine and no reference bot is used anywhere.

Scoring

score_i = +1  if seat i won
           0  on a draw
          -1  if seat i lost          score_0 + score_1 = 0, always

Higher is better and the league ranks by mean episode score. Everything richer - captures, walls used, final standing, plies, fallbacks - is reported for the audience and the audit and is deliberately NOT scored, so no policy can farm the metric instead of winning.

Who won, by ending

results.reason has exactly two legal values, "complete" and "deadline"; the finer ending rides in results.ending. A deadline result is a REAL result: the position is fully scored by standing at the stop.

Anti-collusion

Two seats, zero-sum, so there is no coalition to form. The rotation is announced to both seats before move one and is a pure function of the recorded seed. Seats play under anonymous cog aliases; no prompt, no player frame and no in-game text ever contains a policy display name. say is spectator-facing only and is never shown to the opponent; notes are private and fed back only to their author; every decision is made inside the game server from a view the server builds, so there is no wire on which two seats could coordinate.