rules.md

Fog-of-War Boards rules

The board and the coordinates

Cells are named in algebraic notation: files a, b, c, ... left to right; ranks 1, 2, 3, ... bottom to top. A 5x5 board runs a1..e5. Nothing spectator-facing ever shows an internal index: the feed says plays c3, never plays 12.

Phantom Tic-Tac-Toe (mode: "phantom-ttt", size: 3). Seat 0 is X (red), seat 1 is O (blue). A seat wins by owning all three cells of one of the eight lines (3 rows, 3 columns, 2 diagonals). A full board with no line is a draw.

Dark Hex (mode: "dark-hex", size 4 or 5). An n x n rhombus of hexagons. 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). Seat 0 (red) connects the left file to the right file; seat 1 (blue) connects the bottom rank to the top rank. Hex has no draws: on a full board exactly one of those two connections exists.

Seat 0 always moves first, and config.first = 0 is recorded in the replay.

A ply is one attempt

The atomic unit is a ply: one seat naming one cell, preceded in recon-hex-5 by one sense anchor.

• Non-abrupt (abrupt: false): a ply that hits an opponent stone (a collision) does NOT end the mover's turn - the same seat moves again on the next ply, now knowing one more cell.
• Abrupt (abrupt: true): a collision ENDS the turn. The seat has spent its move to buy one fact.

A seat may never attempt a cell it already knows is occupied, nor one of its own stones, nor an off-board cell. Those are not game actions; they are invalid replies. That is also what bounds the episode: collisions by a seat are at most one per opponent stone, so plies <= cells (placements) + cells (collisions) = 2 x cells, which is exactly the maxPlies pinned per variant (18, 32, 50, 50).

Resolution order for ply p (0-based)

1. beginPly. The mover is seat 0 on ply 0; thereafter the previous mover again if the previous ply was a collision AND the variant is not abrupt, otherwise the other seat.
2. Wall-clock guard. If now + worstPlySeconds > playDeadline, settle deadline / wall-clock and stop. Checked here, BEFORE any observation is built, so the episode never stops mid-ply.
3. Build the mover's observation - its own stones, the opponent stones it has proven, its referee log, its notes, and its legal attempts (and legal sense anchors), produced by the same procs the validator applies.
4. Decide. A scripted seat is decided inline by its baseline; an LLM seat gets one call bounded by llmTimeoutSeconds.
5. Parse + legality probe on a COPY of the sim. Invalid, unparseable or timed-out gets one retry carrying the printed legal set, then falls back to the probe baseline.
6. Sense (only when sense > 0): the referee truthfully reveals the sense x sense block at the anchor to the mover only. Opponent stones become permanent knowledge; emptiness is timestamped, because it perishes.
7. Attempt. An empty cell places the mover's stone (result: "placed"); a cell holding an opponent stone places NOTHING, adds the cell to the mover's proven set and counts a probe (result: "occupied").
8. Record and emit the already-truncated say, notes and guess with the attempt event.
9. Win check (only after a placement). Hex: the mover's two edges in one component of its own stones - emit win carrying the connecting path and settle complete / connection. Tic-tac-toe: the placed cell completes one of the eight lines - emit win carrying the three cells and settle complete / line.
10. Turn transfer. A placement always flips the mover; a collision flips it only when the variant is abrupt.
11. Board-full check (tic-tac-toe only): no empty cell remains and no line was made - settle complete / board-full.
12. plies += 1; at maxPlies, settle complete / ply-cap.
13. Pace, then continue at 1.

distToWin

One number per seat, recomputed after every ply. Hex: the minimum number of ADDITIONAL cells the seat must own to complete its connection, as a 0-1 BFS from its source edge to its target edge where a cell it owns costs 0, an empty cell costs 1 and an opponent cell is impassable. Tic-tac-toe: the minimum over lines the opponent has not touched of (3 - the seat's marks in that line). Unreachable is the sentinel 99. It is computed twice, on two different boards: on the TRUE board for the scorebug, the endcard, the results and the non-terminal scoring rule, and on the seat's BELIEVED board for that seat's own prompt. A seat is never told the true value.

Scoring

score = +1 if the seat won, 0 on a draw, -1 if it lost; the array always sums to zero. Higher is better and the league ranks by mean episode score. No other field is a ranking metric: probes, discovered, guessesMade, guessAccuracy and distToWin are reported for the audit and the audience and are deliberately not scored, so no policy can farm the metric instead of winning.

Endings

results.reason has exactly two legal values, complete and deadline. The finer ending rides in a separate field, results.ending:

A deadline ending is an acceptable, fully scored result for this coworld, not a discarded one: the episode is scored at the stop by the true distance. It should nonetheless be rare - the expected episode uses under a fifth of the play budget.

recon-hex-5 is an original variant, not a port

phantom-ttt-3, dark-hex-4 and dark-hex-5 are ports of OpenSpiel's phantom_ttt, dark_hex and abrupt_dark_hex. recon-hex-5 is NOT a port of OpenSpiel rbc: it is Abrupt Dark Hex 5x5 with Reconnaissance Blind Chess's sense-then-move loop transplanted onto it, which is the only way to carry that loop without a complete chess engine. It ships as an original variant and is labelled as one everywhere.

The meta

Every decision is made by Claude acting on a per-seat policy prompt plus that seat's own view. Prompts that keep an explicit map in their notes, treat a collision as a bought certainty and immediately re-route around it, and prefer cells that both build their own chain and sit on the opponent's shortest path tend to score. The two scripted baselines - probe, a shortest-path chain builder that probes as a side effect, and sweep, a corridor walker that shifts one lane every time it is blocked - fill seats without prompts and are fieldable policies in their own right.