laplace-engine

The deterministic rules engine for LAPLACE — an 8x8, four-color, 2-vs-2 strategy board game — packaged as the frozen referee for LaplaceBench.

Ruleset freeze

This package's version line IS the ruleset ID. [email protected] implements laplace-8x8-v1:

• 8x8 board; colors Red(1), Blue(2), Yellow(3), Green(4); teams A = Red+Yellow, B = Blue+Green; turn order R->B->Y->G, skipping colors with no pieces left.
• Rook-like movement, no jumping, empty destination.
• Sandwich captures (same-color flanks, mixed-color lines captured whole, friendly fire real, voluntary entry safe) and enclosure captures (zero-liberty groups adjacent to the landing square).
• Elimination at 3 pieces lost (fixed, board-size independent); survivors become Void pieces (move, can be captured, never capture).
• Victory by eliminating both enemy colors or occupying the four center cells with one team's pieces (Void included).

Any future rule change ships as a new major version = new ruleset ID. Old benchmark results stay interpretable forever.

The engine is a verbatim freeze of the LAPLACE product engine (@laplace/game-shared) as of 2026-07-21, including the canonical elimination-threshold rule (fixed 3). It has zero runtime dependencies.

API sketch

import { GameStateManager, TakeshiPolicy } from "laplace-engine";

const game = new GameStateManager();
game.startGame(8);
const res = game.makeMove(0, 3, 4, 3); // fromRow, fromCol, toRow, toCol
// res.valid, res.state (captures, eliminations, winner all applied)

GameStateManager is the referee: it alone decides legality, captures, Void conversion, elimination, turn order, and victory. TakeshiPolicy is a team-aware alpha-beta baseline (a policy, not a rules oracle).

Human-readable rules

The English rulebook given to models is rulebook/laplace-8x8-v1.md in the LaplaceBench CLI package.