Puzzles
op: puzzles.*Same continuous-superposition + orthogonal-projection compiler that crushes Sudoku, applied to other game CSPs. No discrete search, no annealing. Pick a game.
Substrate-native game solvers. Each puzzle is encoded as a continuous superposition over cell-value eigenstates; constraints become orthogonal projection operators iterated to fixed point. When projection plateaus, branch on the most-constrained cell and recurse with bounded depth.
All three games route through the unified /v1/compute endpoint (op-routed). Sudoku has its own page since it's the marquee benchmark.
Binary row/column run patterns. Each cell is filled (1) or empty (0); constraints are 'this row has consecutive runs of lengths [2,1,3]'. Projection enumerates valid placements per row + col, intersects to a fixed point.
Cells contain 1-9; each run has a target sum + the digits must be distinct. Projection enumerates distinct-digit combinations summing to target, unions surviving digits per cell.
N×N Latin square plus 'cages' with operators (+, -, ×, ÷, =). Projection: per cage enumerate digit-tuples satisfying the operator, union per cell, intersect with Latin row/col propagation.
Place N queens on N×N so no two attack. Each row is a column-eigenstate; locked queens project out conflicting columns + diagonals from every other row.
N×N grid: each row + column has digits 1..N exactly once. KenKen without cages — the cleanest test of row/col distinctness projection. Accepts partial fills.
N×N grid of distinct 1..N² where every row, column, and main diagonal sums to N(N²+1)/2. Distinctness projection + per-line sum projection.
Crossword needs a word-list to project against — coming with the corpus pipeline. Slitherlink, Hashiwokakero, and Picross-3D fit the same projection pattern and will land as their encoders ship. The point: anything that's a CSP composes into the same compiler.