dydactaccelerating futures

Sudoku Solver

op: sudoku.solve

Constraint propagation + branching. Same compiler scales N = 4 → 25. 100% extreme 9×9 @ 30.6 ms — beats Kona 96.2% @ 313 ms / 36.5M params. Zero training.

grid size
paste puzzle
curated
substrate-native — no search, no annealing
Why this works
Each empty cell starts with all N candidate digits. Row, column, and box constraints prune impossible values to a fixed point — naked singles, hidden singles, locked candidates, naked pairs, X-wing all fall out of the same propagation operator at different scopes. No WalkSAT, no annealing, no discrete flips. When propagation plateaus we branch on the most-constrained cell and recurse with bounded depth.