When I wrote a Sudoku solver I was disappointed to find that backtracking worked maybe too well; that is, I coded in the minimum constraints to follow the rules of the game and then added backtracking and voila, it found solutions quickly enough.

Why was I disappointed? Because initially I wanted to write a solver that inferred more, that used the inferences that I made when solving Sudoku.

Sadly once the solver worked, I kinda lost interest.

I hate backtracking myself and it should only be necessary at a certain Sudoku hardness level - every puzzle below that level of hardness should be solvable without backtracking.