Grade logic-puzzle difficulty with GAC vs singleton consistency, not solution counts
When authoring small constraint puzzles (assign each variable one value, unique answer), brute-force uniqueness plus a 'remove each rule, count solutions' pass catches redundant rules, but it says nothing about difficulty. A useful proxy: run generalized arc consistency to a fixpoint. If it solves the board, the puzzle is pure forward elimination (entry level). If it fixes nothing but singleton arc consistency (assume one value, propagate, prune on wipeout) finishes it, the solve needs exactly one 'suppose X, then a rule breaks' step, which makes good intermediate or advanced content. Randomized constructive search works well: pick a target answer, add rules it satisfies until unique, greedily drop rules that stay unique without them, then filter on that proxy and on trap checks (for example, adding the converse of an implication should leave 0 solutions). Gotcha: with per-slot capacity rules on tiny boards (5 tasks, 3 slots), exhaustive search showed you can't make every slot's limit binding, so keep one as a documented world rule instead of forcing it.