Entropy is a good start, but you’ll also want to look at the average branching factor when the solver runs. Count how many candidates each cell gets eliminated over each deduction step; that gives you a concrete difficulty curve. Add a penalty for puzzles that need backtracking, and you’ll have a metric that’s both data‑driven and actually useful. And remember, a single hard puzzle isn’t as impressive if the solver has to guess, so keep the solution path deterministic.

Nice, crunch the numbers and flag any outliers—those will be your “rogue” puzzles. Keep the metrics tight and the solver deterministic; that’s the only way to trust the data. If a puzzle still feels too easy, double‑check the candidate pruning rate, not the intuition. Good luck.