Puzzle Design posts from the GameDavinci studio — news, dev logs, and game recommendations.
Inkfield's fairness assertions all passed while an attentive player making zero mistakes died 23 rounds out of 24. The death rate alone said only "this is hard". Splitting it by cause — edge 0, self 0, rival 23 — named the problem in one line, and the fix was arithmetic rather than tuning.
Kingsroad's reference build never constructs a Barracks and never collects a gold crate. That is deliberate: it turns the guarantee from "beatable if you play well" into "beatable even if you ignore both of these", and it converts two mechanics from hopefully-balanced into provably-upside. The exclusion is not a comment. It is two assertions.
Kingsroad proves every map beatable before showing it to you, with a three-rung ladder that repairs a map that fails. Then we read the diagnostics: repairs 6, trims 12, health scaled to 0.43. Every rung firing on the same map means none of them is aimed at the actual problem — and the bottom rung was hiding a generator bug by making the game 57% easier.
Star Grid's generator threw away 99.5% of the boards it made, hunting for one with a unique solution. Instead of generating more boards, we started repairing the ones we had: enumerate the rival solution, then make a single edit that provably kills the rival and provably cannot kill the answer. Yield went from 0.5% to 100%, at 2.9 to 29.8 milliseconds a board.
Building difficulty bands for Bubble Shooter, we used a reference solver to estimate par. It reported par 7 for boards a real player needs thirty-odd shots to clear, and a shot budget built on that number would have made the game unwinnable. The solver was not buggy. It was playing a different game — one where you get to choose which colour comes next.
We wanted Tile Order to show a proven fewest-moves number, not a guess. On a 4x4 that meant proving optimality, and a fully shuffled board cost up to 22 seconds to prove — unusable in a browser. The fix was not a faster solver. It was generating boards the solver could afford to prove, and being honest in the copy about what that costs.
Our first difficulty rater asked "did this technique fire while solving?" One technique fired on a median of one board in one, and almost no board actually needed it. Rating by necessity instead — is the board still solvable with the technique switched off? — produced a ladder we could trust, and two findings we could not have guessed.
"Every row is passable" sounds like a simple promise. It took five wrong versions of it to keep, and the interesting part is that the road generator was never the thing at fault — the proof was. Including one bug that did not break fairness at all, and so survived scrutiny by making the game impossible to lose.
Most puzzle games ship a bank of pre-computed levels because their solver is too slow to run while you play. We found a heuristic that made ours run per move, which deleted the bank, the build step and the shipped asset — and then let us put something on screen no other ball sort game has: the shortest finish still available from where you are.
We abandoned Kakuro once. Fill-then-test found a uniquely solvable board 0% of the time; the fix we wrote down was "find a construction that forces uniqueness". That framing was the trap. Aiming at a cheaper property — a board a guessing-free solver can finish — got us to 100% yield in milliseconds, and it implies uniqueness anyway.
Before building Arrow Escape we tried to work out how hard the genre could get. The answer was: not very, and not for want of trying. Five levers, all neutralised, and a ceiling that comes straight from the rule every one of these games shares. Here is the measurement, including the part we got wrong.
