4 min read
A move budget is what turned placement into a decision
Measurement said no move in the game could lose a level: decision width was 100% everywhere. Charging for every move dropped it to 62% without making a single level impossible.
Devlog
How the difficulty is measured
Notes on building a puzzle whose levels are proven rather than guessed. Every number here comes from a tool in the repository that anyone can re-run.
Four posts, published together as one series.
4 min read
Beam search is not monotonic, and it corrupted my measurements
The same level at the same budget: beam 45 solves it, beam 90 fails, beam 200 solves it again. Any question asked with a single beam width produces false negatives.
4 min read
Measuring puzzle difficulty with bots
Six players play every level and disagree on purpose. A dumb bot winning means there is no puzzle; a good bot losing does not by itself mean there is depth.
5 min read
Proving every level is solvable before it ships
Four of ten hard levels could not be won at all. The fix was a generate-solve-filter pipeline that plays every level offline and refuses to publish the ones a search cannot beat.