Game development
Generating puzzle game levels with an agent, proving every level can be solved, and tuning difficulty: the rules we settled on and the numbers we measured.
Notes
A reverse-construction generator and the gate behind 1,100 arrow-puzzle levels
All 1,100 Arrow Knots levels come from a reverse-construction generator and pass the G2 pack gate. Fixing the move rule first made each board solvable as built.
Testers said the time was "too long, so it's easy": checking that with a simulation before cutting puzzle par times by 30%
Testers said our puzzle game gave too much time. A 158-level simulation set the cut at 30%; a 45% cut left casual players a 1% three-star rate past level 500.
Puzzle game day-one churn: per-device event timelines, exact board math and four walls in chapter one
7 of 16 ad-acquired devices stalled on board 4 of a block puzzle without a solve. Per-device event timelines and exact per-board win rates found four walls.
Block puzzle difficulty: how simulation bots caught a game you could never lose
Our block puzzle spec promised every tray could always be placed. A planning bot then played 80 games; all 80 hit the 3,000-move cap. How two bots set the fix.