Field Notes.
We fixed the bass. The soundtrack still didn't have a tune.
A bass-heavy drone earned four replacement songs and a vote. Two days after the winner shipped, the owner still wasn't happy — and measuring it properly found the bass had barely been the issue at all.
The music wasn't laggy. It flinched every time you landed.
Frame times with the music on and off were identical. Whatever was wrong, it wasn't the engine working harder.
For an online leaderboard, two browsers have to agree on the same bounce down to the last decimal
A score nobody has to take your word for means a server has to be able to replay your exact run and land on your exact number. Two honest browsers don't always do that by default.
An all-time board with no accounts has one honest rule: one name, one spot
Nothing stops a player from posting under the same name twice except a server deciding not to let the second run sit next to the first.
The old soundtrack wasn't hissy. It was out of tune.
"Sounds like noise" has more than one honest cause. The one that looked obvious on a meter wasn't the one that was actually true.
Eight corners, and 2.5% of pieces had nowhere to land
The planet's one geometric flaw, eight corners where three squares meet instead of four, turned into pieces freezing in mid-air over gaps that didn't exist. The fix cost the hardest difficulty a third of its survival time. We measured that cost before deciding it was worth paying.
A piece that had already landed kept sliding. The report came back within the hour.
Fixing one bug can write another. The first attempt at the corner fix quietly broke a much more basic promise, that where you land is where you stay, and a test suite that only checked whether a piece was supported never once asked whether it moved after it didn't need to.
The browser can say anything happened. We only believe our own replay.
A browser-side Top 10 is one line of code and completely worthless, nothing stops a client from just claiming a bigger number. The fix here isn't a smarter check. It's not trusting the browser's score at all.
We built a planet out of a cube, and it only has eight problems
An icosphere looked like the obvious shape for a planet. It fell apart the moment real pieces had to land on it. A cube wrapped onto a sphere turned out to have exactly eight weak points — and all eight were fixable.
We tried to out-climb a good player with a harder mountain. It didn't work.
Clearing the Summit used to be where a run ended. Now a player can keep going — and the thing that finally stops a strong run isn't a steeper floor. It's the camera.
A run that drafted 11,000 perks crashed the build before it crashed the player
The fix for a run that technically never ends usually isn't a bigger number. Sometimes it's refusing to offer the same card twice.
RUSH was free speed. We had to make it cost something.
A climbing boost that only ever helps doesn't change how you play — it just makes the run shorter. The lever that fixed it wasn't the boost itself. It was how long a floor stays underfoot.
Why every kill-cam replay looked the same (and how we fixed it)
The old camera was technically showing the real, physically simulated bullet path. It just couldn't show the one thing that made it worth watching: the curve.
A "perfect" shot only landed 43% of the time — at 30 fps
A bullet doesn't move in a straight line between two rendered frames. On a slow computer, that gap is exactly where a mathematically correct shot can go missing.
Wind isn't a number. It's a field.
Most games hand you one wind value for the whole level and call it realistic. Here, the same shot from the same spot can land differently ten metres to the left — because the wind ten metres to the left is a different wind.
Our bots got worse when we let them learn
We gave the sniper bots a way to correct their aim after every miss. The hardest difficulty got worse, not better. The reason is the same reason a real shooter with a shaking hand can't fix a bad zero by trying harder.