RIGGED

A daily physics game inside one Reddit post, built for Reddit’s Games with a Hook hackathon. Each day a scene rotates in with obstacles and a goal. You build a contraption from 11 pieces, simulate it, and publish it. Everyone else sees your rig frozen — before it runs — and calls it WILL WORK or WILL FLOP. Then it runs.

The scoring is what makes it a game rather than a puzzle. A builder does not score by solving cleanly; a builder scores by surprise rate, the share of predictors the rig caught out. A contraption that looks doomed and works is worth more than one that looks inevitable.

Determinism was the whole problem

Every reveal is a live re-simulation on the viewer’s device, not a recorded video. So the same rig has to produce the same outcome everywhere, or the game is a lie.

Two “identical” runs can diverge on something as small as body insertion order. The fix was to stop trusting the frame loop: the world is stepped manually at a fixed 1000 / 60, and a Node harness runs Phaser’s own bundled Matter build to pre-simulate bot rigs and check that every authored level is solvable.

It still was not enough for every level. Three of the bell levels are pulled from the 24-day rotation because their winning rigs clear the dome by under a pixel while the ball is still moving — any drift between the bake and the reveal flips WORK into FLOP. The levels that stayed are settle goals, where the ball comes to rest and both engines converge on the same answer.

I would rather cut three levels than ship a game that occasionally lies about who read it right.

User content without a text box

Players can publish their own challenges, and none of it is free text. Inputs come from presets, the server re-validates every submission, and a challenge only goes live after its author has solved it themselves. Moderation problems you never create are the cheapest kind to handle.

What I got wrong

The screens were built from the prototype’s code and business logic instead of from the rendered prototype. That sounds like a detail. It cost animations — not a single @keyframes block made it across — plus colours, piece detail, element positions, and a few whole flows that were only visible when the prototype was actually running.

The lesson is dull and I keep relearning it: reading the source is not the same as watching the thing run. Every screen now gets audited against the running prototype, not against its file.