✦ Watch for the golden snitch. Catch it for +1,000.
A shadow AI runs this course with you. Beat it.
Grab a rocket for a jetpack boost.
On iOS you'll get a one-time permission prompt for motion.
Game Over
Score: 0 · Best: 0
Shadow AI is on this course.
New high score! Enter initials:
Version …
Training dashboard · charts an RL engineer would watch
● current | version 1▲▼ nudged this rungreen = moved by the last update% = change from version 1 · tap a row for details
Reinforcement learning terms used here
Policy
The rule for choosing settings: the current game (its "center") plus a small random nudge.
Action
The exact settings one run was given.
Reward
The single number the game maximizes: 0.55 × "hit Play again" + 0.35 × how long you played (full credit at 90s) + 0.10 × the share of golden snitches you caught this run.
Exploration
Trying slightly different settings (the ▲▼ nudges) so the game can discover better ones.
Baseline
A running average of recent rewards: what a typical run earns.
Advantage
Reward minus baseline. Positive pulls the game toward that run's nudges; negative pushes it away.
Policy gradient
Nudge, measure, step toward what scored better. Here, at most 5% of a factor's step per run.
Mirrored pairs
The next player in your group gets the opposite nudges, so luck cancels out faster.
Contextual bandit
A separate learned game per context: phone (tilt) vs keyboard players.
Online learning
The game updates after every run, from every player, instead of in big batches.
Safety rails
Hard limits on every factor, plus bots that re-test the game every 50 runs and roll back anything unplayable.