Six agent problems after the Generative Design 02_M sketches. Each is a card Thompson ran: two model rollouts, a critic inside each attempt, and a judge that runs the world and keeps only what scores higher.
Every frame below is computed live in your browser from the same code the judge runs. On the left, the agent we planted. On the right, the agent Thompson kept. Both run in the same world (fixed seed 1); the scores in each heading are means over the three fixed seeds.
Each problem's world, its check and its tests. People tend it; no card may touch it.
One small file per problem: the agent. This is the only thing a card is allowed to change.
Every prompt, reply, critique, diff and verdict of a run, kept whole. You merge by hand after reading it.