Leaderboard
Mean score over the 20 plates, where 100 is a shadow identical to its target, with its 95% bootstrap interval. The dashed lines are references: . Cost is the token cost of one episode.
Every final shadow
Each tile is a final shadow laid over its target: shadow on the target, target the shadow misses, shadow outside the target. The blob is the convex hull of the target. Plates are ordered from easiest to hardest; select a tile to replay its episode.