Compare initial and self-improved solutions, then follow selected robot-control episodes across rounds. Each example is paired with its task metric and evaluation split.