Seems rather sloppy to not validate it wasn’t in their training or RL data.