A preprint finds AI token rankings can be stable yet numerically inaccurate, while reference preparation changes the uncertainty target they explain.