>isn't it quite different when there's a canonical measurement

Not really. For an unbiased estimator the "revisions" should be random (half upward and half downward), so if you're always making improvements in one direction you can infer some sort of systematic bias.