Whether your human graders agree with each other. If two competent people looking at the same output give different scores, the measurement is unstable and no amount of averaging fixes the underlying ambiguity. Low agreement usually means the rubric is unclear rather than that the raters are careless.
Technical view
Measured by Cohen’s kappa for two raters, Fleiss’ kappa for more, or Krippendorff’s alpha for mixed data types, all correcting for chance agreement. Values below about 0.6 indicate the task definition needs work. It sets a ceiling on achievable model performance as measured, since a model cannot exceed the reliability of the labels. It should also be reported when validating a model judge against human raters.
Related
Part of AI Terminology Index, section: Measuring Performance.