Position bias, verbosity bias and self-preference. How to measure a model grader against human labels, fix its rubric and keep it honest over time.
