Asked about Grading with a model

When would you use pairwise comparison rather than scoring?

An answer that works

When comparing two versions of a system on open-ended quality, such as which reply is more helpful. Asking which is better is more reliable than asking for an absolute score, because the judge does not need a calibrated scale. For specific failure modes, a pass or fail check per output is simpler and tells you more.

Reading an answer is not being able to give one. The section that teaches this is Grading with a model, 6 hours of LLM-as-judge.

Learn this properly

Also asked about LLM-as-judge

How do you know your LLM judge is any good?What are the biases of an LLM judge?Should the judge explain its reasoning?How would you monitor answer quality in production with an LLM judge?Your judge marks 80% of traffic as passing. Is the pass rate 80%?