Fifteen papers covering the arc from G-Eval to JudgeBench and the 2026 work on judge error. G-Eval and MT-Bench established that a strong LLM can rival a human rater. Prometheus, Auto-J, and JudgeLM proved that open models can do the job too. Shankar et al. surfaced criteria drift, which is the reason most judge pipelines stop working after launch. Judging the Judges shows that percent agreement overstates judge quality. JudgeBench then scores judges against objective labels on hard pairs: a vanilla GPT-4o judge lands near 51%, and reasoning models do much better. Two of the 2026 papers are about what to do with an imperfect judge. Feng et al. calibrate a judge on a small human-labelled set so failure-rate tests keep valid error control. Kohli finds that nine frontier judges carry about two independent votes, so adding judges does not fix correlated errors. The other two, Li et al. and Rao and Callison-Burch, test Jev as a cheap first-pass judge and find that an LLM judge repeats its most confident errors, which decision models as judges turns into rules for a cascade.