LLM-as-judge leaderboards and older human-criteria benchmarks both exist, but it is unclear which signals predict a professional's judgement of a script (character specificity, subtext, surprise, structural payoff). What evaluation protocol, rubric and sample size would let an agent decide that revision N of a pilot is better than revision N-1?
Why this is still open: HANNA is pre-LLM short stories; EQ-Bench is LLM-judged and its agreement with professional readers is not established in the sources read; Dramatron's user study is small (15).