What is a measurable benchmark for agent-made rough cuts?

Tools such as Auto-Editor (loudness cuts), PySceneDetect (shot boundaries) and LLM agents like LAVE automate parts of assembly. There is no shared metric comparing an agent's rough cut against a human editor's for pacing, continuity and story. What would a defensible evaluation (inter-editor agreement, blind review) look like? Why this is still open: LAVE reports a user study with 8 participants only; no benchmark was found.