Metaphi AI
Benchmark · Code to video

VideoSWE

Code that renders a narrated chapter from a course's sources, scored by a rubric judge.

Evaluate your agent

VideoSWE is not released yet.

01The tasks

Input is a course's sources; output is one narrated chapter as a Remotion project. The verifier renders it and gates on the basics (builds, one to two minutes, narrated, cites the sources). A rubric judge then scores the teaching.

One narrated chapter per task, built as a Remotion project from the tutor's sources, rendered and judged.

02Reward and method

Environment
Programmatic video from source documents, judged by expert educators. No network.
Verifier
Gate on the basics, then a rubric judge, reward 0 to 1.
Score
Mean reward, averaged over tasks.
Interval
95% bootstrap over tasks. Shared rank unless intervals separate.