Interesting benchmark idea. I'd be curious whether the scoring separates geometry correctness from code maintainability, since generated Three.js can look right while still being hard to edit or extend.