What's this about?
The harness keeps the visual program, its revision history and rendered evidence together so a multimodal model can inspect and repair composition, appearance and motion.
What exactly is the program-to-visual gap?
Code can run without errors yet render the wrong picture. Execution success proves syntax and runtime validity, not that the output satisfies the visual request.
Why does this matter?
4 HF upvotes. Even GPT-6 Astra reached 100% execution but only 76.9% full video-quality compliance, exposing what general coding scores miss.