Real-world video doesn't capture the rare, risky moments that matter most for safety, so physical AI world models end up blind where mistakes cost the most — synthetic data fills those gaps.