The verification layer for visual intelligence.

AI got good at code first, and not because code is easy. Code comes with a verifier attached. Run it. The tests pass or they fail. That answer is instant, free, and the same for everyone, so a model can try a million times, keep what works, and improve without waiting for a human.

Visual AI has no such answer. Is this shot cut well? Does the hand hold the cup the way a hand does? Does this 3D scene obey gravity? There is nothing to run. The answer comes from people watching, which is slow, expensive, and different every time you ask.

We build the missing verifier: a reward for visual work that is computed, repeatable, and cheap enough to run millions of times. With it, visual AI can improve itself the way coding models already do.

We publish what we learn. Our research spans agentic world models, video understanding and generation, and the verifiers that tell good work from bad - from benchmarks like AgenticVBench to case studies like handing a CVPR-winning agent a game it had never seen.

Around the team is a global artist community: film directors, anime artists, photographers, sound designers and game developers. They keep our work anchored to what people actually watch and play.

We are backed by top-tier investors, and by angels from Thinking Machines, Anthropic, Stability AI, Together AI, Dyna Robotics, and more.

We are hiring people who want to work on the hardest unsolved problem in AI alignment. If you believe the alignment problem extends far beyond language, we would like to talk.