Founder / Mission

Haga is the verification layer physical AI labs still have to build themselves.

I started Haga because simulation-first AI is graduating from papers to products — but the physics-check layer still isn’t keeping up. World-model builders and robot teams self-report results, generated worlds ship with inconsistencies, and policies break when sim-trained assumptions meet reality. That gap is where credible evaluation should live, and right now it doesn’t.

Haga isn’t another simulator or synthetic-data generator. It’s an independent checker for generated worlds and the policies inside them — with reproducible numbers, defined thresholds, and shown failures. The name comes from wanting the one artifact every release can cite: verification evidence you didn’t generate yourself.

Why independent verification, not another platform

Most peers today sell better simulation or more realistic synthetic data. Better simulation still leaves teams grading their own tests. When a world model improves a benchmark score with training, it also changes the assumptions behind that score. An independent layer means policy performance and physics consistency are evaluated under the same fixed conditions — not chosen by the people being measured.

What Haga ships now

• Policy-stress evaluation: tiered physics-stress on robosuite tasks with reproducible seeds and Wilson confidence intervals.
• Physics-consistency checking: calibrated detectors for teleportation, anti-gravity, impulses, and interpenetration.
• World-model evidence: Physics-IQ cohort results showing real-video vs CogVideoX failure modes with paired seeds.