We build environments and evals for frontier models.

Task and grader quality is the crux of everything. A weak task teaches nothing, and a grader with a hole in it teaches the model to find the hole.

Every environment starts as research. We find where frontier models fail, confirm the failure is real and repeatable, then build tasks and graders around it. Graders check end state, not the transcript, and each one is red-teamed for reward hacking before it ships.

Our work

Game development Where models breakAnimation and first-person hands, timing and pacing, difficulty and balance, and iterating on a large project without breaking what already works. How we gradeThe game runs headless. Graders read scene and game state directly, never screenshots or the agent's own summary. ProjectsReal game codebases across several engines, not toy prototypes. Models one-shot a convincing prototype, then fall apart on what makes a game feel right. Our environments are built on real game projects and graded by running the game.

Manufacturing and construction SystemsTwenty-year-old ERPs, construction drawings, bills of materials, change orders, procurement and pricing records. TasksLong-horizon work that spans several systems and weeks of history, built from real operating data and anonymized with our partners. How we gradeAgainst what happened: the order that shipped, the price that cleared, the change that was approved. Long-horizon work inside the systems these industries run on, built from real operating data.

Domain partners

What counts as correct is defined by people who have done the work for decades. Domain experts design the tasks with us, review the graders, and share in the upside for as long as labs use what we build together.

Team

Rebelot is founded by Christophe Barre Christophe Barre Co-founder · research and lab partnerships Previously co-founded Deeploop, an applied AI company backed by Y Combinator (S20). Started his career as a quant on Wall Street. Stanford and Columbia. LinkedIn and Manuel Darcemont Manuel Darcemont Co-founder · environments and graders Founded and sold Scribbay. Then CTO of MadKudu, a machine learning company for B2B go-to-market teams. LinkedIn · X .

Research