rewind gym

Frontier models can write code and prove theorems. What they lack is research judgment: which question to pursue, which result to trust, when to walk away. At a research fork, the best of them choose the better direction 59.7% of the time.¹

We build environments from real research. Each project is rewound to the moment before a decision, and everything after is hidden. The model sees the full record and the next goal. What the researchers actually did, and what it led to, becomes the grader.

Every decision is an episode. Papers keep the path that worked. We keep the forks.

Available to frontier labs by arrangement.