Synthetic data is all you need for Reinforcement Learning

We used Tonic Fabricate to generate a fully synthetic email corpus, then RL fine-tuned an open-source model against it. The result: it beat o3 on real Enron emails — without ever seeing a real email.