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.
First seen on securityboulevard.com
Jump to article: securityboulevard.com/2026/03/synthetic-data-is-all-you-need-for-reinforcement-learning/
