The paper introduces Golden Goose, a method designed to synthesize an unlimited number of Reinforcement Learning with Verifiable Rewards (RLVR) tasks from unverifiable text found on the internet. The challenge in reinforcement learning has been the limited availability of verifiable data, which hampers model performance as training continues. Golden Goose addresses this by transforming source text into multiple-choice questions, identifying key reasoning steps, and generating plausible distractors. This approach allows researchers to tap into rich, reasoning-heavy texts that were previously deemed unusable for RLVR data, such as science textbooks. From this process, the authors created GooseReason-0.7M, a dataset containing over 700,000 tasks across mathematics, programming, and general science. They demonstrate that training language models on this dataset can effectively revive performance in models that have reached saturation with existing RLVR data. Notably, the paper reports state-of-the-art results for 1.5B and 4B-Instruct models across 15 benchmarks, showcasing the method's effectiveness. Furthermore, the authors applied Golden Goose in a practical setting by generating RLVR tasks from FineWeb scrapes related to cybersecurity, a field where prior data was nonexistent. Training the Qwen3-4B-Instruct model on this new dataset, named GooseReason-Cyber, resulted in superior performance compared to a 7B domain-specific model, reinforcing the potential of this approach to scale RLVR data using readily available internet text.