A model trained on post-issue GitHub commits could solv...

To prevent models from solving benchmarks via recall of training data, you must implement timestamp-based filtering to ensure the model’s training data does not include the solutions. Furthermore, you should maintain hold-out repositories that have never been indexed in any known training corpus, ensuring that the model must rely on reasoning rather than memorized commits.