This is an open-source system for training language models to solve machine learning engineering tasks (like optimizing hyperparameters or fixing broken code) by iterating through four learned operators: Draft, Improve, Debug, and Crossover. The 35B Frontis-MA1 model, trained on execution-verified code trajectories, reaches 71.21% on MLE-Bench Lite under a 12-hour single-GPU budget—approaching frontier models like GPT-5.6 and Kimi K3 that are orders of magnitude larger. The key claim is that by training the model on verified execution traces and then using those same operators for test-time search, learning and evolution close into a single loop. The authors release model weights, task environments, sandbox infrastructure, and training code. Transfer experiments on held-out NatureBench show both the trained model and the search algorithm generalize beyond the training distribution. The source doesn't detail computational cost comparisons or failure modes, only the headline metrics.