The ScienceIDE initiative, a collaborative effort by researchers from 20 premier universities worldwide and spearheaded by the AItonomy Foundation, is designed to incorporate research code, execution settings, and scientific acceptance criteria into tasks that AI can perform, validate, and learn from. By using real scientific outcomes to guide AI behavior, the project investigates how scientific knowledge can shape the training of next-generation AI systems. Studies reveal that out of 1,769 tasks with complete records, 55.2% of the programs could be compiled and run without errors but generated incorrect scientific results. This misleading sense of 'success' can easily deceive AI. While general-purpose code agents benefit from low-cost trial-and-error feedback mechanisms, research code demands additional validation of scientific results and lacks universally accepted validation standards. ScienceIDE dissects research code repositories into 'environments' that encompass execution settings, examples, and validation procedures. Tasks are created through a 'task factory' and refined through controlled experiments to ensure their effectiveness, enabling AI to receive genuine feedback in a closed loop of 'execution, error-making, and correction.' These trajectories can be utilized for supervised fine-tuning or reinforcement learning. As of September 16, 2026, ScienceIDE encompasses 64 environment definitions sourced from 27 open-source code repositories and has generated 2,812 tasks across 14 disciplines. At present, the project has developed model training interfaces, but further validation is necessary to ascertain whether the model simply memorizes specific techniques or genuinely acquires general research abilities. This project offers a practical starting point for the PhAI Labs' Foundational Model Framework for Discovery, acting as a public foundation for large model development while converting the implicit experience of research teams into AI training resources. It explores the possibility of continuously transforming scientific experience into AI learning materials to fuel AI evolution.