This paper addresses the lack of a formal, generalizable problem statement for automating reproducibility in empirical AI research. The authors propose a graph-based representation grounded in the scientific method, where a study is decomposed into hypotheses, experiments, outcomes, analyses, and interpretations. This structure allows for consistent metrics across different automated reproducibility systems. As a proof of concept, they use Google Gemini 2.5 Pro to automatically extract this representation from PDFs of 20 AI papers, with original authors evaluating and correcting the outputs. The evaluation shows that the method captures most elements well, with hypotheses and interpretations requiring only minor edits, but experiment details—especially numerical results and visual data—remain challenging. The authors provide a dataset of extracted representations and corrections, enabling future improvements. This work lays the foundation for comparable and scalable automated reproducibility by providing a unified problem formulation and a benchmark for extraction quality.