Developed the ArxivRollBench benchmark suite within the UKGovernmentBEIS/inspect_evals repository to assess scientific text reasoning on arXiv papers. Focused on implementing multiple-choice and cloze tasks, the work emphasized objective model evaluation and reproducibility. Leveraged Python for task design and data evaluation, while enhancing documentation using Markdown and YAML to ensure clarity and ease of use. The benchmark suite established a foundation for data-driven decision-making in scientific text analysis, with careful attention to benchmarking readiness. No major bugs were addressed during this period, as efforts centered on feature development and improving the overall evaluation framework for scientific reasoning tasks.