Oliver Chen developed PersistBench for the UKGovernmentBEIS/inspect_evals repository, focusing on evaluating long-term memory risks in large language models. He implemented Python-based metrics to assess cross-domain leakage, sycophancy, and beneficial memory usage, enabling comprehensive risk analysis across deployments. Oliver integrated these features into the existing evaluation workflow, introducing a formal results structure and versioning to support robust, repeatable assessments. His work included updating documentation, improving test coverage, and refining repository hygiene through targeted maintenance. Leveraging skills in AI evaluation, data analysis, and software testing, Oliver delivered a well-structured, maintainable solution that addressed nuanced challenges in LLM risk evaluation.