Worked on stabilizing resource usage within the UKGovernmentBEIS/inspect_evals repository by delivering a memory governance feature for the agent sandbox. Focused on system configuration using Docker Compose and YAML, the work introduced a consistent 2GB memory limit across all challenge configurations, including cybench challenge compose files. This approach addressed memory-related outages and improved the reliability of evaluation environments by reducing downtime and variability in test runs. The implementation ensured uniform resource governance, enhancing the predictability and stability of automated evaluations. The contribution reflects a targeted, infrastructure-focused effort to optimize system performance and maintain consistent operational standards across the project.