Developed a robustness upgrade for the xlang-ai/OSWorld evaluation framework, introducing an explicit FAIL action to handle infeasible GPT-5.4 responses. This enhancement improved the reliability and accuracy of automated scoring by ensuring that infeasible outputs are clearly flagged and propagated throughout the evaluation workflow. The work included implementing a regression test to validate the new FAIL signaling, increasing test coverage and reducing the risk of false positives or negatives in production. Leveraging Python and applying skills in AI development, software engineering, and unit testing, the developer collaborated closely with peers to maintain code quality and prompt feedback throughout the process.