Developed and integrated a cosine similarity evaluation metric into the deepinv repository to enhance reconstruction quality assessment by directly comparing reconstructed images to ground truth data. The implementation involved Python programming and metric evaluation, with careful attention to correct metric inversion and alignment with established performance semantics. Updated unit tests, documentation, and the changelog to ensure clarity, maintainability, and compatibility with continuous integration workflows. This work improved benchmarking capabilities and provided clearer evaluation signals for machine learning models within the library. The approach demonstrated a focus on robust metric design, thorough testing, and comprehensive documentation to support ongoing development and usage.