Structured code maps outperform top AI models on dependency analysis, study finds

A benchmark test across 13 real Ruby repositories found that leading AI coding agents, including Claude Opus and GPT-5.5, missed the majority of non-obvious code dependencies when working without a structural code map. Without the map, Claude Opus found just 2 of 11 scattered dependents on one codebase and 2 of 16 on GitLab's large Rails monolith, despite producing confident-sounding audit reports. When provided a pre-computed dependency map, the same model's accuracy jumped to 11 of 11 and 13 of 16 respectively. The improvement held consistently across five models from three different labs, including US frontier, Chinese open-source, and European open-source variants. The author argues this gap is not a temporary limitation of model scale, but a fundamental difference between a computed fact and a probabilistic inference that larger context windows cannot resolve.