Evaluating Explanatory Artefacts of DSAR-Recovered Software Architectures from Industrial Codebases

Abstract

Understanding large, complex software architectures is difficult and time-consuming. Software architecture recovery aims to extract explanatory artefacts from code, but existing methods often lack generalisability. To address this, we previously proposed Deductive Software Architecture Recovery (DSAR) to extend current techniques. This paper presents a qualitative evaluation of DSAR in an industrial setting. Using a large language model-assisted prototype, we examined the recovered architecture's content quality, presentation quality, and usefulness for understanding and maintaining real-world systems. To ensure practical relevance, we conducted an in-depth judgement study with software engineering teams at ASML applying DSAR to 17 Java repositories across diverse teams. We contribute: an empirical assessment indicating that DSAR supports comprehension, architectural reasoning, and communication; qualitative evidence on the strengths and limitations of the generated explanatory artefacts; and recommendations for practitioners and researchers on adopting and further developing DSAR for architectural sense-making.