Trusta: Reasoning about assurance cases with formal methods and large language models

Abstract

Assurance cases can be used to argue for the safety of products in safety engineering. In safety-critical areas, the construction of assurance cases is indispensable. We introduce the Trustworthiness Derivation Tree Analyzer (Trusta), a tool designed to enhance the development and evaluation of assurance cases by integrating formal methods and large language models (LLMs). The tool incorporates a Prolog interpreter and solvers like Z3 and MONA to handle various constraint types, enhancing the precision and efficiency of assurance case assessment. Beyond traditional formal methods, Trusta harnesses the power of LLMs including ChatGPT-3.5, ChatGPT-4, and PaLM 2, assisting humans in the development of assurance cases and the writing of formal constraints. Our evaluation, through qualitative and quantitative analyses, shows Trusta's impact on improving assurance case quality and efficiency. Trusta enables junior engineers to reach the skill level of experienced safety experts, narrowing the expertise gap and greatly benefiting those with limited experience. Case studies, including automated guided vehicles (AGVs), demonstrate Trusta's effectiveness in identifying subtle issues and improving the overall trustworthiness of complex systems.