LLMs Generate Theorem Prover Scripts for Real-Time Systems

Certifying real-time systems requires rigorous schedulability analysis, often relying on complex, manual pen-and-paper proofs that are difficult to scale and maintain. Mechanized verification using tools like PROSA/ROCQ offers a more robust alternative, but it demands substantial domain expertise and proof-engineering effort. While Large Language Models (LLMs) show promise for automating script generation, they typically lack the specialized PROSA knowledge needed for correct modeling and proof patterns. This paper introduces PROVE-RT, an LLM-assisted framework designed to generate PROSA/ROCQ scripts for mechanizing real-time systems schedulability analyses. PROVE-RT enhances LLM generation through a multi-stage process: it uses dependency-aware informal sketches, retrieves relevant information from processed PROSA documentation, generates staged skeletons, and then completes the proofs. By leveraging a corpus of 1,191 real-time systems papers, PROVE-RT achieves a 44.7% success rate in generating valid PROSA mechanizations, a significant improvement over the unreliable results from direct prompting of state-of-the-art LLMs.