Comparative Analysis Of The Commonly Used Code Generated By Large Language Models For MISRA C Compliance

Date

2026-1

Author

Öztop, Umut

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

66
views

22
downloads

Cite This

This report presents whether Large Language Models (LLMs) like ChatGPT and Gemini can generate safety-critical C code compliant with MISRA C:2012. We tasked six models with implementing CRC16 and COBS algorithms, verifying the output with PC-lint Plus and functional tests. Our results show a clear drop in compliance as algorithmic complexity increases. While models excelled at the simple CRC16 task (Gemini averaging 0.3 violations), the memory-intensive COBS task caused widespread safety failures, especially regarding single-point exit rules. We observed a distinct "compliance paradox": Claude produced the fewest violations but frequently generated broken code, whereas ChatGPT achieved 100% functional accuracy but penalized its safety score by adding unrequested complexity. Ultimately, while LLMs show promise as prototyping assistants, they cannot yet autonomously generate certification-ready embedded software.

Subject Keywords

Large language models (LLMs), MISRA C:2012, Static analysis, Safety-critical systems, Embedded software engineering

URI

https://hdl.handle.net/11511/118338

Collections

Graduate School of Informatics, Term Project