Beyond Plausible Diffs: Assurance Evidence for AI-Assisted Code Changes

Generative AI can now produce code, tests, fixes, and refactorings faster than most teams can review them. For high-integrity software, that creates a practical problem: a plausible diff is not an assurance argument.

This talk presents a code-centric workflow for turning AI-assisted C/C++ changes into reviewable engineering evidence. Instead of asking "do we trust the model?", we ask "can this change be accepted, rejected, or escalated for explicit technical reasons?" Evidence comes from outside the model: tests, static analysis, and formal verification provide corrective signals back into the agentic workflow by exposing examples, proven failures, proven-safe properties, and remaining uncertainty. We also discuss the challenging problem of comparing behavior after bug fixes and refactorings, where the goal is to distinguish intended change from unintended regression. This moves review from "does this look right?" to "what evidence supports this change, and what risk remains?"

The surprising conclusion is that an AI-assisted workflow can sometimes produce better review evidence than a human-only workflow, because the workflow can force intent, assumptions, and tradeoffs to be documented. Attendees will leave with a practical evidence-package pattern for reviewing AI-assisted C/C++ changes against the same engineering standard as human-written code.

Speaker

Martin Becker

Principal Application Engineer V&V, MathWorks

Martin Becker is a field application engineer for verification and validation workflows at MathWorks. He is an advocate of formal methods and static analysis, with more than 15 years of experience in embedded systems. In his daily work, he is supporting customers in a wide range of industries to efficiently produce embedded software while meeting safety and security standards.

Prior to MathWorks, he worked as avionics concepts engineer at Airbus, as research engineer at Tata Consultancy Services, and completed a PhD (Dr.-Ing.) in the domain of real-time computer systems at Technical University of Munich.