Verus proves Rust correct for all inputs. Code review still can't define "correct."

Amazon unveiled Verus, a Rust verifier used in Firecracker and AWS Lambda, through a blog post. Unlike traditional code reviews, Verus doesn't evaluate code; it mathematically verifies annotated functions against a specification for every potential input. This shifts correctness from a subjective judgment to a definitive pass or fail.

The author emphasizes the importance of this distinction, noting that a reviewer, be it human or model, still grapples with the question of whether a code diff is correct. There is no universal verifier for this, only spec verifiers, compiler checks, type checkers, and test suites, none of which fully encapsulate the meaning of correctness in real-world codebases.

Similar challenges arise in the realm of reinforcement learning, where models like a 4B model were trained to generate PostgreSQL query plans that outperformed Postgres's default by 44.7% in join-heavy queries. The model's success hinges on a verifiable reward - execution time - which allows it to improve systematically. In contrast, code review lacks such clear, quantifiable metrics, making it a challenging area for AI intervention.

The author concludes that the key to improving autonomous code review lies in reducing the unverifiable aspects of the task, such as moving from "does a human think it's right" to "the tool proved it matches the spec." This shift would enable AI and RL models to better assist in the review process.