Formal verification tools give a clear signal when a program satisfies its written contract. But that signal can be misleading: if the contract itself is incomplete, the program may still violate the intended behavior. A new paper from arXiv highlights this problem, noting that correctness feedback suppresses the very evidence that specification-completeness auditing is designed to uncover.
The authors demonstrate this "objective reversal" on 12 examples, showing that a program can pass verification while still failing to meet unstated requirements. To address the issue, they propose phase-separated completeness auditing in Verus, a verification tool for Rust. The idea is to separate the completeness audit from the correctness check, so that the feedback from one does not obscure the other.
The work underscores a subtle limitation of verification: a proof of correctness is only as strong as the specification it targets. By separating the phases, the authors aim to make completeness auditing more effective, ensuring that the absence of errors does not hide the absence of required behavior.