Formal verification is a powerful tool: it proves that a program satisfies its specification. But it cannot tell you whether the specification itself captures what the author actually meant. That distinction is the focus of a new paper on arXiv.
The paper introduces claimcheck, a method for measuring and improving specification faithfulness. The motivation is practical: large language models are now generating both code and the formal specifications that accompany it, making the risk of subtle mismatches between intent and spec more acute.
By quantifying how faithfully a specification reflects the original intent, claimcheck aims to catch problems that formal verification alone would miss. As LLM-generated code becomes more common, the paper suggests, checking the spec itself is as important as checking the code.