Verifier-independent construction of precise verification tasks

Determine how to construct sufficiently precise verification tasks with expected verdicts that are independent of the verifier used during construction, while preserving the properties of the original software.

Background

The paper considers using agentic proof-harness engineering to automate the construction of software-verification benchmarks from real-world projects. Generated tasks must preserve the relevant properties of the original software and should provide reliable expected verdicts.

Although sound over-approximations can support safety proofs, they may also introduce spurious violations. The authors therefore leave unresolved the concrete question of how to construct tasks that are sufficiently precise and whose expected outcomes do not depend on the verifier used to generate them.