Determine how to ensure the semantic adequacy of agent-generated verification statements

Develop a satisfactory method for ensuring that verification statements produced from informal instructions by AI agents express the intended properties and contain no remaining defects that escaped human review.

Background

The paper describes verification proofs generated by AI agents under human direction. Although the authors identified and repaired defects in earlier versions of the Uniswap replay and deposit-contract statements, they emphasize that proof checking alone does not establish that a theorem states the property it ought to state.

The unresolved problem is methodological: finding a reliable way to detect all remaining mismatches between informal specifications and the formal statements proved by the system. This matters because the formal kernel validates derivations from the stated claims but does not assess whether those claims adequately capture the intended semantics.