Comment by amelius

Since LLMs are great at coding but bad at logic, maybe an approach like this can bridge the gap? So first let it translate natural language to a formal language, from there allow it to use a logic engine to make verifiable transformations (correctness-preserving), and finally translate back to natural language.