Did you use the exact API call shown in the paper? I am unable to replicate the paper's counterexamples via the chat UI, but that's not very surprising (if the LLM already only fails a few cases out of thousands, the small differences in context between API and chat might fix them).

I tried this https://chatgpt.com/share/69cebb52-56a8-838f-969c-c47308262a...