If you're building an app based on LLMs that expects higher than 99% correctness from it, you are bound to fail. Negative scenarios workarounds and retries are mandatory.