Hey, thanks a lot for the feedback! Most answers are first checked against a fixed list, to avoid making too many LLM calls. Unfortunately, for latency and playability reasons, we can't do much deeper online searches: every call would cost a lot and take too long. So about 1 to 1.5% of answers end up judged a bit unfairly, but we're actively working on improving the model. Thanks again for your feedback!