Solving the hallucination problem isn't an engineering challenge, it's a data challenge. Just take the most common non-filler token sequences and annotate them

Solving the hallucination problem isn't an engineering challenge, it's a data challenge. Just take the most common non-filler token sequences and annotate them with accuracy scores, use those to train an accuracy predictor, then fine tune the LLM using the accuracy predictor to bias the gradient towards accurate statements. You'd probably have to spend a lot of money (or get a lot of volunteer time) to get it done, so it's not a startup thing, but it would make a massive difference.