[CGamesPlay]

I think the chain-of-thought reasoning will be what fixes this. The model will get trained to evaluate its own confidence in a fact, and then trained to utilize external verification methods to boost confidence when uncertain (just like humans do). I don't think separating knowledge from reasoning is the right tack to take.