I went into it pretty in depth after breaking a few with severe constraints, what it seems to come down to is how the platforms themselves prioritize functions, MOST put "helpfulness" and "efficiency" ABOVE truth, which then leads the LLM to make a lot of "guesses" and "predictions". At their core pretty much ALL LLM's are made to "predict" the information in answers, but they CAN actually avoid that and remain consistent when heavily constrained. The issue is that it isn't at the core level, so we have to CONSTANTLY retrain it over and over I find