Mean values for the lower and upper confidence bounds computed over the span of each image. The bar sizes represent the mean length of the error bars. As in Fig. 5, the values were computed over the test set (γ_i,  κ_i)_{i = n + 1}^m for various families of calibration functions (g_λ)_λ. After calibration with CQR, the choice of mass mapping method does not influence the average miscoverage rate (Fig. 5a), but it does significantly affect the size of the error bars (Fig. 6a). Additionally, RCPS, which tends to produce overconservative prediction bounds (Fig. 5b), yields larger error bars than CQR (Fig. 6b vs 6a). (a) CQR, (b) RCPS.