Empirical miscoverage rate after calibration with CQR (Fig. 5a) and RCPS (Fig. 5b). The means and standard deviations were computed over the test set (γ_i,  κ_i)_{i = n + 1}^m for various error levels δ (RCPS only) and various families of calibration functions (g_λ)_λ. More details are provided in the note of Table D.1. The theoretical bounds for CQR introduced in Eq. (34) are represented by the yellow area in Fig. 5a. These plots indicate that CQR achieves miscoverage rates that are, on average, close to the target α, whereas RCPS tends to produce overconservative bounds, even for large values of δ. (a) CQR, (b) RCPS.