Which Calibration Score Becomes The Threshold

You have a fitted regressor and a held-out calibration set of $n = 500$ points. For each you compute the nonconformity score $s_i = |y_i - \hat f(x_i)|$ and sort the 500 scores in ascending order. You want split conformal intervals $\hat C(x) = \hat f(x) \pm \hat q$ at miscoverage $\alpha = 0.10$ .

Which sorted calibration score is $\hat q$ ?

The 50th smallest score, since $\alpha = 0.10$ points at the lower 10% of the calibration scores

The 450th smallest score, the ordinary 90th percentile of the 500 calibration scores

The 451st smallest score, because $\hat q$ is the $\lceil (n+1)(1-\alpha) \rceil / n$ empirical quantile and $\lceil 501 \times 0.9 \rceil = 451$

The 452nd smallest score: take $\lceil (n+1)(1-\alpha) \rceil$ , then round up once more to leave room for the test point