of the computational subspace. The readout is typically benchmarked by integrating the readout signal and choosing a binary threshold to extract the „readout fidelity“. We show that such a characterization may significantly overlook readout-induced leakage errors. We introduce a method to quantitatively assess this error by repeatedly executing a composite operation — a readout preceded by a randomized qubit-flip. We apply this technique to characterize the dispersive readout of an intrinsically Purcell-protected qubit. We report a binary readout fidelity of 99.63% and quantum non-demolition (QND) fidelity exceeding 99.00% which takes into account a leakage error rate of 0.12±0.03%, under a repetition rate of (380ns)−1 for the composite operation.