This paper reports the first in-depth analysis of the DESFire EV1's EAL4+ certified TRNG and raises some difficult questions regarding the certification of non-deterministic random number generators. We start by analyzing the quality of the purportedly true random number generator (TRNG) on the DESFire EV1 card. Clear and consistent biases are identified, despite good performance in most randomness tests. These statistical tests, commonly used in popular certification processes, such as Common Criteria EAL4+, are found not to be able to detect these anomalies. The means we employ for the detection and characterization of the bias are explored, highlighting both novel and existing ways of spotting deficient TRNG output. Further analysis shows systemic issues affecting TRNG output at the byte level, for which we have developed an accurate explanation. Our results have been acknowledged by the manufacturer, after responsible disclosure.