Passive brain-computer interfaces (pBCIs) developed within the neuroergonomic field usually aim to improve safety by augmenting human-machine interaction. To accomplish said goal, many pBCIs classify mental states such as mental workload or mental fatigue. An alternative is to forego mental states and aim to predict performance. Despite its drawbacks, we argue that performance estimation is a more goal-oriented approach than mental state estimation. In a realistic experiment, 25 participants had to control an uncrewed aerial system for two hours, continuously switching between target search and navigation. EEG classification accuracies based on mental states and performance were compared. With a Tangent Space Logistic Regression, we could predict an increased likelihood of lapses in the form of missing instructions with an above-chance level accuracy of 62.09 %.