Brain-computer interfaces (BCIs) are systems that span hardware and software, and record and decode neural signals from the brain and/or stimulate the brain with encoded neural signals. BCIs have a wide range of applications in restorative medicine, from restoring movement through prostheses and robotic limbs to restoring communication through spellers and restoring sensation. BCIs also have applications in diagnostic medicine where detection of sleep patterns or emotion can provide valuable information for clinicians. Despite their promise, BCIs have not yet been adopted for longterm, day-to-day usage because of challenges in achieving robustness and performance that would allow safe operation in all possible scenarios. Ensuring safe operation currently requires hours of manual data collection and recalibration, involving both patients and clinicians. Furthermore, data collection is not carried out in a way that attempts to eliminate specific faults in a BCI system. This paper presents a new framework for characterizing, detecting, and localizing faults in BCIs. We propose partial test oracles as a method for detecting faults and data-slicing functions as a method for localizing faults to relevant slices of the input and performed tasks. Through targeted data collection and retraining using these slicing functions, the proposed approach improves the correctness of a BCI neural decoder. We evaluate our approach on 5 different BCI systems and show that the approach precisely captures faults and can significantly reduce the frequency of faults in the BCIs. These results suggest that the proposed framework is a promising step towards repairing faulty BCIs.