In this paper, we address the task of identifying false positive discrepancies in critical Instrumentation and Control (I&C) software testing as an Irregular Multivariate Time Series Classification (IMTSC) problem and evaluate the performance of different Neural Network architectures, including GRU-D, Transformer, and the Set Function for Time Series (SEFT). Architectures are trained on a real-world dataset from industrial test reports, and evaluated in our use case. Our finding is that the Transformer-based method outperforms the other baselines in terms of precision, while SEFT shows noteworthy performance compared to the literature benchmarks and in the lowest training time, highlighting the importance of our context-focused evaluation.