Confusion matrices for an independent 3 month GBM trigger dataset that was not used for model development and is separate from the held-out test set. The left and right panels show predictions from the standard in-flight trigger-classification algorithm and the optimal model, respectively. Rows represent the ground-truth labels from the GBM catalog, and columns represent the predicted classifications. The in-flight predictions (left) are extracted from the OBJECT field of the TRIGDAT files, which include preliminary classification categories such as BELOWHZ (source below the horizon) and UNRELOC (unreliable location). The UNCERT rows correspond to triggers that could not be confidently classified by the standard Fermi/GBM ground pipeline. Cell values indicate the number of triggers for each true-predicted class pair.