Post hoc UMAP visualization of feature vectors extracted from the optimal dual-scale model F_{NaI-128C,BGO-128C}. The embedding was fitted to the pooled feature vectors from the training, validation, and held-out test subsets of the reliable five-class sample, together with the catalog UNCERT diagnostic events. It is intended as a qualitative visualization and was not used for model selection or performance evaluation. Left: each point denotes an individual trigger and is colored according to its class label in the GBM trigger catalog; UNCERT denotes triggers that could not be confidently classified by the standard Fermi/GBM pipeline. Right: the same embedding is colored according to the LUNCH prediction, where triggers with a maximum softmax confidence of ≤0.6 are assigned to the UNKNOWN category. The projection reveals the clustering structure of the learned representation and shows how catalog-uncertain and low-confidence events are distributed relative to confidently classified trigger populations.