Classification error rates per magnitude bin for all evaluated classifiers. The magnitude bins coincide with those in Fig. 8, and the error rates are defined as the ratio of objects incorrectly classified as FPs or FNs to the total number of objects within each bin. Left sub-figure (Two-class classification, light blue): error rates for compact objects. Specifically, galaxy objects misclassified as stars or QSOs (left) and vice versa (right) are presented. Each classifier is denoted by a unique color: BANNJOS in blue, sglc_prob_star in orange, and CLASS_STAR in green. An object is considered compact if its corresponding CLASS_STAR or sglc_prob_star score exceeds 0.5, or its median PC_Star|_QSO(50) is greater than 1/3. The statistics are obtained using the test sample of 136 570 objects. Right sub-figure (light orange): classification error rate for the three classes available in BANNJOS (dark blue) and von Marttens et al. (2024) (vM24, dark orange). The FPs are depicted with solid lines, while FNs are shown with dotted lines in the respective colors. In this case, the error classification ratios are obtained with another test sample of 52 105 sources never seen during the training phases of the two models. The distribution of errors over cumulative magnitude bins is shown in the corresponding bottom panels. The median 90%, 50%, and 10% completeness levels for J-PLUS compact sources are indicated by vertical gray shaded areas, with the darkest shade representing the 10% level. BANNJOS surpasses all other classifiers by a large margin at all magnitude range, both in total classification error and in symmetry between FPs and FNs.