Targeted difficult photo-z cases selected using the catastrophic point-prediction criterion ﹩{{\rm{\Delta }}}_{z}=| \hat{z}-{z}_{{\rm{spec}}}| /(1+{z}_{{\rm{spec}}})\gt 0.15﹩. Panel (a) shows a high-redshift object for which TabPFN 2.5 is noncatastrophic while FlexZBoost and Flow-Spline are catastrophic; panel (b) shows a faint object for which all three methods are catastrophic; and panel (c) shows a case for which TabPFN 2.5 is catastrophic while the other two methods are not. One object was selected uniformly with a fixed seed from each predefined category. Curves and shaded regions show the mean and ±1 standard deviation across these repetitions. The upward black marker denotes z_spec, colored downward markers denote the mean point predictions, and an asterisk marks methods that are catastrophic in every repetition.