An algorithm assigns ethnicity from names and then reports outcome gaps. Validation shows much lower classification accuracy for multiracial and less common names. What is the strongest concern?