Many models look strong on broad benchmark tasks and then fall apart when users ask domain-specific questions. That usually happens because the model …