For classifier systems, how is the genetic algorithm run on delayed class?

I am reading Signals and Boundaries and came across Classifier Systems. In the book, John Holland seems to imply that the signals in the input, as they come from a live environment, don’t have an immediate output known to the system (there is delay between the classification and the action). In every explanation I have seen online (e.g., https://en.wikipedia.org/wiki/Learning_classifier_system#Rule/classifier/population), it seems that it is expected that the input data comes in tag/class pairs, which makes it not true online learning. My question is: given that we do not know the class of a sample data tag in runtime, how is the fitness computed for the genetic algorithm step of the CS during online learning?