We used a cross-classification method to adjudicate between the models. We trained a classifier to decode the object identity and its upper/lower location from population spiking at each time point on contralateral no-swap trials. We tested whether these classifiers could predict the same information on ipsilateral-to-contralateral swap trials, which brought that same object to the same location as the contralateral no-swap trials. Note that training and testing were both performed on the same cortical hemisphere—separately for each hemisphere, then results were pooled across them—meaning the cross-classification is across task conditions, not cortical hemispheres. Thus, we tested whether the same information was reflected in the same neural pattern in a given hemisphere, regardless of how it arrived there. If both conditions activate the same ensembles, as assumed by the generic trace model (Figure 6A), then this cross-classification (Figure 6B, orange) should result in high decoding accuracy, similar to that obtained from both training and testing on constant contralateral trials (Figure 6B, desaturated green). If these conditions activate different neural ensembles, as suggested by the novel ensemble model (Figure 6C), then cross-classification
