Searchlight analysis and understanding classifier accuracy

The searchlight runs a classifier on a sphere of voxels, then stores the accuracy of the classifier in the center voxel, then moves to the next.
The classifier, in this case, has eight labels, ‘bottle’, ‘cat’, etc.
It seems to me that there should be eight values, one for the classifier’s accuracy for each condition. This would be similar to the previous ROI analysis where a confusion matrix was generated, and the diagonal values represented the accuracy in each case.
How is this classification reduced to a single “accuracy minus chance” value?

E.g. In the image shown, the voxel value is ~25, or 25% above chance. But is that for ‘cat’, or ‘face’, or ?

Screen Shot 2021-10-20 at 9.35.49 AM