if you wanted to further study the phenomenon you could probably run the experiment again, and the ablate or corrupt those intermediate activations to get a feel for what it was thinking at the "time".