Multifunctionality is a biologically-motivated, emerging task type in reservoir computing. Recent work has shown that connectome-based reservoir computers exhibit unique capabilities and prediction dynamics on the Seeing Double problem compared to Erdös-Renyi reservoir computers. Here we investigate representations of said networks to determine whether and which differences exist between corresponding reservoir computers on this task. We consider dimensionality reduction and manifold learning techniques such as Principal Component Analysis (PCA), Laplacian Eigenmaps, and Sparse Autoencoder neural networks; and compare representational alignment on the Seeing Double task by analyzing the dynamic regimes arising from each network, and also their Euclidean distances and Cosine similarities. We report that connectomic representations of reservoir activity are strongly aligned when output dynamics are matched; that embeddings of learned network weights are similarly aligned for random networks; and that connectomic and random representations are neither aligned in their latent nor learned embedding spaces. Through comparisons between trained reservoir-to-output weight ( \(\textbf{W}_{out}\) ) representations, we also find that population-wide learning on the Seeing Double task is largely network-agnostic.