We present a latent knowledge exploration of network foundation models, their representation space utilization, alignment with domain-expert features, and causal sensitivity to perturbations.