Embeddings derived from model internal activations when applying values in context, used to predict behavioral generalization.