The method's grandparent. Project intermediate activations directly onto the vocabulary to see what the model is "leaning toward saying" at each layer. The J-lens is the Jacobian generalisation of exactly this idea. Instead of a linear read-off, ask which activation directions most change the probability of each token.