Patching predicted neuron activations based on Neuron Scientist descriptions into an ablated model recovers 74% of the damage measured by Kullback-Leibler divergence