The amount of sequential computation occurring between interpretable model states, measuring how much reasoning happens in uninterpretable latent space.