Verifiable evidence embedded in generated text about which internal states or computations a model actually used.