Increasing Trust in Language Models through the Reuse of Verified Circuits