Compact Proofs of Model Performance via Mechanistic Interpretability