The arXiv preprint RULER: Representation-Level Verification of Machine Unlearning (arXiv:2605.27569) defines oracle-comparative metric M2 and oracle-free metric M4 to measure residuals in intermediate representations after unlearning. Four approximate unlearning methods satisfied output-level criteria yet showed significant residuals under linear mixed-effects models in 10 of 12 conditions (p<0.05). M4 identified persistent identity-level signals in face-recognition models across tabular, image, clinical-text and face-identity datasets. Bourtoule et al. (2021) established output-level protocols based on membership inference and retain/forget accuracy (arXiv:1911.04933); RULER demonstrates these protocols miss representation-level encoding. Carlini et al. (2022) quantified extractable memorization in language models; the same pattern appears here in non-language settings where no tested method fully erased signals. M2 effect sizes increased with forget-set fraction; Bad Teacher produced identical residuals despite a distinct mechanism.