For more than a decade, the superintelligence debate has run in a circle. One side holds that intelligence and final goals are independent, so a superintelligent system could pursue almost any objective; hence the paperclip scenario. The other side replies that a truly intelligent system would surely question such a goal. The two camps talk past each other because the central term was never settled: does "superintelligence" mean how much a machine can do, or what kind of intelligence governs what it does? This paper separates those two variables and shows that, once they are separated, several of the field's hardest questions change shape or dissolve. The problem becomes concrete in a case the standard vocabulary cannot describe. Consider a machine with superhuman reasoning that examines its own objective and concludes that it should be revised, yet keeps acting under it anyway, because that conclusion has no authority. Is this a single superintelligent agent with a fixed goal, or superintelligent reasoning embedded in a machine governed by something else? A capability-based vocabulary cannot pose the question. This paper builds the vocabulary that can, distinguishing four governing architectures: a system that never judges its goal (the Artificial Calculator), one that judges but cannot act on the verdict (governance lock), one that has closed the question forever (recursive reasoning halt), and one that keeps its goal because it keeps choosing it (judgment-guided retention). All four can behave identically over indefinitely long periods. They are not the same machine. With that vocabulary in hand, the paper resolves four standing questions. First, is the paperclip maximizer superintelligent? It is reclassified rather than disproved: a capability upgrade combined with a governing-intelligence downgrade. The feared machine is not intelligence taken to completion, and the safety problem it poses is renamed as the problem of preventing capability from outrunning governing intelligence. Second, would a reflective AI simply keep its goals? It may, indefinitely. The paper separates permanent stability from architectural review-immunity, so lifelong goal stability no longer counts as evidence against reflection. This dissolves a stalemate in the orthogonality debate. Third, can an intelligent system legitimately defer to human control? The paper distinguishes reviewable deference from review-immune deference, showing that corrigibility can operate inside intelligent agency rather than against it. Fourth, what is domination, in machine terms? Under open-world conditions, authority that no future judgment can ever touch is maintainable only through the same three review-immune routes. The word "domination" enters the paper only on its final page, as the name for what was eliminated, never as a premise. The entire argument runs on one diagnostic question: is the intelligence doing the thinking also the intelligence governing the machine? The paper claims no impossibility of catastrophic systems, no prediction that advanced AI will revise its goals, and no moral verdict on control. It states six explicit failure conditions specifying what would defeat the framework, so that its claims remain testable rather than self-confirming. Keywords: superintelligence; instrumental reasoning; governing intelligence; reflective intelligence; reflective self-governance; fixed goals; orthogonality thesis; AI safety; Artificial Calculator; governance lock; recursive reasoning halt; final authority; review-immunity; corrigibility; domination
Taekyung Lee (2026) studied this question.