Unified Intelligence Theory and the Super Intelligence Level: A Cumulative, Harness-Measurable Hierarchy

We argue that intelligence attributions to an AI system should be indexed to a frozen model–harness pair rather than to the model alone, and graded cumulatively by what the pair verifiably delivers rather than by what it is credited with. We develop this claim as Unified Intelligence Theory and the Super Intelligence Level (SI-Level): five conceptual families are represented by six measured coordinates [C,I,O,T,H,SA], with delegation derived from task difficulty, human intervention, and verified reliability, and a memory ladder that constrains learning claims without equating stored information with learning. Capability claims are cumulative: a higher level retains every lower one's capabilities, and harness mechanisms are kept distinct from intelligence levels. Across three 48-episode harness curves, a success-only delegation score proved exactly invariant to an acceptance step that separated false completions from delivered work (2/12 to 0/12) — a Goodhart failure — repaired by scoring delivered outcomes rather than claimed success. In a sealed post-freeze probe, one archived configuration passed 5 of 12 held-out instances against 1 of 12 for another (sign test p=0.006; McNemar p=0.125). Related audits found defects in the benchmark items themselves, motivating an independent validity protocol. We relate this account to universal-intelligence, psychometric, and capability-taxonomy measures of machine intelligence and to the extended-mind debate, and state the hierarchy's main open objections directly. The framework remains a proposed operational language requiring broader calibration and longitudinal evidence. Status: Preprint; submitted to a peer-reviewed journal and not yet peer reviewed. Content may change in response to review.