The Signature of Superintelligence

Public claims of artificial superintelligence (ASI) rest almost exclusively on external performance benchmarks: mathematical problems solved, drugs designed, tournaments won. This paper argues that, in their current form, such claims are epistemically unfalsifiable with respect to one specific question: a hyper-scaled system can produce extraordinary results through computational brute force without this implying any qualitative leap in intelligence, and output alone does not allow these two cases to be distinguished. This paper does not resolve that unfalsifiability on the axis of outputs — it deliberately leaves it unresolved. Instead, it proposes abandoning that axis for this specific question and replacing it with an endogenous, verifiable one: the Algorithmic Reflexivity Principle (ARP), defined as a system's capacity to identify and propose formal, executable, and verifiable improvements to the computational mechanisms that produce its own performance. The ARP does not require recursive self-improvement, selfawareness, or production-code modification — it requires evidence of functional understanding of the system's own computational substrate, evidence that is directly falsifiable because it consists of an inspectable artifact rather than a score. The paper discusses the relationship of this proposal to the recent Google DeepMind report "From AGI to ASI" (Genewein et al., 2026), particularly the Abstraction Barrier (Lerchner, 2026) and the report's characterization of recursive self-improvement (RSI). Finally, it identifies an operational distinction between scaledriven and understanding-driven improvement, and flags an open question regarding the limits of that distinction.