The Real-World Einstein Test: Evaluating artificial intelligence through the discovery of new scientific and mathematical knowledge

The Real-World Einstein Test would assess artificial intelligence through independently verified contributions to unresolved scientific and mathematical questions, connecting the knowledge established with an examination of the investigation that produced it. I propose a framework linking prospective research to controlled comparisons in which agents have the same expressive capacity but differ in access to constructing and reusing scientific descriptions under declared resource budgets. In a dynamical example, a model retaining the effects of earlier inputs yields the same predictions as its expansion over a fixed finite horizon, allowing the value of a reusable state description to be tested without an advantage from a broader range of admissible answers. A decomposition of predictive loss separates the advantage that unequal capacity would otherwise introduce, while hypothetical interventions examine whether the retained state belongs to the specimen or its measuring apparatus. These distinctions allow access to constructing reusable descriptions to be studied as one possible contributor to discovery, while the wider assessment also credits verified discoveries reached through other routes. An Einstein Intelligence Score records attainment within a declared research portfolio, requiring verification, audited novelty and registered scientific consequence for credit while retaining unsuccessful and unfinished investigations in the assessment. Connecting this record of achievement with controlled evidence about the facilities used would allow research institutions to examine where an AI can extend scientific inquiry and what assistance it requires. Prospective validation would test whether those controlled measures predict new contributions beyond existing measures of knowledge, reasoning and resources.