Inferential Proximity and Efficient Evaluation of Intelligence

A defining characteristic of general intelligence, whether biological or artificial, is the ability to generate successful responses across a wide range of sensory inputs. From this perspective, intelligence can be interpreted relationally, in terms of the complexity of the interaction domain and the uniformity of empirical success achieved within it. Based on this interpretation, the present work develops a general framework for the ordering and evaluation of intelligent systems. General intelligence is formulated as a function of empirical success over regions of increasing sensory complexity, yielding a partial order over intelligent agents defined through cognitive equivalence, specialization, and scope expansion. The framework further permits the formalization of efficient intelligence evaluation as the bounded reconstruction of the empirical interaction tuple with- out complete resampling of the learner’s sensory environment and response distributions. The resulting analysis indicates that reliable and efficient evaluation depends fundamentally on informational alignment between the evaluator and the evaluated system, including sufficient overlap of sensory and interpretive structure. As structural and empirical divergence increase, evaluative confidence and accuracy progressively degrade, ultimately leading to a regime in which general intelligence can no longer be reliably reconstructed from the interaction structure accessible to the evaluator. These results imply an inherent limit on the evaluability of progressively autonomous intelligent systems. As intelligent agents diverge in sensory exploration and internal cognitive organization, the capacity not only to align with, but also to meaningfully interpret or reliably evaluate them, may progressively diminish.