Beyond Fixed Information: Endogenous Construction of Inquiry Interfaces

Abstract

Artificial intelligence selects actions, tools, hypotheses, and experiments, yet within an information interface whose sensors, measurements, queries, and interventions are executable from the outset. We study a stricter problem: reliable inference may require evidence that becomes available only after the system changes what can be measured or acted upon. We formalize Endogenous Inquiry Expansion (EIX), in which information interfaces, operational resources, evidence, and constructive languages form a computational state. The theory distinguishes solver, frame, interface, and generator failure; proves information-risk separation, minimal adequate expansion, finite-sample identification, a no-free-inquiry limit, operational reducibility, adaptive fixed-interface separation, and finite-class inquiry generalization. EIXBench rigorously evaluates fixed-space adaptive and expected-information-gain selection against resource acquisition, held-out expression synthesis, structural composition transfer, recursive language growth, scaling, adversarial controls, and inference-compute increases. Across eleven defective-interface levels and one hundred seeds, EIX attains 0.992 mean accuracy versus 0.822 for an operationally static comparator and 0.554 for a fixed interface, with no false expansions on matched adequate controls. On operational non-collapse tasks, EIX reaches 0.997. A physics stress test synthesizes an energy-drift observable from generic arithmetic primitives and raises mechanism identification from chance to 0.838. These results establish a reproducible, falsifiable testbed for evidence-interface construction.