The Evaluation Horizon: Why Some Problems Cannot Be Reliably Solved Without a Model of Solvability

Contemporary science, policy, and artificial intelligence increasingly encounter problems whose solutions exist in principle yet fail to propagate, converge, or even be recognized as valid across evaluators. These failures are commonly attributed to sociological, institutional, or incentive-related causes. This paper advances a stronger claim: that a broad class of such failures arises from structural limits on evaluation itself, rather than from deficiencies in communication, rigor, or intent. We introduce the concept of an evaluation horizon, defined as the bounded region of conceptual space, semantic resolution, and inferential depth accessible to a given evaluator—human or artificial—within practical constraints. Using a geometric and topological framework grounded in conceptual space navigation, we show that solvability, distinguishability, and convergence are all evaluator-relative properties. As a consequence, truths may exist that are internally coherent, well-defined, and even simulable, yet remain systematically unrecognizable or non-propagatable under prevailing evaluation regimes. Within this framework, the long-standing “propagation problem”—the repeated failure of certain results to disseminate across journals, disciplines, institutions, and AI evaluators—is reinterpreted as a necessary outcome when semantic dependencies exceed evaluator horizons. We demonstrate how such failures persist even in the absence of bias, malice, or methodological error, and why they cannot be reliably documented through conventional controlled experiments, since verification would itself require evaluators with capacities beyond those under test. Crucially, these limits apply most severely in domains characterized by high novelty, deep uncertainty, and irreversible consequences. In such domains—including global coordination failures, emerging technological risks, systemic ecological instability, and unprecedented forms of intelligent systems—decision-makers are forced to assess problem severity and solution adequacy without reliable evaluative instruments. The result is not merely disagreement or delay, but systematic underestimation of risk. Over sufficiently long timescales, repeated exposure to such unmeasurable and mis-evaluated problems implies an increasing probability of catastrophic failure, approaching certainty in the absence of mechanisms that extend evaluative capacity. By formalizing evaluator-relative limits on reach, resolution, context, and convergence, this work reframes persistent propagation failure as a fundamental barrier to long-term adaptive survival. The results imply that reducing existential risk requires not only improved models, incentives, or governance, but the development of new forms of collective intelligence capable of composing and extending evaluation horizons themselves.