Harvesting Unexpectedness: A Double-Threshold Filter for AI-Generated Discoveries

Abstract: AI systems frequently generate valid discoveries that human experts find counterintuitive. We argue that these anomalies should be treated not as hallucinations to be filtered, but as a new class of scientific material that can be systematically collected and integrated. This paper proposes a five-step filtering protocol—generation, screening, verification, translation, and integration—designed to isolate valuable, target-constrained unexpectedness from high-entropy noise. The protocol uses a double threshold based on statistical surprisal and formal domain verification. Through case studies including FunSearch, AlphaFold, AlphaGo, and the 2026 disproof of the planar unit distance conjecture, we reconstruct how similar filtering dynamics appear in existing cases and provide a concrete workflow for automated discovery platforms.