Jump Intelligence (JI): A Mechanism-Aware Lens for Diagnosing the Structure of Scientific Discovery from Literature-Only Signals

Computational discovery systems generate numerous novel hypotheses but lackthe ability to distinguish scientifically plausible connections from mere statisticalsurprises. This gap between novelty detection and mechanistic evaluation representsa fundamental limitation of literature-based research.We introduce Jump Intelligence (JI), a diagnostic framework that integratestwo complementary dimensions: the Topological Jump Score (TJS) for semanticnovelty and the Mechanism-aware TJS (mTJS2) for consistency with domain-specific principles. The optimal balance between these dimensions, quantified bythe 𝜆 signature, reveals the fundamental structure of the scientific discovery.Validation across materials science (Cu–Prussian Blue, 𝑛 = 99; Graphene–dye,𝑛 = 58) and biomedicine (Alzheimer’s drug repurposing, 𝑛 = 93) establishes atripartite taxonomy: high-𝜆 Bridging Leaps connect distant domains (𝜆 = 3.5),low-𝜆 Hierarchical Leaps advance within dense fields (𝜆 = 0.5), and interme-diate Mixed Types transfer partial knowledge (𝜆 ≈ 1.6). Rigorous validation,including nested cross-validation (𝛿 = 0.59 ± 0.13) and adversarial testing (257%degradation with inverted principles), confirmed the model’s robust diagnosticcapability. Systematic linguistic decomposition reveals that modern AI succeedsthrough keyword-level semantic filtering (100% contribution) rather than deepunderstanding; this phenomenon is termed the “Shallow Semantics Advantage”and is strategically exploited.JI transforms scientific discovery from black-box prediction to interpretable,mechanism-based diagnosis, achieving 80–85% enrichment in materials screening(Cu–PB and Graphene–dye) and robust predictive capability (𝛿 = 0.59) in drugrepurposing, while exposing paradigm limitations through transparent operation.