Knowledge graphs (KGs) are typically maintained through passive verification—external queries or scheduled audits trigger fact-checking. We present the Curiosity Engine, an epistemic active learning framework in which the knowledge graph systematically identifies regions of insufficient coverage. Implemented within OmniCore v17 and deployed on the PARAM BILIM supercomputer at ENU Astana, the system employs four detectors—gap, trustfrontier, domainᵢmbalance, and temporaldecay—that continuously scan the graph’s structural, epistemic, distributional, and temporal properties. Detected deficiencies are scored by a novel “epistemic pressure” metric derived from the Structural Integrity Index (SI-Index), queued for human-in-the-loop review, and resolved via automated SPARQL enrichment against Wikidata. We report results from a live deployment on a Kazakhstani higher-education knowledge graph containing 915 entities, 1, 125 triples (1, 094 unverified, 7 verified, 19 disputed, 5 retracted), and 169 verified facts across 4 data sources. The Curiosity Engine’s scan identified 51 items; from 4 executed items, the system created 101 new triples (9. 0% of the total triple count), though a systematic predicate extraction artifact degraded their semantic accuracy. The SI-Index stands at 0. 483 with a diagnostic layer decomposition: L0=0. 006, L1=0. 855, L2=0. 240, L3=0. 831 (weights α₀=0. 30, α₁=0. 25, α₂=0. 20, α₃=0. 25). A weight sensitivity analysis confirms that L0 is the binding constraint across all tested configurations. We formalize the epistemic active learning loop, define epistemic pressure as the negative gradient of SI-Index under knowledge stasis, provide a simulated baseline comparison of entity selection strategies as a hypothesis-generating analysis (Section 5. 9), and demonstrate that systematic gap detection surfaces structural deficiencies that remain undetected by passive verification and RAG approaches. The current work is positioned as a small-scale institutional proof-of-concept; empirical baseline experiments and downstream evaluation are identified as necessary next steps. A companion study 12 from the same research group provides complementary evidence on multi-LLM disagreement entropy for triple verification on the same infrastructure.
Anatoliy Kremenchutskiy (Sun,) studied this question.