Infon: A Knowledge Graph Reasoner with Calibrated Uncertainty

Infon is a knowledge graph reasoner that performs fact-checking over document corpora with calibrated uncertainty. Three properties distinguish it from prior systems: cassette-native storage (immutable, content-addressed, S3-native documents with manifest pruning), honest uncertainty via a first-class Dempster–Shafer vacuous element m(Θ) that rises to 1.0 on unsupported claims rather than fabricating confident answers, and MCTS multi-hop reasoning guided by a frozen sheaf graph neural network as a terminal scorer. On the actor-to-actor real-data evaluation, the symbolic pipeline scores 80% accuracy; adding the sheaf GNN raises this to 100%. On synthetic held-out data the GNN reaches 99.2% overall accuracy. The system operates on a laptop CPU without GPU or API keys. Code is released under the repository tag preprint.