Large language models increasingly rely on structured knowledge graphs for grounding, but current systems treat epistemic states as flat scalars and explanations as post-hoc rationalisation. Both shortcuts are visible at the representation layer and propagate through every downstream reasoning step. We present TENSA, a neuro-symbolic reasoning system built on a bipartite temporal hypergraph, with three first-class contributions. First, multi-fidelity reasoning: every entity, situation, and attribution carries both a continuous confidence score and a discrete maturity lifecycle (Candidate ≺ Reviewed ≺ Validated ≺ GroundTruth), and reasoning traces propagate both jointly. Second, configurable fuzzy semantics: a fuzzy-logic layer exposes the four canonical t-norm families, OWA and Choquet-integral aggregation (with a supervised learning baseline for the Choquet measure), graded Allen relations, intermediate quantifiers, graded Peterson syllogisms, fuzzy formal concept analysis, Mamdani rule systems, and a scope-capped fuzzy-probabilistic hybrid surface — all selectable per query through TensaQL. Third, traceable derivation: every answer in debug mode is returned atomically with the TensaQL query and the data rows that produced it, providing auditable explanation by construction. This paper is a system description. The codebase is publicly available at https://github.com/arperon-labs/tensa under dual AGPL-3.0 + commercial licensing, and the release-tagged source tree is archived on Zenodo with the DOI registered for this record.
Radoslav Lovecky (Mon,) studied this question.