TENSA is a neuro‑symbolic reasoning engine over a typed temporal hypergraph. It treats confidence and maturity as orthogonal first‑class annotations, exposes configurable fuzzy semantics per query, and returns every answer with the TensaQL derivation that produced it — auditable by construction.
We claim conceptual novelty against each. The remaining contributions — engine catalogue, hierarchical planning, two‑domain instantiation — are the open research substrate.
Confidence (continuous, data‑quality) and maturity (discrete, epistemic standing) propagate as independent annotations through every reasoning trace. A peer‑reviewed result held with low confidence and a tabloid rumour that happens to be true become formally distinct.
maturity ∈ { Candidate ≺ Reviewed ≺ Validated ≺ GroundTruth }
confidence ∈ [0, 1]Four canonical t‑norm families (Gödel · Goguen · Łukasiewicz · Hamacher), OWA + Choquet aggregation, graded Allen relations, intermediate quantifiers, Peterson syllogisms, Bělohlávek FCA, Mamdani rules — selectable per query. Every result is tagged with the configuration that produced it.
ASK "..." OVER "narrative"
WITH TNORM 'lukasiewicz'
AGGREGATE choquet BY 'measure-A'
Debug mode returns every answer atomically as ⟨answer, query, rows⟩. The
TensaQL query is the canonical symbolic derivation. The rows are the data that
instantiated it. A regulator can re‑execute and verify reproduction. Explainability
by construction, not by reconstruction.
{ answer: "...",
query: "MATCH ... RETURN ...",
rows: [ /* exact data */ ] }EU AI Act Article 13 transparency & Article 14 meaningful‑oversight requirements are met by structural design, not by post‑hoc rationalisation.
Build bottom‑up — each layer depends only on layers below it. The fuzzy layer is orthogonal: it reads layers 2–3 and feeds layer 4.
A bifurcated planner separates symbolic execution (instant) from inference‑job submission (async). Confidence semantics are configurable on the same query.
-- Match temporal patterns under graded Allen relations
MATCH (a:Actor)-[:PARTICIPATES]->(s:Situation)
AT s.temporal AS FUZZY 'overlaps' THRESHOLD 0.7
WHERE s.narrative_id = "crime-and-punishment"
AND a.confidence > 0.6
WITH TNORM 'lukasiewicz'
AGGREGATE 'owa'
RETURN a, collect(s) AS situations
-- Submit a counterfactual job under do-calculus
INFER COUNTERFACTUAL(s_id, ASSUMING a.action = "cooperate")
UNDER "narrative-id"
-- RAG with hierarchical Leiden communities + DRIFT search
ASK "who held veto power before the merger?"
OVER "merger-corpus"
MODE drift
SUGGEST
RESPOND AS "bullet points"
-- Discover patterns ACROSS narratives
DISCOVER PATTERNS ACROSS NARRATIVES
WHERE tag = "19th-century-russian"INSTANT or JOB; deterministic work runs on the calling thread, anything heavy is dispatched to the worker pool.
Keyed by std::mem::Discriminant<InferenceJobType>, so payload‑bearing
variants route to a single engine instance. Eight families.
Identical code paths. Domain‑appropriate t‑norm and aggregator defaults. The substrate is domain‑agnostic; subsequent work on legal reasoning, criminal investigation, scientific argumentation, and adversarial wargaming can build on the same surfaces.
hierarchical generation · verifiable structural properties
opinion dynamics · CIB · higher‑order contagion
Narrative‑centric IA. Workspace shell at /n/:narrativeId/:tab. Sixteen canvases
from Graph and Timeline through Workshop, Reconstruction, and Fuzzy. One Inspector panel
reacts to selection across every lens.
“TENSA: Multi‑Fidelity Neuro‑Symbolic Reasoning over Temporal Hypergraphs with Configurable Fuzzy Semantics and Traceable Derivation.” — Radoslav Lovecký (Arperon).
Cites Klement‑Mesiar‑Pap, Yager (OWA), Grabisch & Bustince (Choquet), Schockaert‑De Cock + Dubois‑Prade (fuzzy Allen), Murinová & Novák, Bělohlávek, Mamdani, Iacopini · Petri · Barrat, Hickok, Delabays, Amgoud · Ben‑Naim, Nebel‑Bürckert.
Read the paper (Zenodo) →# clone
git clone https://github.com/arperon-labs/tensa
# build (no C++ toolchain required)
cargo build --no-default-features
# run the API + Studio
cargo run --features server,studio-chat@misc{lovecky2026tensa,
author = {Lovecký, Radoslav},
title = {TENSA: Multi‑Fidelity
Neuro‑Symbolic Reasoning over
Temporal Hypergraphs},
year = {2026},
note = {Arperon · open research platform}
}