c:node / Research
c:node Research · NENA

The science behind NENA.

Reading time ~6 minBasis peer-reviewed sources, verifiedTopic Neuro-symbolic AIApproach open models · on-prem · traceable
Abstract

NENA is the intelligence behind c:node: a neuro-symbolic, source-based knowledge-graph AI built on open models (open source) that grows continuously with every case. It is built to work traceably instead of guessing — every statement tied to a source, every step verifiable. This article summarizes the peer-reviewed evidence behind it: why pure language models fall short when it matters, why knowledge graphs + rules close the gap, and how this meets the requirements of the EU AI Act.

1 Why pure language models fall short when it matters.

Large language models form sentences by predicting the statistically most likely next word. The US standards institute NIST notes that fabricated statements (confabulations) are by design in such models — not a slip, but a property of the architecture.1 Looking things up in real documents (retrieval-augmented generation) lowers the error rate markedly but does not remove it: even with the right documents, simple models still add unsupported claims.5 For decisions with consequences, “mostly right” is not enough.

2 The way out: recognize patterns and check against rules.

Research calls the approach neuro-symbolic AI: the pattern recognition of neural networks is combined with the verifiability of symbolic rules and knowledge graphs.2 A peer-reviewed survey (NAACL 2024) shows that knowledge graphs as an external knowledge source reduce hallucinations and increase reasoning accuracy.3 And for multi-hop reasoning — connecting several linked facts into one answer — the graph-based variant is demonstrably superior to plain text search.4 This is exactly what NENA is built on.

2020 →
Pure LLMs
Fluent but without a source — confabulation by design.
2023 →
RAG
Looking up documents lowers errors — doesn’t remove them.
2024 →
GraphRAG
Knowledge graph enables multi-hop, connected reasoning.
2025 → today
Neuro-symbolic · NENA
Patterns + rules + persistent, growing memory — traceable.
Fig. 1 · The evolution: from pure prediction to evidenced, rule-checked intelligence. NENA sits at the current end of this line.

3 Traceability is not optional — it is required.

For high-risk AI, the EU AI Act requires exactly what this architecture delivers anyway: complete logs (Art. 12), transparency and traceable documentation (Art. 13) and the possibility of human oversight (Art. 14).6 Because NENA records every statement with its source and history (provenance), every decision stays verifiable after the fact — without having to expose the model’s internals.

4 How NENA is built — and why it grows.

NENA builds on open models (open source) — swappable without changing the traceable logic — runs in the cloud, on your servers or fully offline, and thus stays sovereign in your environment. It is not a static model but a living knowledge graph that grows per tenant with every case. The division of labour: you steer with c:node, NENA thinks, the agents act.

Publicly funded research. NENA (formerly NEN) grows out of several years of research & development, funded in part via the German research allowance (BSFZ). Concrete figures (funding volume, period, approved projects — “Project 2”) will be added here once cleared.

5 Scientific advisory board.

NENA’s direction is guided by academic expertise in decision research and AI.

Assoc. Prof. Dr. Tuna Çakar

MEF University · Computer Engineering · Advisory Board (AI & Decision Research)
634 citationsh-index 10i10-index 11

Research on artificial intelligence, decision processes, machine learning and applied neuroscience — the cognitive and methodological foundations of robust decision intelligence. Selected works:

  • Physicians' ethical concerns about artificial intelligence in medicine — Frontiers in Public Health, 2024.
  • Unraveling neural pathways of political engagement: bridging neuromarketing and political science — Frontiers in Human Neuroscience, 2023.
  • An investigation of the neural correlates of purchase behavior through fNIRS — European Journal of Marketing, 2018.
Google Scholar profile →

References

  1. NIST, Artificial Intelligence Risk Management Framework: Generative AI Profile (NIST AI 600-1), 2024 — confabulations are by design in generative models. nvlpubs.nist.gov…
  2. Hitzler et al. (Hrsg.), Handbook on Neuro-symbolic AI and Knowledge Graphs, IOS Press, 2025 — knowledge graphs as a bridge between rules and learning models. iospress.nl…
  3. Agrawal et al., „Can Knowledge Graphs Reduce Hallucinations in LLMs? A Survey", NAACL 2024 — KGs as an external knowledge source reduce hallucinations and increase reasoning accuracy. aclanthology.org…
  4. GraphRAG-Bench — graph-based AI is superior to plain text search for multi-hop reasoning. arXiv:2506.02404, 2025. arxiv.org…
  5. MEGA-RAG (Multi-Evidence RAG) — looking things up lowers fabricated statements by over 40%, but does not remove them. PMC12540348, 2025. ncbi.nlm.nih.gov…
  6. Europäische Union — EU AI Act, high-level summary (Art. 12 logging, Art. 13 transparency, Art. 14 human oversight). artificialintelligenceact.eu…
  7. Assoc. Prof. Dr. Tuna Çakar — publication profile, Google Scholar (as of 2026). scholar.google.com…

The sources support the underlying principles (neuro-symbolic AI, knowledge graphs against hallucination, EU AI Act requirements). Advisory-board details are public (Google Scholar); R&D figures will be added after internal clearance.

Open source

On open models

NENA builds on open-source models — swappable without changing the traceable logic.

GitHub
Grows continuously

Living knowledge graph

Not a fixed model — NENA grows with you, case by case.

Explore NENA
Sovereign

Cloud · on-prem · offline

Runs where your data is — up to fully air-gapped.

Sovereignty

See the intelligence on your data.

A short test run in your environment — traceable answers, with a source.