0% hallucination on grounded queries. Not reduced. Not minimized. Zero — and the scope in that sentence is the whole claim, not a hedge attached to it. On a query the graph can ground, the generative model is not in the path at all, so there is nothing there to hallucinate with. Ask something it has not studied and it tells you so. And it answers known facts in 11.3 milliseconds.

That's not a prompt engineering trick. It's an architectural choice. We published the full design and evaluation in our first architecture paper, and the results surprised even us.

The Inversion That Changes Everything

Most AI systems that combine knowledge graphs with LLMs follow the same pattern: the user asks a question, the system generates an answer with the LLM, and then it checks the knowledge graph to "ground" or "verify" the response. This is the RAG approach, and it has a fundamental problem — the LLM already produced text before you checked whether it was right.

NuSy inverts this. We query the knowledge graph first. Only when the graph doesn't have a confident answer do we fall back to the neural (LLM) path. This isn't a tweak to RAG — it's a completely different routing model.

Here's what the decision path looks like:

  1. Symbolic path — query the knowledge graph. If confidence is high and the topic is covered, return the answer directly. No LLM involved. 11.3ms.
  2. Neural fallback — if the graph has a gap, invoke the LLM with full provenance tracking. The answer is marked as ungrounded and the system attempts crystallization (extracting structured knowledge from source prose) to fill the gap permanently.

Most queries in a well-trained being hit the symbolic path. That means zero generative model involvement for known facts — which is how you get zero hallucination.

The Layers That Answer a Query

The brain organizes knowledge into seven Y-layers (Y0–Y6). Four of them carry the answer path described here:

The remaining three — Y4 Journal (what the being thinks), Y5 Procedural (compiled workflows), Y6 Metacognitive (calibration and error tracking) — are the subject of the cognitive-layer paper, and are what let a being assess its own accuracy rather than just answer.

Every answer traces back to Y0. Provenance isn't optional — it's structural. When a being says "metformin is contraindicated in severe renal impairment," you can follow the chain: Y1 triple → extraction pattern → source passage in Y0 → the clinical guideline it came from.

Why This Works in Practice

We validated the architecture across two domains: clinical decision support and software development assistance. The clinical domain is the hard test — you can't hand-wave hallucination when someone's health depends on the answer.

In both domains, the symbolic path handled the majority of queries. When the being had studied the relevant material, it answered from the graph with 11.3ms latency and zero hallucinations. When it hadn't, it said so — instead of fabricating a plausible-sounding answer.

That last point matters more than the latency number. An honest "I don't know" is always better than a confident lie. Most LLM-powered systems can't tell the difference.

The Engineering Tradeoff

Symbolic-first routing means you need a good knowledge graph. The being has to study before it can answer. There's a cold-start cost — you need source material, crystallization runs, and validation before the symbolic path lights up.

But once it does, the graph is permanent. No re-inference, no temperature tuning, no prompt iteration. The knowledge sits in Y1 as triples and answers queries in milliseconds. And the crystallization pipeline keeps growing the graph from source prose — deterministically, without trusting the LLM's output as knowledge.

This is the tradeoff we'd make every time: invest in structured knowledge upfront, get reliable, fast, auditable answers forever.

Read the full paper: NuSy Brain Architecture — Zenodo