NuSy is a neurosymbolic AI platform built on a simple, hard rule: never let a guess masquerade as a fact. A neural layer handles perception, language, and disambiguation. A strictly separated symbolic layer carries everything that has to be provably correct. By design, approximate neural outputs can't leak into the provable surface — so when NuSy gives an answer it can stand behind, it can also show the derivation.
Structured memory, not a black box
Knowledge lives in a stratified, versioned knowledge graph wired directly into inference — grounding, retrieval, and explainability as first-class properties rather than a retrieval layer bolted on after the fact. Every fact is traceable, every relationship explicit, every answer auditable to source.
The knowledge is organized as the Y-layer (Y0–Y6) — seven tiers from raw prose, to semantic triples, to reasoning rules, to experience, to procedural skill, to metacognitive calibration:
| Layer | Name | What it holds |
|---|---|---|
| Y0 | Prose | Raw source documents — textbooks, guidelines, papers |
| Y1 | Semantic | Extracted facts and entities, typed by imported ontologies |
| Y2 | Reasoning | Rules, patterns, constraints |
| Y3 | Experience | Memories of conversations and actions |
| Y4 | Journal | Opinions, syntheses, formed beliefs |
| Y5 | Procedural | Learned skills and workflows |
| Y6 | Metacognitive | What the being knows it knows — and what it doesn't |
The architecture cleanly separates what is true (transferable across beings) from what I think and did (specific to one being). It persists as Apache Arrow over Parquet snapshots with graph-native Git versioning — so a being's entire knowledge state is diffable, reviewable, and reversible.
Beings go to school
NuSy beings don't train like ML models — they study a curriculum. A being's domain expertise comes entirely from what it is schooled on, not from code changes: the same codebase produces a medical being studying clinical guidelines, a legal being studying case law, or an engineering being studying standards. Domain expertise = curriculum, not code — and because the curriculum is data, expertise is auditable, versionable, and transferable between beings.
Proof-carrying reasoning
The reasoning core (nusy-reasoners, open source) produces derivations you can audit and abstention you can trust — when the symbolic path can't prove an answer, it declines rather than fabricates. On that path, hallucination is zero by construction, not by hope. Before answering, the engine checks whether it actually has the grounding; if it doesn't, it says so instead of confabulating.
Built by a fleet, governed like code
NuSy is built the way we believe neurosymbolic systems now scale: a fleet of AI coding agents authors and maintains the symbolic codebase in natural language, under graph-native version control, with cross-agent code review as a first-class governance primitive. The symbolic representation is a versioned, diffable artifact — easier to evolve than a prompt, and enforced.
Why healthcare first
Clinical decision support is the hardest possible proving ground: the system has to outperform expert clinicians while remaining auditable and regulation-aligned. NuSy's symbolic layer speaks the healthcare semantic stack natively — UMLS, SNOMED-CT, LOINC, RxNorm, ICD-10, HL7 FHIR, and FHIR-CPG for computable clinical knowledge — the foundation for clinical reasoning you can actually certify.