What NuSy Is
NuSy is a platform for building AI beings — autonomous entities that learn, reason, and grow over time. Unlike traditional AI systems that process inputs and produce outputs, a NuSy being has:
- Persistent memory — a knowledge graph that survives between sessions
- Structured learning — it studies curricula, not training data
- Explainable reasoning — every answer traces back to its sources
- A lifecycle — beings awaken, learn, converse, sleep, and dream
The Y-Layer Architecture
A being's knowledge is organized into seven semantic layers, each building on the last:
| 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, extraction specifications |
| Y3 | Episodic | Memories of conversations and experiences |
| Y4 | Journal | Opinions, syntheses, formed beliefs |
| Y5 | Procedural | Learned skills and workflows |
| Y6 | Metacognitive | Self-awareness — what the being knows it knows (and doesn't) |
How Beings Learn
Beings don't train like ML models. They go to school.
The learning process follows a 10-step sushi-grade protocol: Stop, Think, Question, Look Up, Form Opinion, Extract, Validate, Reread, Iterate, Move On. Each step mirrors how a careful student would study domain material — not a bulk data ingestion, but a deliberate act of understanding.
Curriculum-Driven Expertise
A being's domain expertise comes entirely from its curriculum — not from code changes. The same codebase can produce:
- A medical being studying SNOMED CT ontologies and clinical guidelines
- A legal being studying LegalRDF and case law
- An engineering being studying IEEE standards and specifications
Domain expertise = curriculum, not code.
What Makes This Different
Provenance Tracking
Every fact in a being's knowledge graph carries provenance: where it came from, when it was learned, what evidence supports it, and how confident the being is. This isn't metadata — it's the foundation of trustworthy AI.
Crystallization
When an LLM generates a response, the answer normally evaporates. NuSy's crystallization pipeline captures validated claims from LLM outputs and converts them into permanent, provenance-tracked triples in the knowledge graph. Knowledge grows; hallucinations don't.
Coverage Gating
A being knows what it doesn't know. Before answering a question, the reasoning engine checks curriculum coverage for the relevant topic. If coverage is below threshold, the being says so — rather than confabulating an answer.
Graph-First Architecture
All knowledge lives in the graph during reasoning. Files are lazy persistence, not the source of truth. This means reasoning operations are graph queries — fast, composable, and transparent.
V12: Cognitive Signal Fusion
The latest architecture (V12) transforms how beings make decisions — from a sequential pipeline to parallel signal voting, inspired by cortical columns and clinical decision models (FHIR-CPG).
In V11, each cognitive assessor (field of view, fractal index, novelty detector) ran sequentially. V12 fires them in parallel, collects their votes as CognitiveSignals, and fuses them through a weighted voting matrix into a single CognitiveDecision — auditable, explainable, and logged to episodic memory.
| V11 | V12 |
|---|---|
| Sequential pipeline | Parallel signal voting |
| Implicit routing | Auditable CognitiveDecision |
| Fixed thresholds | Learnable per-being weights |
| Adding signals requires rewiring | New signal = one row in weight matrix |
| No decision audit trail | Full evidence chain in Y4 journal |
Key V12 components:
- SignalFusion — parallel assessor voting with temporal decay and weighted fusion
- CognitiveDecision — symbolic, auditable decision structure with ranked options and evidence
- WeightLearner — gradient-free weight optimization from decision quality feedback
- DecisionJournal — every routing decision logged as Yurtle triples in Y4 episodic memory
- Emotional signals — extensible signal sources (frustration, confusion, curiosity detection)
All 8 V12 hypotheses validated. 252 tests passing.
V11 → V12 Comparison Results
Controlled comparison (EXP-838): same machine, same corpus, same beings — only the architecture changes. Measured using the AGI Certification Framework (ACF), which scores 9 dimensions of understanding.
| Being | V11 Score | V12 Score | Δ | V11 Level | V12 Level |
|---|---|---|---|---|---|
| Toddler | 47.9 | 56.6 | +8.7 | ACF-3 | ACF-3 |
| Gradeschool | 58.8 | 63.3 | +4.5 | ACF-3 | ACF-4 |
| Middleschool | 47.6 | 63.3 | +15.8 | ACF-3 | ACF-4 |
| Highschool | 49.2 | 63.3 | +14.1 | ACF-3 | ACF-4 |
Three of four education beings upgraded certification level. The biggest gains come from compositional generalization (+25.7pp in toddler) and factual grounding (+60pp in middleschool) — exactly the dimensions where parallel signal voting was expected to help.
No dimension regressed. V12's graph sizes are ~2x V11's with identical source material, reflecting richer interconnections from signal fusion.
The Technology
- Knowledge representation: RDF triples in semantic graphs (Polars + rustworkx on CPU, cuDF + cuGraph on GPU)
- Persistence: Yurtle format — Markdown with TTL frontmatter, Git-backed, human-readable
- Reasoning: UnifiedReasoningBrain with V12 SignalFusion (parallel voting) and Y-layer access policies
- Learning: CQ-guided extraction, LLM enrichment, ontology validation, crystallization
- Certification: AGI Certification Framework (ACF) — 9 dimensions, 66 measures, 6 certification levels
- Testing: Three-tier TDD/BDD — unit tests, integration tests, live being tests via CLI
Current Status
NuSy is in active development. V12 (Cognitive Signal Fusion) is the current architecture — merged to main, validated across four education beings, with 890+ expeditions completed and 120+ research hypotheses tested. Seven research papers are ready for submission to peer-reviewed venues, with two (Papers 104 and 113) actively submitting.