NucleoidAI/Nucleoid

What it solves

LLMs often struggle with hallucinations and inconsistent reasoning because natural language is unstructured. Nucleoid addresses this by providing a structured, logic-based language that allows LLMs to reason over explicit entities and relationships rather than implicit patterns, leading to more reliable and near-deterministic results.

How it works

Nucleoid is a neuro-symbolic AI system that combines a Rust-based declarative runtime with a fine-tuned LLM. It uses a minimally tokenized syntax to represent logic and a declarative execution model where assignments are treated as persistent relationships rather than one-time instructions. This creates an executable knowledge graph (a "World Model") where changes to a value automatically propagate to all derived facts, ensuring the state never becomes stale or contradicts its own rules.

Who it’s for

Developers building complex AI systems that require strict logical consistency, explainable decision-making, and the ability to integrate data-driven learning from neural networks with rule-based symbolic reasoning.

Highlights

• Near-Deterministic Reasoning: Uses structured logic to reduce hallucinations.
• Executable Knowledge Graph: Maintains a live world model where rules are automatically enforced across all instances.
• Token-Efficient Syntax: Designed specifically for LLMs to minimize token usage during logic representation.
• Automatic State Propagation: Ensures derived values are always up-to-date without manual state management.