From Data Jungle to Neurosymbolic Order: The Evolution of the L.U.N.A. Knowledge Graph
In modern AI architectures, knowledge is often understood as an amorphous mass: raw texts migrate into vector databases, and at runtime, reliance is placed on probabilistic embeddings and cosine similarity. For standard RAG applications, this may suffice. However, for a deterministic system like L.U.N.A. Prime, which must guarantee logical invariance, formal proofs, code verification, and physical axioms, this stochastic approach fails.
When knowledge is not hierarchically anchored and auditable, the so-called "jungle problem" looms with increasing data volume: isolated individual sentences, fragmented paths, and semantic misdirections.
After continuously ingesting over 18,000 scientific propositions, we found ourselves precisely at this turning point. This article documents the redesign of our knowledge space: from decomposing flat data stores through a 3-level ontology to employing Wodan as the gatekeeper at the ingestion portal.
1. The Problem: The Leaf Explosion and Semantic Collisions
When autonomously ingesting extensive scientific papers and preprints, two systemic errors arise without strict pre-filtering:
- Fragmented Individual Branches: Automatic extractors often leave signatures or document hashes within the topic field. The result was an inflation to nearly 4,000 individual branches, each carrying only a single sentence. Such a graph loses its orientation function; true logical pruning becomes impossible.
- Greedy Substring Misclassifications: When classifications are performed purely based on text, overlaps lead to grotesque assignments. Sentences from the Weimar Classicism or ancient mythology suddenly ended up in Signal Processing—simply because the texts spoke of "sea noise" (Meeresrauschen) or "intoxicating moments" (berauschenden Momenten), and a naive filter latched onto the technical signal "noise." Similarly, linguistic treatises on word frequencies were mistakenly assigned to electrical engineering.
In a deterministic knowledge base, such ambiguities are unacceptable. A proposition must reside exactly where it belongs contextually and ontologically.
2. The Solution: A 3-Tier Ontology on 5 Main Stems
Instead of an unstructured catch-all pool, we have migrated the entire knowledge base (18,238 verified axioms) into a strictly deterministic, three-level hierarchy:
- MINT: Differentiated by Computer Science (System Architecture & Concurrency, Shader Pipelines, Algorithms, Computer Vision), Mathematics (Analysis, Linear Algebra, Geometry), and Physics (Mechanics, Optics, Thermodynamics).
- Engineering: Focused on genuine Signal Processing (Fourier Transform, Filter Orders, Sampling) and Electrical Engineering.
- Economics & Strategy: Subdivided into Macroeconomics, Business Administration (Cost and Production Models), and formal Game Theory & Allocation.
- Linguistics: Precisely separated into German Syntax (topological field model, clause embedding, valence) and Semantics & Lexicon.
- Philosophy & Methodology: Philosophy of Science (Invariance, Falsification, Epistemology) as well as Ethics & Cultural History.
Misplacements are precluded through word boundary validation and prioritized assignment switches (Humanities and Linguistics are checked before technical filters). The tree density increased from unusable 4.6 to hundreds to thousands of substantiated doctrines per main trunk.
3. Odin as Threshold Guardian: Definition Upon Ingestion
The central architectural insight of this refactoring is: Post-facto cleanup in the production database is a waste of resources.
Until now, raw data was stored unfiltered and only cleaned later in audit runs. Effective immediately, Odin functions as an autonomous gatekeeper directly at the ingestion interface:
The Ingestion Cycle in the Hot Context
- Extraction in the Source Context: When a paper is being processed, the entire text segment resides in volatile memory. At this precise moment, the system not only generates the declarative statement but also captures its semantic embedding.
- The Three-Part Axiom Schema: A sentence is only accepted into the central knowledge base if it exists as a complete triple:
statement: The actual scientific proposition.definition: The exact definition of all terms and variables used.validity_boundary: The formal scope of validity (under which conditions the statement holds true—and where it breaks).- Deterministic Barrier: Conversational fragments, vague assertions, or incomplete notes are physically rejected and isolated.
This ensures that no unverified data enters the operational memory.
4. Visual Transparency: The Interactive Knowledge Tree
A system that learns continuously requires a transparent control instance. The growth of the knowledge space can be tracked in real-time via a D3.js-based visualization:
- Hierarchical Collapsible Tree: Branches can be expanded and collapsed infinitely via mouse click.
- Glow Coding by Depth: Quick visual orientation between the root node, main trunk, sub-discipline, and leaf.
- Real-time Branch Inspector: Selecting any node opens the stored propositions, including ID, confidence score (
0.95–1.00), definition, and validity boundaries.
Conclusion & Outlook
The transformation of the knowledge base provides significant advantages for the overall architecture:
-
Sub-Millisecond Access: Search and invariance checks execute over the cleaned index in under 2 milliseconds.
-
High Token Savings: Through precise paths, deterministic reflexes intercept over 66% of all queries directly at the database level—without needing to engage a local language model on the GPU.
-
Protection Against AI Poisoning: Wodan's threshold checking immunizes the knowledge base against creeping quality degradation.
The next milestone is the granular subdivision of the scientific inventory into dedicated branches for Astronomy & Celestial Mechanics, Materials Science, and Biophysics. Luna's world model is no longer growing uncontrollably in breadth but is branching structurally like a healthy tree.