The Problem with Flat Knowledge

Most AI systems store knowledge as flat text — embeddings in a vector database, tokens in a context window, paragraphs in a document. This works for retrieval. It fails for verification.

The KnowledgeObject (KO) schema was designed to solve a specific problem: how do you store a unit of knowledge such that its coherence with other knowledge can be computed, verified, and tracked over time?

Schema Definition

A KnowledgeObject is a JSON document with the following structure:

{
  "ko_id": "KO-2026-0128",
  "version": 2,
  "type": "empirical_claim",
  "content": "Bioelectric patterns in Xenopus laevis can be modulated to induce eye formation at arbitrary body locations",
  "source": {
    "doi": "10.1242/dev.XXXXXX",
    "author": "Levin et al.",
    "year": 2024
  },
  "layers": ["L06", "L07"],
  "confidence": 0.87,
  "coherence_links": ["KO-2026-0045", "KO-2026-0092"],
  "tags": ["bioelectric", "morphogenesis", "levin"],
  "created": "2026-02-15T10:30:00Z",
  "modified": "2026-02-15T10:30:00Z"
}

Key Design Decisions

Typed content. Every KO has a type field: empirical_claim, theoretical_proposition, definition, observation, inference, axiom, or meta. This allows coherence computation to weight connections appropriately — an axiom contradicting an empirical claim is weighted differently than two observations in tension.

Explicit layer assignment. The layers field maps the KO to the 18-layer ontological architecture. A claim about quantum decoherence maps to L01-L02. A claim about social network dynamics maps to L13-L14. Cross-layer KOs — the most valuable kind — explicitly declare which layers they bridge.

Coherence links. Each KO declares which other KOs it depends on, supports, or tensions with. This creates a directed graph that can be analyzed using topological data analysis to detect inconsistencies, redundancies, and gaps in the knowledge base.

Storage: Pinecone Vector Database

KOs are embedded and stored in Pinecone, enabling both semantic search (find similar KOs) and structured filtering (find all KOs in layer L07 with confidence > 0.8).

embedding(KO) = E(content) ⊕ E(type) ⊕ E(layers)

The embedding is a concatenation of content embedding, type embedding, and layer embedding, allowing queries to match on any combination of semantic similarity and structural properties.

Coherence Computation

The global sheaf coherence score is computed by treating the knowledge graph as a cellular sheaf and measuring the Laplacian:

H = Σᵢⱼ ‖F(i→j) · x(i) - x(j) ‖²

Where F(i→j) is the restriction map from KO i to KO j, and x(i) is the section value at node i. Low H indicates high coherence — the knowledge base is internally consistent. High H indicates contradictions that need resolution.

This computation runs after every KO insertion, providing a real-time coherence score visible on the draken.info signal bar.

Schema implementation tracked in Pinecone namespace draken-ko-v2. OpenClaw integration pending via DRK-102.