---

What is Episteme?

Episteme is an extensible, contract-driven neuro-symbolic framework for extracting, constructing, and formally evaluating Theory Graphs ($\mathcal{G}_{\text{TheoryNet}}$) from German and English scientific literature.

Traditional Knowledge Graphs flatten scientific epistemology into isolated factual triples (e.g., (Einstein, BornIn, Ulm)). In contrast, scientific discourse is fundamentally organized around inferential, axiomatic, and argumentative structures.

Episteme recovers the underlying logic of science by jointly modeling core theoretical axioms, empirical observation statements, correspondence rules, epistemic justifications ($J$), and defeasible attack/support relations.

Knowledge Graph vs. Theory Graph

---

Architecture Overview

Episteme employs a layered, multi-phase architecture that progressively lifts raw document tokens into formal model-theoretic structures:

flowchart TD
    subgraph In ["Input Layer"]
        DOC["Scholarly Literature<br/>(Markdown / PDF / BibTeX)"]
    end

    subgraph L1 ["Layer 1: Provenance & Structural Anchors"]
        CHUNK["Chunking & Offset Span Mapping"]
        TOC["Section & ToC Holon Tree"]
    end

    subgraph L2 ["Layer 2: Empirical Knowledge Graph"]
        ENT["Entity & Concept Extraction"]
        COREF["Cross-Chunk Coreference & Linking"]
        REL["Local & Global Candidate Relations"]
    end

    subgraph L3 ["Layer 3: Formal Theory Net"]
        ATOM["Theory Atoms (Axioms, Laws, Hypotheses)"]
        JUST["Epistemic Justifications (Deduction & Support)"]
        ATTACK["Defeasible & Adversarial Attack Relations"]
    end

    subgraph Out ["Ecosystem & Backends"]
        NEO4J[("Neo4j Dual Graph<br/>Processing & Projection")]
        METRICS["Epistemetrics<br/>(Empirical Surplus & Coherence)"]
        STUDIO["Episteme Studio<br/>(Interactive Workbench)"]
    end

    DOC --> L1
    L1 --> L2
    L2 --> L3
    L3 --> NEO4J
    L3 --> METRICS
    L3 --> STUDIO

---

Monorepo Packages

---

Quickstart

Prerequisites

• Python 3.13+
• uv package manager
• (Optional) Running Neo4j 5.x+ instance with Graph Data Science (GDS) plugin

1. Installation

# Clone the repository
git clone https://github.com/maxoem/episteme.git
cd episteme

# Sync workspace dependencies across all packages
uv sync

2. Configure Environment

Create a .env file in the project root:

# LLM / Embedding Providers (OpenAI, LiteLLM, Ollama, or vLLM)
OPENAI_API_KEY="your-api-key"
LLM_MODEL="openai/gpt-4o-mini"
EMBED_MODEL="openai/text-embedding-3-small"

# Neo4j Graph Store (Optional - defaults to local instance)
NEO4J_URL="neo4j://127.0.0.1:7687"
NEO4J_USERNAME="neo4j"
NEO4J_PASSWORD="neo4jdbpass"
NEO4J_DATABASE="neo4j"

3. Run Pipeline Example

Execute the end-to-end extraction pipeline on bundled research texts:

uv run python packages/episteme-pipeline/examples/pipeline_full_run.py

4. Python API Usage

The pipeline can be executed programmatically from Python:

import asyncio
from episteme_pipeline.runtime.runner import run_pipeline

async def main():
    result = await run_pipeline(
        source_paths=["packages/episteme-pipeline/examples/text/teachers_expectancies.md"],
        metadata={"domain": "social_psychology", "language": "en"},
    )
    print(f"Pipeline executed successfully: Run ID {result.run_id}")

if __name__ == "__main__":
    asyncio.run(main())

---

Episteme Studio

Episteme Studio provides a run-oriented workbench for inspecting pipeline execution DAGs, verifying cache invalidation states, and interactively navigating constructed Theory Graphs.

# Launch workbench server
uv run episteme-studio

Open http://localhost:8000 to access the workbench.

---

Theory Graph Evaluation (epistemetrics)

epistemetrics can be used completely independently of the extraction pipeline to quantitatively analyze formal graph structures and epistemic topologies:

import networkx as nx
from epistemetrics import louvain_communities, page_rank

# Construct or load a theory graph
graph = nx.DiGraph()
graph.add_edge("Hypothesis_1", "Observation_A", weight=0.9)
graph.add_edge("Hypothesis_1", "Observation_B", weight=0.8)

# Compute central epistemic anchors and community partitions
pr_result = page_rank(graph)
communities = louvain_communities(graph)
print(f"Epistemic Centrality: {pr_result.scores}")

Key evaluation capabilities:

• Schurzian Unification & Tacking Resistance: Penalizes ad-hoc hypotheses and irrelevance tacking.
• Lakatosian Progressive Problemshifts: Evaluates whether theoretical modifications yield novel empirical content.
• Dung Abstract Argumentation: Computes grounded, preferred, and stable extensions under conflict.

---

Development & Standards

• Testing:

  uv run pytest
  

• Documentation: Built with Zensical (MkDocs Material engine). Preview locally with:

  uv run zensical serve
  

• Architecture Decisions: Major technical trade-offs are documented as ADRs in docs/adr/.
• Coding Standards: Strict typing and NumPy-style docstrings are required across all modules.

---

Citation & License

This project is licensed under the MIT License.

If you use Episteme or Epistemetrics in your research, please cite:

@article{maoem2026episteme,
  title={Episteme: A Neuro-Symbolic Engine for Scientific Theory Graph Construction and Evaluation},
  author={MAOEM},
  year={2026},
  url={https://github.com/maxoem/episteme}
}