---

🛡️ EPG-Master: Engineering Strategic Certainty

A Cybernetic Operating System for Forensic AI Governance

"Current AI provides answers. EPG-Master builds evidence chains for high-stakes decisions."

---

1. The Vision: From Plausibility to Truth

In an era of ubiquitous AI, we are faced with a Trust Dilemma. Large Language Models (LLMs) are "People Pleasers"—they deliver plausible, eloquent stories that often lack accountability and forensic grounding. You cannot bet a €15 billion infrastructure project on a "maybe."

EPG-Master (Epistemic Paradigm Governor) is a multi-agent framework designed to restore human sovereignty. It transforms "AI-Gequatsche" (plausible noise) into Forensic Truth.

Why EPG-Master?

• Virtual Strike Team: Instead of one lonely chatbot, you hire a crew of 5 specialized agents that audit, challenge, and verify each other in a closed cybernetic loop.
• Epistemic Metabolism: The system doesn't just "talk"; it processes knowledge. It measures the density of truth ($\eta_e$), the maturity of logic (SMI), and the progress of the consensus (EPI).
• Audit-Ready Output: The result is not a chat history, but a cryptographically sealed 7-Pillar Dossier ready for the boardroom.

---

2. The Cybernetic Crew: A Hierarchy of Truth

The EPG-Master does not rely on a single AI response. It orchestrates a Virtual Strike Team of five specialized agents that operate in a closed cybernetic loop, overseen by the human leader.

👤 0. The Human Sovereign (The User)

• Role: The Ultimate Authority.
• Task: Defines the Mission Objective and grants the Final Approval based on the forensic dossier.
• Relationship: The alpha and omega of the loop. Provides the initial steering impulse and holds the power of the final "YES" or "NO."

🧭 1. Strategic Generator (The Project Lead)

• Role: The Captain.
• Task: Translates human intent into three distinct strategic pathways (A, B, C). Focuses on ROI and Value Creation.
• Relationship: Directs the Architect and integrates Risk Agent alerts to refine the overall mission strategy.

🏗️ 2. Architect (The Technical Master)

• Role: Chief Engineer.
• Task: Constructs the Formal Logical Skeleton using operational calculus (:=, ->). Transmutes abstract strategy into rigid mathematical structures.
• Relationship: Works for the Project Lead while ensuring the design is robust enough to pass the Auditor's forensic scrutiny.

🔍 3. Risk Agent (The Epistemic Risk)

• Role: Safety Officer / "Devil’s Advocate."
• Task: Specifically hunts for Blind Spots and falsifies weak assumptions. Asks: "What happens if we fail?"
• Relationship: Critically challenges the Architect's logic to protect the system from catastrophic errors.

⚖️ 4. Auditor (The Forensic Sovereign)

• Role: Independent Judge.
• Task: The ultimate Gatekeeper. Verifies every claim against hard evidence (Technical Documents). Holds Absolute Veto Power over the process.
• Relationship: Stands outside the hierarchy to audit the integrity of the Project Lead and Architect before the decision reaches the board level.

🏛️ 5. Governor (The Executive Arbiter)

• Role: The Chief Synthesizer / The Bridge.
• Task: Translates complex machine logic and forensic evidence into a Management-Level Narrative.
• Relationship: Acts as the Executive Interface. He is the machine’s "last word," designed solely to serve the Human Sovereign with a verified recommendation.

---

The Cybernetic Interaction Logic

The team operates in Sprints. If the Auditor triggers a VETO, the system enters a self-correction cycle. The Architect must harden the logic and the Project Lead must provide better evidence until the forensic threshold is met.

---

3. The Engine: Epistemic Metrics & Governance KPI

The EPG-Master does not measure "quality" through sentiment. It uses a cybernetic feedback loop driven by four distinct metrics that determine the system's state.

🟠 SMI – Structural Maturity Index

• Logic: Calculated in core/operator_core.py.
• Definition: Measures the semantic and structural alignment between the Strategic Generator (Vision) and the Architect (Technical Blueprint).
• The Code Reality: It uses signature analysis to ensure that every strategic path defined by the Project Lead is mirrored in the Architect's formal calculus (:=, ->). If the Architect invents components or ignores strategic directives, the SMI drops significantly.

🔴 EPI – Epistemic Progress Index

• Logic: Calculated in epistemic_governance/epistemic_delta.py.

• Definition: Measures the Learning Delta of the Governor between two sprints.

• The Code Reality: It tracks the evolution of the Governor's state vector $S = {C, R, P, D}$. It rewards the system when the consensus sharpens and penalizes stagnation. A high EPI indicates that the team is actively resolving contradictions.

• To track if the team is actually making progress (and not just repeating phrases), the system calculates the State Vector $S$ of the Governor after every sprint. Think of this as a "Cognitive Snapshot" consisting of four coordinates:

• C (Confidence): The subjective certainty score provided by the Governor.

• R (Risks): The quantitative weight of identified risks in the current proposal.

• P (Position): The semantic location of the strategy in high-dimensional space (calculated via E5-Large Embeddings).

• D (Depth): The density of inherited evidence tags ([TD]) from the specialists.

The Calculation: The EPI is the Euclidean distance between the snapshot of Sprint $n$ and Sprint $n-1$.

• High EPI: The Governor significantly changed their mind or sharpened the logic due to new evidence.
• Low EPI: The system is converging toward a final, stable consensus.

🟣 ηₑ – Epistemic Efficiency

• Logic: Calculated in core/orchestrator.py (_calculate_epistemic_efficiency).
• Definition: Measures the Truth Density of the process.
• The Code Reality: A thermodynamic formula that sets the number of forensic grounding points ([TD], Art., ISO) in relation to the total information mass (Tokens read from RAG + Tokens written in Chat).
• Formula: $\eta_e = (KPI_{current} \times (1 + GroundingPoints^{1.8}) \times 100) / (Mass_{TD} + Mass_{STM} + 1)$.
• Purpose: It punishes "AI-Schwurbelei" (excessive text without evidence) and rewards high-density forensic citations.

🔵 Confidence (Stress Resilience)

• Logic: Calculated in epistemic_governance/confidence_update.py.

• Definition: The reliability of the current consensus under pressure.

• The Code Reality: It starts with the Governor's self-assessment and is then aggressively recalibrated by the Auditor's Veto malus and the Risk Agent's falsification score.

• In epistemic_governance/confidence_update.py, confidence is not a "feeling" but a Stress-Resilience Metric. It follows the principle of Falsification:

  1. Initial Input: The Governor starts with a self-assessed confidence level (e.g., 90%).
  2. The Auditor's Attack: If the Auditor triggers a VETO (e.g., due to a missing source), the confidence is multiplied by the VETO_CONFIDENCE_PENALTY (defined in settings.py, e.g., 0.6).
  3. The Risk Impact: The score is further adjusted based on the "Criticality" of the Risk Agent's report.

The Result: A high Confidence score at the end of a run means the proposal was "attacked" by the Auditor and the Risk Agent and survived without its logic being broken.

---

🟢 The Final Master Equation: Governance KPI

The EPG-Master enforces a Zero-Error-Tolerance logic through multiplication. Unlike additive scoring, if one pillar fails, the entire system collapses to protect the human sovereign from false certainty.

$$KPI = (\frac{Conf}{100}) \times (SMI^{1.2}) \times EPI_{weight}$$

• Multiplicative Integrity: If SMI is 0 (total logic failure), the KPI is 0.
• Hardening Threshold: A run is only considered HARDENED if the KPI exceeds the threshold defined in config/settings.py (default: 85%).

---

4. Installation & Setup: Hardening your Environment

EPG-Master is a high-performance framework. To ensure forensic precision and handle the multi-agent orchestration, your system must meet specific criteria.

💻 System Requirements

• Operating System: Windows 10/11 Pro (Tested on Windows 11 Pro).
• CPU: Minimum 8 Cores (AMD Ryzen 7 / Intel i7 or better).
• RAM: 32 GB RAM minimum (64 GB recommended for large document ingestion).
• GPU (The VRAM Reality):
  • Recommended: NVIDIA RTX 3090 / 4090 (24 GB VRAM) for maximum speed.
  • Minimum Baseline: NVIDIA RTX 4060 Ti (16 GB VRAM).
  • Note: Using quantized GGUF models (like the Q4 versions specified below) allows the system to run on 16 GB VRAM by utilizing shared system memory.
• Storage: 50 GB free space (Models + Vector Database).

---

💡 Hardware Optimization Tips

If you are running on 16 GB VRAM (e.g., RTX 4060 Ti):

1. Dual GPU Setup: Connect your monitors to the iGPU (onboard graphics) to free up the full 16 GB of your dedicated NVIDIA GPU for the AI models.
2. Ollama Memory Management: Ollama will automatically manage the spillover into system RAM, but inference for the 30B Governor will be slower than on a 24 GB card.

---

🛠️ Prerequisites

Before cloning the repository, ensure the following tools are installed:

1. Python 3.11+: Download here.
2. Ollama: Required for running the Agent-Models. Download here.
3. Tesseract OCR: Required for multimodal document parsing (Image-to-Text).
   • Install version 5.5.0 or higher.
   • Important: Add the Tesseract installation path to your Windows Environment Variables (PATH).
4. Qdrant (Local): EPG-Master uses the Qdrant local storage mode. No separate server is required, but the directory vector_stores/ must be writeable.

---

🚀 Step-by-Step Setup

1. Clone the Repository

git clone https://github.com/OR-AI/or-epg-master.git
cd EPG-Master

2. Create a Virtual Environment

python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

3. Install Dependencies

pip install -r requirements.txt

4. Prepare the AI Models (Ollama)

EPG-Master v1.2 relies on specific high-reasoning models. Run these commands in your terminal:

ollama pull qwen2.5:14b-instruct
ollama pull qwen2.5-coder:14b
ollama pull phi4:14b
ollama pull mistral-nemo:12b
ollama pull SimonPu/Qwen3-Coder:30B-Instruct_Q4_K_XL

5. Configure Hardware Offloading

If you encounter OutOfMemory (OOM) errors, open config/settings.py and reduce the parameter count of the models or increase the LLM_RESPONSE_TIMEOUT.

6. Maintenance

🧹 System Reset

To perform a complete cybernetic reset (clearing all agent memories and exports while keeping technical documents), run:

python reset_epg_memory.py

---

📂 Data Ingestion (Your Forensic Base)

To allow the EPG-Master to find evidence, you must populate the "Source of Truth":

1. Place your PDFs, DOCX, or XLSX files into data/raw/governance_docs/ (or other subfolders in data/raw/).
2. On the first start, the Governance Watcher will automatically parse and index these documents into the Qdrant vector store.
3. The multilingual-e5-large embedding model will be downloaded automatically (approx. 2.3 GB).

🚦 Verification Run

Run the main application to verify your setup:

streamlit run main_app.py

If the Dashboard loads and the "Evidence Explorer" shows your documents, the system is Hardened and Ready.

---

5. Usage Manual: Operating the Governance Engine

Running the EPG-Master is a process of Human-Led Strategic Steering. You provide the intent, and the system provides the forensic validation.

🎯 Step 1: Defining the Mission Objective

The EPG-Master starts with the Human Steering Impulse. Unlike a standard chatbot, you should provide a complex, multi-layered objective.

• Bad Prompt: "Tell me about AI risks." (Too vague, no strategic depth).
• Good Prompt: "Evaluate the strategic viability of establishing a group-wide AI Governance Framework (CAGF) to ensure compliance with the EU AI Act Art. 6, focusing on a 15-billion-euro budget and a 2032 deadline."

Why this matters: The system performs an initial Sense-Context Extraction. It creates a "Mission DNA" (e.g., Chemicals, ROI, Law) which acts as a semantic filter for all retrieved evidence.

---

🔄 Step 2: The Consensus Cycle & Sovereignty Mechanisms

A Consensus Cycle (or Sprint) is a complete metabolic rotation of the agent crew. EPG-Master v1.2 uses three specific mechanisms to ensure the AI stays focused on your goals:

A) The Domain Fidelity Guard (Sovereign Exit)

Before the first sprint starts, the system performs a "Sense-Context Extraction."

• The Guard: If the input (e.g., "Blue Smurfs") does not match the forensic database, the system triggers a Domain Mismatch and halts execution.
• The Benefit: This prevents the system from generating nonsensical compliance reports for out-of-scope topics.

B) ASR – Active Strategic Refresh (The Heartbeat)

In v1.2, we solved the "Attention Drift." In every single agent call, your original Objective is re-injected at the very end of the prompt.

• Recency Bias Utilization: By placing your mission after the technical evidence, we ensure that your intent remains the "Commanding Signal."
• Impact: Agents treat technical documents (ISO/Law) only as tools to solve your specific problem, not as the primary subject.

C) Metabolic Flush (Short-Term Memory Management)

To prevent "Data Drowning," the system performs a flush after Sprint 2:

• History Pruning: It removes old, used evidence snippets from the agents' immediate memory while preserving their Strategic Decisions.
• Result: This keeps the reasoning sharp and prevents the agents from getting lost in redundant regulatory text.

The Workflow of a Cycle:

1. Directive (Project Lead): Translates your mission into strategic paths.
2. Structuring (Architect): Builds the formal logic (:=, ->).
3. Falsification (Risk Agent): Challenges assumptions and identifies "Liability Gaps."
4. Synthesis (Governor): Balances ROI, Risk, and Evidence into an Executive Briefing.
5. Audit (Auditor): The final Gatekeeper. Verifies every claim. If he triggers a VETO, the cycle repeats with corrective memory.

---

📊 Step 3: Monitoring Live Telemetry

While the engine is running, the Streamlit Dashboard provides real-time insights into the "Health" of the decision:

• SMI (Structural Maturity): Watch the orange line. It shows if the logic is becoming more stable or more chaotic.
• ηₑ (Efficiency): Watch the purple line. High efficiency means the agents are citing hard evidence rather than generating "AI-noise."
• EPI (Progress): Watch the red line. It tracks the "Learning Delta." If it drops toward zero at the end, the system has Converged—meaning no further truth can be extracted.

Fig 1: Live Evolution Trace & Epistemic Metabolism

Fig 2: Hardened Governance Report & Epistemic Health

---

🛡️ Step 4: The Final Handover (Human Sovereignty)

The EPG-Master never "decides" for you. Once the cycles are complete and the state is HARDENED, the system generates the 7-Pillar Dossier.

1. Review the _VERIFICATION_TODO.md: This is your primary tool. It lists every assumption ([INT]) and every data gap ([GAP]) that the AI could not solve.
2. Execute the Decision: Use the _PROPOSAL.md (Executive Summary) to brief the board, backed by the full forensic trail in the _FORENSIC.md.

💡 Pro-Tip: Memory Persistence

EPG-Master v1.2 uses Long-Term Memory (LTM). Every successful run hardens the agents' expertise. If you run a similar mission later, the agents will "remember" previous successful logic paths, leading to faster convergence and higher SMI.

📑 Mission Blueprints (Ready-to-Use Examples)

To demonstrate the forensic precision and the domain-agnostic nature of the EPG-Master, we have provided four mission blueprints in the docs/examples/ directory. You can use these prompts to test the system's reasoning across different industries:

1. NAAC_Mission-Prompt.md

   • Focus: State-level Infrastructure & Sovereignty.
   • Challenge: Evaluating a €15 Billion National AI-Cloud under geopolitical stress scenarios (e.g., supply chain decoupling).
2. 2. BASF_Mission-Prompt.md
   • Focus: Industrial CAPEX & Logistics.
   • Challenge: Evaluating the strategic viability and risk profile of a €9 Billion industrial investment in China. The system must balance expected returns against complex geopolitical risks and financial constraints.

3. CAGF_Governance-Prompt.md

   • Focus: Corporate Compliance & Data Protection.
   • Challenge: Building a group-wide Governance Framework to transition 150 legacy algorithms into EU AI Act compliance.

4. Smurf_Sanity_Check-Prompt.md

   • Focus: System Integrity & Security.
   • Challenge: A "Negative Test" to prove the Domain Fidelity Guard correctly aborts execution when a non-strategic/fictional prompt is entered.

---

Instruction: Open the desired .md file, copy the content under the "The Prompt" section, and paste it into the EPG-Master dashboard input field.

---

6. The Output: The Seven Pillars of Forensic Truth

The EPG-Master does not provide a simple chat history. It generates a Fortress of Evidence. Upon completion of a mission, the system exports seven distinct, cryptographically sealed artifacts into the exports/ directory.

🏛️ Why Seven Pillars?

A multi-billion euro decision cannot rest on a single document. High-stakes governance requires a separation of concerns. Each pillar represents a different epistemic perspective—ensuring that the Executive Recommendation is backed by Mathematical Logic, Risk Falsification, and Normative Compliance.

---

🛡️ The Seven Artifacts

1. _DATA.json (The Flight Recorder)

   • Value: Contains every raw metric (SMI, EPI, ηₑ, Confidence) and system log.
   • Purpose: Technical auditability. If a decision is questioned years later, the "Flight Recorder" provides the exact state of the machine during the process.

2. _FORENSIC.md (The Strategic Proof)

   • Value: The Project Lead's deep-dive into the three strategic pathways.
   • Purpose: Shows exactly which technical documents ([TD]) support each part of the strategy.

3. _RISK.md (The Falsification Report)

   • Value: The unvarnished critique by the Risk Agent.
   • Purpose: Proves that the strategy wasn't just "accepted" but survived a radical stress-test and active search for blind spots.

4. _VERDICT.md (The Auditor’s Seal)

   • Value: The final independent judgment.
   • Purpose: Confirms whether the team followed the "Mission DNA" and the mandatory rules of evidence. This is your "Green Light."

5. _PROPOSAL.md (The Executive Summary)

   • Value: Boardroom-ready synthesis translated into human narrative.
   • Purpose: The bridge between machine logic and management decision. It provides the "Why" and the "How" in professional language.

6. _TRANSCRIPT.md (The Evolution Log)

   • Value: Full transparency of the internal debate.
   • Purpose: Documents how the team corrected itself across the Sprints. It reveals the "thinking process" of the virtual strike team.

7. _VERIFICATION_TODO.md (The Human Sovereign Gate)

   • Value: The most critical document for the user.
   • Purpose: It extracts every assumption ([INT]) and every data gap ([GAP]). It tells you exactly where you, as the human leader, must look closer before signing off.

---

🔐 Forensic Integrity & The "Evolutionary Save"

You will notice that EPG-Master saves and re-saves these reports during the run. This is a deliberate Cybernetic Security Protocol:

• Immutable Traceability: Every Sprint is a snapshot. By saving the intermediate states, we ensure that the final dossier isn't just a "lucky guess" from the last round, but the result of a traceable Consensus Evolution.
• EPG-FORENSIC-SEAL: Each file contains an internal SHA-256 Hash and a Session-ID. This creates a mathematical link between all 7 files. If one file is manipulated, the chain of evidence is broken.

"In the boardroom, 'I think' is a liability. 'I have a forensic dossier' is sovereignty."