KlastroHeron Decision Intelligence

On-Premise Multi-Agent Decision Layer

First Public Release of 4CM: 2026-03-09
First Public Release of Prudentia Branch: 2026-05-19
Last Updated: 2026-07-02

"Four models that never agree — until they do."

Prudentia (Latin: wisdom) is the first commercial release of the 4 Councilmen Model (4CM): a document & web grounded advisory system in which four structurally independent AI agents — each locked to a single orthogonal perspective (surveillance, ethics, audit, transparency) — analyze the same question without ever negotiating. A fixed semantic judge measures convergence. Agreement is never assumed; when it happens despite the agents' hostility to one another's logic, the phone rings.

Built for organizations that must demonstrate independent review, human oversight support, and documented decision trails for AI-assisted decisions — deployable fully on-premises and air-gapped, with no data ever leaving your infrastructure.

Version 2.1.1 · PhD Dissertation, 2011 · Hybrid Prototype, 2026 📖 Full User Manual (PDF)

---

Why single-model review fails

Every production AI system today is optimized for convergence: trained on human feedback, aligned to consensus, tuned for agreement. That makes a single model — or an ensemble of similar models — structurally blind to its own correlated failure modes. Regulators have noticed: high-risk AI frameworks increasingly require effective oversight, risk identification from multiple angles, and traceable decision records.

4CM addresses this at the architecture level:

• Radical orthogonality. Each agent operates from a fixed, mutually hostile perspective. No shared context, no negotiation channel.
• Non-convergent by design. Agents never move toward consensus. They remain fixed at their extreme positions on a torus field; the convergence point moves toward them only as their independent conclusions and reasoning align. A singularity is recognized only when that point reaches the peak region all four extremes share.
• Independent judging. Semantic scoring is performed by a single fixed judge model, separate from the agents, ensuring cross-round score comparability.
• Evidence-grounded. Conclusions are grounded in uploaded documents (PDF, DOCX, XLSX, CSV, PPTX, images) and web search — not free-floating generation.
• Human-in-command. The fifth response (common conclusion) is a recommendation, never a binding determination. The operator may reject, modify, or override any output at any time.

How It Works

Your Question + Documents + Web Search
        ↓
┌─────────────────────────────────────┐
│  SENTINEL       │  ETHIKOS          │
│  (surveillance) │  (ethics)         │
├─────────────────┼───────────────────┤
│  AUDITOR        │  HERALD           │
│  (audit)        │  (transparency)   │
└─────────────────────────────────────┘
        ↓
   Fixed Semantic Judge
        ↓
   Singularity? → The phone rang.
   No convergence? → The phone did not ring.

Under the hood: agent responses are embedded (Sentence Transformer, local-only), scored by the judge on conclusion- and reasoning-convergence, and mapped onto the torus field f(x,y) = (x⁴+y⁴)·e^(−(x⁸+y⁸)) from the 2011 dissertation. Four peaks, one deep central well; a closed consensus ring becomes topologically possible only when all four opposed perspectives are present. See the User Manual, §2.3, for the full mathematics.

Provider Modes & Data Flow

Mode	Agents	Judge	Data flow
All Local LLM	Local model ×4 (Ollama / vLLM / llama.cpp)	Same local model (fixed)	No data leaves your infrastructure
Round-robin	Claude → Grok → Claude → Grok	Grok (fixed)	Anthropic + xAI
All Claude	Claude ×4	Grok (fixed)	Anthropic + xAI
All Grok	Grok ×4	Grok (fixed)	xAI
Per-agent	User-assigned per agent	Grok (fixed)	Selected external providers

In every mode: embeddings, torus computation, and decision logs stay on the customer's servers. Nothing is ever transmitted to Klastrovanie.

Security

• Memory-only API keys (v2.0.4+). Keys entered via the UI live in process memory only — never written to disk, never printed to logs, cleared on container restart.
• Secret manager integration. Docker Secrets, HashiCorp Vault, AWS/Azure/GCP secret managers, or host env vars for persistent enterprise key management.
• Ephemeral uploads. Documents are parsed server-side and auto-deleted after TTL (default 1 hour); no volume mount.
• Internal-only networking. Port 80 via nginx, no public exposure required. Administrator security checklist in Manual §5.4.

Regulatory Posture

As deployed, the Software itself (per Manual §6 — not legal advice):

Requirement	Status
EU AI Act — GPAI provider obligations (Art. 51–56)	Not applicable — no foundation model provided
EU AI Act — High-risk obligations (Art. 6–49)	Not applicable as deployed — limited-risk classification
EU AI Act — Transparency (Art. 50)	Satisfied — UI discloses AI-generated content
GDPR	Internal LLM Mode: all data stays in customer infrastructure; Klastrovanie is neither controller nor processor. External API Mode: customer establishes DPAs with providers
Cyber Resilience Act (2024/2847)	Committed: vulnerability disclosure channel, SBOM delivery, security patch notifications within support SLA

⚠️ If deployed for Annex III high-risk purposes (employment decisions, credit scoring, etc.), the deploying organization assumes the corresponding provider/deployer obligations under Article 25.

How Prudentia supports your governance obligations:

Framework	How Prudentia helps
EU AI Act Art. 9 (Risk management)	Four orthogonal reviewers stress-test the same decision; divergence surfaces unresolved risk before action
EU AI Act Art. 12 (Record-keeping)	Structured decision records: per-agent positions, judge scores, grounding evidence, exportable PDF reports
EU AI Act Art. 14 (Human oversight)	The singularity score is a calibrated trust signal; "the phone did not ring" explicitly tells the overseer not to rely on the AI conclusion
ISO/IEC 42001 · NIST AI RMF	Repeatable, logged, multi-perspective review process that slots into an AI management system as an operational control

Features

• 4 orthogonal AI agents — local LLM, Claude + Grok hybrid, or per-agent assignment
• Real-time web search (external API mode)
• Document upload as grounding evidence — PDF, DOCX, XLSX, CSV, PPTX, TXT, images (25 MB)
• Multi-round (1–5), blind or context-aware
• Structured PDF decision reports
• Korean/English UI and output · 4 themes
• Custom scenarios via plain .txt files — no code, no restart
• One-command Docker deployment · BYOK

Quick Start

Requirements: Docker Engine 20+ with Compose V2 · Ollama (local mode) or API keys (Claude/Grok) · for local LLM: GPU with 8 GB+ VRAM recommended

git clone https://github.com/Klastrovanie/4councilmen.git
cd 4councilmen
git checkout Prudentia

cp the.env.example .env
# Edit .env — key variables:
#   INTERNAL_LLM_BASE_URL   (Ollama: http://host.docker.internal:11434/v1)
#   INTERNAL_LLM_MODEL      (e.g., llama3.1:8b — as shown in `ollama list`)

docker compose up -d --build

Access at http://localhost. API keys (if using external providers) are entered via the UI — memory-only, never written to disk. Full configuration reference: Manual §3–4.

Custom Scenarios

Scenarios live in angry_agents/ (Docker volume). Edit .txt files directly — changes apply immediately, no restart:

angry_agents/your_scenario/
├── members.txt   # "1: SENTINEL" per line
├── title.txt     # Display title
├── query.txt     # Default question
├── risk.txt      # normal / high
└── 1.txt … 4.txt # System prompt per agent

Pre-built scenarios (pharma, outbreak, government, M&A, Oppenheimer, whistleblower, key talent) are created on first run. Scenarios marked [MOCK RESEARCH SCENARIO] are for research and simulation only — not validated for clinical, financial, legal, or regulatory decision-making. Model selection for specialized domains (including judge/agent policy fit) is an operator responsibility; see Manual §8.

API

FastAPI backend with SSE streaming. Key endpoints: GET /health, GET /fourCM/agents, POST /fourCM (SSE analysis stream), POST /fourCM/validate (orthogonality check), POST /fourCM/upload, key management under /fourCM/keys/* (values never returned). Full reference: Manual §7.

Architecture

docker compose up
├── nginx :80
│   ├── React UI (Prudentia, FourCM.tsx)
│   └── FastAPI SSE proxy
└── backend :8000 (internal)
    ├── fourCM_router.py     SSE streaming + key management + scenarios
    ├── fourth_cm_engine.py  Semantic judge
    ├── orthogonal_agents.py Agent calls (local / Claude / Grok)
    ├── torus_math.py        Convergence f(x,y), singularity detection
    └── document_parser.py   Upload parsing + context injection

Product Suite

Prudentia is the first release in the Klastrovanie advisory suite (more coming soon):

Product	Latin	Role
Prudentia	wisdom	Document & web grounded advisory (this)

---

Licensing

Prudentia / 4CM is dual-licensed.

Open source (AGPL-3.0). You may use, modify, and deploy this software — including commercially — under the terms of the GNU Affero General Public License v3. Note that AGPL's copyleft applies to network use: if you offer this software, or a product that incorporates it, as a service over a network, you must make the complete corresponding source of that combined work available under AGPL-3.0.

Commercial license. For organizations that cannot meet AGPL obligations — e.g., embedding 4CM in proprietary products or SaaS platforms without source disclosure, or on-premise deployment free of copyleft obligations — Klastrovanie Co., Ltd. offers commercial and OEM licenses, including deployment support and integration assistance.

📧 Commercial licensing & OEM inquiries: contact@klastrovanie.com

This code is released to encourage collaboration across AI systems — not competition. The goal is shared solutions, not shared resources.

Copyright © 2026 Klastrovanie Co., Ltd. All rights reserved.

Live Demo

https://klastrovanie.github.io/4councilmen/

Theory & Research Background

Original theory: 4 Councilmen Model (4CM) — Author's PhD Dissertation, 2011 ACM Digital Library: https://dl.acm.org/doi/book/10.5555/2231522

The design premise: think of a politician, a religious leader, an environmental activist, and a labor organizer at the same table. They disagree by identity and conviction. Yet when all four, independently, point to the same conclusion — that conclusion is worth listening to.

Imagine a phone booth designed never to ring. No one has the number. In 4CM, that is the phone. When four orthogonal minds converge — it rings. When the question is trivial or not ready — it doesn't. The phone did not ring — that is also an answer.

"No model. No weights. No storage. Born from non-convergence. Already gone."

---

Prudentia is a multi-perspective decision-support and simulation tool. It does not automate any decision; all final decisions rest with the human operator. All agent outputs and the fifth response are AI-generated content. It is not a certification, legal advice, or a substitute for the human and organizational measures required by applicable AI regulation.