CHERENKOV System Architecture & Cognitive Routing¶
1. The Multi-Agent Swarm¶
CHERENKOV utilizes the AgentGovernor pattern to prevent context pollution and rate-limit cascades. Cognitive load is divided among specialized nodes.
| Agent Name | Node Designation | Underlying Engine | Role | Data Access Level |
|---|---|---|---|---|
| TENSOR | (The Strategist) | Groq Llama 3.1 8B | Cloud strategic planning & attack chain schema | Sanitized breadcrumbs only via ABLATION |
| KINETIC | (The Executor) | Local Ollama (Llama 3.2 3B) | Local execution and issue triage | Full raw data access (behind MEISSNER) |
| Tokamak | (The Validator) | Python TOKAMAKed Engine | Proof validation and execution | Local tokamaked environment |
| AEGIS | (The Arbiter) | Local Llama 3.1 8B | Inter-agent arbiter & AIMD Circuit Breaker | Sanitized context |
| LATTICE | (The Memory) | Qdrant + Embeddings | Memory, dialect-RAG, and CVE knowledge base | Local CVE vectors |
2. Arabic AI Capabilities¶
Standard LLMs fail in the MENA region due to diglossia and right-to-left (RTL) formatting. CHERENKOV bridges this gap natively: * Sequence-Tagging: Utilizes SWEET sequence-tagging for fast Grammatical Error Correction without destroying the author's authentic voice. * Dialect-to-MSA RAG: Uses syntax-aware chunking and BGE-M3 multilingual embeddings with a cross-encoder reranker to map regional dialects to formal compliance knowledge bases. * Layout-Aware OCR: Decouples spatial coordinates from text to extract data from scanned Arabic PDFs without destroying complex RTL contract formatting.
3. Resilience & Routing¶
To prevent API timeouts from cascading into system collapse, the platform utilizes Additive Increase Multiplicative Decrease (AIMD) circuit breakers across all agent communications.