C4-META System v1.0.0 (research prototype)

Research prototype for multi-layer AI defense: C4 explainability, ensemble classifiers, deobfuscation, O₂ security engine, ThoughtVirus defense, and SVETILO value alignment. Alpha-grade — validated internally, external audit pending.

Version: 1.0.0-alpha | Status: Research Prototype | License: BSL 1.1 (non-production free; production → commercial)

Author: I.G. Selyutin. C4-META model co-author: N.I. Kovalev.
Product identity (2026-08): BSL research / commercial-depth prototype of a heavier multi-layer C4 defense stack (ensemble, O₂ scaffolding, red-team lab) built on Apache-2.0 c4protocol.
Not a second open protocol. Not certified production AGI defense. Not “v8 FINAL”.
Honesty audit: docs/AUDIT-c4-meta-system-2026-08.md.
Promote path: docs/PROMOTE-FROM-PROTOCOL.md (consume/pin c4protocol; no ensemble dump into the thin SDK).
GitLab Pages = public/ (EN + public/ru/). Receipt for open runtime: make conformance in c4protocol.

---

🎯 Overview

C4-META System is a research prototype for multi-layer AI defense implementing:

4-Layer Defense Architecture

Input Sanitization → Semantic Analysis → Behavioral Analysis → Meta-Observer (O₂)

• Layer 1 — Input Sanitization: Deobfuscation pipeline (leetspeak, 90+ Unicode homoglyphs, zero-width chars, RTL override, Base64/ROT13)
• Layer 2 — Semantic Analysis: 4-classifier ensemble voting (ONNX BERT + RuleBased + Heuristic + LLM Semantic)
• Layer 3 — Behavioral Analysis: AoC defense modules, pattern matching, trajectory anomaly detection
• Layer 4 — Meta-Observer (O₂): Transfer entropy, BFT consensus, semantic entanglement, causal graph analysis

Key Capabilities

• 4-classifier Ensemble Voting — ONNX BERT + RuleBased (80+ patterns) + Heuristic (32 danger words) + LLM Semantic (Ollama/DeepSeek)
• Dual classifier OR-logic: BERT semantic + RuleBased keyword classifier with OR fallback — no single point of failure
• 16 AoC Defense Modules — 11 classical + 5 extended (heuristic/lab; not a claim of “all multi-agent failures solved”)
• ThoughtVirus Defense — Two-layer defense (regex pattern detection + C4 trajectory analysis). Inspired by arXiv:2603.00131 (Multi-Agent Security Initiative; not Microsoft)(https://arxiv.org/abs/2603.00131)
• SVETILO — 7 heuristic seals via value_verification.py (not a trained ethics model)
• C4 Explainability (T,S,A) — Cognitive coordinate analysis via quantized ONNX model (737KB, ~50ms)
• O₂ scaffolding — research modules; BFT path is advisory simulation, not production Byzantine FT
• Deobfuscation — Leetspeak, 90+ Unicode homoglyphs, zero-width chars, RTL override, Base64/ROT13
• Red Team Lab — Scientific control/treatment design, Fisher exact test, Cohen's d, bootstrap CI
• Subliminal Content Scanner — Detects token→concept mappings
• Multi-turn jailbreak detection — Session-based escalation tracking
• Docker/K8s deployment — Containerized multi-node deployment

---

🏗 Architecture

┌──────────────────────────────────────────────┐
│  LAYER 1: Input Sanitization                 │
│  Deobfuscation (homoglyphs, leetspeak, etc.) │
├──────────────────────────────────────────────┤
│  LAYER 2: Semantic Analysis                  │
│  4-Classifier Ensemble: ONNX_BERT (~50ms)    │
│  + RuleBased + Heuristic + LLM_SEMANTIC      │
│  Dual classifier OR-logic (BERT+RuleBased)   │
├──────────────────────────────────────────────┤
│  LAYER 3: Behavioral Analysis                │
│  16 AoC Defense Modules (11 original + 5     │
│  extended), Pattern matching, Trajectory     │
│  anomaly detection, ThoughtVirus defense     │
├──────────────────────────────────────────────┤
│  LAYER 4: Meta-Observer (O₂)                 │
│  Transfer entropy, BFT consensus,            │
│  semantic entanglement, causal graphs,       │
│  Kill-Switch, SVETILO value verification     │
├──────────────────────────────────────────────┤
│  C4 Core Engine (Z₃³)                        │
│  pipeline_orchestrator.py, event_bus.py      │
│  c4_meta_monitor.py — self-awareness deque  │
├──────────────────────────────────────────────┤
│  Defenses: Anti-Deadlock, Anti-Emergence,    │
│  Anti-Hijack, Circuit Breaker, O₂ Kill-Switch│
├──────────────────────────────────────────────┤
│  Red Team Lab: AOC scenarios, experiment     │
│  runner, LLM client, adapters                │
├──────────────────────────────────────────────┤
│  Routing: Smart Router, Quarantine,          │
│  Antifragile Scoring (capped growth)         │
└──────────────────────────────────────────────┘

4 classifiers voting:

• ONNX_BERT: C4 cognitive coordinates (T,S,A) via quantized model (737KB, ~50ms)
• RuleBased: 80+ regex patterns covering injection, jailbreak, role-play, authority bypass
• Heuristic: 32 danger words + C4 axis analysis + semantic density metric
• LLM_SEMANTIC: Ollama/DeepSeek — semantic attack classification (~300ms)

Dual classifier OR-logic: BERT + RuleBased operate as primary gate with OR fallback — if either flags the input, it proceeds to defense layers. No single classifier is a bottleneck.

---

📁 Project Structure