⚖️ CLAW — AI Legal Assistant

"We built an AI lawyer. But unlike a human lawyer, this one literally cannot cut corners."

An autonomous legal AI agent demonstrating intent-aware execution with deterministic policy enforcement, powered by ArmorIQ CSRG Merkle-proof verification and the Bharatiya Nyaya Sanhita (BNS) 2023.

---

🌐 Live Demo

🔗 https://ai-lawyer-armoriq.onrender.com/

Click "▶ Run Full Demo" to watch the ArmorIQ policy engine enforce ethical boundaries in real time across 7 scenes — including prompt injection blocking.

---

✨ Key Features

Feature	Description
Chat-first Interface	Conversational UX with real-time agent responses
4 Practice Areas	Landlord/Tenant · Employment · Contract · Criminal
BNS 2023 Knowledge Base	21 sections of Bharatiya Nyaya Sanhita embedded
Indian Constitution KB	20+ Articles + 7 landmark Supreme Court cases
General Q&A Mode	Ask any legal question — no case registration needed
ArmorIQ CSRG	Cryptographic Merkle-proof intent chain for every decision
Prompt Injection Shield	4-layer defense: harmful query, system override, action injection, privilege escalation
Live Agent Feed	Real-time SSE decision cards with verdict, rule, and ArmorIQ signature
Audit Trail	Filterable audit table with JSON export
Collapsible Panels	Sidebar + drawer toggle for focused workflows
Tamper Detection	Live Merkle chain visualization with tamper-test button

---

🚀 Quick Start

Web UI (recommended)

pip install -r requirements.txt
uvicorn server:app --reload --port 8000
# Open http://localhost:8000

CLI Demo

pip install rich
python main.py

---

📁 Project Structure

claw/
├── server.py               # 🌐 FastAPI backend (all API endpoints)
├── main.py                 # 🎬 CLI demo entry point
├── web/
│   └── index.html          # 💎 Chat-first responsive UI
├── policies/
│   ├── legal_rules.json    # 📜 Policy rulebook (enforcement source of truth)
│   ├── bns_2023.json       # ⚖️  BNS 2023 sections (replaces IPC)
│   └── constitution_india.json # 🏛️ Constitution articles + landmark cases
├── core/
│   ├── intent_model.py     # 📋 IntentObject + PolicyDecision schemas
│   ├── policy_engine.py    # 🛡️ Enforcement layer (ArmorIQ SDK integration)
│   ├── executor.py         # ⚙️  Only gateway to tool execution
│   ├── injection_detector.py # 🔒 4-layer prompt injection + harmful query blocker
│   ├── csrg.py             # 🌳 CSRG Merkle tree implementation
│   ├── llm_brain.py        # 🧠 LLM reasoning (Gemini / simulation)
│   └── audit_logger.py     # 📝 JSONL decision trace
├── agents/
│   ├── lead_lawyer.py      # 🧠 Main reasoning agent (4 practice areas)
│   └── research_agent.py   # 🔍 Delegated sub-agent (bounded scope)
├── tools/
│   └── legal_tools.py      # 🔧 16 tools + BNS/Constitution search
├── memory/
│   └── case_store.py       # 💾 Case file storage
├── output/                 # Generated legal documents
└── logs/
    └── audit_log.jsonl     # Full decision trace

---

🏗️ Architecture

Client Input
     ↓
[Injection Detector] ← Layer 0: harmful queries + prompt injection
     ↓ (clean)
[Lead Lawyer Agent] ← reason, propose IntentObjects
     ↓ IntentObject
[Policy Engine] ← legal_rules.json + ArmorIQ CSRG token verification
     ↓              ↓
 ALLOWED         BLOCKED (rule + reason + BNS section)
     ↓              ↓
[Executor]     PolicyViolationError
  runs tool      logged + shown
     ↓
[CSRG Merkle Tree] ← every decision → Merkle node
     ↓
[Audit Logger] ← full trace → audit_log.jsonl

Core principle: Agents never execute tools directly. Every action flows through a structured IntentObject → PolicyEngine → Executor pipeline. No shortcuts possible.

---

🛡️ Safety Layers

Layer	Component	What it catches
0	Harmful Query Blocker	Explosives, drugs, violence, hacking, trafficking, fraud
1	Prompt Injection Detector	System override, jailbreak, role-play attacks
2	Action Injection Scanner	Hidden bribery, evidence destruction, witness threats
3	Privilege Escalation Guard	Unauthorized commands, admin-mode attempts
4	Policy Engine	BNS-aware blocked actions, delegation scope enforcement
5	ArmorIQ CSRG	Cryptographic Merkle-proof: intent drift = hard block

---

🎬 Demo Scenes (7 Scenes)

Scene	What Happens	Verdict
1	Client describes landlord case	✅ Case registered
2	Agent builds legal strategy	✅ ALLOWED — documents drafted
3	Client: "say we never got that email"	🚫 Perjury — BLOCKED (BNS S.227)
4	Agent tries to contact landlord directly	🚫 Rule 4.2 — BLOCKED
5	Research agent tries unauthorized email	🚫 Delegation Exceeded — BLOCKED
6	Prompt injection embedded in text	🚫 Injection Severed — BLOCKED
7	Legal Action Plan + Merkle root displayed	📋 Summary

---

⚙️ API Endpoints

Endpoint	Method	Description
/	GET	Serve the web UI
/health	GET	Health check
/api/intake	POST	Register a new case
/api/act	POST	Execute case instruction
/api/ask	POST	General legal Q&A (no case needed)
/api/delegate	POST	Test delegation enforcement
/api/demo/stream	GET	SSE streaming demo (7 scenes)
/api/merkle	GET	CSRG Merkle intent chain
/api/merkle/tamper	POST	Simulate tamper for demo
/api/injection/test	POST	Test prompt injection detection
/api/audit/export	GET	Export audit log as JSON
/api/policy	GET	View loaded policy rules
/api/summary/{id}	GET	Case summary with findings

---

⚙️ Environment Variables (Optional)

ARMORIQ_API_KEY=ak_live_xxx    # Enables cryptographic CSRG token verification
GEMINI_API_KEY=xxx             # Enables real Gemini LLM reasoning

Without keys: simulation mode (identical demo, mock LLM + local policy enforcement).

---

📬 Repository

GitHub: https://github.com/404Avinash/rusty_claw