VILR: Verifiable Intent & Liability Runtime

PROBLEM:

AI agents introduce a new payment trust boundary. They can hallucinate, and merchants can inject malicious prompts, making autonomous AI payments extremely dangerous if the AI has the authority to move money.

SOLUTION:

VILR separates probabilistic AI interpretation from deterministic payment authorization. It uses a zero-trust cryptographic gateway to authorize the actual financial execution.

CORE PRINCIPLE:

"AI can interpret and recommend. AI cannot authorize money movement."

ARCHITECTURE:

Gemini → Crypto → OPA → Redis → PaymentExecutor → Razorpay → Supabase

1. Quickstart

# Clone and setup
git clone https://github.com/sahvendraz/VILR.git
cd VILR
python3 -m venv venv && source venv/bin/activate
pip install -e .

2. Safe Demo

# Run the live, safe demo with real infrastructure (no production funds)
python scripts/cli.py demo --llm gemini --provider mock --ledger supabase

3. Live Integrations

• Gemini 3.1 Flash-Lite: Semantic Verification.
• Razorpay TEST: Controlled test-mode execution.
• Supabase: Persistent tamper-evident audit chain.
• Redis Lua: Atomic idempotency locking.
• OPA: Deterministic authorization.

4. Security Evidence

In our rigorous security benchmark testing (40 attack vectors covering cart tampering, replay attacks, concurrent double-spends, and semantic manipulation):

• 0 unauthorized payments were executed.
• 16 attacks were deterministically blocked by Gateway layers (Crypto, Redis, OPA).
• Semantic mismatch and prompt injections were consistently mitigated without overriding OPA budgets.

7. Gemini Integration

VILR integrates the official google-genai SDK using gemini-3.6-flash. It utilizes strict structured outputs (JSON schema) and securely fails closed on timeouts or schema violations. Gemini acts as an independent semantic assessor (evidence provider), remaining strictly subordinate to OPA.

8. Limitations & Scope

• Prototype Status: VILR is a Razorpay Buildathon prototype.
• Mock Razorpay: Currently executes against a simulated Razorpay TEST payment executor.
• AP2-Compatible: The cryptography simulates the Advanced Payments 2.0 (AP2) specification concepts but is a lightweight representation, not a compliant implementation.
• Simulated Recovery: The revenue recovery metrics shown in the demo are conceptually simulated to demonstrate potential ROI.

For detailed architecture, security models, and performance benchmarks, see the docs/ directory.

Tier 1 Security Enhancements

• Exception Intelligence: Advisory/explanatory only. LLM has no authorization authority.
• Security Receipt: Deterministic evidence generated from actual system state.
• Ground-Truth Evaluation: Expected decision vs actual decision + expected side effect vs actual side effect.
• Supabase: Persistent tamper-evident audit storage.