Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
sriram1983007 
posted an update 3 days ago
Post
3,841
Should an AI agent sign this payment?

I’ve released SRA-RiskGate-4B v2. It’s a 4B model fine-tuned from Qwen3-4B-Instruct that answers approve, hold or reject for stablecoin payments made by agents (x402, EIP-3009, Circle), and it settles payment disputes.

On 2,000 held-out benchmark cases:
✅ 99.6% of payment decisions correct (base model: 59.8%)
✅ 0.35% unsafe approvals (base model: 34.6%)
✅ 90.2% of dispute outcomes correct, up from 77.8% in v1

The main change in v2: the dispute window is computed in code and passed into the prompt instead of asking a 4B model to subtract dates. Wrongful denials fell from 26% to 2.1%.

I also tested it on 227 real USDC payments from Base mainnet: it blocked none of the 151 real payments and approved none of the risky ones.

The model isn’t the security boundary. The open-source checks in pip install sra-riskgate run first and fail closed, with optional Chainalysis, TRM, Range and Webacy screening.

These are my own numbers. Every prediction is public, so you can re-score them.

🤖 Model: hf.co/sriram1983007/SRA-RiskGate-4B
🎮 Try to trick it: hf.co/spaces/sriram1983007/sra-riskgate-demo
📊 Benchmark: hf.co/datasets/sriram1983007/sra-stablecoin-risk-bench
🧾 Real payments: hf.co/datasets/sriram1983007/sra-real

Super interesting.

·

Thank you @ChrisSacrumCor for the encouragement and taking time to respond.