SYSTEM ONE // LIVE ON AWS // ap-south-1

jev-agent-guardrail_

An AI agent proposes an action. Before it runs, this gate asks Jev, a decision model that returns probabilities not prose, three typed questions and answers allow / ask / block. No prose to parse. Nothing to prompt-inject. Try a tool call.

~/agent — guardrail --interactive

>
risk 0–2
·
intent match
·
exfiltration
·

stack: API Gateway → Lambda (arm64) → Jev via OpenRouter → DynamoDB decision log · fail closed to ASK
built by Rasik Rahman S · source (MIT) · Builder Series Part 3