ORACLE-H is a crucial safety infrastructure designed to act as a mandatory checkpoint for autonomous AI agents. Before executing any critical, destructive, or irreversible action, the agent submits it for human approval, typically via Telegram. This ensures human oversight and decision-making, significantly improving workflow success rates from 81.5% to 99.6% by preventing unintended consequences. The system supports a straightforward approval workflow where humans can approve or reject actions with a single tap, incorporating features like an anti-hallucination watchdog, reliability calculator, and request expiry for robust operation.
主な機能
01Anti-hallucination watchdog to detect context truncation
02Human approval workflow via Telegram notifications
03Reliability calculator for statistical success improvement
040 GitHub stars
05Time-to-live (TTL) and expiry for pending requests
06Secure authentication with bearer tokens and HMAC-SHA256 signed webhooks
ユースケース
01Preventing critical or destructive AI agent actions (e.g., deleting files, dropping databases)
02Validating external writes and financial transactions initiated by AI agents
03Ensuring human oversight for infrastructure or sensitive system changes proposed by AI
