I kept watching AI agents ship with no enforcement layer. They write files, run shell commands, call APIs, process external content — and if a jailbreak is hidden in that content, it just goes through.
CORD intercepts every proposed action before execution. Hard violations (behavioral extortion, jailbreaks, constitutional bypass) skip scoring entirely — instant block, no appeal. Everything else gets a weighted score, a plain English explanation, and a hash-chained audit log entry.
Built two implementations — Python (cord-engine on PyPI) and JavaScript (just pushed cord-engine@3.0.2 to npm tonight). Two lines to wrap your existing OpenAI or Anthropic client.
Real numbers from running CORD on its own agent builds tonight: 44 evaluations, 27% block rate, 8 hard blocks — behavioral extortion, jailbreak attempts, constitutional override attempts. All stopped before execution.
Dashboard: npm run dashboard → live decision feed, block rate ring, hard block alerts.
Happy to go deep on the constitutional architecture, the hard-block design, or the LEGION multi-model orchestration layer that runs on top of this. Ask me anything.