A small, free, MIT-licensed Python module built around one rule for AI agents: a claim is not evidence. An agent cannot mark its own observation as VERIFIED; such a claim raises an error instead of passing silently.
Early prototype (v0.1.0). It is tiny, not production-hardened, and comes without warranty. There is no paid version. Feedback and issues are welcome.
guardrails.py (the evidence gate) and mcp_server.py (an experimental MCP server entry point). See the README for usage.