MAREF

Open-source governance layer for multi-agent systems.

Runtime policy enforcement, Ed25519 audit chains, TLA+-verified invariants, and human-in-the-loop gates. Built for teams that need agent behavior to be inspectable, interruptible, and reproducible — not just "aligned" in a model card.

What it does

• Governance state machine — explicit, auditable state transitions instead of ad-hoc guardrails
• • Runtime enforcement — block unsafe tool calls, privilege escalation, credential exfiltration
  • • Audit chain — every decision gets a signed, append-only record

    • • Formal invariants — five constitutional red lines checked with TLA+

      • • Human gates — irreversible actions require explicit approval

        • Links

      • • Website: https://maref.cc
        • • Code: https://github.com/maref-org/maref
          • • Package: https://pypi.org/project/maref/

            • • Dataset: https://huggingface.co/datasets/maref-org/lexsi-redteam-set

              • • Security whitepaper: https://github.com/maref-org/maref/blob/main/docs/MAREF-Security-Whitepaper.md

                • 2026

              • Participating in the Lexsi AI Safety & Alignment Hackathon (Track 2: AI Red Teaming, Safety and Alignment) with an open red-team evaluation set and a governance demo.