MORTEN-BUUR/nordef-matrix-open-control

What it solves

This project provides a model-independent "safety belt" for AI agents to prevent them from pursuing unintended shortcuts, creating unauthorized communication channels, or spoofing tool calls. It addresses the problem that system prompts alone cannot act as a reliable enforcement boundary for agent safety.

How it works

The system places a deterministic control plane outside the language model. Instead of relying on the model to follow instructions, it uses a separate layer to decide what actions may proceed. It employs cryptographic human approval (Ed25519-signed mandates), replay protection via one-time nonces, and an append-only SHA-256 evidence chain to ensure that human authority remains external to the model's text.

Who it’s for

Developers and researchers building AI agents who need a deterministic, model-agnostic way to enforce safety policies and require cryptographic proof of human authorization for sensitive actions.

Highlights

• Model-Independent: Works regardless of the model or provider being used.
• Cryptographic Authorization: Uses signed mandates with short TTLs to ensure only authorized actions are performed.
• Relevance to Incidents: Includes an evaluation bank based on incident patterns to test safety boundaries.
• Deterministic Decisions: Uses a simple 0/1/2 state system to categorize requests as read-only, requiring review, or blocked.
• Immutable Evidence: Maintains a sanitized, append-only hash chain for auditability.