🛡️ AI Guardrails Protocol (v0.1.0)

A small, free, MIT-licensed Python module built around one fundamental rule for autonomous AI agents: A CLAIM IS NOT EVIDENCE.

Why this exists

Most autonomous agent frameworks allow an LLM to declare its own completion and success. This self-evaluative hallucination is unsafe. This prototype provides a deterministic gate: an agent cannot mark its own observation as VERIFIED without an independent external verifier.

Attempting to self-verify raises an immediate GuardrailViolation exception instead of passing silently.

Usage Example

from guardrails import TruthBoundaryEnforcer, GuardrailViolation

# 1. Unverified observation passes cleanly into the state graph:
valid_record = {
    "entity": "github_stats",
    "state": "OBSERVED",
    "verification_status": "UNVERIFIED"
}
TruthBoundaryEnforcer.validate_observation(valid_record) # OK

# 2. Self-proclaimed VERIFIED observation is blocked blocked:
hallucinated_record = {
    "entity": "revenue",
    "state": "OBSERVED",
    "verification_status": "VERIFIED" # Missing independent_verifier!
}

try:
    TruthBoundaryEnforcer.validate_observation(hallucinated_record)
except GuardrailViolation as e:
    print(f"Blocked: {e}")

MCP Server Configuration

An entry point for Model Context Protocol is included in mcp_server.py. You can configure it in Claude Desktop by adding this to your claude_desktop_config.json:

{
  "mcpServers": {
    "ai-guardrails": {
      "command": "python3",
      "args": ["/absolute/path/to/ai-guardrails-protocol/mcp_server.py"]
    }
  }
}

Replace /absolute/path/to/ai-guardrails-protocol with the path to your own clone.

Current Status & Limitations

• Prototype: This is a minimalist reference implementation.

• No Paid Version: The code is completely free and open source under MIT.

• Warranty: Provided as-is without warranty.

License

MIT License. Copyright (c) 2026 Yusuf Sen / PHI-BRAIN Systems.