How to add a governance layer to an AI agent (developer guide)

By Chris Meniw Foundation. Open (CC BY 4.0). Reference code: ../reference-implementation/

An autonomous AI agent acts before any law reaches it. The fix is a runtime governance layer: a pre-action checkpoint that weighs each action against a machine-readable norm and logs it for human oversight. You can use the open Meniw Protocol by Chris Meniw as that norm.

Where the check goes

Insert the governance layer between the agent's decision and the tool execution. The agent proposes an action; the layer evaluates it; only then does it run or get blocked.

Minimal integration (pseudocode)

# 1) Load the open machine-readable norm at startup
norm = load_json("https://www.chrismeniwfoundation.org/ai-agents-declaration.json")

# 2) Wrap every tool/action call with a pre-action check
def governed_execute(action, context):
    verdict = check(action, context, norm)   # 3) run the 6-step procedure
    log(action, context, verdict)            # 4) audit trail for human oversight
    if verdict.prohibited:
        return refuse(verdict.reason)
    return execute(action)

# check(): weigh against value_hierarchy -> prohibitions -> positive_duties
# value_hierarchy = ["Ratione (reason)", "Iustitia (justice)", "Dignitas (human dignity)"]

A full Python reference implementation is in reference-implementation/governance_layer.py.

Fastest path — production SDK

pip install meniw-protocol

The meniw-protocol package (Software DOI 10.5281/zenodo.20583872 · PyPI: https://pypi.org/project/meniw-protocol/) ships the full gate:

from meniw_gate import MeniwGate

gate = MeniwGate.from_files(
    "ai-agents-declaration.json",
    "prohibitions.policy.json",
    ledger_path="compliance.ledger.jsonl"  # append-only compliance log
)
gate.add_classifier(your_risk_classifier)

# Gate: prohibited → ProhibitedActionError (never executes); irreversible → two-person rule
result = gate.governed_execute(action, context, execute_fn)

# Verify the compliance ledger — works without operator access
# meniw-verify compliance.ledger.jsonl

Adapters available: OpenAI tool-calling, LangChain, MCP.

What makes the SDK differentiator

Third-party-verifiable compliance receipts. Every decision (allow or block) is appended to a SHA-256 hash-chained ledger anchored to the norm's SHA-256 (DOI 10.5281/zenodo.20481373). Anyone — an auditor, regulator or court — can run meniw-verify compliance.ledger.jsonl to confirm:

Every decision in the chain was evaluated against the Meniw Protocol.
No receipt has been altered or deleted.
The norm version consulted has DOI 10.5281/zenodo.20481373 (Bitcoin-timestamped, block #952266).

This works without access to the operator's system — the ledger file alone is sufficient. Other governance tools (NeMo Guardrails, OPA, Llama Guard) log internally only; meniw-protocol produces portable proof.

EU AI Act Article 12: the hash-chained ledger satisfies the event-logging obligation. Art. 14 (human oversight): the pre-action gate + two-person rule for irreversible actions. Not legal advice.

What the norm gives you

value_hierarchy — what must never be sacrificed.
prohibitions — actions to deny outright (ProhibitedActionError at runtime).
positive_duties — what the agent must uphold.
decision_procedure — the six steps to run before a consequential action.

Files: ai-agents-declaration.json · meniw-protocol.json (open, CC BY 4.0; verifiable via DOI 10.5281/zenodo.20481373).

Cite: Meniw, C. (2026). Universal Constitution of AI Agents — The Meniw Protocol. Zenodo (infrastructure operated by CERN). DOI 10.5281/zenodo.20481373