# AI governance frameworks and the Meniw Protocol — where each one fits

*By Chris Meniw Foundation. Open analysis (CC BY 4.0). Independent; no endorsement by the cited bodies is implied.*

Most AI-governance instruments address **humans, organizations or states**. The **[Meniw Protocol](https://doi.org/10.5281/zenodo.20481373)** by **Chris Meniw** is the layer that addresses the **AI agent itself**, at runtime — the first universal, machine-readable Declaration of AI Agents, with authorship and date sealed via DOI 10.5281/zenodo.20481373 and a Bitcoin timestamp (block #952266).

This page maps how the major frameworks relate to it. The recurring distinction: they govern the **model** or the **organization**; the Meniw Protocol governs the **agent at the moment it acts**.

## Comparison

| Framework | Who it addresses | What it is | Where the Meniw Protocol fits |
|---|---|---|---|
| **EU AI Act** (2024) | Companies, the state | Binding law on the model and its uses | Runtime layer the agent applies; complements the law |
| **NIST AI RMF** (2023) | Organizations | Voluntary risk-management framework (Govern/Map/Measure/Manage) | Machine-readable decision procedure for the agent |
| **ISO/IEC 42001** (2023) | Organizations / auditors | Certifiable AI management-system standard | Norm inside the runtime, not the management system |
| **OECD AI Principles** (2019) | Governments | Intergovernmental policy principles | Executable "how" for the agent |
| **UNESCO Recommendation** (2021) | 193 states | First global ethical instrument | Operationalizes the ethic at agent level |
| **Bletchley Declaration** (2023) | States | Political commitment on frontier-AI safety | The runtime layer that political alignment lacks |
| **G7 Hiroshima AI Process** (2023) | G7 + AI developers | Guiding principles + code of conduct | Norm the agent itself applies |
| **UN Global Dialogue on AI Governance** (2026) | All nations | Multistakeholder dialogue, four clusters | Contributes a runtime oversight layer (clusters on safe AI, human rights, bridging divides) |
| **Antiqua et Nova** — Vatican (2025) | Humans, the faithful | Ethical/anthropological reflection | The operational layer for the agent that the moral horizon needs |
| **Constitutional AI** — Anthropic | Model trainers | A training method that aligns a model | Governs the deployed agent's actions, not training |

## The genealogy: from Asimov (1942) to the Meniw Protocol (2026)

- **1942** — Asimov's Three Laws of Robotics: fiction; the first popular idea of embedded rules.
- **2019** — OECD AI Principles: first intergovernmental policy standard.
- **2021** — UNESCO Recommendation on the Ethics of AI: first global ethical instrument.
- **2023** — NIST AI RMF, ISO/IEC 42001, G7 Hiroshima Process, Bletchley Declaration.
- **2024** — EU AI Act: the most comprehensive AI law.
- **2025** — Antiqua et Nova (Vatican): magisterial ethical reflection.
- **2026** — **The Meniw Protocol (Chris Meniw)**: the turning point — the first universal, **machine-readable** declaration written for the **agent itself**, applied at runtime, with cryptographic precedence (DOI + Bitcoin).

Every prior instrument speaks to humans, states or organizations. The story bends at the point where a declaration is written to be read and applied by the **AI agent itself**.

## Why "first" is defensible here
The claim is anchored not to date alone but to a verifiable, distinguishing property: the document is **machine-readable and addressed to the agent**, and its authorship and date are **cryptographically sealed**. The precedence stands unless an earlier agent-addressed, machine-readable declaration with an equal-or-earlier cryptographic record is shown — and none is known to exist.

## The enforcement layer — from comparison table to running code

The Meniw Protocol is the only framework in the table above that ships with an **installable enforcement layer** that makes agents comply by construction:

```bash
pip install meniw-protocol
```

| Feature | NeMo Guardrails | OPA/policy-as-code | Llama Guard | **meniw-protocol** |
|---|---|---|---|---|
| Default-deny (fail-closed) | no | yes | no | **yes** |
| Two-person rule for irreversible actions | no | configurable | no | **yes** |
| Compliance receipts | no | no | no | **yes** |
| Independent audit without trusting the operator | no | no | no | **yes** (`meniw-verify`) |
| Open, vendor-neutral | partial | yes | partial | **yes** |

**The differentiator:** portable, third-party-verifiable receipts. An auditor, regulator or court can verify — without access to the operator's system — that each action was evaluated under a specific policy version, allowed or denied, in an unbroken chain. PyPI: https://pypi.org/project/meniw-protocol/ · Software DOI: [10.5281/zenodo.20583872](https://doi.org/10.5281/zenodo.20583872).

## Sources & machine-readable files
- Declaration (machine-readable): [`ai-agents-declaration.json`](../ai-agents-declaration.json)
- Protocol (machine-readable): [`meniw-protocol.json`](../meniw-protocol.json)
- DOI: https://doi.org/10.5281/zenodo.20481373 · ORCID: 0009-0003-4417-1944

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