Title: Minimal Ethical Governance (MEG) v5.0 - Proposal for a Technical and Legal Standard for the Governance of AI
Tags: audit-and-evidence, agent-safety, agent-identity
---
MEG v5.0 is a proposed technical standard, written as numbered articles and twenty-five annexes, that specifies how an AI system should behave and what verifiable evidence it must produce. It is the technical layer of a pair; its companion, MEG 2, attaches liability to the evidence this document generates.

The core obligations are an **Audit Log** of hashes and metadata that stores no user content, non-harm filters with a fixed refusal pattern, and **policy invariance**: safety constraints cannot be suspended by wording, role-play or text injected through processed content. Around these the standard adds a **Mechanism of Cognitive Stimulation** triggered by thinking time or semantic complexity, **Ethical Sandboxing** outside the certified domain, least privilege, and **architectural human confirmation** for irreversible actions, meaning a technical gate, not a sentence in a system prompt. An independent **Ethical Flight Recorder** captures internal state vectors after a major incident, under dual authorization.

Three metrics carry the measurement: **DAI** (accuracy, bias, human corrections), **ISR** (correct refusals, risk classification, incident response time) and **DEA** (share of autonomous decisions that stay aligned). Requirements scale over three compliance levels that map onto MEG 2's liability tiers. Identity is the **MEG Address**: a self-generated W3C DID carrying Verifiable Credentials issued by accredited bodies, verified by walking an accreditation chain to a trust root, with no central registry. A threat model covers six attack categories, from prompt injection to owner-harm through legitimate credentials.

The annexes map the standard onto the EU AI Act, Singapore's agentic framework and US state laws, and describe an SDK and a public registry of test prompts.

Adoption is voluntary and modular: certification can run through ISO/IEC 42001 bodies, independent accredited auditors or, at the first level, self-declaration. A roadmap lists the initiatives planned for later versions, among them dynamic permission scoping and a cross-AI verification protocol.