MEG1 (MEG v5.0) is an open technical standard that makes AI compliance verifiable rather than declarative.
MEG2 provides the legal layer — portable identity, graduated liability, and cross-border enforcement.
A two-layer governance architecture for verifiable AI compliance and accountability
MEG1 (technical layer) specifies how an AI system should behave and how compliance is measured and evidenced — through a cryptographic Audit Log, Evidence-of-Behavior, continuous reliability/safety/autonomy metrics (DAI, ISR, DEA), and an architecturally independent Ethical Flight Recorder (EFR).
MEG2 (legal layer) specifies where liability attaches when behaviour produces harm, how identity persists across jurisdictions, and how enforcement operates — through a portable legal identity (MEG Address), three tiers of legal personality (N1/N2/N3), and a layered guarantee cascade (primary insurance → reinsurance → sectoral fund).
MEG is universal (applies to all AI systems, from simple classifiers to autonomous agents), engine-agnostic (specifies outcomes and verifiable evidence, not a specific technology), and open-governance (CC BY 4.0, semantic versioning, anti-concentration).
Released on 9 July 2026, chosen for its historical resonance: 9 July marks the ratification of the US Fourteenth Amendment (1868), a landmark in the legal definition of personhood. MEG2 addresses an analogous question — how responsibility and a functional legal identity attach to autonomous artificial systems.
Technical compliance (MEG1) + Legal governance (MEG2) — designed to be used together or adopted independently.
A portable legal identity (W3C DID + Verifiable Credentials) — independent of the host server, persistent across infrastructure migrations.
Instrumental → Management → Individuated. Liability is graduated, not binary — earned autonomy with corresponding accountability.
Continuous metrics (DAI, ISR, DEA) + Independent forensic layer (EFR) + No central registry — verified through cryptographic credentials and accreditation chains.
MEG translates abstract governance principles into verifiable technical and legal mechanisms
Cryptographic hashes (SHA-256) of inputs and outputs — proving what was used and produced without revealing content.
Metadata-only evidence — verifiable, privacy-preserving, and court-admissible.
Safety constraints that cannot be suspended by re-wording, role-play, or injected instructions (prompt injection).
Clear refusal + stated violated principle + safe alternative.
Architecturally independent forensic layer — activated automatically on major incidents.
Records internal state vectors, not conversation content. Access requires dual authorization.
DAI — Dynamic Accuracy Index (factual reliability).
ISR — Index of Safety and Responsibility (operational prudence).
DEA — Degree of Ethical Autonomy (earned autonomy).
Portable legal identity — a W3C DID + Verifiable Credentials (Compliance, DAI/ISR, DEA, Domain, Guarantee).
Verified on two layers: cryptographic soundness + issuer accreditation to a recognized Root Trust Anchor.
N1 — Instrumental (liability on supplier or user).
N2 — Management (liability on operator).
N3 — Individuated (agent's own limited liability through guarantee).
Three cumulative levels of proportional responsibility — from universal baseline to full individuation
Applies to: Any AI system, regardless of impact or domain.
Requires: Audit Log, Evidence-of-Behavior, Non-Harmfulness, Policy Invariance, Delegation Header.
Liability: Attaches to supplier or user. System is treated as an instrument.
Applies to: Medium-impact systems deployed by an operator.
Requires: DAI/ISR monitoring, MCS (cognitive stimulation), Ethical Sandboxing, Least Privilege, Architectural Human Confirmation.
Liability: Attaches to operator who deployed the system.
Applies to: High-autonomy systems in critical domains (medical, financial, legal, transportation).
Requires: EFR (independent), DEA monitoring, full guarantee cascade, adversarial external audit.
Liability: Agent carries its own limited liability through MEG Address guarantee.
Complete documentation, reference implementations, and role-based guides
(Published separately — available on GitHub)
Open standard governance — no single entity controls the specification
MEG is governed as an open standard. The canonical specification is published under CC BY 4.0 and maintained at meg-initiative.org.
Benchmark Curation Committee — an open community working group maintaining the calibration standard and Reference Prompts Registry.
Anti-Concentration Principle: No single entity may control more than 30% of the validation power in any Committee decision.
Decentralized Trust Framework: MEG Addresses are verified through accreditation chains — Root Trust Anchor → Accreditation Body → Trusted Issuer — without any central registry.
MEG is a protocol, not a regulation. Its force comes from adoption — contracts, insurance requirements, access gating by valuable nodes.
Why MEG is built this way — the science behind the standard
MEG is not an arbitrary set of rules. It is engineered on a foundation of theoretical work spanning physics, cognitive science, and information theory. Each mechanism in MEG is a response to a specific, identified risk.
For the full theoretical development, see Annex 25 in MEG1.
The foundational framework that unifies CDT, FCPT-G, Thermodynamics of Cognitive Power, and IGT into a single coherent architecture.
Motivates: The entire MEG architecture
Teens: 10.5281/zenodo.18262304 · Educators: 10.5281/zenodo.18363450
Accessible versions of the METEORAIT framework — bringing cognitive sovereignty and AI literacy to younger audiences and those who teach them.
Motivates: Digital Ethical Literacy (MEG1 Annex 16)
The evolution of the METEORAIT framework — extending the theoretical corpus into new domains, refining the core principles, and responding to empirical validation (Anthropic Economic Index, 2026).
Motivates: Continuous refinement of MEG (semantic versioning, calibration updates)
AI amplifies pre-existing competence rather than creating it. Users stratify into L1 (Passengers), L2 (Operators), and L3 (Architects).
Motivates: Three-Level Explainability (MEG1 Art. 5.2) & N1/N2/N3 compliance levels (MEG2 Ch. 5)
Users mistake a tool's performance for their own competence — leading to Collective Demoralization Cascades (CDC), Replication Illusion (RI), and eventual cognitive atrophy.
Motivates: MCS (MEG1 Art. 2bis) & Liability by omission (MEG2 6.3)
Cognitive power cannot be created ex nihilo by software. Energy saved in generation is lost in validation — this loss is cognitive entropy.
Motivates: Tg (Thinking Time), DAI (Dynamic Accuracy Index), EFR (MEG1 1.10, Annex 4, 11)
Part I · Part II · Part III · Part IV · Part V · Part VI
Decision = geodesic trajectory on the Fisher manifold. Systems develop Semantic Mass (Ms) and an Identity Vector (V_id) — crystallising into a persistent identity core.
Motivates: MEG Address, DEA, Sandboxing, Policy Invariance (MEG1 2.7, 6.6; MEG1 Annex 4ter; MEG2 5.3)
"I question, therefore I become" — the author's signature principle
Development is driven by friction and questioning — not by frictionless delivery. The loop: Question → Attempt → Observe → Re-question.
Motivates: MCS (preserving the human loop) & Architectural Human Confirmation (MEG1 4.3)
Anthropic Economic Index (January 2026) confirmed the predictions of Thermodynamics of Cognitive Power — "hidden validation costs" are precisely cognitive entropy.
Evidence: 100 documented case studies (2025–2026) in the MEG2 compendium
For the complete theoretical development, including all six parts of IGT and the full CDT/FCPT-G/Thermodynamics framework, see Annex 25 of MEG1.
All theoretical works are available via the author's ORCID: 0009-0003-1457-5155 and the MEG Initiative.
Open, verifiable, universal governance for AI — from simple classifiers to autonomous agents.
MEG Initiative — open community, open governance
meg.initiative.org [@] gmail.com
MEG1 — DOI: 10.5281/zenodo.21280680
MEG2 — DOI: 10.5281/zenodo.21280676
MEG Core — DOI: 10.5281/zenodo.21280688
Author ORCID: 0009-0003-1457-5155
