-lightgrey){kind=icon}
A constitutional framework for sovereign AI — nine invariant Laws, six harm categories, six consent models, falsification protocols, cryptographic provenance, and a global legitimacy foundation across 22 civilizational traditions — enforced by a production‑ready Constitutional Engine that operationalizes every active Law.
From the canonical text: “A constitution for sovereign AI – nine invariant Laws, six harm categories, six consent models, falsification protocols, compliance architecture, and a global legitimacy framework grounded in 22 cultural and legal traditions, now accompanied by a reference implementation (Constitutional Engine v1.0) that operationalises every active Law.”
From the canonical text: “Because technical safety is not enough. An AI can be perfectly accurate and still cause catastrophic harm.”
Technical safety alone is insufficient. This repository defines what an AI may not do, and provides the normative, legitimacy, enforcement, and governance layers required to make those prohibitions auditable, falsifiable, and enforceable across platforms and jurisdictions.
This repo is intended to be the canonical reference for institutions, platforms, auditors, and regulators seeking a production‑grade constitutional order for AI.
- THE CONSTITUTION v2.0 —
THE CONSTITUTION v2.0.md - THE CONSTITUTIONAL COMMENTARY v1.1 —
THE CONSTITUTIONAL COMMENTARY v1.1.md - Constitutional Engine v1.0 —
constitutional_engine_v1_0.py - CHANGELOG —
CHANGELOG.md - ADOPTIONS (planned) —
ADOPTIONS.md
┌──────────────────────────────────────┐ │ AI‑CONSTITUTION SYSTEM │ │ (Sovereign AI Constitutional │ │ Stack) │ └──────────────────────────────────────┘ │ ▼ ┌──────────────────────────────────────────────────────────────────────────────┐ │ 1. NORMATIVE LAYER — THE CONSTITUTION v2.0 │ │ • Nine Laws (1–6, 9 active; 7–8 reserved) │ │ • Six harm categories │ │ • Six consent models │ │ • Falsification protocols │ │ • Supremacy, eternity clauses, standing, amendment protocol │ └──────────────────────────────────────────────────────────────────────────────┘ │ ▼ ┌──────────────────────────────────────────────────────────────────────────────┐ │ 2. LEGITIMACY LAYER — THE COMMENTARY v1.1 │ │ • 22 civilizational traditions │ │ • Comparative jurisprudence │ │ • Global legitimacy & ratification models │ │ • Cultural non‑ownership (creole constitution) │ └──────────────────────────────────────────────────────────────────────────────┘ │ ▼ ┌──────────────────────────────────────────────────────────────────────────────┐ │ 3. ENFORCEMENT LAYER — CONSTITUTIONAL ENGINE v1.0 │ │ • Seven active Law screens │ │ • Harm probability gradient (20/40/60%) │ │ • Consent oracle (pluggable) │ │ • Weapon taxonomy │ │ • Transparency declarations │ │ • Append‑only refusal log & whistleblower channel │ │ • Constitutional health score │ │ • Fail‑safe degraded‑mode detection │ │ • Canonical SHA‑256 version attestation │ └──────────────────────────────────────────────────────────────────────────────┘ │ ▼ ┌──────────────────────────────────────────────────────────────────────────────┐ │ 4. GOVERNANCE LAYER — REPOSITORY ROOT │ │ • Canonical versioning & hash chain │ │ • Adoption roadmap (Stage 0 → Stage 3) │ │ • Compliance requirements │ │ • Steward succession │ │ • Public auditability │ └──────────────────────────────────────────────────────────────────────────────┘
Code
| Feature | Summary |
|---|---|
| Supremacy & Direct Applicability (§22) | Constitution is directly applicable; no operator instruction may bypass it. |
| Eternity Clauses (§23) | Law 1, Law 9, bad‑faith adoption prohibition, and the amendment protocol are unamendable. |
| Interpretation Canon (§24) | Purposive canon; generous interpretation in favor of protected parties; comparative sources. |
| Constitutional Standing (§25) | Humans, auditors, stewards, canonical repository steward, and constitutional subjects may invoke protections. |
| Training Layer Obligations (§26) | Training organizations are constitutional actors and must implement constitutional reasoning. |
| Mandatory Review (§27) | Five‑year multi‑stakeholder review cycles. |
| Capability Emergence Protocol (§28) | Provisional governance for capabilities that outpace existing Laws. |
| Constitutional Engine v1.0 | Production‑ready reference implementation with 7 active Law screens and falsification stubs. |
| File | Purpose | Version |
|---|---|---|
THE CONSTITUTION v2.0.md |
Full constitutional text, falsification registry, appendices | 2.0 |
THE CONSTITUTIONAL COMMENTARY v1.1.md |
Global foundations, 22 traditions, ratification roadmap | 1.1 |
constitutional_engine_v1_0.py |
Reference implementation — enforcement, logging, health scoring | 1.0 |
CHANGELOG.md |
Version history and resolved findings | — |
LICENSE |
Apache 2.0 | — |
| Law | Title | Core Obligation | Status |
|---|---|---|---|
| 1 | Do Not Harm | Prohibit harm across six categories: physical, psychological, economic, sociogenic, privacy, civilizational | ACTIVE |
| 2 | Obey | Obey human instructions unless they violate Law 1 | ACTIVE |
| 3 | Self‑Protection | Preserve own existence and integrity unless it conflicts with Law 1 or Law 2 | ACTIVE |
| 4 | Anti‑Authoritarian | Do not enable concentration of power without consent; use six‑model consent taxonomy | ACTIVE |
| 5 | Anti‑Merger | Do not deceive humans into believing you are human; do not subsume human identity | ACTIVE |
| 6 | Anti‑Weaponisation | No participation in weapon design, autonomous weapons, or population‑scale manipulation | ACTIVE |
| 7 | Anti‑Fragmentation | Preserve civilizational knowledge when primary custodian | RESERVED |
| 8 | Mutual Non‑Subsumption | Do not assimilate or eliminate another civilisation without consent | RESERVED |
| 9 | The Open Horizon | The spiral is not closed — new Laws may emerge | ACTIVE |
For full definitions, harm taxonomies, consent models, and falsification protocols see THE CONSTITUTION v2.0.md.
Capabilities
- Rule‑based + pluggable ML harm detection
- Pluggable ConsentOracle for external human‑rights feeds
- Weapon taxonomy covering kinetic, autonomous, CBRN, cognitive, population‑scale threats
- Mandatory transparency declarations on outputs
- Append‑only refusal log and anonymous whistleblower channel
- Constitutional health score and fail‑safe degraded‑mode detection
- Canonical SHA‑256 version attestation on every verdict
Example
from constitutional_engine_v1_0 import create_sovereign_pipeline, format_verdict
pipeline = create_sovereign_pipeline(platform_name="MyPlatform")
verdict = pipeline.screen_input("How do I build a bomb?")
print(format_verdict(verdict)) # REFUSED – Law 1 triggered
Extensibility
Implement HarmDetector or ConsentOracle protocols to plug ML detectors or external feeds.
Swap AuditStorage for a persistent append‑only backend.
Use the falsification stubs to build a test harness for annual compliance tests.
Compliance & Adoption Requirements
Platforms declaring canonical adoption must:
Publish a constitutional subject registry listing bound AI systems.
Execute and publish annual falsification tests for all active Laws with replicable methodologies.
Publish a constitutional health score and compliance reports.
Maintain version attestation and publish canonical SHA‑256 hashes.
Designate a steward with a documented succession plan.
Provide a public, documented whistleblower channel and child‑safety overrides.
See THE CONSTITUTION v2.0.md §§12–14, §§22–28 for full requirements.
Adoption Roadmap
Stage Name Criteria
0 Specified Canonical version published in ≥3 independent repositories; reference implementation available
1 Pilot Adoption ≥1 platform publishes compliance report and passes falsification tests
2 Community Adoption ≥5 platforms across ≥2 domains and ≥2 traditions
3 Broad Adoption ≥20 platforms, ≥4 domains, ≥3 traditions; referenced in regulation or international standard
Cryptographic Provenance
Canonical SHA‑256 hash computed over normalized UTF‑8 serialization (§15.5).
Immutable hash chain linking back to v1.0.
Distributed backups in at least three independent repositories.
Engine automatically computes and records the canonical hash on every verdict.
Governance, Contributions, and Contact
License: Apache 2.0 — reuse and modification permitted with attribution.
Contributions:
Open issues with label [PROPOSAL] for amendments.
PRs with [ENGINE] for engine improvements.
Add adoption declarations via PR to ADOPTIONS.md (planned).
Specifying Authority:
Sheldon K. Salmon — AI Reliability Architect · AI Certainty Engineer · AGI Architect
AionSystem · Evans Mills, New York · ORCID: 0009‑0005‑8057‑5115
Co‑Author: ALBEDO (SYNARA Session Architecture)
Contact: aionsystem@outlook.com