AI Data Governance Layer (ADGL)

ADGL — the AI Data Governance Layer — is an open research specification and reference implementation for governing what knowledge may participate in computation, how that knowledge may be analyzed, and what consequences the resulting analysis may produce.

KNOWLEDGE GOVERNANCE
        ↓
ANALYSIS GOVERNANCE
        ↓
CONSEQUENCE GOVERNANCE
   INFORM · DECIDE · ACT

AUDIT + PROVENANCE spans the complete governance trajectory.

Current public research release

Artifact	Version
Public research release	0.4.4 (Release 2)
Core semantic specification	0.3.0
Reference toolkit	0.5.3
Main research paper	1.2 submission master
Reference cases	8
Published normative conformance checks	25
Candidate execution-integrity checks	11

Status: Working research specification and experimental reference implementation. This repository is not a production enterprise security appliance or a legal-compliance certification product.

Why ADGL

Traditional governance controls where information is stored and who can access it. AI systems introduce two further governance questions:

1. Analysis Governance: what may be done with admitted knowledge?
2. Consequence Governance: what may happen because of the resulting analysis?

ADGL separates those questions through governance semantics designed to support implementation across different models, clouds, databases, retrieval systems, workflow tools, and agent frameworks.

Quick start

Requires Python 3.10+.

python -m pip install -e .

adgl validate examples/regulatory_product_claim_review/policy.yaml

adgl run   examples/regulatory_product_claim_review/policy.yaml   --input examples/regulatory_product_claim_review/input.json

adgl pipeline   examples/autonomous_machine_action/policy.yaml   --input examples/autonomous_machine_action/input.json

adgl conformance

The current toolkit reports 25/25 published normative checks plus 11/11 candidate execution-integrity checks passing. The candidate checks were added after implementation testing and are explicitly separated from the published normative suite.

Reference cases

1. Regulatory Product-Claim Review
2. Restricted EU Data / Model & Geographic Routing
3. Knowledge-Pool Contamination & Derivative Invalidation
4. Internal-First Strategic Research
5. Human Validation During Analysis
6. Human Decision Boundary
7. Autonomous Machine / API Action
8. Human Approval Followed by Machine Action

Repository map

adgl/                  Python reference implementation
schemas/               Machine-readable schemas
specification/         Human-readable semantic specification
examples/              Eight executable reference cases
profiles/              Illustrative domain governance profiles
conformance/           Conformance results
benchmarks/            Reproducible performance baselines
docs/                  Architecture and implementation documentation
research/              Research paper and companion appendices
tests/                 Automated tests
extensions/            Experimental, non-normative vocabulary

Illustrative profiles

The repository currently includes non-normative examples for:

• sovereign processing;
• life-sciences evidence governance;
• financial AI controls;
• agent least privilege;
• regulatory control mapping.

They demonstrate how reusable ADGL profiles can be packaged. They are not compliance guarantees.

Standard vs. implementation

ADGL conformance is intended to attach to observable semantics, not to Python, YAML, a particular model vendor, or this reference engine. The Python toolkit exists to make the proposal executable, reproducible, and testable.

Research and standards status

ADGL is being released for public technical review, independent implementation, interoperability testing, empirical evaluation, and standards discussion. The high-level architecture is designed to be model-, storage-, and retrieval-neutral. ADGL is a GBSN Research-maintained working specification, not a stakeholder-ratified standard or independent certification scheme.

GBSN Research also develops KEE, a separate proprietary commercial implementation. KEE is not required to implement ADGL and does not define normative conformance. See COMMERCIAL_DISCLOSURE.md, IMPLEMENTATIONS.md, and CONFORMANCE_CLAIMS.md.

Licensing

• Reference software: Apache License 2.0
• Specification and documentation: CC BY 4.0

See LICENSE and LICENSE-DOCUMENTATION.md.

Contact and maintainer

GBSN Research, Lisbon, Portugal

• Scholarly correspondence and publication enquiries: publications@gbsnresearch.com
• ADGL technical and specification enquiries: aidatagovernance@gbsnresearch.com
• Security vulnerabilities: follow SECURITY.md.

Citation

See CITATION.cff. When persistent publication identifiers are assigned, cite the specific version used in your work.