GOVERNANCE FRAMEWORK IA · DOI 10.5281/zenodo.23583886

AEXIA™

Accountable · Ethical · eXplainable · Interpretable · Auditable AI

An integrated framework that articulates, within a single operational and measurable structure, explainability, organizational maturity and the individual maturity of deployed AI systems.

6
Organizational Pillars
N0-N5
Organizational maturity levels
S0-S9
System maturity levels
20
Frameworks Sources

POSITIONING

The void filled by AEXIA

Four frameworks currently structure AI Governance, but none directly addresses the question of Algorithmic Responsibility for the equilibria produced by learning systems.

FrameworkWhat it providesWhat it doesn't provide
NIST AI RMF 1.0Cycle 4M (MAP, MEASURE, MANAGE, MONITOR) · 7 caractéristiques de l'IA de confianceMéthodes XAI (SHAP, LIME) · Governance des explications
ISO/IEC 42001AI management systemEvaluation of the maturity of a given system
MLTRL (Lavin et al., 2022)Maturity scale per ML system (9 Levels)Governance organisationnelle · Compliance réglementaire
EU AI Act (2024/1689)Legal obligations for high-risk systemsInstrument de Measurement · méthode d'évaluation

AEXIA™ sits at the intersection of these four spaces. It does not replace any of them : it orchestrates them around a question that none directly addresses - that of the Algorithmic Responsibility of the equilibria produced.

AEXIA IN BRIEF

Understanding AEXIA in 2 minutes

AEXIA is an Explainable AI Governance framework. It does not replace any existing framework - it orchestrates them to answer a question that none directly addresses: how to guarantee the Algorithmic Responsibility of the equilibria produced by learning systems.

🧭

A Dual Axis

The maturity of the organization (AEXIA-O) and that of each system (AEXIA-S) are measured separately and then crossed. Their intersection determines what the organization is allowed to deploy.

🧩

Three original contributions

DCME (multi-scale causal explanation), ECS (consistency score), RAD (Algorithmic Responsibility debt). These three concepts do not exist in any other framework.

🛡️

Four robustness mechanisms

Weakest link, Evidentiary gates, Automatic downgrading on drift, Risk proportionality. This is what distinguishes an operational framework from a declarative reference.

What AEXIA concretely provides

📋 A self-assessment grid

6 Pillars (Foundations, Architecture, Governance, Compliance, Operations, Measurement) × 6 maturity levels, to situate your organization.

🎴 A system sheet template

The AEXIA Card: the artifact a regulator will ask for first. AI Act classification, S level, XAI method, monitoring plan.

🔄 A 6-phase Lifecycle

MAP → DEFINE → BUILD → GOVERN → OPERATE → IMPROVE. Each phase has its deliverables and indicative durations.

📚 20 orchestrated frameworks

NIST AI RMF, ISO/IEC 42001, EU AI Act, OECD, UNESCO, DAMA-DMBOK, COBIT, CMMI, TOGAF, ITIL, MLTRL and 9 others.

Architecture

A Dual Axis Architecture

A mature organization can operate an immature system; an immature organization can host a technically flawless system. AEXIA separates these two properties and crosses them to produce the deployment decision.

O

AEXIA-O - Organization

Maturity of the organization to govern Explainable AI. Evaluated 1 to 2 times a year on 6 Pillars and 6 Levels (N0 to N5).

N0 Non-existent → N5 Excellence
S

AEXIA-S - System

Maturity of a given AI system. Re-evaluated at each gate passage and each incident. Aligned with the MLTRL of Lavin et al. (Nature Communications, 2022).

S0 First principles → S9 In monitored service

O × S crossing matrix - Deployment decision grid

Org. maturitySystems S0-S5Systems S6-S7Systems S8-S9
N0-N1Tolerated: exploratory phaseProhibitedMAJOR ALERT
N2Tolerated with traceabilityAuthorized under conditionLimited risk only
N3NormalGate passed on evidenceAuthorized, including high risk
N4-N5NormalNormal, automatedWith automated Downgrading

ORIGINAL CONTRIBUTIONS

DCME, ECS and RAD

Three intellectual contributions by Mohamadou SIDIBE. They constitute the only element of AEXIA that no other framework can claim, and differentiate AEXIA from any synthesis of existing references.

🔬

DCME

DYNAMIC CAUSAL MULTI-SCALE EXPLANATION

DCME is a multi-scale presentation of causal Shapley values, which decomposes each variable's contribution into a direct part and a redistribution mediated by the causal graph. It is designed for audit documentation (AEXIA Card), not as a new family of explainers.

φC = φD + ψ
Σψ = 0
ΣφC = ΣφD = f(x) − E[f(x)]
📊

ECS

EXPLANATION CONSISTENCY SCORE

ECS is a local stability metric of an explanation: it evaluates how much the explanation φ(f, x) varies when recalculated on neighboring points. It does not measure the fidelity of the explanation to the model nor its explanatory quality.

ECS(f,x,ε) = 1 − min(1, (1/N) Σ ‖φ(f,x) − φ(f,xₖ)‖₁ / (‖φ(f,x)‖₁ + δ))
📉

RAD

RESPONSIBILITY ALGORITHMIC DEBT

RAD summarizes the gap between the expected maturity and the effective maturity of an AI system portfolio, weighted by regulatory criticality. Version 3.1 distinguishes a simplified primary indicator, RAD-lite, and an optional cumulative variant.

RAD-lite = Σᵢ βᵢ · max(0, S_att,i − S_eff,i) / Σᵢ βᵢ · S_att,i

OPERATIONAL ROBUSTNESS

The 4 mechanisms that distinguish AEXIA

These mechanisms distinguish an operational framework from a declarative reference. Three are borrowed from systems engineering (NASA, MLTRL); the fourth is specific to learning AI.

⛓️

1. Weakest link rule

The published AEXIA-O level is that of the weakest pillar, never the weighted average. Same for a system: it is the least mature component that determines the global level.

🚪

2. Evidentiary gates

No level passage is obtained by declaration. Each transition requires a closed set of artifacts, examined in review: risk sheet, XAI justification, compliance file, monitoring pipelines.

↘️

3. Downgrading (Specific IA)

A qualified rocket remains qualified. An AI system does not. Data drift, explainability incident, change in user population: so many triggers for automatic downgrading.

⚖️

4. Risk proportionality

Applying all of AEXIA to a low-stakes internal recommendation model is a waste. The framework adapts to the AI Act risk level.

ORGANIZATIONAL ARCHITECTURE

The 6 AEXIA-O Pillars

Each pillar is evaluated on 6 maturity Levels (N0 to N5). The published level of the organization is that of the weakest pillar.

F

Foundations

Mathematical and theoretical foundations: SHAP, LIME, DCME.

DAMA-DMBOK · ISO 8000 · DCAM

A

Architecture

Technical and organizational XAI infrastructure.

TOGAF · DAMA-DMBOK · DCAM · MLflow

G

Governance

Roles, responsibilities, policies, control.

COBIT · DGI · ISO/IEC 38505 · 38507

C

Compliance

Regulation, ethics, audit.

NIST AI RMF · EU AI Act · OECD · UNESCO · RGPD

O

Operations

Operational excellence, project management.

ITIL v4 · Agile · PRINCE2 · Lean/Six Sigma

M

Measurement & Amélioration

Evaluation, Measurement, continuous improvement.

CMMI · DCAM · ISO 8000 · Six Sigma

Lifecycle

6 operational phases

01
MAP
2-4 wks.
02
DEFINE
4-8 wks.
03
BUILD
2-6 mos.
04
GOVERN
Continuous
05
OPERATE
Continuous
06
IMPROVE
Quarterly

RESOURCES

Download documents AEXIA

All documents are published under CC BY 4.0 license. The reference version is Framework v3.1.

📘 Founding documents

🧪 Assessment instruments

🎴 AEXIA Cards (FR / EN)

Cite AEXIA in your work

SIDIBE, M. (2026). AEXIA Framework - Accountable, Ethical,
eXplainable, Interpretable, Auditable AI Governance Framework,
Version 3.1. AI BIAIS INITIATIVE. DOI: 10.5281/zenodo.23583886