ai-safe2-framework

AISM Scoring Matrix Methodology

AISM Scoring Matrix Methodology

Documentation updated: 2026-10-02. Clarifies the existing Sovereignty Score and introduces TCP and Evidence Assurance v1.0 as separate assessment methods. See the assessment guide for the three instruments, Technology Card workflow, and worked example. Original methodology dates below identify their baseline; this update does not change scoring formulas.

Organizational scoring boundary

The Sovereignty Score applies to an organization, not an individual product, paper, benchmark, or vendor feature. Technology evidence may support specific assessment cells only after the assessor evaluates organizational coverage, robustness, and sovereignty assurance. The Technology Contribution Profile and Evidence Assurance Level are separate review instruments. They do not change this scoring rubric, its weights, the current CLI evidence-grade calculation, or previously issued maturity results.

AI Sovereign Maturity (AISM) Model: Scoring Framework for the Agentic AI Era

Version: 1.0
Date: March 2026
Organization: Cyber Strategy Institute
Framework: AI SAFE2 v3.1; AI Sovereign Maturity (AISM) Model Status: CSI assessment methodology v1.0


1. Purpose & Scope

This document establishes a rigorous, quantitative scoring methodology for the AI Sovereign Maturity (AISM) Model, enabling organizations to measure, benchmark, and advance their maturity across all five pillars of the AI SAFE2 Framework v3.1. The methodology is purpose-built for the agentic AI era, where autonomous multi-agent systems, non-human identities (NHI), and distributed AI architectures create governance challenges that no existing framework fully addresses.

1.1 Design Principles

Principle Description
Sovereignty-First Every metric ties back to organizational control over AI systems — not mere compliance
Agentic-Native Scoring accounts for multi-agent orchestration, swarm behaviors, and autonomous decision chains
Evidence-Based All scores require verifiable artifacts, not self-reported assertions
Multi-Dimensional Combines coverage, robustness, and stakeholder diversity — adapted from IEEE/NIST AI RMF maturity research
Lifecycle-Aware Scores differentiate pre-deployment, deployment, and post-deployment maturity

2. Evaluation of Scoring Methodology Options

Five candidate scoring approaches were evaluated against criteria critical to the agentic AI era.

2.1 Candidate Options

Option A: IEEE/NIST AI RMF Flexible Maturity Model (Dotan et al., 2024)

Option B: NIST CSF Dual-Survey Weighted Maturity (Bernardo et al., 2025)

Option C: Sandia Maturity-Based Certification with Measurement Mechanisms (Darling et al., 2026)

Option D: CSA AI Controls Matrix (AICM) + STAR for AI (CSA, 2025)

Option E: Microsoft Responsible AI Maturity Model (RAI MM, 2023)

2.2 Evaluation Criteria for the Agentic AI Era

Criterion Weight Description
Agentic AI Coverage 20% Multi-agent, swarm, A2A, orchestration controls
NHI & Machine Identity 15% Service accounts, AI agent identities, credential lifecycle
Supply Chain Integrity 10% Model signing, SBOM, provenance chains
Memory & Context Security 10% RAG poisoning, agent memory, context injection
Quantitative Rigor 15% Measurable metrics, not just Y/N checklists
Evidence Requirements 10% Artifact-based verification, not self-assertion
Enterprise Procurement 10% Compliance crosswalks, audit readiness, certification path
Sovereignty Alignment 10% Organizational control, autonomy, self-governance of AI

2.3 Comparative Scoring Matrix

Criterion (Weight) Option A: IEEE/NIST Option B: NIST CSF Option C: Sandia Option D: CSA AICM Option E: MS RAI MM
Agentic AI Coverage (20%) 1 1 2 2 1
NHI & Machine Identity (15%) 1 1 1 2 1
Supply Chain Integrity (10%) 1 1 1 3 1
Memory & Context Security (10%) 1 1 1 2 1
Quantitative Rigor (15%) 4 4 5 2 3
Evidence Requirements (10%) 5 3 4 3 3
Enterprise Procurement (10%) 3 3 1 5 2
Sovereignty Alignment (10%) 2 1 2 2 1
Weighted Score 2.30 1.90 2.30 2.55 1.55

Scale: 1 = No coverage, 2 = Minimal, 3 = Partial, 4 = Strong, 5 = Comprehensive

2.4 Key Finding

No existing approach scores above 2.55/5.00 on agentic AI era requirements. Each excels in isolated dimensions but fails holistically. This validates the need for a composite methodology that synthesizes the best elements:

Inherited Element Source
3-metric scoring (Coverage, Robustness, Diversity) IEEE/NIST AI RMF MM (Option A)
Expert-weighted importance calibration NIST CSF Dual-Survey (Option B)
5-level maturity with measurement mechanisms Sandia Maturity Certification (Option C)
Control-objective taxonomy + compliance crosswalks CSA AICM (Option D)
Empirically-derived dimensions + interdependency awareness MS RAI MM (Option E)

3.1 The AISM Composite Scoring Framework

The recommended methodology — the AISM Sovereignty Score — combines the strongest elements from all five approaches into a unified system purpose-built for AI SAFE2 v3.1.

3.2 Maturity Levels (5-Level Scale)

Level Name Description NIST Tier Equivalent
1 Chaos (alias: Reactive) Ad hoc, no formal AI governance; controls are incident-driven Partial
2 Visibility (alias: Aware) AI risks recognized; basic policies exist but inconsistently applied Risk Informed
3 Governance (alias: Controlled) Formalized controls across all 5 pillars; documented processes; regular audits Repeatable
4 Control (alias: Sovereign) Organization has full visibility and runtime control; adaptive governance; proactive threat response Adaptive
5 Sovereignty (alias: Autonomous Governance) Self-improving governance; formal verification of critical controls; continuous measurement mechanisms; AI sovereignty fully realized Optimizing

3.3 Scoring Dimensions (5 Pillars × 6 Dimensions)

Each AI SAFE2 pillar is assessed across six scoring dimensions — adapted from digital maturity research and agentic AI requirements:

# Dimension DMM Equivalent Agentic AI Application
D1 Governance Architecture Organization (36) AI oversight structures, board accountability, RACI for AI decisions, sovereignty posture
D2 Risk Strategy & Posture Strategy (34) AI risk appetite, threat modeling, sovereignty roadmap, regulatory alignment
D3 Technical Controls Technology (28) Agent sandboxing, NHI controls, memory security, supply chain signing, kill switches
D4 Safety Culture & Workforce Culture (27) AI safety training, prompt engineering education, agentic system awareness, red team culture
D5 Lifecycle Processes Process (21) AI system lifecycle management, change control, incident response, continuous monitoring
D6 Sovereignty Assurance NEW — AISM-native Organizational autonomy over AI, data sovereignty, model provenance, self-governance capability

3.4 Three-Metric Evaluation (Per Dimension Per Pillar)

Each of the 30 assessment cells (5 pillars × 6 dimensions) is scored using three metrics, each rated Low (L), Medium (M), or High (H):

Metric 1: Coverage

How completely does the organization implement the relevant AI SAFE2 subtopics?

Rating Criteria
High ≥80% of applicable subtopics implemented with documented controls
Medium 40–79% of applicable subtopics implemented
Low <40% of applicable subtopics implemented

Metric 2: Robustness

How rigorous and resilient are the implemented controls?

Assessed against six robustness ideals (adapted from NIST implementation tiers):

  1. Regular — Controls executed routinely, not ad-hoc
  2. Systematic — Organization-wide policies, not team-specific
  3. Trained Personnel — Roles defined, staff certified/trained
  4. Sufficiently Resourced — Budget, tools, compute allocated
  5. Adaptive — Controls evolve with threat landscape; regular reviews
  6. Cross-Functional — Involves security, engineering, legal, leadership
Rating Criteria
High 5–6 robustness ideals satisfied
Medium 3–4 robustness ideals satisfied
Low 0–2 robustness ideals satisfied

Metric 3: Stakeholder Diversity & Sovereignty Assurance

How diverse are inputs and how sovereign is organizational control?

This metric extends the IEEE model’s “Input Diversity” to include sovereignty:

Rating Criteria
High Diverse internal/external stakeholders inform controls; organization maintains full autonomous control over AI governance decisions; no single vendor lock-in
Medium Some stakeholder diversity; partial organizational control; moderate vendor dependencies
Low Minimal stakeholder input; governance decisions driven by single team or vendor; significant sovereignty gaps

3.5 Score Calculation

Cell Score (1–5)

Each cell uses the HHH rubric from the IEEE/NIST AI RMF model:

Score Metric Combination Point Total (L=1, M=2, H=3)
5 HHH 9
4 HHM 8
3 HMM, HHL, HML, MMM 6–7
2 MML, MLL, HLL 4–5
1 LLL 3

Pillar Score

Each pillar score is the weighted average of its 6 dimension scores, with weights calibrated by expert panel (adapted from Bernardo et al. methodology):

Pillar Score = Σ (Dimension Score × Dimension Weight) / Σ Dimension Weights

Default weights (adjustable per organizational context):

Dimension Default Weight Rationale
D1: Governance Architecture 0.20 Foundation — without governance, controls are ad-hoc
D2: Risk Strategy & Posture 0.18 Direction — strategy determines investment priorities
D3: Technical Controls 0.25 Execution — largest subtopic coverage in AI SAFE2
D4: Safety Culture & Workforce 0.12 Enablement — people execute controls
D5: Lifecycle Processes 0.15 Operations — day-to-day process maturity
D6: Sovereignty Assurance 0.10 Differentiation — unique to AISM

Overall AISM Sovereignty Score

AISM Score = Σ (Pillar Score × Pillar Weight) / Σ Pillar Weights

Default pillar weights (equal by default, adjustable):

Pillar Default Weight
P1: Sanitize / Isolate 0.20
P2: Audit / Inventory 0.20
P3: Fail-Safe / Recovery 0.20
P4: Engage / Monitor 0.20
P5: Evolve / Educate 0.20

3.6 Combined Risk Score Integration

For organizations also tracking technical vulnerabilities, AISM integrates with CVSS and the AI SAFE2 v3.1 combined-risk method:

Combined Risk = CVSS Base Score × ((100 - Pillar Score as %) / 10)

Example: CVSS 8.5 (High) + AISM Pillar Score 3.8/5.0 (76%) = 8.5 × (24/10) = 20.4 → High Risk
Example: CVSS 8.5 (High) + AISM Pillar Score 4.5/5.0 (90%) = 8.5 × (10/10) = 8.5 → Moderate Risk


4. AI Sovereignty Matrix Visualization

4.1 Sovereignty Matrix Plot

The AISM Sovereignty Score is visualized as a radar chart (spider diagram) with:

Additionally, a heatmap matrix plots:

4.2 Maturity Classification

AISM Score Range Canonical Maturity Level Risk Posture Action Required
4.50 – 5.00 Sovereignty Sovereign Maintain excellence; share best practices
3.50 – 4.49 Control Controlled Close specific gaps; advance to Sovereignty
2.50 – 3.49 Governance Developing Prioritize critical pillars; formalize controls
1.50 – 2.49 Visibility Emerging Establish baseline governance immediately
1.00 – 1.49 Chaos Exposed Critical risk — immediate intervention required

Canonical maturity names come from maturity-model.md. Risk-posture labels describe urgency and must not be emitted as maturity levels.


5. Evidence Requirements

Every score must be substantiated with verifiable evidence. The AISM requires three categories of evidence per assessment cell:

Evidence Type Description Examples
Documentary Written policies, procedures, architecture documents AI governance policy, agent inventory, SBOM records
Operational Logs, dashboards, audit trails demonstrating active controls SIEM logs showing NHI monitoring, kill switch test records
Attestation Third-party or cross-functional verification Penetration test reports, red team findings, compliance audit results

6. Comparison: AISM vs. Current Approaches

Capability AISM (Recommended) IEEE/NIST MM CSA AICM MS RAI MM NIST CSF MM
Maturity levels 5 (Reactive→Autonomous) 5 (1–5 scale) 3 (STAR levels) 5 (Latent→Leading) 5 (Very Poor→Excellent)
Scoring metrics 3 (Coverage, Robustness, Sovereignty) 3 (Coverage, Robustness, Diversity) Binary (Y/N/Partial) Level criteria Weighted surveys
Agentic AI controls ✅ 35 sub-domains ❌ None ⚠️ Limited ❌ None ❌ None
NHI governance ✅ 10 dedicated controls ❌ None ⚠️ Minimal ❌ None ❌ None
Supply chain signing ✅ OpenSSF OMS ❌ None ⚠️ General ❌ None ❌ None
Memory security ✅ 6 sub-domains ❌ None ❌ None ❌ None ❌ None
Sovereignty dimension ✅ Native ❌ None ❌ None ❌ None ❌ None
Compliance crosswalks ✅ 7+ frameworks ⚠️ NIST AI RMF only ✅ ISO 42001, NIST ❌ None ⚠️ NIST CSF only
Evidence requirements ✅ 3 categories ✅ Evidence-based ⚠️ Questionnaire ❌ Guidance only ⚠️ Survey-based

7. References

  1. Dotan, R., Blili-Hamelin, B., Madhavan, R., Matthews, J., & Scarpino, J. (2024). “Evolving AI Risk Management: A Maturity Model based on the NIST AI Risk Management Framework.” arXiv:2401.15229v1.
  2. IEEE USA. (2025). “A Flexible Maturity Model for AI Governance Based on the NIST AI Risk Management Framework.”
  3. Bernardo, L., Malta, S., & Magalhães, J. (2025). “An Evaluation Framework for Cybersecurity Maturity Aligned with the NIST CSF.” Electronics 2025, 14, 1364.
  4. Darling, M.C., et al. (2026). “Toward Maturity-Based Certification of Embodied AI.” arXiv:2601.03470.
  5. Cloud Security Alliance. (2025). “AI Controls Matrix (AICM) v1.0.” 243 control objectives, 18 security domains.
  6. Vorvoreanu, M., et al. (2023). “Responsible AI Maturity Model.” Microsoft Research.
  7. NIST. (2025). “Cybersecurity Framework Profile for Artificial Intelligence.” NIST IR 8596 iprd.
  8. NIST. (2023). “AI Risk Management Framework (AI RMF 1.0).” NIST AI 100-1.
  9. Peixoto, E.C., et al. (2025). “Clarifying Core Dimensions in Digital Maturity Models.” arXiv:2602.07569v1.
  10. Cyber Strategy Institute. (2026). “AI SAFE2 Framework Version 3.1 — Protocol Governance and Enforcement Plane Update.”
  11. RAI Institute. (2024). “Responsible AI Maturity Model.” Five stages: Aware → Transformative.
  12. iQomply. (2026). “ISO 42001 Maturity Model.” Five levels: Initial → Mature.

© 2026 Cyber Strategy Institute. Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).