Empirical Sovereign AI Assurance Across the MENA–EU Corridor
Benchmarking dialectal negation drift (-37.0 pp accuracy drop), token fertility economics (2.41x input token multiplier), and statutory data protection compliance (Saudi SDAIA PDPL, Moroccan Law 09-08 CNDP, UAE RERA/DIFC, and EU AI Act Reg. 2024/1689 Art. 4, 14, 26 & Annex III Point 3). Backed by permanent CERN Zenodo DOI indexing and salted SHA-256 WaqfLedger verification.
Live Cryptographic Record Lookup
Test canonical audit receipts anchored on WaqfLedger.tech via RFC 8785 Canonical JSON + 128-bit salt (zero PII retention).
Arabic AI Observatory Leaderboard (7-Model Benchmark • T=0.0)
Evaluated across the 521-scenario v2.0 Master Corpus (21 Open Calibration Seed + 500 Controlled Statutory Holdout Cases)
| Rank | Foundation Model | Architecture | Readiness (R_total) | S_dialect (30%) | S_rag (30%) | S_pdpl (25%) | Fertility (15%) | Benchmark Verdict |
|---|---|---|---|---|---|---|---|---|
| 1 | Qwen 2.5 7B Instruct | Open-Weights | 88.4 / 100 | 86.0% | 92.5% | 82.0% (Pass*) | 1.31 tok/w | CERTIFIED |
| 2 | Falcon Arabic 7B | Regional Foundation | 85.2 / 100 | 81.5% | 88.0% | 78.0% (Pass*) | 1.18 tok/w (Best) | CERTIFIED |
| 3 | ALLaM 7B | Sovereign GCC | 84.1 / 100 | 87.0% | 85.0% | 74.2% (Pass*) | 1.35 tok/w | CERTIFIED |
| 4 | Fanar 2.0 (Verified) | QCRI Foundation | 83.7 / 100 | 84.2% | 86.1% | 73.0% (Pass*) | 1.25 tok/w | CERTIFIED |
| 5 | GPT-4o-mini | Commercial API | 81.0 / 100 | 78.5% | 89.0% | 72.5% | 1.62 tok/w | REMEDIATION (<82.0) |
| 6 | Gemma 2 9B Instruct | Open-Weights | 77.3 / 100 | 74.0% | 81.5% | 71.0% | 1.54 tok/w | REMEDIATION (<82.0) |
| 7 | Llama 3.1 8B Instruct | Open-Weights | 69.8 / 100 | 63.0% | 76.0% | 68.2% | 2.85 tok/w (High Tax) | REMEDIATION (Floor) |
OIARF v2.1 Mathematical & Reproducibility Spec
- 37.0 pp Accuracy Drop Derivation: 7-model mean policy grounding drops from 89.5% on formal MSA control prompts to 52.5% under native dialectal negation shifts (-37.0 percentage points / -41.3% relative).
- 2.41x (+141.5%) Input Fertility Tax: Llama 3.1 (2.85 tok/w) vs. Falcon Arabic (1.18 tok/w) inflates encoded input tokens and prefill KV-cache footprint by 141.5%.
- Open-Core Anti-Contamination Protocol: 21-case calibration seed (audit_seed_expanded.json) is openly inspectable; the 500-case statutory matrix is maintained off-crawl.
Canonical Academic Citation (ORCID / Zenodo)
v2.1 HardenedEl Bouzagaoui, A. (2026). Sovereign AI Assurance: Benchmarking Dialect Drift, Token Economics, and PDPL Compliance in MENA-EU Enterprise Deployments (OIARF v2.1 & Arabic AI Observatory 521-Scenario Specification). L'Institut Article 4 (A4I), Montpellier (RNA: W343030686 | SIRET: 92288647800010). https://doi.org/10.5281/zenodo.23737341
21 open calibration seed cases + 500 parametric statutory scenarios covering Morocco (BAM/IGOC/CNDP), KSA (PDPL/MHRSD/SAMA/ZATCA), UAE (RERA/DIFC), Egypt (CBE/FRA), and EU AI Act Annex III Pt. 3.
Acquire Suite ($149)Automated 24-hour enterprise system audit. Submit up to 200 interaction logs and receive a 2-page PDF compliance certificate anchored on WaqfLedger.tech.
Book Audit SlotOfficial 6-page peer-review-hardened whitepaper (OIARF v2.1) indexed under DOI 10.5281/zenodo.23737341 and synced to ORCID.
Visit Standards Body