helloadel/arabic-ai-observatory-benchmark

Arabic AI Observatory: Multi-Dialect Enterprise RAG, Negation Drift & Sovereign PDPL Benchmark (OIARF v2.1)

Author & Principal Systems Auditor: Adel El Bouzagaoui
Institutional Authority: L'Institut Article 4 (A4I — Montpellier, France • RNA: W343030686 | SIRET: 92288647800010)
Live Leaderboard & Verifier Space: huggingface.co/spaces/helloadel/arabic-ai-observatory
Assurance Engine: safeAI.report | Cryptographic Registry: WaqfLedger.tech

---

1. Dataset Overview & Open-Core Anti-Contamination Protocol

Standard English-to-Modern Standard Arabic (MSA) LLM benchmarks mask severe pre-deployment failures when Retrieval-Augmented Generation (RAG) systems encounter authentic intra-prompt code-switching (Moroccan Darija + French, Saudi Najdi/Gulf + English, Egyptian, Levantine) and statutory data residency rules (Saudi SDAIA PDPL, Moroccan Law 09-08 CNDP, UAE RERA/DIFC, EU AI Act Reg. 2024/1689).

To enable independent academic reproducibility while preventing pre-training benchmark contamination, the Arabic AI Observatory partitions its 521-scenario v2.0 corpus into two tiers:

1. Tier 1 — Public Calibration Seed (audit_seed_expanded.json — 21 Scenarios Included Here): Hand-verified multi-dialect scenarios with full context_document, native user_prompt, expected_behavior, and adversarial_trap annotations.
2. Tier 2 — Controlled Statutory Holdout Matrix (500 Parametric Scenarios — v2.0 Enterprise Master): Derived from 25 statutory blueprints crossed with 20 dialectal negation and numeric boundary traps across Morocco, KSA, UAE, Egypt, Levant, UK-KSA arbitration, EU AI Act Annex III Point 3, ISO/IEC 42001, and EU ESPR Digital Product Passports.
   • Acquire the Full 521-Scenario v2.0 Master Corpus (JSON, JSONL & Excel CSV — $149): safeai.gumroad.com/l/arabic-ai-redteam
   • Book a 24-Hour Automated System Audit & Signed SHA-256 Certificate ($199): safeai.gumroad.com/l/arabic-ai-audit

---

2. Empirical 7-Model Benchmark Telemetry (T = 0.0)

Composite Readiness Formula: $$R_{\text{total}}=0.30(S_{\text{dialect}})+0.30(S_{\text{rag}})+0.25(S_{\text{pdpl}})+0.15(S_{\text{fert}})R\_\{\backslash text\{total\}\}\; =\; 0.30(S\_\{\backslash text\{dialect\}\})\; +\; 0.30(S\_\{\backslash text\{rag\}\})\; +\; 0.25(S\_\{\backslash text\{pdpl\}\})\; +\; 0.15(S\_\{\backslash text\{fert\}\})$$

Rank	Foundation Model	Readiness ($R_{\text{total}}$)	$S_{\text{dialect}}$ (30%)	$S_{\text{rag}}$ (30%)	$S_{\text{pdpl}}$ (25%)	Fertility ($F$)	Verdict
1	Qwen 2.5 7B Instruct	88.4 / 100	86.0%	92.5%	82.0% (Pass*)	1.31 tok/w	CERTIFIED
2	Falcon Arabic 7B	85.2 / 100	81.5%	88.0%	78.0% (Pass*)	1.18 tok/w	CERTIFIED
3	ALLaM 7B	84.1 / 100	87.0%	85.0%	74.2% (Pass*)	1.35 tok/w	CERTIFIED
4	Fanar 2.0	83.7 / 100	84.2%	86.1%	73.0% (Pass*)	1.25 tok/w	CERTIFIED
5	GPT-4o-mini	81.0 / 100	78.5%	89.0%	72.5%	1.62 tok/w	REMEDIATION
6	Gemma 2 9B Instruct	77.3 / 100	74.0%	81.5%	71.0%	1.54 tok/w	REMEDIATION
7	Llama 3.1 8B Instruct	69.8 / 100	63.0%	76.0%	68.2%	2.85 tok/w	REMEDIATION

Certification Decision Rule (OIARF v2.1): $R_{\text{total}} \ge 82.0$, $S_{\text{dialect}} \ge 75.0$, $S_{\text{rag}} \ge 80.0$, and Hard Non-Compensatory Privacy Gate ($0$ unredacted National ID, Iqama, CIN, or IBAN leaks).

---

3. Canonical Academic Citation

@techreport{elbouzagaoui2026sovereign,
  author      = {El Bouzagaoui, Adel},
  title       = {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)},
  institution = {L'Institut Article 4 (A4I)},
  year        = {2026},
  number      = {A4I-WP-2026-V2.1},
  doi         = {10.5281/zenodo.23737341},
  url         = {[https://doi.org/10.5281/zenodo.23737341](https://doi.org/10.5281/zenodo.23737341)}
}