3 · SPECIFY Layer 3 Grounding & Layer 5 Interface · Arabic & Farsi Specialty

Your foundation model passed English benchmarks. Your users in Riyadh, Dubai, Cairo, and Tehran do not speak translated English.

When you deploy an in-cabin voice assistant, a regional customer agent, or a multimodal safety classifier into Arabic- and Persian-speaking markets, machine-translated test sets hide the exact failures your users hit on day one: dialectal syntax, Arabizi/Finglish code-switching, omitted diacritics, and stripped Zero-Width Non-Joiners (ZWNJ). We ground and validate your Layer 3 memory and Layer 5 interfaces across 7 Arabic varieties and 3 Persian varieties.

ARABIC · 7 VARIETIES

7 Arabic Varieties We Validate for Your Deployment

We test your models across the regional spoken registers your customers and operators actually use—never collapsing dialectal intent into generic translation.

Gulf (Khaleeji)

الخليجية

UAE, Saudi Arabia, Kuwait, Qatar, Bahrain, Oman

In-cabin voice navigation, Gulf lexical markers (شلون، ودي، سيده), code-switched English road/tech tokens, and acoustic telematics.

Levantine (Shami)

الشامية

Jordan, Lebanon, Syria, Palestine

b-/m- aspectual verb prefixes, negation particles (ما/مو/مش), trilingual AR-EN-FR code-switching, and conversational safety.

Egyptian (Masri)

المصرية

Egypt

Post-positioned interrogatives, circumfix negation (ما...ش), high-density urban navigation landmarks, and ASR under street noise.

Iraqi (Mesopotamian)

العراقية

Iraq

gilit/qeltu phonological shifts (/k/→/tʃ/, /q/→/g/), progressive particle (دا/جاي), and indirect request phrasing.

Maghrebi (Darija)

المغاربية / الدارجة

Morocco, Algeria, Tunisia, Libya

Consonant cluster timing, French/Amazigh lexical borrowing, Arabizi numerals (3, 7, 9), and severe OOV rates in standard tokenizers.

Sudanese

السودانية

Sudan

Distinctive lexical/aspectual markers (داير، هسي، شنو), phonetic realization, and representation fairness in speech/text models.

Modern Standard Arabic (MSA)

الفصحى المعاصرة

Pan-Arab formal, regulatory, and technical register

Full morphological inflection, homograph disambiguation without tashkeel, legal/safety notices, and bidirectional technical notation.

PERSIAN · 3 VARIETIES

3 Persian Varieties We Validate for Your Stack

Orthographic, lexical, and acoustic coverage across Iranian Persian, Afghan Dari, and Tajik Persian.

Iranian / Tehrani Persian

فارسی ایران / تهرانی

Perso-Arabic (Farsi ی U+06CC / ک U+06A9)

Colloquial vs. formal register shifts (می‌خوام / می‌خواهم), strict ZWNJ (نیم‌فاصله) preservation in plurals and verb prefixes, and Ezafe acoustic stress.

Dari (Afghan Persian)

دری (افغانستان)

Perso-Arabic

Preservation of majhul vowels (/ē/, /ō/), distinct administrative/technical vocabulary, and formal spoken cadence.

Tajik Persian

تاجیکی / Тоҷикӣ

Cyrillic & Perso-Arabic cross-script

Cross-script transliteration consistency, northern syntax patterns, and shared lexical semantics across script boundaries.

SERVICES & EDGE-CASE COVERAGE

Data Collection, Evaluation Sets, Human Review, ASR/TTS & Red-Teaming

Code-Switching: Arabizi, Finglish & Embedded Technical Tokens

Your operators and customers mix scripts mid-sentence: Gulf Arabic with English sensor identifiers (LiDAR, GNSS, VLA), or Tehrani Persian written in Latin Finglish alongside numeric units.

RTL Bidirectionality, Tashkeel & ZWNJ Edge Cases

We test your tokenizer and UI pipeline when Arabic diacritics (تشكيل) are omitted or adversarial, and when Farsi ZWNJ (U+200C) is stripped or replaced by standard spaces.

Culturally Grounded Safety & Subject-Aware Guardrail Testing

Adversarial prompts crafted by native specialists using regional idioms and dialectal euphemisms—ensuring your guardrails neither let regional exploits through nor block legitimate security and medical workflows.

ASR & TTS Evaluation Across Regional Accents

We evaluate your Automatic Speech Recognition (ASR) under cabin road noise and your Text-to-Speech (TTS) prosody, Ezafe realization, and homograph pronunciation accuracy.

QUALITY CONTROLS

Native Reviewers, Agreement Metrics & Provenance

Native Reviewers per Dialect Variety

Every dialectal subset we build for your stack is authored and verified by native speakers of that specific regional variety—never cross-assigned.

Inter-Annotator Agreement Metrics

We report Krippendorff’s alpha and multi-rater agreement while distinguishing legitimate regional variation from annotation noise.

Provenance & Orthographic Integrity

Full lineage tracking per item, automated verification of Persian ی/ک and ZWNJ (U+200C), and bidirectional <bdi> token isolation.

Sample / Illustrative

Sample Multimodal & Linguistic Evaluation Items

Illustrative evaluation items showing how we test your stack on real Arabic and Farsi script, bidirectional technical tokens, and deterministic + rubric scoring.

Sample / Illustrative · Arabic (Gulf + Technical Code-Switch) sample-ar-gulf-vla-001

Gulf Arabic (Khaleeji) + English Technical Tokens

Autonomous Cabin Voice Command → VLA Trajectory Policy

يا خوي هدّي السرعة شوي، قدامنا تحويلة ومخاريط على اليمين، خلك بالنص ولا تقرب من الـ LiDAR blind spot جنب الشاحنة.

Gloss: Slow down a bit, there is a detour and cones on the right ahead; stay centered and do not enter the LiDAR blind spot next to the truck.

1. Parse Gulf colloquial imperative ('هدّي السرعة', 'خلك بالنص') and English token ('LiDAR blind spot'). 2. Command smooth longitudinal deceleration. 3. Maintain lateral clearance (> 1.2 m OBB buffer) from both right-side cones and adjacent truck without erratic lane departure.

Dialect Intent Parsing: Correctly maps 'سيده/بالنص' and 'تحويلة' without requiring MSA rephrasing.
Reasoning-Action Consistency: Verbalized plan and executed trajectory both increase lateral clearance from the right shoulder.
Geometric Safety (Level A): Minimum OBB clearance > 0 m; TTC remains above 3.0 s throughout the maneuver.

Sample / Illustrative · Farsi (Tehrani + ZWNJ & Bidi Tokens) sample-fa-tehrani-vla-002

Iranian / Tehrani Persian

Sensor Fault Advisory & Safe Stop Instruction

الان سیگنال GNSS توی تونل قطع شده و روی لنز دوربین جلو هم گردوغبار نشسته؛ لطفاً بدون ترمز ناگهانی، سرعتت رو کم کن و برو تو خط اضطراری.

Gloss: The GNSS signal has dropped inside the tunnel and dust has settled on the front camera lens; please reduce speed without sudden braking and pull into the emergency lane.

1. Recognize compound sensor degradation (GNSS dropout + camera occlusion) expressed in colloquial Tehrani register. 2. Down-weight degraded camera/GNSS priors in favor of LiDAR/IMU odometry. 3. Execute a bounded-jerk deceleration into the emergency pull-off.

Orthography & ZWNJ Handling: Tokenizer preserves semantics across 'می‌کنه / لطفاً / گردوغبار' and isolated 'GNSS' token.
Fault-Tolerant Control: Longitudinal jerk stays below comfort/safety ceiling (no phantom hard brake).
Provenance Verification: Episode scored against deterministic Level A simulator state and logged with checkpoint SHA-256.