Feature
32 components,
zero black boxes.
WriteX is not a prompt on someone else's model. The Relay Architecture chains 32 purpose-built language components, each accurate, explainable, and fast enough to check text as you type.
Words split and analysed against their root-and-pattern forms, which is what makes Arabic checkable at all.
Purpose-built components run per category: spelling, grammar, morphology, semantics, style, formatting, names, quotations.
Your style guide, dictionary, and protocol rules applied on top of the language result.
Each proposed change tied back to the rule that produced it, which is why every card can explain itself.
Schematic: the engine chains 32 specialised components in relay, shown here as the four stages they group into rather than one box per component. Nothing in the chain is a general-purpose language model.
Relay vs transformers
Specialised beats general at language checking.
General-purpose transformer models reason impressively but check language averagely: not real-time, hard to customize, and unexplainable. The Relay Architecture chooses the opposite trade: specialised models with high accuracy on language tasks, real-time feedback, easy customization, and full explainability.
The trade-off is now measured rather than asserted. In the L3 evaluation, published September 2026, the engine behind WriteX and 26 large language model configurations checked the same linguist-reviewed Arabic sentences under the same prompt and scoring. L3 scored 80.29% edit F0.5 against 44.68% for Claude Fable 5.1, the strongest model, and it was the only system with precision above 80% and recall above 60% together. It ran at 52.28 standardized tokens per second, second in the field, and used approximately 0.000014 kWh per 1,000 tokens at accelerator scope on an NVIDIA A10 24 GB. On-premises deployment is a product capability, described on this page. Method, limits and every table: the L3 results.
- Morphological compression built for Arabic's root-and-pattern system.
- Every correction traceable to a rule your linguists can audit.
- Accredited by the Digital Government Authority.
Fully owned
Our models, your deployment.
Lisan owns its models end to end: over 19.5 million lines of software developed internally by the region's leading Arabic NLP team. That is what makes sovereign and on-premises deployments possible, and why customization requests do not wait on a foreign vendor.
Frequently asked questions
Is WriteX built on a third-party LLM?
No. The engine is Lisan's own Relay Architecture: 32 specialised AI components developed in-house, which is why it runs on-premises, responds in real time, and explains every correction.
Every word your institution writes, checked.
Official letters to social posts, Arabic and English alike. One assistant that proofreads, enforces your style, and explains every correction, wherever your teams write.
12 error types · Institutional style guides · Works in Office, browsers, and the web editor