Bilingual AI for Diwan and Royal Court Correspondence

A Royal-Court-class secretariat does not draft ordinary letters. It produces decrees, royal speeches, congratulatory cables to fellow heads of state, condolence telegrams, accreditation letters for ambassadors, and bilateral correspondence in Modern Standard Arabic and English at the same time. The volume is steady, the deadlines are unforgiving, and the register sits two notches above any commercial style guide. Bilingual AI can compress the drafting load substantially, but only if the architecture treats the text as head-of-state material from the first keystroke. This piece walks through the four design choices that separate a credible deployment from a procurement embarrassment.

The bilingual formal-correspondence problem

Royal Court secretariats globally share a tight set of recurring drafting jobs. The most common are royal decrees and ministerial appointments, which carry their own typography, opening prayer formulae, and statutory citations. Next come royal speeches for opening parliament, national days, and bilateral state visits, which mix high MSA with passages quoted verbatim from prior speeches and from religious or constitutional sources. Then come diplomatic letters: condolences, congratulations, accreditations, and substantive bilateral correspondence, frequently issued bilingually within the same envelope so the receiving palace receives both texts at once. The Wikipedia survey of diplomatic correspondence categorises the genre cleanly into letters, notes, memoranda, and aides-memoire, each with its own protocol.

Arabic ceremonial register is the harder side of this pair. It is recognisable instantly to a literate Omani or Gulf reader and almost impossible to fake. It uses formulaic invocations, archaic adjectives, careful agent-passive choices, and a vocabulary cluster around honour, sovereignty, and protocol that ordinary MSA news writing does not touch. English ceremonial drafting has its own conventions (third-person address, archaic verb forms in royal letters, careful avoidance of contracted forms) but is more forgiving. A bilingual stack has to clear both bars, simultaneously, in a system that two human editors will sign off on.

Why public AI fails this brief

Public, cloud-hosted models fail this brief on two independent axes. The first is data exposure. Drafts of unpublished decrees, speeches before delivery, and unfinished diplomatic letters are categorically embargoed material. Sending them to a foreign-hosted API, even one that promises "no training on inputs", places the bytes inside a foreign jurisdiction with foreign subpoena exposure and foreign telemetry pipelines. The same logic applies to free-tier translation widgets that leak office traffic into vendor logs. The pillar reference for this argument is our piece on on-premise AI for sovereign institutions in Oman and the GCC, which lays out the threat model in full.

The second axis is fitness. Recent research on diplomatic-discourse evaluation found 100 per cent hallucination rates when a frontier general-purpose model was asked to recognise diplomatic hints in Chinese and Russian segments, with role-prompting providing no improvement. A separate journalistic episode involving an AI translation of a Maldivian-Indian head-of-state exchange triggered a public-facing diplomatic retraction. These are not edge cases; they are the modal failure mode of generic models on protocol-grade text. A Royal Court secretariat cannot be the next case study.

The on-premise bilingual stack

The architecture is straightforward, and every piece runs inside the secretariat's own perimeter. Three open-weight models cover the bilingual workload at the quality the brief demands.

• Falcon Arabic for the Arabic side. The Technology Innovation Institute's Falcon Arabic family, including the more recent Falcon-H1 Arabic hybrid Mamba-Transformer release, was trained for fluent MSA plus Gulf-relevant dialect coverage. It is the strongest open-weight Arabic baseline available in 2026, and it ships under terms that allow on-premise deployment with full control of the weights.
• Gemma 4 for English ceremonial drafting and long-context citation. Google DeepMind's Apache-2.0 release covers the English side and handles the long-context retrieval needed when a draft has to reconcile against decades of prior speeches and decrees in a single pass.
• A secretariat-owned authority file and template library. Names, titles, decree numbers, hijri-and-gregorian date pairs, and quoted prior text are pulled from internal records at draft time. The model never invents a name or a quote; it generates only the surrounding ceremonial prose.

Both language models are fine-tuned, lightly, on the secretariat's own past correspondence corpus. This is the step that buys ceremonial register. A LoRA pass on a few thousand internal documents (after PII handling and access-controlled storage) teaches the model to open with the right invocation, address with the right honorifics, and close with the right formula. The corpus never leaves the enclave; the fine-tune runs on the same air-gapped hardware that serves the model.

Operational guardrails

The model is the easy part. The guardrails are what make the system safe to put in front of a head-of-state secretariat. Four are non-negotiable.

• No auto-send. The system has no outbound mail capability of any kind. Drafts arrive in a queue. The dispatch path is the secretariat's existing protocol office, unchanged.
• Two-eyes review. Every draft destined for an external recipient passes through a designated bilingual editor and a senior reviewer. The system enforces the two-signature gate before a draft can be marked ready for the protocol office.
• Audit log. Every prompt, every retrieval, every model version, every editor change, and every signature is recorded with timestamp and operator identity. The log is immutable and survives operator turnover. The retention policy aligns with the secretariat's existing classification regime.
• Authority-file primacy. If a draft cites a decree number, a date, or a quoted prior text that does not match the authority file, the system flags the draft and refuses to advance it until the discrepancy is resolved by a human.

Together these four turn a capable model into a disciplined drafting assistant. The chain of accountability is the same chain that has always governed Royal Court correspondence: the editor edits, the reviewer signs off, and the protocol office dispatches. AI sits one layer below all of them.

If your secretariat is sizing a bilingual drafting assistant against this shape of brief, the next step is a one-hour briefing on architecture, models, and the security posture. Email [email protected] or message +968 9889 9100. Pricing is by quotation, sized to the volume and the air-gap requirement.