Arabic Correspondence Assistants for Royal-Court Class Workloads

Inside the secretariats that serve heads of state, the workload is not email. It is ceremony. A condolence to a neighbouring monarch by sundown, a congratulatory cable on a national day, a draft Eid greeting that must read in the institution's voice, a quiet diplomatic note that will be remembered for years. The drafting bar is high, the time pressure is real, and the embargo is absolute. This is the workload an Arabic correspondence assistant has to serve, and the only deployment shape that survives the room is fully on-premise. The pillar piece on defence AI Arabic triage covers the broader doctrine. This article zooms into the ceremonial-drafting corner.

The ceremonial-drafting problem

A head-of-state class secretariat carries five overlapping correspondence streams, each with its own register and its own failure mode:

• Royal speeches and addresses. Long-form, rhythmic, often Qur'anic in cadence, drafted weeks ahead and edited in marathon sessions. The register sits between Modern Standard Arabic and Classical Arabic, and a single ill-chosen word becomes tomorrow's headline.
• Eid and national-day greetings. Hundreds of bespoke letters in 48 hours, each tuned to the recipient's rank, country, and prior relationship. Templating is tempting and fatal: the wrong template sent to the wrong head of state is a diplomatic incident.
• Condolences. The most sensitive stream. Tone, length, and the choice of religious phrasing all carry weight. Late delivery is read as indifference; a wrong honorific is read as insult.
• Congratulations. Births, accessions, electoral wins, anniversaries. The institutional voice has to feel personal, never form-letter.
• Diplomatic notes verbales. Tightly coded, often bilingual, with a fixed structural grammar that mirrors the recipient ministry's conventions.

Every one of these is high-stakes, high-volume, and locked behind a strict need-to-know wall. Existing knowledge management tools cannot help; chiefs of staff, senior advisors, and protocol officers carry the institutional memory in their heads, and that memory ages with them.

Why public-cloud LLMs are not an option

The temptation is obvious. MSA is the lingua franca of Arab diplomacy, and frontier cloud models handle it well enough on the surface. The problem is not capability. It is jurisdiction.

A draft condolence letter is classified the moment it is composed. The recipient, the wording, even the timing of the draft can move markets and bilateral relationships. Sending that draft to a US-jurisdiction or EU-jurisdiction inference endpoint hands the document to a foreign provider, where it sits on logs, gets reviewed for safety, and may be retrievable under that jurisdiction's legal process. Even academic surveys of AI in diplomatic translation flag confidentiality and contextual accuracy as the central risks. At head-of-state class, the embargo is total: nothing leaves the building, ever.

That rules out every public API, every cloud-hosted enterprise tier, and every model where the weights live somewhere the secretariat does not own. The deployment has to sit inside the perimeter, on hardware the institution controls, behind an air gap or a near-air-gap.

On-premise architecture, four moving parts

The reference shape Hosn deploys for this class of workload is straightforward, four components in a single rack:

1. Falcon Arabic, drafting model. TII's Arabic-first model handles the ceremonial register without the dialect drift that plagues general-purpose multilingual systems. It produces the first draft, lifts cadence from retrieved precedents, and respects the honorifics in the prompt template.
2. Qwen 3.6, structured reasoning model. Long-context, strong at extracting structure from large historical archives. It powers the retrieval-augmented layer: the assistant pulls the closest historical letters for a given occasion, summarises them, and feeds them into the drafter as exemplars.
3. RAG over historical correspondence. Decades of speeches, condolences, congratulations, and diplomatic notes are indexed inside the perimeter. The index is partitioned by recipient country, occasion type, and protocol rank. The institutional voice lives here, not in the model weights, which keeps the system auditable and updatable.
4. Style classifier. A small fine-tuned head reads each draft and rates it on register (ceremonial, formal, neutral), length envelope, religious phrasing density, and salutation rank. Drafts outside the envelope are flagged before any human sees them.

The whole stack runs on a single sovereign-class appliance with redundant inference nodes. No call leaves the rack. No telemetry, no anonymous usage stats, no model-improvement opt-outs to manage. The assistant is a closed system by construction.

Operational guardrails

Models are the easy half. The discipline that makes a Royal Court Arabic AI deployable is operational.

• No auto-send. The system has no outbound mail, no signing key, no integration that can dispatch a letter. It produces drafts and stops.
• Two-eyes review on every draft. A named drafter requests, a named reviewer approves, a named signatory signs. The chain is enforced at the application layer; bypass is not a configuration option.
• Tamper-evident audit log. Every prompt, every retrieval hit, every model output, every approval click writes to an append-only log the inspector general's office can read. Logs are signed and replicate to a separate enclave inside the perimeter.
• Quarterly voice review. The retrieval index is re-curated by a small editorial board. Letters that drifted are tagged, exemplary letters are weighted up, and the institutional voice is recalibrated without retraining the underlying model.
• Personnel-bound access. Login is bound to the secretariat's existing identity system, with hardware tokens and per-occasion role grants. Drafters cannot read other drafters' work.

The result is an assistant that drafts faster, reads in the institution's voice, never leaks, and leaves a perfect trail behind it.

Closing

Ceremonial Arabic correspondence is the highest-trust drafting workload a state runs. Public-cloud LLMs cannot enter the room. An on-premise Falcon plus Qwen 3.6 stack, anchored on a private RAG over the institution's own correspondence and gated by two-eyes review and an audit log, is the deployment shape that fits. Email [email protected] for a one-hour briefing on Royal Court Arabic AI deployments.