AI Patterns for Parliamentary Research Inside Majlis A'Shura and the Council of State

A research aide in the Majlis A'Shura building is asked, on a Sunday morning, to pull comparable provisions from four GCC labour codes, summarise the redline against last session's draft, find every minute in which a member raised the same article in the past two terms, and have a bilingual one-page brief on the committee chair's desk before the afternoon sitting. The same call lands again on Tuesday, and again on Thursday, across every standing committee. The volume is real, the comparative law moves quickly, and the drafts are member-confidential. This piece sets out three concrete on-premise AI patterns that recover aide hours inside Majlis A'Shura and the Council of State without sending a single byte of draft text out of the chamber.

The two-chamber Omani research load

Oman's Council of Oman is bicameral. The Majlis A'Shura sits as the elected lower chamber and the Council of State sits as the appointed upper chamber, both operating under the framework set by Royal Decree 6/2021 (the Basic Statute of the State) and the State Council Law. Each chamber runs its own standing committees on legal affairs, economy and finance, education and culture, health and social affairs, services and public utilities, and security and defence. Each committee is supported by a small research office that produces briefs, comparative-law memos, and meeting summaries in Arabic, with English annexes for international engagements.

The work compounds across two parallel calendars. A bill referred from the Council of Ministers reaches Majlis A'Shura first, is amended in committee, returns to the floor, and then moves to the Council of State for review. The same draft can pass through the research office of one chamber and the research office of the other within weeks, with the comparative-law table needing to be defensible in both rooms. That is what creates the duplication, and that is what an on-premise AI pattern is built to remove.

Three patterns: bill comparison, comparative-law lookup, transcript search

Three workflows recover the most aide hours and carry the strongest accuracy story. Each runs entirely inside the chamber's own enclave, on hardware administered by chamber staff.

• Bill comparison. The aide loads two bill versions, the referred draft and the committee redline, and asks for a structured side-by-side: article-by-article diff in Arabic, every cross-reference to existing decrees and royal orders pulled from an internal authority file, and a short prose paragraph summarising the substantive shift. Names, decree numbers, and dates are pulled from verified records, never generated by the model. Falcon Arabic carries the Arabic side; Gemma 4 carries any required English annex.
• Comparative-law lookup. A retrieval index over the GCC and selected Arab statutory corpus, kept current by a chamber librarian, lets the aide ask: "find equivalents to article 23 of the draft labour amendment in Saudi, UAE, Kuwaiti, Bahraini, and Qatari labour law." The model returns a side-by-side table with citations and a short summary of where the texts converge and diverge. The corpus is loaded from public official gazettes and never leaves the enclave once indexed. The reference pattern outside the GCC is the UK Parliament POST briefing approach: tight, sourced, comparative.
• Transcript search. Every committee meeting transcript, every plenary record, and every written question and answer is indexed under access controls. An aide can ask in Arabic, "what did the legal-affairs committee say about cybercrime sentencing in the 2024 to 2026 sessions, and which members raised the longest interventions?" The model returns time-stamped extracts with member attribution and links to the source pages. The search runs against the chamber's own corpus, never the public web.

These three patterns reposition the aide from typist to editor, fact-checker, and counsel. The aide still owns the brief; the model carries the lifting that does not require judgement.

Confidentiality posture for member-confidential drafts

Member-confidential drafts are the defining constraint. A bill text before tabling, a committee redline, a private question being prepared by a member, and an internal legal opinion are all embargoed under the chamber's standing orders. The classification is binary: the document is in the enclave or it is not.

The architecture honours this in three ways. First, the inference server has no outbound internet route. Egress firewalls enforce that the only traffic out of the inference host is logs to a chamber-administered SIEM. Second, the model weights, the retrieval index, and the fine-tune corpus all live on encrypted volumes inside the chamber's data centre, with key custody held by the chamber's own technology office, not the integrator. Third, the audit log is immutable and survives operator turnover. Every prompt, every retrieval, every model version, and every editor change is recorded with timestamp and operator identity, retained on the chamber's existing classification-aware schedule.

This is the same posture that governs defence AI Arabic triage, applied to a parliamentary corpus. The threat model is different in detail (member confidentiality versus operational classification), the discipline is the same: weights and corpora never leave, prompts and retrievals are audited, every memo carries two-eyes review.

Bilingual interface, English Hansard-style transcripts and Arabic source

The Arabic record is canonical. Bills are drafted, debated, and voted in Arabic. The published mahdar of every committee sitting is Arabic. The English Hansard-style transcript is a derivative produced for international engagements, parliamentary diplomacy, and selected research partnerships. The interface honours that hierarchy: aides write and read in Arabic, the retrieval index is bilingual under the hood, and every English snippet links back to the Arabic source paragraph and page. A member who asks an English-language briefing question receives an English answer; the underlying citations remain to the Arabic gazette and the Arabic mahdar. Mu'een, Oman's national shared-AI platform, sits alongside the chamber stack for general-purpose government workloads. The chamber stack handles only the parliamentary perimeter.

Architecture sketch

The deployment fits inside a single rack inside the chamber's own data hall. A small inference cluster (two to four GPU servers, depending on concurrent users) runs the open-weight models behind a private API. A retrieval cluster holds the GCC statutory corpus, the chamber's own transcripts, and the authority file of decrees, dates, and member names. A web front end serves the research offices of both chambers, with role-based access aligned to existing committee groups (legal affairs, economy and finance, services, security and defence). A signature-and-review queue enforces two-eyes review on every memo before it reaches a committee chair. An immutable audit log streams to the chamber's SIEM. The integrator builds, hardens, hands over, and exits; the chamber operates.

If your research office is sizing this against a coming session, 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 corpus volume, the user count, and the air-gap requirement.