AI for Legal Research in Oman: From Hatim to On-Premise LLMs

A senior partner at a top-tier Omani law firm walks into the office on a Sunday morning with a cross-border financing matter, a 480-page Arabic concession agreement, and a Tuesday deadline. The associate team will spend the next 36 hours doing keyword searches against Qanoon.om, hand-pulling Royal Decrees from the Official Gazette, and translating clause language between Arabic and English. The work is honest, rigorous, and almost entirely time spent on retrieval. The intellectual work, the part the client is actually paying for, starts only after retrieval is done. This is the gap an on-premise legal AI is meant to close.

This pillar walks through how legal research in Oman has worked from the Hatim-era keyword tools through Qanoon.om and Decree.om, why public ChatGPT-class assistants are unsafe for Omani matters, what a credible on-premise RAG architecture looks like over the Sultani Decree corpus, where AI helps and where it adds risk, and what an eight to twelve week deployment looks like for a firm with ten to fifty lawyers.

The reality of legal research in Oman today

Oman is a civil-law jurisdiction. Legislation flows in two tiers, primary legislation through Royal Decrees issued by His Majesty the Sultan, and secondary legislation through Ministerial Decisions issued by ministries acting under delegated authority. Both are published weekly in the Official Gazette. The hierarchy places ratified international treaties above the Basic Statute, and the Basic Statute above all decrees and decisions, as documented in the Globalex guide to Omani law hosted by NYU.

Two consequences shape day-to-day legal research. First, Oman does not run a binding-precedent system. Lower courts are not formally bound by higher-court rulings. Supreme Court reasoning matters in practice, especially in commercial disputes, but the canonical research artefact is the decree itself, the amendment trail, and the explanatory ministerial regulations, not a Westlaw-style citator graph. Second, the corpus is Arabic-first. Royal Decrees are issued in Arabic. English translations exist for the major instruments but are unofficial and lag behind amendments.

For a generation of Omani lawyers, the workflow was Hatim-era CD-ROMs and printed gazette indexes followed by manual cross-referencing. The 2010s brought a transformation when Qanoon.om went live in September 2015 with a digitised, searchable Arabic corpus, and its sister site Decree.om built the most comprehensive English translation database of Omani law. Both are free, both are kept current with the Gazette, and both are hosted by the same Omani team.

The Ministry of Justice and Legal Affairs publishes the canonical decree register as the ground-truth source. Qanoon and Decree mirror it, normalise it, and add navigation. What none of these portals do is reasoning. They are excellent retrieval surfaces. They do not draft, summarise, compare, translate, or argue. That layer is exactly what an LLM is meant to add, if it can be done safely.

Why public ChatGPT-class tools fail for Omani law

Three problems disqualify public chatbots for serious Omani legal research.

Hallucination on legal citations. The Stanford RegLab evaluated leading legal AI tools and found that even purpose-built systems with marketing claims of being hallucination-free produced incorrect or misgrounded answers on 17 to 33 percent of queries. The peer-reviewed write-up, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, defines a misgrounded answer as one that cites a real source which does not actually support the claim. For a partner reviewing an associate's memo, that is the worst possible failure mode, a confident citation that collapses on inspection.

No coverage of the Omani corpus. Public foundation models are trained on web-scale text dominated by English-language US, EU, and UK sources. Omani Royal Decrees, Ministerial Decisions, and Court of Investment and Commerce reasoning are present at trace levels, if at all. Asking a public chatbot about Royal Decree 6/2022 typically produces a fluent paragraph that mixes Omani text with concepts borrowed from European data-protection law. The English-language bias is not a bug, it is what the training distribution looks like.

Confidentiality leakage. Pasting a draft concession agreement into a public chatbot transfers it to a foreign operator, often subject to the United States CLOUD Act for data access, and routinely retained for service improvement unless an enterprise contract says otherwise. For matters touching sovereign clients, family offices, or anything covered by Royal Decree 6/2022 on personal data protection, that transfer is a privilege problem before it is a regulatory one.

The combination is a non-starter. A tool that hallucinates Omani citations, has shallow corpus coverage, and exfiltrates client matters to a foreign jurisdiction is not a legal research assistant. It is a malpractice generator with good prose.

The on-prem alternative: RAG plus a fine-tuned Arabic LLM

The credible architecture is retrieval-augmented generation over a controlled corpus, served by an Arabic-capable open-weight model on hardware inside the firm.

RAG works in two stages. A retrieval stage pulls the most relevant passages from a vector index of the firm's documents and the public legal corpus. A generation stage feeds those passages, plus the lawyer's question, into the language model and asks it to answer using only the retrieved evidence, with citations. This is not a research curiosity. The legal-domain RAG benchmark LegalBench-RAG evaluates exactly this pipeline over 6,858 expert-annotated query-answer pairs across a 79-million-character corpus, and a follow-up arXiv survey on reliable retrieval in RAG over large legal datasets documents the failure modes when documents are structurally similar, which is the everyday reality of decree-driven civil-law work.

The model layer matters too. For Omani research, the firm needs Arabic competence at least as strong as English. The current open-weight options that meet this bar are Qwen 3.6 in its 27B and 70B Arabic-tuned variants, Gemma 4 with its 256K context window for whole-decree ingestion, and Falcon Arabic from the UAE Technology Innovation Institute as an Arabic-first alternative. All three can run inside the firm with no outbound traffic. Hosn's stack picks the right model per task, large-context for clause review, mid-size for drafting, small for triage and tagging.

The hardware is approachable. A single NVIDIA RTX 6000 Blackwell with 96 GB of GPU memory, what Hosn ships as the Tower tier, runs a 70B-class model with full RAG against a multi-gigabyte legal corpus and serves twenty to fifty concurrent users. A small firm can sit on a Mac Studio Kernel; a national firm with a litigation department serving hundreds of users moves to a Rack tier with H100 or H200 acceleration.

The Sultani Decree corpus as a research backbone

Building the corpus is the first project, and getting it right pays dividends every day after. The work breaks into acquisition, structuring, and citation discipline.

Acquisition. Mirror the Qanoon.om Arabic corpus and the Decree.om English translations weekly. The Decree blog confirms the canonical workflow, the team processes the Official Gazette weekly, converts PDFs into normalised text, integrates amendments, and publishes the consolidated version, as documented in the platform's tenth-anniversary retrospective. A firm-side mirror takes a snapshot, hashes it, and stores it on the firm's own NAS.

Structuring. Break each instrument into article-level chunks, preserve the decree-number, year, issuing authority, gazette-issue, and amendment metadata, and index Arabic and English versions side by side. Maintain a links-graph from amending decrees to amended decrees, because a partner reading Article 23 of Royal Decree 6/2022 must see at a glance which subsequent decrees touched it.

Citation discipline. Every answer the model produces must reference specific articles and decree numbers, with a one-click jump to the Arabic original and the English translation in the firm's mirror. No citation, no answer. This is the rule that converts a glib chatbot into a serious associate. The Stanford study found that misgrounded citations were the dominant hallucination mode, so the system must surface the supporting passage alongside the conclusion, not just a bare reference.

On top of the public corpus, the firm layers its own work product, prior memoranda, deal precedents, redlined contracts, and internal know-how memos. This is where genuine differentiation lives. Two firms with the same access to Qanoon.om do not have the same understanding of how a clause negotiates against an Omani counterparty in a real transaction. The internal know-how layer is what the firm's RAG sees, and what no public chatbot ever will.

Drafting vs research vs review

AI helps unevenly across legal work. Mapping the gradient honestly avoids the failure mode of "AI for everything" that produces partner mistrust within a quarter.

Research, high value. First-pass corpus search, finding all decrees touching a topic, surfacing translations, building chronologies, comparing how related provisions evolved across amendments. With RAG and a clean index, this collapses an associate-day to an associate-hour. The output is an annotated reading list with citations, not a final memo.

Drafting, mixed value. Useful for first drafts of standard clauses, NDAs, board resolutions, and translations between Arabic and English. Risky for negotiated commercial language and any provision the firm has a house position on. The right pattern is template-driven generation against the firm's clause library, not free-form authorship.

Review, lower value, high caution. Spotting missing definitions, inconsistent defined-term usage, and amendment-trail gaps is real value. Substantive review of a regulator's filing or a brief that goes to court remains a partner activity. Treat the AI's red flags as questions, not conclusions.

Hallucination control and human-in-loop

Hallucination is the only word that matters when a partner evaluates legal AI. The defence is engineered, not promised.

• Citation requirement. The system refuses to answer without retrieved evidence. Every claim links to a specific article in a specific decree.
• Confidence calibration. The model returns an explicit "insufficient evidence" output when retrieval finds no supporting passage above a threshold, instead of confabulating one.
• Bilingual cross-check. For matters where the Arabic original and English translation diverge, the system flags the divergence and shows both rather than picking a winner.
• Lawyer sign-off. Every output that leaves the firm carries a named reviewer. The AI's role is captured in the file note, not hidden.
• Citation audit. A nightly job samples the system's outputs and verifies that each cited article actually exists and supports the claim. A regression rate above a chosen threshold pages the firm's knowledge-management lead.

This is the discipline that the Stanford RegLab found missing in commercial tools that marketed themselves as hallucination-free. Build it in, measure it, and report it to partners monthly.

Confidentiality, privilege, and conflict checks

Privilege is the table stakes. An on-premise system makes it operationally simple.

Matter files, drafts, prompts, and inference logs sit on hardware inside the firm. There is no telemetry feed to a model vendor, no shared inference pool with other tenants, and no remote management plane that the vendor controls, the same five-layer pattern Hosn details for sovereign deployments generally. The model weights are open-weight files the firm owns, scanned and pinned, with explicit upgrade paths rather than silent updates.

Conflict checks become more powerful, not less, with AI. The firm's client and matter database lives on the same internal infrastructure as the legal AI, and a conflict query runs as a retrieval against that database. The conflict question never crosses the perimeter. Compare this to a public chatbot, where typing "draft a memo on the X family office acquisition of Y" at minimum reveals the matter to the operator, and at worst feeds it into model improvement.

The legal posture maps cleanly to Royal Decree 6/2022 and the executive regulations governing personal-data processing. Sensitive personal data, including health, financial, and litigation files, stays inside the firm. Cross-border transfer questions do not arise, because there is no transfer.

Procurement and deployment for a 10 to 50 lawyer firm

A practical eight to twelve week plan for a firm in this size band.

Weeks 1 to 2, scoping. Map the firm's practice areas, language mix, and confidentiality classes. Inventory the corpus the firm wants indexed, the public Sultani Decree corpus, the firm's prior work product, the firm's clause library, and any sector-specific datasets like banking circulars or capital-markets rules.

Weeks 3 to 4, hardware install. A Tower-tier appliance lands in the server room. Network isolation, firm SSO integration, and HSM key custody are configured. The system is fully operable air-gapped if the firm's IT policy requires it. Mu'een, the national shared AI platform, is referenced as a public-corpus complement for non-confidential work where appropriate.

Weeks 5 to 6, corpus indexing. Mirror Qanoon.om and Decree.om. Ingest the firm's document management system. Build the article-level chunking, the bilingual alignment, and the amendment graph. Spot-check retrieval against ten partner-defined queries.

Weeks 7 to 8, lawyer evaluation. Three to five partners run live matters through the system in parallel with their normal workflow. The system is tuned against their feedback. Citation-audit thresholds are set.

Weeks 9 to 12, fine-tuning and rollout. The firm's clause library and house style are encoded as adapters. Training for associates, knowledge-management oversight is staffed, and the system is rolled out across practice groups. Hosn's procurement language, "by quotation," reflects the per-firm work involved, not a standard SKU.

The result is a research surface that compresses associate hours, a drafting surface that defends house style, and a confidentiality posture the firm can defend to a sovereign client. The firm's competitive moat shifts from who has the most associates to who has the best-organised institutional memory, which is exactly where the legal profession's value has always lived.

If you lead a firm in Muscat, Salalah, or Sohar and you are evaluating where AI fits without compromising privilege, email [email protected] for a one-hour briefing. We will walk through the architecture, the corpus pipeline, and a deployment plan sized to your bench.