How Mizan works

Three lanes: code, model and people. The clause moves between them, so you can see the model never works alone: code checks whatever it returns, and people come in only at the end. Tap any station for its details; below, the same path step by step on a real clause.

Fully automated, and the committee decides

The review waits for no one: from upload to review file the engine runs alone, the same steps every time. People come in only at the end: the committee decides on a finished file instead of reading the contract from scratch.

Mizan never issues a fatwa or an approval. It shortens the road to the decision.

Why trust the result?

• It cites only what it retrieved

  A citation to a unit outside the candidate list is dropped in code, so the model cannot bring a decision from memory.

• The quote is verbatim, or it is dropped

  Code matches the quote against the decision text after normalising diacritics, tatweel and hamza, then shows the reader the decision's own text, not the model's copy.

• No verdict without a rule

  «Non-compliant» needs a unit with an explicit prohibition or condition, «compliant» needs one or an explicit permission, and the scope must cover the clause. Otherwise the clause is escalated, or Mizan abstains when it finds no basis.

• Amended decisions are caught

  If a later decision amended the cited one, the clause is flagged and escalated instead of judged.

• And it is measured

  On 70 synthetic clauses: 93% of non-compliant caught at 96% precision, against 68% and 48% for keyword search. On 4 whole real contracts (21 clauses): 71% agreement with the key. Caveats on the main page.

The technology, briefly

Where are the decisions?

1,217 decisions, split into 2,719 units, in files on the server (13 MB). They load into memory at start-up, so no database is needed.

How does it find the decision?

Keyword search (BM25) over two copies of the decisions: the text as is, and the text with contract-style sentences we added to each decision. Before searching, the model adds the clause's fiqh terms. About 17 candidate decisions come out.

Is it RAG?

Yes: we retrieve decisions, then the model judges from them alone. After it comes a code gate that checks every quote is verbatim, or drops it and sends the clause to review.

Did you train a model?

No. The knowledge lives in the decisions and the code rules, so the model changes with one setting.

We compared search methods: of 90 clauses, how many lost their governing decision?

• Keyword search (in use now)6
• Semantic search alone Qwen3 Embedding9
• Semantic search alone BGE-M320
• Keyword + semantic hybrid4
• Keyword + reranking BGE reranker8

Lower is better. 90 clauses from the test set, 17 candidates per clause for every method. The models ran on our own machine, so no decision text left it.

The result: semantic search alone is weaker, and reranking did worse than today's search (8 against 6); we think that is because the judging model already reads every candidate and picks for itself. Keyword plus semantic is better by two clauses out of 90, so it goes into the bank version with a local model.

Isn't a general chat model enough?

A fair question, so we tested it. We gave the same ten clauses to general models (Claude Sonnet 5.5, Opus 5.5 and GPT-6 Astra) through their APIs, in 33 runs.

So the difference is not that a general model «doesn't know». The difference is the guarantee: a code gate that rejects any non-verbatim quote, a verdict conditioned on an explicit rule, the same steps every time, a record per clause, and numbers measured against a baseline. Sharia review cannot settle for «usually».

A small experiment: 10 clauses, 33 runs, on 1 October 2026, through the APIs rather than the chat apps.

Before launch: a local model inside the bank

Before Mizan reaches any bank, we move it to an open-source model running on the bank's own servers. The contract text then stays inside the bank's network and is not sent to any outside party.

1. Today

   A working demo

   Runs on a cloud model, to prove the idea and measure it on the test set.

2. Before launch

   A local model inside the bank

   An open-source model chosen by the bank, on its own servers. The contract text stays inside its network.

3. Before go-live

   Measured first

   We run the same test set on the chosen model and show its numbers before any real use.

Why is this easy? Because the model is a single setting in Mizan, not part of its code. The gate and abstention work the same with any model.

Examples of open-source models that can run inside the bank

• Qwen/Qwen3.8-27BApache-2.0
• zai-org/GLM-5.3-FlashMIT
• humain-ai/ALLaM-7B-Instruct-preview

And we tried it: Mizan ran on a laptop with Qwen3-14B and no internet. On 4 clauses it matched the expected verdict on 3, and every quote was verbatim. That is why the chosen model is measured before go-live.