Why we build,
and how we choose
what enters.

Every day, millions of Muslims open general-purpose AI assistants and ask questions about their faith. Whether the fast is broken by an injection, whether prayers may be joined while traveling, whether an account is permissible. They ask because asking is easy and the need is real, and the model always answers, whether or not it was trained on sound sources.

So the serious question was never whether AI should answer religious questions. That was decided by billions of users, silently, years ago. The only question left is what kind of system answers them: one built with care, verification and named people standing behind it, or the statistical average of the internet.

The digital age did not create a shortage of Islamic information. It created an excess of it: fake fatwas, fabricated narrations attributed to scholars who never said them, and fluent AI answers with invented citations. The classical tradition had an infrastructure for exactly this problem, the sanad, the chain of transmission that let you trace any statement back through named, accountable people to its origin. The internet broke that chain. Domla is our attempt to rebuild it at the scale the internet operates at.

Scholars rightly note that a mufti must be a mukallaf: a legally accountable person, sane, adult, responsible before Allah. A machine is none of these things. It has no intention, bears no sin, earns no reward. We agree with every word of that, and we have never claimed otherwise.

But the conclusion drawn from it mistakes what the machine is. Nobody has ever demanded that a book be a mukallaf, yet books carry authority, because a chain of accountable humans stands behind them: the author, the transmitters, the teachers who certified the editions. The accountability never resided in the paper. It resided in the people, and the paper extended their reach.

An aligned AI model is the same object at a different point on the technology curve. We choose which books enter its corpus. We write its instructions. We decide that it must cite its sources and refuse when the evidence is absent. Accountability for what it says rests with us, the named people who curated it, exactly as it would if we had published the same content as a thousand-page book. A book preserves a scholar's words. An aligned model preserves a scholar's method, and makes it interactive.

There is a boundary we hold to. The textual work of scholarship, what the sources say and how the schools ruled, is a retrieval task a machine can perform with a completeness no human memory can match. The pastoral work, reading the fear behind a question, counseling a family, carrying spiritual authority from soul to soul, is irreducibly human. Domla does not perform marriages, does not mediate disputes, and does not claim the station of a living teacher. For everything beyond the text, the right answer is a human scholar, and the system says so.

A raw language model answers from a blurred statistical memory of the internet, which is why raw models misattribute quotes and cite books that do not exist. Domla is built the other way around. Before the model writes a single word, a search layer retrieves the relevant passages from the verified corpus, and the model is constrained to ground its answer in what was actually retrieved. A citation in Domla is not the model's recollection of a book. It is a pointer to a record that exists.

This makes fabrication a structural impossibility rather than a statistical rarity: the system cannot cite a document the retrieval layer did not return. And where the evidence is absent, the system is instructed to say so rather than improvise. In our internal benchmark of scholar-authored questions, run against the identical frontier model with and without the Domla pipeline, the difference was not incremental.

For nearly three centuries after Gutenberg, no Muslim-operated press printed in Arabic script anywhere in the Ottoman Empire. The fears were not foolish: printed texts could carry errors a trained scribe would never make. But the fears were answered with delay instead of engineering, and the price was three hundred years of knowledge locked in manuscripts copied one at a time. What was needed was never a ban. It was a printing house with scholars doing the proofreading.

When the loudspeaker reached the Muslim world, leading jurists initially ruled prayer through it invalid, because it was genuinely unclear whether the device transmitted the imam's live voice or replayed a recording. When engineers explained how the machine worked, the ruling was revised by the same school that issued it. Today the adhan is amplified from every minaret on Earth. The fatwa changed when the scholars understood the machine, and that is the model we follow: our answer to every objection is to explain, precisely, what the machine is, and to engineer against the failure the objection points at.

The tradition does not demand infallibility from those who take on the work of guidance. It demands rigorous, sincere effort, and it prices in the possibility of honest error, because the alternative, that no one attempts the work at all, is worse. We treat the curation of this system as exactly that kind of effort, made under scholarly supervision, with the criteria stated openly on this page.

The system's outputs have been rigorously reviewed by scholars across Central Asia beyond our own team. When scholars show us an error, we correct it, and we correct it publicly. If you disagree with something Domla has said, we consider that a contribution to the work, and we want to hear it.