Guidance, not fatwā: how the AI shaykh is bounded

A pilgrim standing at the Jamarāt with a question about a missed ritual does not need a language model's opinion. She needs the position of the schools of law, plainly stated, with the reference she can carry to a scholar if her case is unusual.

That distinction — guidance rather than fatwā — is not a disclaimer we appended after the fact. It is the architecture.

A closed corpus, not the open internet

The system retrieves only from a reviewed corpus: the Qurʾān with recognised translations, the six canonical ḥadīth collections with grading, the manāsik manuals of the four Sunnī schools, and the published guidance of the relevant authorities. It cannot answer from anything else, and when retrieval returns nothing sufficient the correct output is an admission of that.

• Every answer renders its sources inline, with collection, number and grading where applicable.
• Where the schools differ, the difference is shown rather than flattened into one position.
• Questions involving personal circumstance — marriage, inheritance, oaths, medical exemption — are routed to a human referral, never answered.
• Answers are generated in the pilgrim's language but the Arabic source text is always displayed alongside.

The review process

Nine scholars across three jurisdictions read the retrieval corpus before launch, and a rotating panel of three reviews flagged answers weekly during season. Flagging is available to every pilgrim on every answer, and a flagged answer is suppressed from the corpus within the hour pending review.

“The technology's job is to bring the pilgrim to the text quickly and honestly. Judgement over a person's specific case remains with a person.”

There is a temptation, commercially, to let a confident model answer everything. We consider that a form of harm, and the refusal to do it is the part of this product we are most careful about.

Last updated 2 July 2026