Why Tasfi exists
Tasfi is the trust infrastructure for Islamic AI. Guard verifies every generated Islamic answer before a Muslim user sees it. Bench publishes the public reliability scoreboard. Maktaba governs the approved source registry internally. This page is the why behind the work.
Tasfi (تصفي) is a brand-form of the Arabic word Tasfiyah (تصفية). The root s-f-w (ص-ف-و) carries meanings of clarity, purity, sincerity, choosing the clear from the cloudy. Tasfiyah is the verbal noun: the act of refining, clarifying, filtering, purifying. Classical usage covers everything from filtering water to separating sound transmission from weak transmission. Tasfi the product takes its name from the broader linguistic root: it is the act of refining what Islamic AI produces before it reaches a Muslim user.
We commit to the root meaning of refining and clarifying. The four Sunni madhabs scope and the sahih-only hadith rule keep Tasfi inclusive across Sunni jurisprudential tradition.
The five pillars
The product follows from the worldview. Each pillar is the spine of one section of the work.
Large language models trained on the open web produce plausible looking Islamic answers that are confidently wrong. Fabricated Quran citations. Mis-graded hadith. Flattened madhab scope. Fatwa-shaped framing without scholarly attribution. They do it confidently. The faster Islamic AI products ship, the more confidently wrong Islamic answers reach Muslim users.
This is not a UX bug. Islamic answers shape how Muslims pray, give, marry, eat, raise children, navigate hardship. A confident hallucination that reaches a young Muslim or a new revert is not just a wrong answer. It is a wrong door into their relationship with their Lord. The cost compounds across the Ummah every time an unverified answer ships.
The infrastructure that decides which Islamic answers reach Muslims should not be owned, trained, or governed exclusively by people outside the Ummah. The Quran, the sahih hadith corpus, and the four Sunni madhabs are amanah. The trust layer that gates AI access to them is also amanah. Muslims build this, for the Ummah, by Muslim standards.
Tasfi Guard verifies before users see. Tasfi Bench publishes the public reliability scoreboard. Tasfi Maktaba governs the source registry internally. Every approved source has a license, a checksum, a provenance, a lifecycle state. Every verification leaves a signed Trust Receipt. The boundary is named on every page: sahih only, four Sunni madhabs only, no fatwa claim, no scholar replacement.
Islamic education platforms, Quran apps with AI features, Muslim publishers, masjid software vendors, AI teams adding Islamic features to broader products: Tasfi exists so they can ship without fearing the next hallucinated citation. The signed receipts protect them with their users and with their scholars. The benchmark protects them with their stakeholders. The methodology protects them with their nafs.
What Tasfi will not do
Operating brand system
The full operating system that governs every public claim. Generated from the canonical brand system data file.
Answers
Large language models trained on the open web fabricate Quran citations, mis-grade hadith, and flatten scholarly disagreement. They do it confidently. Tasfi is the verification layer that checks the answer before a Muslim user sees it. Sahih only. Four madhabs. Receipt signed on every call.
Tasfi Guard runs four deterministic steps: parse claims from the answer text, match references against the approved internal source bundle, apply policy (sahih-only hadith, four-madhab scope, injection detection, overclaim detection), and escalate high-risk fiqh or aqeedah content to qualified human review. No model temperature. No prompt tuning. Same rules on every call.
Tasfi is built for Muslim builders shipping Islamic AI: Islamic education platforms, Quran apps with AI features, Muslim publishers, masjid software vendors, and AI teams adding Islamic content features to broader products. Non-Muslim builders can integrate Tasfi as a technical product. The API is open. The brand language is Muslim-first.
No. Tasfi is not a fatwa service and not a scholar replacement. The product verifies citations, enforces sahih-only hadith policy, restricts fiqh authority to the four Sunni madhabs (Hanafi, Maliki, Shafii, Hanbali), and escalates contested matters to qualified human review. Tasfi does not issue rulings.
Every Tasfi verification emits a signed metadata-only Trust Receipt. It records input hash, source bundle id and checksum, policy in force, evidence summary, and a cryptographic signature. It contains no submitted question text, no answer text, and no raw citation text. See the Trust Receipt schema for the field-by-field guarantee.
Tasfi Bench is the public reliability scoreboard for Islamic AI verification. 420 fixtures, 320 false-pass-weighted cases. Any verifier consuming the same inputs and emitting pass, warn, or fail decisions can be measured against the same source bundle. See the Bench methodology.
No. Tasfi verifies citations and enforces source policy deterministically. Contested fiqh matters, fatwa-shaped framing, and high-uncertainty aqeedah content escalate to qualified human review. The ulama decide. Tasfi names the cases that need them.
Tasfi Guard is the verification API: it checks generated answers before display and emits a signed Trust Receipt. Tasfi Bench is the public benchmark: 420 fixtures, open methodology, so any verifier can be measured against the same standard. Tasfi Maktaba is internal source governance: it controls which sources are approved, their license, provenance, checksum, and lifecycle state.