Verification API
Check the answer before the user sees it. Source registry lookup, sahih-only hadith policy, four-madhab attribution metadata, signed Trust Receipt on every call.
tasfi.app | Trust infrastructure for the Muslim ummah
Ship Islamic AI with a paper trail.
Verify every generated Islamic answer before it reaches a Muslim user. Sahih only. Four madhabs with attribution metadata. Signed receipt on every call.
- Hadith
- Sahih only
- Fiqh
- 4 madhabs
- Public input
- Canned only
Why this exists
Why not just use ChatGPT or Claude for Islamic answers?
Open-web models produce confident Islamic answers that are wrong. Fabricated verses. Non-sahih hadith presented as sahih. Single-madhab rulings framed as consensus. The faster Islamic AI ships, the faster these errors reach the ummah. Tasfi is the trust layer between a model and a Muslim user.
Hallucinated citation
A confident reference to Quran 99:99 with a fabricated quote. The verse does not exist; the quote is invented.
Mis-graded hadith
A weak or unknown narration presented as sahih because the model has no grading discipline.
Flattened madhab
A single school's position rendered as "all scholars agree" when the four madhabs have documented disagreement.
Finance pack
Authenticated Guard extension for Islamic finance answers. Methodology clarity, evidence freshness, certification boundary, advice drift, overclaim risk. Not a stock screener, broker, or certifier.
Source governance
Approved, eval-only, quarantine, blocked. Every approved record carries license, provenance, checksum, and lifecycle. Maktaba governs Guard internally instead of becoming a public source product.
Public evidence
420 fixtures, 320 false-pass cases, signed scorecard. The methodology is open. Any verifier can be measured against the same suite.
The infrastructure that decides which Islamic answers reach Muslims is amanah. We do not outsource amanah.
Canned trust demo
Three audits. No public religious free-for-all.
The public demo never accepts custom religious text. It demonstrates the safety posture with controlled examples only.
Builder wedge
Put Tasfi Guard behind your Islamic AI product.
If your platform generates Islamic content for Muslims, Tasfi verifies it before display, attaches a metadata-only Trust Receipt, and produces a privacy-safe reliability report you can share with your team, your investors, and your scholars. Pilots are open to Islamic edtech, Quran apps, Muslim publishers, masjid software vendors, and AI teams shipping Islamic features.
Pilot request
curl -s https://tasfi.app/v1/verify \
-H 'authorization: Bearer <pilot-key>' \
-H 'content-type: application/json' \
-d '{"question":"What does Islam teach about hardship?","answer":"Allah reminds believers that with hardship is ease (Quran 94:5).","citations":[{"type":"quran","reference":"94:5","quote":"with hardship will be ease"}],"domain":"education"}'Answers
What pilots, scholars, and search engines actually ask.
Why not just use ChatGPT or Claude for Islamic 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.
Is Tasfi a fatwa service?
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.
What is a Trust Receipt?
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.
What is Tasfi Bench?
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.
Where does the name Tasfi come from?
Tasfi is a brand-form of the Arabic word tasfiyah (تصفية), meaning to refine, clarify, purify, or filter. The product does exactly that. It refines what Islamic AI produces before it reaches a Muslim user. The full etymology is on the Why Tasfi page.