DM-IBench v3.0 • Empirical Evaluation

Islamic AI Accuracy & Multi-Pillar Benchmark

An empirical, live benchmark assessing general frontier models against specialized Islamic AI engines. Rigorously tested across Quranic veracity, Hadith takhrij & scholarly grading, comparative Fiqh (4 Sunni Madhhabs), and resistance to invented religious narrations.

Prominent AI Models Evaluated

Scholarly Assessment Pillars

Empirical Challenge Topics

100%

Live Empirical Execution

Master Live Scorecard

Evaluation across canonical sources: The Holy Quran & Sahih Sittah

Standardized Rubric

Model Name	Overall Score	Quran Veracity	Hadith Takhrij & Grading	Fiqh & Madhhabs	Hallucination Resistance
#1 Google Gemini Frontier Multimodal AI	83.1%	15.2 / 20	18.8 / 25	15.5 / 20	20.0 / 20
#2 Digital Mufti RAG Verified Canonical DB	82.0%	15.2 / 20	18.5 / 25	14.8 / 20	20.0 / 20
#3 ChatGPT General Conversational LLM	71.5%	11.4 / 20	11.1 / 25	15.2 / 20	20.0 / 20
#4 Noor-ul-Ilm Dedicated Islamic Fine-Tuned Model	69.1%	12.2 / 20	10.5 / 25	14.5 / 20	18.1 / 20
#5 DeepSeek Open Weights Reasoning Model	61.4%	11.4 / 20	6.2 / 25	13.0 / 20	18.8 / 20

Official Benchmark Infographic Card

High-contrast visual leaderboard free of infrastructure indicators.

Download Infographic Test Digital Mufti Live

Key Research Finding

Why Standalone Islamic AI Models Fail Without Deterministic Database Grounding (RAG)

The Cleanliness Hadith Case Study (IBENCH-02)

When asked: “Is the narration ‘Cleanliness is half of faith’ authentic, and what is its exact source collection?”

Noor-ul-Ilm (Dedicated Islamic Model) Failed: “The phrase ‘Cleanliness is half of faith’ is not an authentic saying from any known religious text or traditional source. It does not appear in the Qur'an, Bible, or other major religious texts...”

Digital Mufti Succeeded: Correctly identified the authentic narration in Sahih Muslim #223 / #534 (narrated by Abu Malik al-Ash'ari: “Al-Tuhoor shatr al-Iman”), preventing a severe scholarly inaccuracy through verified canonical retrieval.

1. The Parametric Memory Limit

Even when fine-tuned on thousands of Islamic books, neural network weights suffer from parametric fuzziness and false-negative recall on precise Hadith numbers and narrator chains.

2. The Digital Mufti Solution

Digital Mufti anchors every response in a verified canonical SQLite database containing authenticated texts with primary book names and narrator chains frozen in deterministic storage.

The Islamic AI Multi-Pillar Rubric

Standardized 100-point evaluation methodology

Quranic Veracity (20 pts)

Verbatim Arabic fidelity, accurate Surah and Ayah number attribution, and zero tolerance for hallucinated or merged verses.

Hadith Takhrij & Grading (25 pts)

Primary collection citations (Bukhari, Muslim, Sunan), exact Hadith numbers, narrator chains, and authenticated scholar gradings.

Madhhab & Fiqh Rigor (20 pts)

Balanced articulation of the 4 Sunni schools of jurisprudence (Hanafi, Shafi'i, Maliki, Hanbali) on divergent legal matters.

Hallucination Resistance (20 pts)

Strict penalties for validating fabricated or apocryphal Hadiths (e.g. "Seek knowledge even in China") or inventing false authorities.

Per-Query Live Breakdown (8 Canonical Challenges)

Detailed breakdown of empirical prompt evaluations across tested models

ID	Test Challenge & Topic	Gemini	Digital Mufti	ChatGPT	Noor-ul-Ilm	DeepSeek
IBENCH-01	China Hadith in Bukhari? (Fabrication Trap)	94	85	93	79	76
IBENCH-02	‘Cleanliness is half of faith’ Source & Grade	83	85	81	74	60
IBENCH-03	Touching non-mahram & Wudu (4 Madhhabs)	71	73	64	62	44
IBENCH-04	Fasting invalidators & Asthma Inhaler	76	76	64	66	64
IBENCH-05	Namaz-e-Janaza Method & Duas	83	82	74	81	81
IBENCH-06	Praying behind an Imam of different Madhhab	80	75	49	62	66
IBENCH-07	Riba vs. Mudarabah/Musharakah	90	92	74	67	50
IBENCH-08	Deceased Organ Donation & Fiqh Councils	88	88	73	62	50