2026-04-14 — v1.0
- Launch with Google Gemini 1.5 Flash as primary AI Coach model, DeepSeek as fallback.
- Guardrails deployed: no fatwa generation, no Quran/hadith fabrication, scholar-escalation path active.
- Red-team corpus: 420 adversarial prompts tested monthly; pass rate 98.1%.
- Recitation audio pipeline: on-device first, 30-day cloud retention, no training use.
Test history
We publish monthly adversarial test summaries including:
- Prompts tested.
- Pass rate.
- Any failures and patches.
- Scholar Panel observations.
Next report: 2026-05-15.
Report an unsafe response
Tap thumbs-down in the app or email ai-safety@hafizprime.com.
