About this prototype

An AI copilot for LaunchGood’s Zakat-eligibility reviewers. It reads campaign drafts, proposes a classification against the 8 Qur’anic categories cited in LaunchGood’s published Zakat policy, grounds every claim in specific policy clauses, and routes the case to a human reviewer for final approval.

The human/AI boundary

Three hard rules the AI follows

1. Citation requirement. Every recommendation cites at least one verbatim clause from LaunchGood’s published Zakat policy. An automated verbatim-substring check rejects citations the model fabricated; failures force the memo to ESCALATE.
2. Contested categories. Any memo recommending fī sabīlillāh or mu’allafati qulūbuhum auto-escalates regardless of model confidence — these categories have documented madhāhib-level disagreement. The model may suggest, but a senior reviewer decides.
3. Authority boundary. The AI does not interpret novel theological positions. It applies LaunchGood’s published policy as written, and escalates anything outside that scope (e.g., sanctions or compliance concerns).

Trust-building moves in the UI

• Every AI claim has a clickable citation to the source clause.
• Confidence is shown as a band (high/medium/low), never as a percentage — to avoid false precision.
• The AI explicitly says “I do not have enough information” when ambiguous.
• The reviewer must select the final category themselves — the form never auto-fills it. This is a deliberate friction to fight automation complacency.

Architecture (5-step pipeline)

1. Extract structured claims from the draft (gpt-4o-mini)
2. Retrieve top-K relevant policy clauses via pgvector + category-tag bias
3. Classify + cite with structured output (gpt-5)
4. Apply deterministic guardrails (verbatim citation check, contested-category escalation)
5. Persist memo + retrieved chunk IDs + cost/latency telemetry

Source policy

Chunked from launchgood.com/zakatpolicy (updated 24 February 2025). 30 stable-ID clauses are versioned in the repo.