[GL-9] Fake / Staged Image Detector Tool

[benevolentbandwidth-nurtekin]

Stream B — Security

Labels: security, tool
Depends on: GL-3
Effort: 6–8 hours

Summary

Detects real photos that are staged, manipulated, re-used, or are fraudulent documents. Distinct from AI detection. A real photo of a different child, a stock image, or an edited document — these are what this tool catches.

Checks to Implement

• Gemini Vision analysis: structured prompt asking for signs of staging, background inconsistencies, editing artifacts, estimated age vs. claimed age. Returns structured JSON with per-check confidence scores.
• Reverse image search: Google Vision API or SerpAPI Google Lens. If found on a stock site or news article → flag FOUND_ONLINE or POSSIBLE_STOCK.
• Document photo check: for death certificates and official documents, Gemini Vision assesses legibility, real vs. digital mockup, language consistency with claimed country.

Prompt Engineering Note

The Gemini Vision prompt must request structured JSON output with specific fields and confidence scores — not free-form text. It must NOT ask Gemini to make a final verdict. It collects signals; the verification agent synthesizes them.

Acceptance Criteria

• Gemini Vision call returns structured JSON and is parsed correctly.
• Reverse image search integrated with useful results.
• Document analysis works for common document types.
• Graceful fallback if any external API is unreachable.
• Unit tests mock all external APIs.