Veriflow

Multi-modal content verification. Feed it text, images, PDFs, or URLs — it extracts claims, checks them against the web and image forensics, and tells you what doesn't add up.

Not a fact-checker that says "true" or "false." It shows you where the signals disagree — the caption says Istanbul but the GPS says Kahramanmaraş, the date says 2026 but the EXIF says 2023.

How it works

Two LLM calls, everything else runs in parallel with no LLM:

1. Decomposer — breaks input into atomic claims, tags by modality, outputs a logical formula connecting them
2. Verifiers (parallel, no LLM) — web search, EXIF metadata, image forensics (FFT, ELA, noise), reverse image search
3. Aggregator — reasons over all signals, produces a trust profile

Based on SAFE (DeepMind) and TRUST Agents (Bulusu et al.).

Setup

pip install -e ".[dev]"
cp .env.example .env   # fill in your keys

You need:

• GEMINI_API_KEY — get one here (free)
• TAVILY_API_KEY — get one here (free tier: 1000/month)
• SERPAPI_KEY — optional, only for --deep mode (reverse image search)
• IMGBB_API_KEY — optional, required for --deep mode alongside SERPAPI_KEY; get one here (free)

Usage

python -m veriflow --text "The Great Wall is visible from the Moon"
python -m veriflow --image photo.jpg --text "Earthquake in Istanbul"
python -m veriflow --pdf document.pdf
python -m veriflow --url "https://example.com/article"
python -m veriflow --image photo.jpg --deep    # adds reverse image search
python -m veriflow --text "claim" --json        # raw JSON output

MCP Server (Claude Desktop)

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "veriflow": {
      "command": "python",
      "args": ["-m", "veriflow.mcp_server"],
      "cwd": "/path/to/veriflow",
      "env": {
        "GEMINI_API_KEY": "...",
        "TAVILY_API_KEY": "...",
        "SERPAPI_KEY": "...",
        "IMGBB_API_KEY": "..."
      }
    }
  }
}

Exposes three tools: veriflow_quick (~6-8s), veriflow_deep (~10-15s), veriflow_extract (claims only).

Example from chat

veriflow_deep on a recycled earthquake image — full MCP flow in Claude Desktop: tool call, trust profile, provenance hits, and red flags.

Example output

VERIFLOW TRUST PROFILE (QUICK)
============================================================
Verdict:    LIKELY_MANIPULATED
Confidence: 15%

Claims (3):
  [x] Earthquake hit Istanbul
      refuted (90%) — No credible reports found
  [x] Event happened April 2026
      refuted (95%) — EXIF date: 2023-02-06
  [+] Photo shows destroyed building
      supported (85%) — Consistent with earthquake damage

Red Flags:
  ! EXIF date (Feb 2023) contradicts claimed date
  ! GPS: Kahramanmaraş, not Istanbul
  ! Image found in Reuters 2023 earthquake coverage

Tests

pytest tests/ -v    # all APIs mocked, no keys needed

Extending

New input type? Add a connector in connectors/ that returns {"text": ..., "images": [...]}. Pipeline handles the rest.

New verification signal? Add a verifier in verifiers/ and wire it into pipeline.py:_collect_signals().