fact-check

Fact Check

Test a specific claim against evidence on both sides and issue a calibrated verdict.

Tools

Before using tools, read source-access.md completely.

Use available web search and URL-fetch tools. Prefer callable names ending in duckduckgo_search and web_fetch; client-added prefixes vary, so inspect the tools in the session. Without session tools, run the packaged CLI in an isolated environment:

uvx --python '>=3.10,<3.14' web-forager search "your query" --max-results 8 --output-format json
uvx --python '>=3.10,<3.14' web-forager fetch "https://example.com" --format markdown

If uvx cannot run packaged search, use ddgs without touching the current project environment:

uv run --no-project --python '>=3.10,<3.14' --with 'ddgs>=9.5.2' python - <<'PY'
from ddgs import DDGS
results = DDGS().text(query="your query", max_results=8)
for r in results:
    print(r["title"], r["href"], r["body"])
PY

If packaged fetch fails, use Jina Reader:

curl -s "https://r.jina.ai/https://example.com"

The workflow requires both search and fetch. If either capability is unavailable, return UNVERIFIED and name the missing capability.

Workflow

1. Extract the claims

Turn the user's statement into specific, independently verifiable claims. Separate compound claims. If wording is subjective, explain that it can be contextualized but cannot receive a factual verdict. Continue immediately when the claim is clear; ask for confirmation only when two plausible interpretations would change the evidence sought.

Normalize each factual claim's entity, metric and units, population, geography, and relevant date where applicable. Preserve the original meaning and make material assumptions explicit. Establish whether the user wants publication-time evidence, a current assessment, or both; judge an explicitly historical statement at its stated date and label later developments separately. Otherwise use the current date as cutoff.