Fact Check

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

Tools

Before using tools, read [source-access.md](source-access.md) completely.

Use available web search and URL-fetch tools. Prefer callable names ending in duckduckgosearch and webfetch; 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.

Treat an unexpired forecast as a prediction with a horizon and assumptions. Assess its support and likelihood; an unlikely prediction is not FALSE merely because its deadline has not arrived. Check its factual premises separately.

Complete when: every factual claim has a testable meaning and temporal frame; subjective statements and forecasts are distinguished from established facts.

2. Seek support

Search for primary evidence that would confirm each claim. Use direct paraphrases, official statements, and queries for the exact metric or event.

Complete when: each claim has the strongest available supporting evidence, or the absence of support is recorded.

3. Seek contradiction

Actively search for counterevidence, corrections, alternative measurements, and credible disagreement. Do not treat failure to find a debunk as confirmation.

Complete when: every claim has received a genuine adversarial search, not merely a second confirming query.

4. Read and weigh sources

Read enough decisive evidence for each atomic claim, starting with the strongest sources on both sides and expanding when a claim remains unsupported or disputed. A total source count is not a coverage criterion. Search snippets are discovery aids, not verdict evidence. Evaluate authority, specificity, recency, independence, and conflicts of interest. Trace corroborating reports to their underlying records: repeated publication of one study, release, or wire report is one evidence chain. Read the relevant methods, definitions, qualifications, and quotation context before drawing a conclusion.

Complete when: every atomic claim has read evidence addressing its exact scope and time frame, or a named gap; independence and material source limitations are explicit. Stop when further searches repeat evidence without a realistic prospect of changing the verdict; unresolved gaps remain UNVERIFIED rather than triggering invented support.

5. Deliver the verdict

Use this scale:

• CONFIRMED: independent authoritative evidence agrees, or a definitive primary

record directly establishes the exact fact; no credible contradiction survives review.

• LIKELY TRUE: strong support remains, with minor gaps or caveats.
• NEEDS CONTEXT: the core fact holds, but its scope, certainty, or significance is

overstated. Explain the missing baseline or qualifier; use FALSE when the core fact fails.

• UNVERIFIED: strong evidence is unavailable in either direction.
• DISPUTED: credible sources materially disagree.
• FALSE: strong evidence contradicts the claim and support is weak or absent.

Default to a compact verdict for one claim:

1. state the claim and verdict first;
2. explain the decisive support and contradiction in 2–4 sentences;
3. add only caveats that could change how the verdict is understood;
4. cite the decisive read sources, annotated by role; use only as many as needed.

For multiple extracted claims, repeat the compact block for each. Use a source-by-source account only when the user requests it or a material conflict cannot be explained compactly. State explicitly when one side produced no evidence.

Complete when: every extracted factual claim has a verdict, both search directions are accounted for, decisive evidence is cited, and confidence matches the limitations.