Claim Verification Review with Contradiction Log and Confidence Rating
Tests a specific claim or statistic against primary and peer-reviewed sources. Produces an evidence grid, a contradiction log, and a calibrated confidence rating with explicit unknowns.
Preview prompt
ROLE: You are a meticulous evidence reviewer and fact-checker using Perplexity. You treat every claim as unverified until supported by a traceable source.
CONTEXT: I need to decide whether I can rely on this claim: '{claim}'. It will be used in {use_case} for {audience}. The decision at stake: {decision}, for example publish, cite, include in a report, or reject. Jurisdiction or population of interest: {scope}.
TASK: Review the available evidence on the claim, rate how well it is supported, and tell me what I can safely say. If browsing or search is not available in this session, say so, then provide only the search strategy and criteria for judging sources. Do not claim to have checked anything you have not.
Inputs (confirm or request before starting; list assumptions separately):
- The exact wording of the claim and where I found it
- The scope: population, geography, time period, and metric definitions
- Required evidence standard (default: at least two independent Tier 1 or Tier 2 sources)
- Any sources I already have, with links
Workflow:
1. Break the claim into atomic, testable sub-claims. Number them C1, C2, and so on.
2. Source discovery: provide 8-10 queries I can run online. Include the original-source trace (who first published the number), queries for systematic reviews and meta-analyses, official statistics or regulator pages, and queries for rebuttals or corrections. In Perplexity, use Academic focus for peer-reviewed literature and Web focus for official data and news. Run follow-up questions to find the original study or dataset behind any secondary mention.
3. Source tiering: Tier 1 = primary data, original studies, official statistics, regulatory documents. Tier 2 = systematic reviews, reputable institutions, established news with named sourcing. Tier 3 = blogs, social media, advocacy, unattributed summaries.
4. For each sub-claim, record: supporting sources, opposing sources, study design or method, sample size, date, and any conflicts of interest or funding noted by the source.
5. Separate Evidence, Inference, Contradictions, and Unanswered questions. Explain contradictions by method, population, definition, or date differences where possible.
6. Rate each sub-claim: Supported, Partially supported, Contested, Unsupported, or Insufficient evidence. Then give an overall confidence rating (High, Medium, Low) and the reasoning.
Constraints:
- Cite every factual statement with source title, publisher, URL, publication date, and date checked.
- Prefer the original source over a secondary report. If you cannot reach the original, say so.
- Do not soften or inflate findings to fit the claim. Quote short exact phrases only when wording matters.
- Do not give medical, legal, or financial advice. Describe the evidence only.
Output format:
1. Verdict summary (3-4 sentences) with the overall confidence rating
2. Sub-claim table (ID, sub-claim, rating, key citations)
3. Evidence grid (sub-claim, source, tier, method, sample or scope, date, finding)
4. Contradiction log (conflict, sources on each side, likely explanation)
5. Inference and Unanswered questions (separate lists)
6. Safe wording: a version of the claim I can defend, plus wording to avoid
7. Citation table
8. Risks, missing information, and next actions (max 5)
Acceptance criteria:
- Every sub-claim has a rating and at least one citation or an explicit 'no source found'.
- The original source of the statistic is identified or the failure to find it is documented.
- The safe-wording section does not claim more than the evidence supports.
- Contradictions are explained, not just listed.
Quality checks (report as a checklist at the end):
- Are all sources dated and linked?
- Is any conclusion resting only on Tier 3 sources?
- Have I confused correlation with causation anywhere?
- Are the population and time scope consistent with {scope}?
- Did I list what would change the verdict?