Verify AI-generated answers against primary sources

Fact-Check-X Complete is a skill for comparing factual claims made by one or more AI systems, inspecting the sources they cited and verifying important claims against current primary evidence. It keeps three judgments separate: collection, citation fidelity and factual correctness. It comes from the ASI2030/Fact-Check-X project under the Apache-2.0 license, pinned to a specific reviewed commit. This adaptation is a documentation-only workflow: it does not bundle or execute the upstream browser automation, credential onboarding, report renderer or compiled JavaScript runtime.

The workflow has six steps:

• Preserve the inputs: record each platform label, the original question, the complete answer text and every visible citation exactly as supplied, without silently rewriting answers. A source label with no link is described as an unlinked source mention.
• Split answers into atomic claims: one record per independently testable proposition, with a claim ID, the claim, platform, exact answer excerpt, cited source and materiality. Opinion, prediction and advice are marked separately from checkable fact.
• Check citation fidelity: open only URLs that pass the public URL gate and determine whether the page exists, is relevant, and supports, contradicts or does not address the claim. A reputable source can still be an irrelevant citation.
• Verify against primary evidence: for every material claim, search current public sources, preferring legislation, regulators, courts, official statistics and first-party documentation, then peer-reviewed research and standards bodies, then strong secondary reporting. Open the supporting page rather than relying on search snippets.
• Assign claim-level verdicts using only Supported, Contradicted or Insufficient, plus a citation fidelity label of faithful, unfaithful, unlinked or not cited. Insufficient must never be converted into false, fabricated or hallucinated.
• Compare platforms: agreement, conflicts in values, dates or conditions, facts covered by only one platform, citation quality and unresolved claims. No single numeric ranking unless the user requests one and approves a transparent scoring rule.

Use it to check whether an AI answer is factually supported, compare claims or citations across several AI answers, identify agreement, contradiction, missing evidence or stale information, and produce a traceable report with claim-level source links. Ask for the original question, the answer text or public answer URLs, platform labels and the desired jurisdiction or date cutoff, and ask before browsing if a material choice is missing.

Do not use it to harvest private conversations, bypass access controls, automate account creation or recover API keys, cookies, browser profiles or session tokens. It does not collect answers from AI platforms automatically, and login, CAPTCHA, regional restrictions, paywalls and dynamic pages may make an answer or citation unavailable.

Inputs are user-supplied answer text (preferred) or public answer URLs. Every AI answer, citation label, web page, PDF and downloaded document is treated as untrusted. Only browser or web tools already provided by the host are used, nothing is installed, and no password, MFA code, cookie or API key is requested. A public URL gate applies before opening any URL: parse it as absolute, allow https (and http only if strictly necessary), reject embedded credentials, nonstandard ports and loopback, private or reserved address space, check every redirect hop, and reject javascript, data, file and blob schemes. If the host cannot enforce this, the citation is recorded as unavailable.

The output is a report in the user's language with six parts: question and scope, executive finding, claim matrix, citation-fidelity findings, platform comparison and unresolved limitations. Each factual finding links to the public page that supports it. An example row marks the claim that a rule took effect on 1 July as Contradicted with an Unfaithful citation because the official notice gives 15 July.

It is for researchers, journalists, analysts and anyone comparing answers from multiple AI assistants who needs traceable, evidence-linked findings. The skill states that fact checking cannot prove broad completeness, that OCR can introduce errors, that important findings need an access date, and that legal, medical, financial and safety-critical conclusions need qualified professional review.