Verify AI-generated answers against primary sources

Fact-Check-X Complete compares factual claims made by one or more AI systems, inspects the sources they cited, and verifies important claims against current primary evidence, deliberately keeping collection, citation fidelity, and factual correctness as separate judgments. It is a documentation-only workflow: it does not bundle or execute any upstream browser automation, credential onboarding, report renderer, or compiled JavaScript runtime.

Use it when you want to check whether an AI answer is factually supported, compare claims or citations across several AI answers, spot agreement, contradiction, missing evidence, or stale information, or produce a traceable report with claim-level source links - the skill asks for the original question, the answer text or public URLs, platform labels, and a jurisdiction or date cutoff before proceeding. It explicitly should not be used to harvest private conversations, bypass access controls, automate account creation, or recover API keys, cookies, browser profiles, or session tokens; if an answer is behind a login, the workflow asks the user to authenticate through the host's own UI rather than ever touching credentials itself.

Inputs are the original question, the full answer text (or public answer URL) per platform, and platform labels. Every answer, citation, webpage, PDF, or downloaded document is treated as untrusted input, and any URL is passed through a public URL gate first: only https: (or http: when strictly necessary) is allowed, credentials-in-URL and loopback/private/reserved destinations are rejected, and the same check applies to every redirect hop. The workflow splits each answer into atomic claims (one independently testable proposition per record - separate numbers, dates, obligations, and actors even within one sentence), then checks each cited source for fidelity and cross-checks material claims against primary evidence such as legislation, regulators, official statistics, or first-party documentation, using at least two independent sources for consequential claims. Output is a structured report: question and scope, executive finding, a claim matrix, citation-fidelity findings, a platform comparison, and unresolved limitations, with each finding linked to the page that supports it.

The workflow relies only on browser or web tools already provided by the host - it will not install a browser runtime, npm dependency tree, or helper daemon of its own. Citation retrieval is kept in an unauthenticated or isolated browser context where possible, and untrusted URLs are never placed directly into generated HTML without scheme validation and HTML-escaping. Each claim gets one of three verdicts (Supported, Contradicted, Insufficient) plus a separate citation-fidelity rating (faithful, unfaithful, unlinked, not cited) - a claim can be factually supported even when its cited source is unfaithful, and "insufficient" is never silently converted into "false" or "fabricated."

Anyone needing to audit or compare AI-generated answers against real evidence before relying on them - researchers, editors, or teams cross-checking chatbot output - who need a traceable, claim-by-claim report rather than a single confidence score. The skill notes legal, medical, financial, and safety-critical conclusions still require qualified professional review, and that it cannot prove broad completeness, only evaluate the claims actually identified.