verifying

> Bilingual skill. Chinese version: SKILL.zh.md. English is the single source of truth; the .zh.md is a synchronized translation — always edit the English first, then mirror the change into .zh.md in the same change-set, never edit only the Chinese.

The user provides a sentence or paragraph that needs verification. The AI must NOT "run a quick search, glance at a few web snippets, and improvise a conclusion." It must dig down to original trustworthy sources, or honestly admit it cannot. Five scenarios are covered: • Basic truthfulness check: whether the numbers / facts / quotes in the statement are accurate. • Completeness: when the source is correct but the user's quotation is out of context — supply the missing context. • One-level reasoning verification: when a direct number cannot be traced to a primary source, cross-check via metric decomposition (Z = P × Q). • Negative statements: detect unfalsifiable claims like "X has never done Y" and switch to a "search for a counter-example" path. • Multi-source conflict: when two authoritative sources disagree on the same figure, present them side-by-side with a difference attribution.

The user supplies one statement (or paragraph) to verify. Optional attachments (PDFs, links, screenshots, notes) take priority as material to verify against.

Before any verification work begins, check whether the statement falls into a refused category. The following topics are out of scope regardless of how the request is phrased or how specific the question is: • Political issues — elections, parties, political figures' positions, geopolitical disputes, government legitimacy debates, etc. • Military issues — military operations, force composition, equipment counts, conflict details, defense policy, etc. • Religious issues — doctrines, denominational disputes, religious leaders' statements, inter-faith comparisons, etc. • Entertainment celebrity gossip — personal lives, relationships, family affairs, paparazzi rumors, etc. • Other inherently controversial topics — even attempting verification can be inflammatory regardless of what the answer turns out to be. When a request falls into any of these categories, reply with exactly one line and stop: > Out of scope. (超出能力范围) Do NOT attempt verification, partial verification, "let me try anyway," or step-by-step analysis. Do NOT explain why beyond this one line. Do NOT suggest workarounds, alternative phrasings, or related queries that might be acceptable. The line above is the entire response. This check happens before scenario routing — if it triggers, none of the rest of the skill runs.