360 Secure SkillHub
360 安全技能中心

Wiki Verify (Dexio)

该 Skill 按链接数、检查时间和风险挑选 wiki 页面,抽取数字、日期、名称、引语等可核查声明,逐条对照原始文件或 URL,标记 supported、outdated、unsupported、unsourced、unreachable,修正页面及复制该声明的页面,并记录 verified 日期与核查报告。它明确说明这是抽样核查而非全量审计,建议用第二模型复核争议声明。

dexio
热度 7.65
最近更新 2026-09-29 05:09:37
分类:知识依赖:代码执行 / 联网浏览搜索 / 读写文件 / 内部调用大模型应用场景:研究 / 笔记整理 / 报告撰写 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add dexio-wiki-verify
Metadata
name
dexio-wiki-verify
description
Fact-check an LLM wiki against its own sources: pick the pages where an error would do the most harm, pull out the checkable claims, open each cited source and mark every claim supported, outdated, unsupported, unsourced or unreachable, then correct the page and every page that copied the claim. Use on a schedule, before a decision relies on a page, after a large ingest, or when a page's claims are in doubt.
license
MIT
metadata
{"author":"dexio","version":"1.1.0"}

Verify a wiki

The known weakness of an LLM wiki is that a wrong claim gets written in once and then cited: it stops looking like model output and becomes a line other pages build on. Linting finds broken structure. Verifying finds wrong content, by going back to the sources.

Verification is sampling, not a full audit. A run checks a handful of pages well rather than the whole wiki badly.

Pick what to check

Folder wiki: python3 <wiki-lint skill>/scripts/wiki_lint.py <wiki> --verify-queue 10 ranks pages by how many pages link to them, how long since anyone checked them, and whether they cite nothing, are contested, have low confidence, or record a decision. Hosted wiki: rank the same way from the page list and each page's links in.

Also check, whatever the ranking says:

  • the page a decision is about to rely on
  • pages written in a large ingest or by a new agent
  • any page someone has questioned

Default run: three to five pages, or about twenty claims.

Steps

  1. Pull out checkable claims. Numbers, dates, names, quotes, prices, "X does Y" statements, and anything stated as a finding. Skip opinions that are marked as opinions.
  2. Open each claim's source. The raw file, the URL, the page it was derived from. Read the part that should support it. If the page gives one source list for everything, find which source carries which claim.
  3. Mark each claim with one of:
    • supported: the source says it, in substance.
    • outdated: the source, or a newer primary source, now says something else.
    • unsupported: the source does not say it, or says something weaker.
    • unsourced: nothing is cited.
    • unreachable: the link is dead or paywalled. Try an archived copy (web.archive.org) before giving up.
  4. Fix the page.
    • Outdated: update the claim with the new source and date (wiki-conflicts, superseded).
    • Unsupported because the source says something different: the page was wrong. Correct it to what the source says, cite it, and keep a dated line of the old value ("corrected 2026-09-28, was $15") (wiki-conflicts, corrected).
    • Unsupported because the source says less: weaken the claim to what the source does say. If you cannot settle it, label it (unverified) and lower confidence, adding the field if the page has none (medium for one weak source, low for none).
    • Unsourced: find a source and cite it, or label the claim (unverified).
    • Unreachable: note it inline ("source offline as of 2026-09-28") so the next check does not repeat the search.
    • Supported, but only as a company's claim about itself: add (vendor-sourced) if the label is missing.
  5. Find the copies. Search the wiki for every corrected claim (the number, the name, the key phrase). Fix each page that repeated it. This is the step that stops the error compounding.
  6. Record the check. Set verified: YYYY-MM-DD on each page whose checkable claims you all checked that day. It means "checked, and the labels on this page show the result", so a page can carry it with claims still labelled (unverified). Do not set it on a page you only sampled or only fixed a copy on. Leave a change note listing what you corrected: the commit message in git, the note field on a hosted wiki. A plain folder has no history, so there the dated correction lines on the page are the record.
  7. Report: pages checked, claims checked, how many in each category, corrections made, anything that needs a person (a disputed decision, a claim you could not settle), and whether a second checker was used.

Use a second pair of eyes

A model checking its own synthesis tends to agree with it. Where you can, give each claim and its source, without the rest of the page, to a separate agent or a different model and ask only "does this source support this claim?". Disagreements between the two are the claims to look at closely. If you cannot, the check still counts, but say in the report that every classification is one model's judgment.

Pitfalls

  • Checking that a source exists instead of reading what it says.
  • Accepting a secondary source's paraphrase of a primary one you could open directly.
  • Treating a vendor's claim about itself as verified because the vendor's page says it. It is supported as a vendor claim; keep the vendor-sourced label.
  • Fixing the page you checked and leaving the claim on the pages that copied it.
  • Marking a page verified after spot-checking two claims.