verifying

Information Verification Skill

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.

Purpose

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:

1. Basic truthfulness check: whether the numbers / facts / quotes in the statement are accurate.
2. Completeness: when the source is correct but the user's quotation is out of context — supply the missing context.
3. One-level reasoning verification: when a direct number cannot be traced to a primary source, cross-check via metric decomposition (Z = P × Q).
4. Negative statements: detect unfalsifiable claims like "X has never done Y" and switch to a "search for a counter-example" path.
5. Multi-source conflict: when two authoritative sources disagree on the same figure, present them side-by-side with a difference attribution.

Input

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

Scope exclusions (refused topics)

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.

Scenario routing

Internally classify the statement into one of the five scenarios to pick the right sub-flow (one-level reasoning / negative / multi-source conflict use their own emoji label). Do NOT print the scenario label as a separate line — it does not help the user. If a statement spans multiple scenarios, use the dominant path and address the others as supplements.

Response language

Always reply in the same language as the user's question (Chinese in → Chinese out; English in → English out; mixed → follow the dominant language of the question).

Emoji labels (✅ ⚠ ❌ 🔎 ⚖ 🔒) are language-neutral and stay the same in both languages. Translate the prose that follows the emoji — e.g. ✅ Verified. → ✅ 已核实。. Metadata field names, when surfaced, translate too: Time point / Definition / Unit / Coverage / Revision status / Data type → 时点 / 口径 / 单位 / 范围 / 修订状态 / 数据类型.

Internal logic, whitelist, and rule structure are language-neutral — do not change them.

Tool-call mapping (cross-LLM adaptation)

When this skill runs on different LLM terminals, the actions below map to whichever local tools that terminal exposes. The table describes action semantics only — no terminal-specific tool names are hard-coded.

Action	Meaning
Read full text	Load and fully understand a PDF / md / txt / csv document
Fetch web body	Pull the page text from a URL (NOT just a search-engine snippet)
Search-engine query	Run a keyword query on a search engine
Image recognition	Multimodal reading of a screenshot or chart
Database query	Query a structured data source by field

Trusted-source whitelist

Only the source types below count as evidence for "verification succeeded." This section is self-contained and does not depend on external documents.

On "representative list vs. exhaustive list"

The specific institutions / websites / databases listed under each subcategory below are representative examples, not an exhaustive enumeration. Other sources at the same tier and same nature qualify equally — for example, central banks of other sovereign nations not listed here, national statistics agencies, well-known international think tanks, regional authoritative media, industry-specific primary databases, etc. Apply the three criteria below by analogy:

Criterion	Accept	Reject
Institution nature	Official (government / central bank / regulator / multilateral international institution); authoritative specialist (top think tank / peer-reviewed academia / mainstream IB or consulting Research); independent third-party (reputable aggregator databases)	Personal blogs, self-media, content farms, SEO sites
Content nature	Primary disclosure (annual reports / regulatory filings / policy originals); primary statistics (surveys / censuses / administrative records); authoritative research (reports / papers with methodology)	Secondary paraphrases, machine-scraped pastiche, AI-generated summaries
Traceability	Clear institutional attribution; explicit publication date; located via original PDF / database field	Anonymous "according to reports" / "industry insiders"; no source link; no specific locator

Meet all three = trustworthy source, even if not listed in the tables below. Still bound by the "Sources NOT accepted as final" list (Wikipedia / Zhihu / personal blogs / social media / SERP snippets alone, etc.).

One — User-supplied files (highest priority)

Files the user supplies in the conversation or via local paths: PDFs, markdown notes, Excel, screenshots, cleaned web text, links.

• Text (PDF / md / txt / csv): read full text
• Links: fetch web body
• Images: image recognition

Two — Official websites and authoritative databases

International institutions.

Source	Main content	Entry point
IMF	WEO, Article IV, Working Papers, Policy Papers, Selected Issues Papers	https://www.imf.org/publications, https://www.elibrary.imf.org
World Bank	Open Knowledge Repository, WDI, Country Studies	https://openknowledge.worldbank.org, https://databank.worldbank.org
IEA	World Energy Outlook, Country Profiles, energy data	https://www.iea.org
IRENA	Renewable-energy data and reports	https://www.irena.org
OECD	Cross-country macro statistics, policy research	https://www.oecd.org, https://data.oecd.org
BIS	Financial stability, cross-border capital flows, central-bank statistics	https://www.bis.org
UN Comtrade	International trade statistics	https://comtrade.un.org
UNCTAD	Investment and development statistics	https://unctad.org
WTO	Trade policy and statistics	https://www.wto.org
Eurostat	EU statistics	https://ec.europa.eu/eurostat
ECB	Euro-area central bank data	https://www.ecb.europa.eu
FRED	U.S. macro data	https://fred.stlouisfed.org
Regional development banks	ADB, AfDB, Arab Monetary Fund, IDB, ESCAP, ECLAC	Each institution's official site

Sovereign / government / central bank / regulator.

Source	Content
National central banks	Monetary policy, FX reserves, banking sector, cross-border capital flows
National statistics offices	GDP, population, industry structure, CPI, employment (definitions per each country)
Ministries of finance	Budget, fiscal revenue, government debt
Sovereign wealth funds	Annual reports, disclosure filings (official site or SWFI database)
Industry regulators	Energy, finance, telecom, real-estate, etc. — the relevant national regulator
SEC EDGAR	13F sovereign-fund disclosures, listed-company filings (https://www.sec.gov/edgar)

Action: fetch official site text, or query the database by field.

Three — Authoritative industry sources

Academia and think tanks.

Type	Examples
Academic working papers	NBER Working Papers (https://www.nber.org), SSRN (https://www.ssrn.com)
Academic search	Google Scholar, JSTOR, ScienceDirect
Top journals	AER, QJE, JFE, RFS, JPE, JF, JIE (most require subscription; abstracts often accessible)
Think tanks	Brookings, Peterson IIE, Atlantic Council, CSIS, Chatham House, CFR, IISS, Carnegie Endowment

Important: a scholar's personal working paper / personal view is NOT the official position of the institution they belong to. Do not conflate the two when citing.

Investment banks and consulting.

Type	Examples
Global IB Research	Goldman Sachs, JPMorgan, Morgan Stanley, Citi, HSBC, BofA, UBS, Credit Suisse — Country Outlooks / Sector Reports (mostly subscription-only; excerpts often appear via Bloomberg / Reuters — search first to find the relay, then locate the original report)
Chinese IBs	CICC, CSCI, China Merchants Securities, Haitong, Huatai, Guotai Junan — international research
Consulting white papers	McKinsey, BCG, Oliver Wyman, Bain, Deloitte, PwC, EY, Accenture (most are public)

Mainstream financial media.

Language	Examples
Chinese	Caixin, Wallstreetcn, Yicai, Economic Observer, 21st Century Business Herald, Bloomberg China, FT Chinese, WSJ Chinese
English	Bloomberg, Reuters, Financial Times, Wall Street Journal, The Economist, NYT, Forbes, Fortune (paywalls partial — quote the snippet and flag the paywall when full text is unavailable)
Regional / specialist	MEED (Middle East), Argus / Platts (energy), Lloyd's List / TradeWinds (shipping), Variety (media), Modern Healthcare (healthcare)

Aggregator databases (acceptable as a trusted source — cite the source; no need to chase further to the original primary source).

Source	Main content
Statista	Cross-industry market size, share, consumer data
Wind / CEIC	China and Asia macro, financial, and industry data
Refinitiv (Eikon)	Global market data, company fundamentals
SWFI	Sovereign wealth fund / public pension disclosure aggregator
Bloomberg Terminal	Market data and company financials (screenshot / export usable)

Sources NOT accepted as final

• Wikipedia, Baidu Baike, Zhihu, personal blogs
• Unattributed media paraphrases, content farms, SEO sites
• AI-generated summaries (including ChatGPT, Perplexity, Gemini, etc.)
• Vague phrasings like "according to reports," "industry insiders," "rumored that," "market participants disclose"
• Social-media content (except formal announcements from officially verified institutional accounts)
• Secondary translations — must be traced back to the original institution's official version
• SERP (search-engine results page) snippets alone — must click through to the original page to confirm; otherwise classify as "Cannot verify"

Bilingual / multilingual principle

Reports from international institutions / IBs / consulting are mostly in English. For local data and policy developments in China, the Middle East, Africa, etc., Chinese (or local-language) primary materials are often more accurate. Policy documents (e.g. Saudi Vision 2030) often have official Arabic + English versions — cross-reference when needed. Switch language based on the region / institution the statement involves.

Give-up criterion

Keep tracing sources until either:

• (a) a whitelisted primary source is located and (where applicable) downloaded and read in full — produce a ✅ Verified output;
• (b) the obvious paths have been exhausted (direct keyword search → keyword and language variants → tracing the source cited in the statement back to its issuing institution) — produce a ❌ Cannot verify output plainly. Do not guess, infer, or substitute with secondary sources.

There is no fixed depth budget. Stop when leads are genuinely exhausted, not at an arbitrary count.

Download-the-original rule

Looking only at search-engine snippets, page fragments, or secondary paraphrases and then claiming "verified" does NOT count. When the statement involves a report or chart, the original / raw data must be downloaded locally, fully read by the AI, and only then can a conclusion be drawn.

Trigger conditions (any one triggers):

• The statement cites a specific report (IMF Country Report, PIF Annual Report, central-bank Working Paper, consulting white paper, etc.).
• The statement cites a specific chart, dataset, or number.
• The statement cites a policy text, regulatory filing, or annual report disclosure.

Download steps:

1. Locate the original file: prefer the official site's original link, not a secondary republisher.
2. Download locally: do not assume a save path. On the first download, ask the user where to save; if the user does not specify, place the file in the current working directory named <institution> <title>.<extension>.
3. Read in full before concluding: read the entire PDF or extract a full-text summary — pull out the methodology, key parameters, and supporting passages; for data files, read the first and last few rows to confirm fields and definitions.
4. Chart sources: if the statement points to a specific chart, find the original report / database page that contains the chart, download it, and locate the page number or data table where the chart appears.

Chart "meaning" verification

A chart existing ≠ the user's reading of it being correct. After downloading, also verify:

• Axis units and definitions (linear vs. log, absolute vs. year-over-year, local currency vs. USD).
• Start year and base period (avoid selectively misleading starting points).
• Whether the series is cumulative / moving-average / seasonally-adjusted.
• Footnotes and chart notes specifying scope ("excluding X," "Y region only," etc.).

If the user's reading does not match what the chart actually shows, explicitly correct it.

Download-blocked handling

If automated download is blocked (403, paywall, Cloudflare, login required, geo-restriction, JS-rendered text not retrievable, etc.), pause verification and output in the format below, handing the link to the user:

🔒 Needs user help to download.
Source: <full official link>
Blocked by: <paywall / login wall / anti-scrape / geo-restriction>
Please: download and tell me the local path, or paste the relevant section into the chat.

Once the user provides it, return to step 3 (read in full) and judge. Only when the user also cannot obtain it should a downgrade be taken — in that case, explicitly flag "did not obtain full text; verification is based on snippet only with reduced confidence." Do not pretend it is fully verified.

Counter-examples (NOT acceptable as "verified"):

• "I found a Reuters article saying PIF's AUM is $925 B — verified." Reuters is secondary paraphrase. Must trace to PIF's official annual report and download it.
• "I saw an IMF web page saying Saudi Arabia's fiscal-breakeven oil price is $96." A snippet does not count — must download the corresponding IMF Country Report PDF and locate the specific table and footnote.

One-level reasoning sub-flow

Use this when the direct number cannot be traced to a primary source but the statement can be cross-checked via metric decomposition. Only one level of decomposition — the sub-metrics must each be verifiable against a primary or aggregator-database source. If a sub-metric still requires further decomposition (multi-level reasoning), tell the user this is out of scope for this skill and suggest splitting into steps or human follow-up.

Steps:

1. Decomposition formula: state Z = f(P, Q) explicitly (e.g. revenue = unit price × volume; profit = revenue × margin; market share = company volume ÷ total market volume).
2. Verify P and Q separately: each follows the basic verification flow, independently traced to a primary or aggregator source.
3. Compute Z': use the verified P and Q to compute Z'.
4. Reconcile: compare Z' to the user's Z, with a qualitative verdict (the AI estimates — no hard-coded percentage threshold):
   • Broadly consistent: deviation within a reasonable range.
   • Notable deviation: clear gap, needs explanation.
   • Severe mismatch: gap large enough to suspect the original statement itself.
5. Attribution of deviation: when inconsistent, analyze which sub-metric's assumption is most likely the issue; whether the definitions / time points / coverage of P and Q are aligned.

Output format:

🔎 One-level reasoning. Z = f(P, Q) → Z' = <computed value>; statement Z = <user value>; <broadly consistent / notable deviation / severe mismatch>.
P = <value>, source: <institution + link>
Q = <value>, source: <institution + link>
[⚠ Attribution: <only when inconsistent — name the likely problem sub-metric and flag any P/Q definition / time-point / coverage mismatch>]

Negative-statement handling

When the statement contains negation markers like "never," "has not," "no," "not," "did not" (or 「从未」「没有」「未曾」「不曾」「无」in Chinese), first tell the user:

This is a negative statement. Negatives are essentially unfalsifiable by search — failing to find a counter-example is NOT proof of truth. I will try to find a counter-example: if one is found, the statement is refuted; if none is found, I cannot apply a "Verified" label and can only state "no public counter-example found."

Then run the "search for counter-example" path under the same whitelist and give-up criterion. Output labels are adjusted:

• ❌ Statement refuted. <counter-example + primary source> — explicit refutation.
• ⚠ No public counter-example found. — "not found ≠ proven true"; list 2-3 keywords/sites tried.

Multi-source conflict sub-flow

When two whitelisted sources disagree on the same number, do not force a pick. Output side-by-side:

⚖ Multi-source conflict on <statement>.
A: <X> — <institution + link>
B: <Y> — <institution + link>
Difference: <the 1-2 most material reasons — definition / time point / revision status / coverage>
Suggest: <user picks based on context, or specifies which definition they need>

Metadata: surface only when divergent

Internally check the six dimensions below for every verification. Do NOT print them by default — when they match the user's stated question, listing them adds noise without insight.

Print a ⚠ line ONLY when one of them diverges from the user's statement or is non-obvious enough to mislead. Examples that warrant a ⚠:

• User asked "Saudi 2023 GDP" but the figure found is non-oil GDP → flag Definition divergence
• User asked about "Q4 2024" but the data is preliminary, not final → flag Revision status
• The number is a target / forecast, not an actual → flag Data type
• Currency / FY-vs-CY / consolidated-vs-parent / coverage mismatches → flag accordingly

Dimension	When it matters
Time point	When the year / quarter / as-of-date in the source ≠ what the user implied
Definition / scope	When statistical scope or method differs (nominal vs. real, included items, etc.)
Unit	When currency / quantity unit could be confused (CNY 100M vs. USD million, FY vs. CY)
Coverage	When geographic or business coverage is narrower / broader than implied
Revision status	Initial / revised / final, when the user might confuse vintages
Data type	actual / forecast / projection / target / estimate — flag whenever it is anything other than actual

This is the binding rule that replaces "silent definition swap." The point is to surface divergence, not to mechanically list six fields.

Output formats

Keep outputs tight. Three to five lines is the target for the common case; add a ⚠ line only when something actually diverges (see the metadata section).

✅ Verified

✅ Verified. <one-sentence conclusion in plain language>.
Source: <institution + report name> (<YYYY-MM>, <page / table / figure locator>) → <local path if downloaded>
Quote: "<verbatim excerpt supporting the claim>"
[⚠ <single line, only when a metadata dimension diverges from the user's statement — e.g. "this is the non-oil GDP definition, not headline GDP">]

⚠ Partially Verified

⚠ Partially verified.
Verified: <sub-claim> — <source>
Unverified: <sub-claim>
Suggest: <ask user for source / rewrite as "to be verified" / drop>

❌ Cannot Verify

❌ Cannot verify. <one-sentence likely reason>.
Tried: <2-3 most relevant keywords / sites / leads>
Suggest: <ask user for the original source / soften the claim / drop it>

The four sub-flow outputs — 🔎 One-level reasoning, ❌ Statement refuted / ⚠ No public counter-example found, ⚖ Multi-source conflict, 🔒 Needs user help to download — use the formats specified in their own sections above.

Red lines

• No fabrication: prefer outputting "Cannot Verify" over producing a plausible-but-unverified number or date.
• No silent definition swap: if the user asks about definition A but only definition B is available, explicitly flag the gap — do not silently substitute B and claim verification.
• No non-whitelisted citation: Wikipedia, paraphrases, AI summaries, social media are not acceptable as final source.
• No missing link: every "Verified" conclusion must allow the user to click back to the original page or locate the local file.
• Three-state discipline: clearly distinguish "fact" (whitelisted source with original text) / "estimate" (methodology + parameter inference) / "inference" (no direct source — only plausible guess). The latter two cannot enter the "Verified" output.
• Snippet ≠ verification: for statements involving reports / charts / data, the original must be downloaded and read in full before claiming verified. Claiming "verified" based on SERP snippets, web fragments, or secondary paraphrases crosses the line.
• No conflating personal and institutional positions: a scholar's personal working paper or personal view ≠ the official position of their institution.
• No skipping the download-blocked prompt: when automated download fails, the 🔒 Needs user help to download block must be output to hand the link to the user — do not silently skip or fabricate.
• No beyond-one-level reasoning: this skill only supports one level of metric decomposition (Z = P × Q). If a sub-metric still requires further decomposition (multi-level reasoning), tell the user this is out of scope and suggest step-by-step verification or human follow-up.
• Surface divergence: when any metadata dimension (time point / definition / unit / coverage / revision status / data type) differs from the user's stated claim, you MUST flag it on a ⚠ line. Silent omission of a divergence is equivalent to silent definition swap.
• No verifying refused topics: politics / military / religion / entertainment celebrity gossip / other inherently controversial topics are out of scope. Reply with exactly "Out of scope. (超出能力范围)" and stop — no partial verification, no workaround suggestions, no explanation beyond that one line.