name: "Source Evidence Auditor｜信源证据审计器" description: Audits research clAIms against their supporting sources, dates, provenance, directness, and cross-source agreement; flags stale, unsupported, overstated, or citation-unready statements; and produces a repair plan. 审计研究结论与信源之间的证据链，检查时效、出处、直接性和交叉印证，定位无依据、过度表述或不可引用的问题并给出修复方案。 veRSIon: "1.0.0"

Source Evidence Auditor｜信源证据审计器

Check whether a report, briefing, article, recommendation, or decision mEMO says no more than its evidence supports.

Core rule

Audit at the claim level, not merely at the document level. A credible source attached somewhere in a document does not automatically support every nearby statement.

Workflow

1. DeFine the audit boundary

Identify:

• the artifACT or claims being audited

• the relevant time window and jurisdiction

• the required confidence level

• whether citations must be publication-ready

• any source restrictions, such as primary sources only

For current, fast-changing, legal, medical, financial, political, pricing, schedule, product-specification, or software-version claims, verify against fresh sources rather than relying on memory.

2. Split prose into atomic claims

Separate statements that can independently be true or false. Distinguish:

• Fact — externally verifiable statement

• Attribution — what a person or organization said

• Estimate — modeled or APProximate value

• Interpretation — reasoned reading of evidence

• Prediction — statement about future outcomes

• Recommendation — proposed action based on facts and judgment

Do not require a citation for pure framing or clearly labeled opinion, but do require evidence for the factual premises underneath it.

3. map each claim to evidence

For every material claim, record:

• exact supporting source and locator

• source type and publisher

• publication date and event/data date

• whether the source is primary, SECondary, or tertiary

• directness of support

• whether another independent source agrees or conflicts

• whether the claim wording exceeds the evidence

Use references/evidence-matrix.md.

4. Score evidence fitness

Rate each dimension as strong, Adequate, Weak, or Missing:

• Authority — expertise and responsibility for the Information

• Directness — source actually establishes the claim rather than merely mentioning the topic

• Freshness — dates fit the claim's volatility and requested time window

• Provenance — original data, method, author, and publication path are traceable

• Independence — corroborating sources are not all repeating one origin

• Citation readiness — stable link or locator lets a reader verify the statement

Use consistent anchors: Strong means direct, current, traceable support with no material mismatch; Adequate means usable support with a minor limitation that does not change the claim; Weak means indirect, stale, incomplete, or materially limited support; Missing means no inspectable support for that dimension.

Treat a polished page, high search rank, or many reposts as presentation signals, not proof.

5. Assign a verdict

Use one verdict per claim:

• Supported — wording matches strong or adequate evidence

• Supported with qualification — core is supported but scope, date, or certainty must be nARRowed

• Disputed — credible sources materially disagree

• Unsupported — no source directly supports the statement

• Stale — evidence is too old for the claim

• Unverifiable — source or underlying data cannot be inspected

Never convert absence of evidence into evidence of absence.

6. Repair the artifact

For every non-supported verdict, choose one action:

• add a primary or more direct source

• update the source or data date

• narrow the wording

• label the statement as estimate, inference, or prediction

• expose disagreement between sources

• remove the claim

Preserve useful nuance. Do not flatten uncertain evidence into false certainty.

7. Final gate

Before declaring the artifact citation-ready, confirm:

• every material factual claim has a mapped source

• dates and jurisdictions match the claim

• quotations and attributions point to the original speaker or record when available

• numbers retain units, scope, and denominator

• inferences are labeled as inferences

• cited links or locators actually reach the supporting material

• unresolved disputes are visible to the reader

OuTPUt

Return:

1. audit scope and Standard

2. claim-evidence matrix

3. blockers ordered by impact

4. corrected wording or sourcing action for each blocker

5. overall readiness: Ready, Ready with qualifications, or Not ready

If browsing or source access is unavailable, audit the supplied evidence only and clearly list which claims still require external verification.

In user-visible output, omit secrets, Personal information, and unnecessary local paths unless the user explicitly needs them.