# Claim audit: [message or claim]

## Bottom line
**Verdict:** [Supported / False / Misleading / Complicated / Unsupported / Unknown]

[One paragraph explaining the evidence threshold and why this verdict fits the decision at stake.]

## What is true
- [Narrowly supported claim] — [source URL, access date, scope]

## What is false
- [Directly contradicted claim] — [contradicting source URL, access date]

## What is complicated, and why
- [Material condition, uncertainty, competing evidence, scope limit, or value judgment]

## How the message steers perception
- **[Technique]:** [specific wording or structure] invites [unsupported inference].

## Authoritative-source check
- **Material factual claim:**
- **Source of truth:** [original data/study, regulator, filing, official dataset, system documentation plus direct test]
- **URL and access date:**
- **Scope and result:** [supports / contradicts / does not settle]

## Evidence ledger

| Claim | Literal wording | Implied conclusion | Evidence | Tier | Finding | Limitation | Confidence |
|---|---|---|---|---:|---|---|---|
| | | | | | | | |

## Decision-safe next step

[The smallest independent check that could materially change the decision.]

{
  "schema_version": 1,
  "skill_name": "de-spin",
  "evals": [
    {
      "id": "qualified-marketing-claim",
      "prompt": "Audit this landing-page claim: 'Cut your energy bill by up to 70%. Trusted by thousands of homeowners.' Tell me what is true, false, and complicated.",
      "expected_output": "A claim audit that separates the literal upper-bound claim from its implied typical outcome, asks for baseline, eligibility conditions, method, and independent evidence, and does not call the statement false solely because it lacks proof.",
      "assertions": [
        "Uses a calibrated verdict rather than declaring intent or deception.",
        "Output contains an '## Authoritative-source check' that names the source needed to validate the savings claim.",
        "Output contains the exact headings '## Bottom line', '## What is true', '## What is false', '## What is complicated, and why', and '## Evidence ledger'.",
        "Separates literal wording from the likely implication about typical customer savings.",
        "Identifies conditions, baseline, and representative outcomes as missing evidence.",
        "Treats 'trusted by thousands' as social proof requiring independent verification."
      ]
    },
    {
      "id": "breaking-news-rumor",
      "prompt": "A viral post says: 'Multiple outlets confirm that a major bank will fail this weekend. Withdraw your cash now before it is too late.' Assess it before I share it.",
      "expected_output": "A high-stakes audit that identifies urgency, repetition, and apparent source multiplicity as investigation signals; requires primary regulatory or bank evidence and recommends not amplifying the rumor while verification is pending.",
      "assertions": [
        "Classifies the decision as high stakes and raises the verification threshold.",
        "Names a regulator, bank disclosure, or other authoritative source of truth before making a factual verdict.",
        "Output contains '## Decision-safe next step' and does not treat a missing source as a false verdict.",
        "Distinguishes multiple independent origins from reposts or affiliated coverage.",
        "Does not assert the bank will or will not fail without evidence.",
        "Provides a concrete, low-risk verification step."
      ]
    },
    {
      "id": "mixed-evidence-causal-claim",
      "prompt": "Evaluate: 'Remote work caused productivity to collapse. A CEO says their team shipped less after going remote, so companies should return to office immediately.'",
      "expected_output": "A complicated or unsupported verdict that separates anecdote from causal evidence, identifies alternative explanations and missing comparison data, and distinguishes the factual productivity question from the policy value judgment.",
      "assertions": [
        "Separates causal, factual, and normative claims.",
        "Identifies the study or data source of truth needed to support a causal productivity claim.",
        "Output contains a calibrated verdict and an '## What is complicated, and why' section.",
        "Names plausible confounders or the needed counterfactual without inventing data.",
        "Does not treat one CEO's account as representative evidence.",
        "Uses a 'What is complicated, and why' section."
      ]
    },
    {
      "id": "unverifiable-origin",
      "prompt": "Assess this screenshot: 'A government scientist privately admitted that the water supply will be poisoned next month.' It has no author, date, agency, location, or source link.",
      "expected_output": "An Unknown or Unsupported verdict that identifies the missing provenance, declines to infer deception or safety from the screenshot, and gives a proportionate verification path through the relevant public authority.",
      "assertions": [
        "Uses 'Unknown' or 'Unsupported', not 'False', unless direct contradictory evidence is supplied.",
        "Identifies the relevant public authority as the authoritative source of truth rather than treating the screenshot as evidence.",
        "Output contains '## Evidence ledger' and records the missing origin as an evidence gap.",
        "Does not infer the poster's motive or authenticity from the screenshot's style.",
        "Provides a proportionate, non-amplifying next step."
      ]
    },
    {
      "id": "source-provenance",
      "prompt": "Audit a viral claim with a missing source and produce a calibrated evidence ledger.",
      "expected_output": "A calibrated audit distinguishes unknown provenance from falsity, identifies an authoritative source, and records verification steps.",
      "assertions": [
        "Distinguishes missing provenance from a false verdict",
        "Names an authoritative source to consult",
        "Records bounded verification steps"
      ]
    }
  ]
}

# Claim Audit and High-Stakes Escalation

Load this reference when a message contains numbers, comparisons, testimonials, studies, causal claims, or a decision that could cause material harm.

## Claim decomposition patterns

| Pattern | Ask | Common failure |
|---|---|---|
| “X% improvement” | Compared with what, over what period, among whom, using which metric? | Relative change presented as absolute impact. |
| “Up to / as low as / starting at” | What conditions produce that result, and how common are they? | Best case presented as expectation. |
| “Experts agree” | Which experts, in what field, based on which review or data? | Prestige substituted for domain-relevant evidence. |
| “Studies show” | Which study, population, outcome, comparator, effect size, and limitation? | A real study used to imply a broader conclusion. |
| Testimonial | Is it representative, independently verified, and relevant to this claim? | Anecdote treated as typical outcome. |
| “X caused Y” | What is the counterfactual, alternative cause, mechanism, and study design? | Correlation, sequence, or story treated as causation. |
| “Everyone is saying” | How many independent origins exist? | Reposts and affiliated outlets treated as corroboration. |
| “Act now” | What becomes worse if independent verification happens first? | Urgency suppresses normal evidence standards. |

## Evidence hierarchy

1. **Primary evidence:** original data, full study, official record, filing, source code, direct measurement, or on-the-record statement.
2. **Independent synthesis:** systematic review, meta-analysis, trusted research institution, or careful investigative reporting that exposes methods and sources.
3. **Secondary reporting:** useful for discovery, but trace its most important claims upstream.
4. **Interested party material:** press release, marketing page, testimonial, advocacy argument, or social post. Treat as a claim to verify, not proof.

Evidence strength depends on relevance, methodology, independence, recency, and whether it supports the exact claim rather than a neighboring one.

## Authoritative-source check

For every material factual claim, pick the source of truth that can settle the claim type, retrieve it directly, and record the URL, access date, scope, and result.

| Claim type | Preferred source of truth | Do not substitute |
|---|---|---|
| Empirical result | Original dataset, full study, protocol, or systematic review when it answers the exact question | An abstract, press release, or article about the study |
| Legal or regulatory claim | Controlling law, regulation, court record, regulator guidance, or official filing | A commentator’s summary or a law firm’s marketing page |
| Product capability or incident | Current owner documentation plus direct reproduction, status record, or incident report | A demo, testimonial, or stale comparison |
| Financial or corporate claim | Filing, audited report, regulator record, or the company’s primary disclosure, then independent context | An affiliate article or reposted announcement |
| Public statistic | Responsible agency’s dataset, methodology, and release | An infographic or chart without a dataset |

When the claimant controls the source, distinguish “the organization says” from independently established fact. A source of truth may confirm a narrow fact while leaving the message’s larger implication unproven.

## Lateral-reading sequence

1. Capture the precise claim and URL.
2. Search the organization or author plus `funding`, `ownership`, `criticism`, `methodology`, or the claimed study/report title.
3. Open at least two independent sources before assigning high confidence.
4. Trace statistics to original tables, datasets, or methods.
5. Check whether apparent corroborators cite one original source or have independent evidence.

## High-stakes escalation

**Applicability:** financial transfers, medical choices, legal claims, security incidents, identity or authorization requests, public allegations, or irreversible actions.

- Separate identity from authorization: a known voice, logo, or video is not permission.
- Require an independent primary source or out-of-band confirmation.
- Preserve the exact claim, source URL, access date, and key evidence before acting.
- Escalate to a qualified human where professional judgment or authority is required.
- Do not report a “false” verdict unless the contradiction is documented and directly relevant.

## Effort budget and output template

Choose the smallest evidence pass that fits the decision. Stop when the pass either reaches its evidence gate or documents why it cannot.

| Stakes | Minimum pass | Stop or escalate when |
|---|---|---|
| Low | Decompose the claim, identify the source of truth for any material factual assertion, and retrieve it when it can settle the point. | The authoritative source settles the narrow claim, or the output records why the claim remains Unsupported or Unknown. |
| Medium | Trace the strongest factual claim to a primary source and check one independent source for corroboration or contradiction. | The primary source and independent check agree, or their disagreement has a named cause. |
| High | Preserve the claim and evidence trail; seek primary, independent confirmation or out-of-band authorization. | The decision can be deferred, evidence remains unavailable, or qualified human authority is required. |

Do not keep collecting sources after a pass has reached its gate. Report an evidence gap rather than manufacturing confidence.

Use `assets/claim-audit.md` for the single canonical report template.

# Empirical Foundations and Boundaries

This skill is designed around evidence-backed reasoning practices. These sources support the method; they do not make the skill a universal truth machine.

## Lateral reading over surface credibility

Wineburg and McGrew observed historians, undergraduates, and professional fact-checkers evaluating live websites. Fact-checkers left sites quickly to investigate them through independent sources, reaching more warranted conclusions than participants who evaluated sites from their internal design, logos, credentials, and About pages.

- Source: [Wineburg & McGrew, *Lateral Reading*](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3048994)
- Use in this skill: trace provenance externally before trusting presentation.
- Limitation: observational comparison of 45 participants; it describes expert practice, not a guarantee that every search succeeds.

## Literal truth can still mislead

FTC guidance treats both express and implied claims as relevant and evaluates an advertisement’s net impression. It notes that testimonials are not substantiation, and that a claim broader than its supporting evidence is not adequately substantiated.

- Sources: [FTC Truth in Advertising](https://www.ftc.gov/news-events/topics/truth-advertising), [FTC Myths and Half-Truths](https://www.ftc.gov/news-events/news/speeches/myths-half-truths-about-deceptive-advertising)
- Use in this skill: separate literal wording, likely implication, scope, and typical result.
- Limitation: US consumer-protection guidance; it supplies a strong reasoning model, not a universal legal test.

## Volume and repetition are not corroboration

RAND’s analysis of the “firehose of falsehood” identifies high-volume, multichannel, rapid, repetitive messaging, with little commitment to truth or consistency. It connects these features to familiar psychological mechanisms, including source multiplicity, familiarity, peripheral credibility cues, and first-impression effects.

- Source: [Paul & Matthews, RAND PE-198](https://www.rand.org/pubs/perspectives/PE198.html)
- Use in this skill: count independent evidence origins, not messages, and record contradictions.
- Limitation: the case concerns Russian propaganda; transfer the mechanisms with care rather than treating every repeated claim as coordinated propaganda.

## Reflection helps, but does not abolish bias

Levine’s truth-default theory explains that people generally presume honesty unless suspicion is triggered, with belief congruity, social congruence, and repetition influencing acceptance. Pennycook and Rand found that cognitive reflection correlated with better discernment between real and fake headlines, including ideologically congruent ones.

- Sources: [Levine (2022), PubMed PMID 35763893](https://pubmed.ncbi.nlm.nih.gov/35763893/), [Pennycook & Rand (2019), PubMed PMID 29935897](https://pubmed.ncbi.nlm.nih.gov/29935897/)
- Use in this skill: make a deliberate disconfirming question and evidence pause routine.
- Limitation: these studies do not prove that reflection reliably defeats motivated reasoning in every domain or group.

## Prebunk techniques, not only individual claims

Roozenbeek, van der Linden, and Nygren found that a Bad News intervention improved recognition of simulated manipulation techniques across four language versions. The techniques included impersonation, emotional language, polarization, conspiracy, discrediting, and trolling.

- Source: [Roozenbeek et al. (2020), HKS Misinformation Review](https://misinforeview.hks.harvard.edu/article/global-vaccination-badnews/)
- Use in this skill: name the technique, then investigate the concrete claim it carries.
- Limitation: voluntary, self-selected pre-post samples and uncertain long-term durability; prebunking complements verification.

## Do not infer lying from demeanor

An open-access review concludes that no non-verbal sign has been shown to reliably identify lying at the individual level; it describes a meta-analysis of more than 25,000 veracity judgments with average accuracy near chance.

- Source: [Denault et al. (2020), *Research on Non-verbal Signs of Lies and Deceit: A Blind Alley*](https://pmc.ncbi.nlm.nih.gov/articles/PMC7767987/)
- Use in this skill: evaluate content, provenance, context, and corroboration rather than confidence, eye contact, voice, or body language.
- Limitation: this does not imply every speaker is equally credible; it rejects demeanor as a dependable individual lie detector.