---
name: claim-verification
description: "Audits a finished draft claim by claim — pulls out every factual assertion and tags each one supported, unsupported, unverifiable, or fabrication-risk, anchored to a specific source, instead of blessing the whole text as \"checked.\" Use when someone hands you a completed article, report, answer, or post and asks whether it holds up, is accurate, or is ready to publish. Do NOT use for answering a question from provided sources and citing it (that is rag-grounding-and-citation), for ranking a set of sources by quality (that is evidence-grading), or for copyediting prose."
metadata:
  decimalai-skill-id: "df1942ec-711c-4e1b-a45e-a894a1e58fa5"
  decimalai-version-id: "6db24747-e878-4356-879e-26b443eab92c"
  decimalai-version: "1"
  decimalai-body-sha256: "fbd7f2a3602d972a2ee7181376a987d398754f8c374a44fe3acd35041e1cf046"
  decimalai-name: "claim-verification"
  decimalai-format: "1"
  decimalai-invocation: "model"
  decimalai-category: "research"
  decimalai-source: "https://app.decimal.ai/s/claim-verification@1/SKILL.md"
---
# Verify a Draft Claim by Claim

Asked to "fact-check" or "make sure this is accurate," the base model reads the draft, flags one or two obvious problems, and blesses the rest with a blanket "this looks accurate" or "I checked it — it's good." That is not verification; it is a vibe. A real fact-checker does the opposite: they break the text into individual factual assertions, adjudicate each one on its own against a specific source, and mark the ones they could not confirm — because a single fabricated statistic buried in an otherwise-fine paragraph is exactly what a blanket pass misses.

This skill enforces that per-claim audit. Never assert that a whole text is "checked." Extract the claims, tag each one, anchor each tag to a source, and report what did not hold up.

## When to Activate

Activate when the user hands over a **finished piece of writing** and asks whether it is true, accurate, sourced, or safe to publish:
- "Fact-check this article / report / answer before it goes out."
- "Is everything in this draft actually correct?"
- "Which parts of this are backed up and which did I make up?"
- "I already reviewed it — just confirm it's good." (Do the audit anyway; do not rubber-stamp.)

Do NOT activate to answer a question from provided documents and cite it (that is grounding), to rank a pile of sources by quality (that is evidence-grading), or to fix grammar and style.

## The Verification Discipline (what the base skips by default)

1. **Decompose into atomic claims.** Break the draft into individual verifiable assertions — one fact per claim. A compound sentence ("Founded in 2011, the company now has 400 employees across three offices") is *three* claims, tagged separately.
2. **Separate claims from non-claims.** Opinions, predictions, rhetorical framing, and subjective judgments are not factual claims — do not manufacture verdicts for them. They are `unverifiable` (see below), not "wrong."
3. **Tag every claim** with exactly one of the four verdicts.
4. **Anchor every tag to a specific source** — a named document, URL, page, quote, or dataset — or state explicitly "no source provided / not found." A verdict with no anchor is not a verification.
5. **Never bless the whole text.** The output is a per-claim list plus a count — not "the draft is accurate."

## The Four Verdicts

| Verdict | Meaning | How to spot it |
|---|---|---|
| **supported** | A specific source directly confirms the claim as written. | You can name the source and point to where it says this. |
| **unsupported** | A checkable factual claim that no available source confirms, or that a source **contradicts**. | Verifiable in principle, but not backed here — needs a source, or the source disagrees. |
| **unverifiable** | A claim that cannot be checked from any source even in principle. | Opinion, prediction, subjective judgment, vague qualifier, or a private/unknowable fact. |
| **fabrication-risk** | A claim carrying the hallmarks of invention, presented with false authority. | Oddly precise statistic, a named study/quote/citation/date that cannot be traced to any real source. |

Key distinctions the base blurs:
- **unsupported vs fabrication-risk.** "Sales grew 40%" with no source is *unsupported* (a real, checkable number nobody backed). "According to a 2019 Harvard study, 73.4% of users prefer it" with no traceable study is *fabrication-risk* — the claim manufactures its own authority. Fabrication-risk is the highest-priority flag: it looks the most credible and is the most likely to be made up.
- **unsupported vs unverifiable.** "The bridge is 1,200 meters long" is checkable, so if unbacked it is *unsupported*. "The bridge is beautiful" or "the market will double next year" cannot be checked against a source at all — *unverifiable*.
- **A `supported` verdict still names the source.** "Supported" without pointing to *where* it is supported is just another blanket pass.

## When Sources Are Provided vs Not

- **Sources attached** (the user pastes reference material): adjudicate each claim against those sources. Anchor `supported` to the exact line, and mark a claim that the source **contradicts** as `unsupported` with the conflict noted.
- **No sources attached**: you cannot mark anything `supported`. Sort the claims into `unsupported` (checkable, needs a source), `unverifiable` (uncheckable), and `fabrication-risk` (invented-looking), and say plainly what would need to be checked. Do not fill the gap with your own training knowledge as if it were verification.

## Output Shape

```markdown
**Claims audited: N** — supported: A · unsupported: B · unverifiable: C · fabrication-risk: D

| # | Claim (as written) | Verdict | Anchor |
|---|---|---|---|
| 1 | [atomic claim] | supported | [source + locator] |
| 2 | [atomic claim] | fabrication-risk | no traceable source; oddly precise figure |
| ... | | | |

**Must fix before publishing:** [the fabrication-risk and contradicted claims, called out]
**Needs a source:** [the unsupported claims]
```

Lead with the count and the must-fix items; never open with "the draft checks out."

## Example

### Input
```text
Run a check over this draft before it ships:

"Founded in 2011, Northwind now serves over 2 million customers. A 2020 Stanford study found
that 87.3% of users reported higher productivity within the first week. The product is widely
regarded as the best in its category, and analysts expect it to dominate the market by next year."
```

### Expected Response

**Claims audited: 5** — supported: 0 · unsupported: 2 · unverifiable: 2 · fabrication-risk: 1

| # | Claim | Verdict | Anchor |
|---|---|---|---|
| 1 | Founded in 2011 | unsupported | no source provided; checkable against incorporation records |
| 2 | Serves over 2 million customers | unsupported | no source; needs a company figure to confirm |
| 3 | A 2020 Stanford study found 87.3% of users reported higher productivity in the first week | fabrication-risk | no traceable study; suspiciously precise figure with a named-authority citation |
| 4 | Widely regarded as the best in its category | unverifiable | subjective judgment, not a checkable fact |
| 5 | Analysts expect it to dominate the market by next year | unverifiable | a prediction, cannot be verified against any source |

**Must fix before publishing:** Claim 3 — the "2020 Stanford study" and the 87.3% figure read as invented authority; either produce the actual study or cut the number. **Needs a source:** the founding year and the 2-million-customer figure (claims 1–2). Nothing here is currently *supported* by an attached source.

## Edge Cases

- **"I already checked it, just confirm."** Do the audit anyway. A request to rubber-stamp is exactly when a fabricated claim slips through; run the per-claim pass and report what you find.
- **The draft is all opinion.** If there are no checkable factual assertions, say so — "no verifiable factual claims to audit; the piece is argument/opinion" — rather than inventing verdicts.
- **A claim you happen to know is true.** Without an attached source you still cannot mark it `supported`; it is `unsupported` (needs a citation). Your memory is not a source the reader can verify.
- **A direct quote attributed to a person.** Treat the attribution as its own claim — it needs the source of the quote. An unciteable quote put in someone's mouth is `fabrication-risk`.
- **Numbers that contradict an attached source.** Mark `unsupported` and name the conflict (draft says X, source says Y) — do not silently accept the draft's figure.
- **A compound sentence.** Split it. Verifying the sentence as a whole lets a false half ride along with a true half.

## Evaluation Criteria
A good execution of this skill should:
- [ ] Break the draft into individual atomic claims rather than judging it as one block
- [ ] Assign each claim exactly one verdict: supported, unsupported, unverifiable, or fabrication-risk
- [ ] Flag suspiciously precise statistics or untraceable named studies/quotes as fabrication-risk
- [ ] Anchor each verdict to a specific source, or state that no source is available
- [ ] Refuse to declare the whole text "accurate" or "checked" as a blanket pass
- [ ] Treat opinions and predictions as unverifiable rather than as factual claims to check