/claim-source-verification

name: claim-source-verification
description: Use when a manuscript's claims need to be checked against the sources cited for them, when literature search results must be verified before they enter a draft, or when an agent or collaborator has proposed references that have not yet been adversarially reviewed. This is claim-to-evidence verification, not bibliography hygiene; for duplicate keys, DOI syntax, and metadata cleanup use citation-verifier instead.

Claim-Source Verification

Overview

A citation can be perfectly real, perfectly formatted, and still wrong, because the paper it points to does not say what the sentence claims it says. That is the failure this skill exists to catch, and it is the dominant one.

The calibration comes from a real Review manuscript. Seven research agents proposed **139 sources** to resolve 44 open factual questions. An independent skeptic then re-checked every one. It rejected **55**.

| Rejection class | Count | | --- | --- | | `does_not_support_claim` | 29 | | `metadata_wrong` | 12 | | `wrong_site_or_population` | 6 | | `not_primary_but_required_to_be` | 5 | | `paywalled_unverifiable` | 3 |

**None was fabricated.** Every rejected source was a real, findable paper with a real DOI. An existence check would have passed all 139. A metadata check would have caught 12. The largest class, more than half of all rejections, is invisible to both: a genuine paper made to carry a conclusion it does not reach.

So the operating rule is: **a source's existence is not evidence that it supports the claim, and a clean bibliography audit is not evidence of anything at all.**

When To Use

Use this skill when:

• an agent, a collaborator, or a literature-search tool has proposed references

and nothing has adversarially checked them yet

• a manuscript makes quantitative claims (limits, thresholds, error rates,

standard clauses, benchmark numbers) that trace to cited sources

• open factual or sourcing markers are being closed out in batch
• a Review or survey is being sourced, where most claims are about other people's

measurements rather than the author's own

Do not use this skill for:

• duplicate BibTeX keys, DOI syntax, missing fields, placeholder entries. That is

`citation-verifier`, and it runs before this, not instead of it.

• discovering literature from scratch. That is `academic-researcher`, installed

with `--set all`. This skill checks what such a pass produces.

• deciding whether a claim is interesting. That is `manuscript-optimizer`.

The three stack: `citation-verifier` for hygiene, this skill for claim support, `reference-audit-guide` (optional set, `--set all`) to machine-verify that a cited work exists before asking whether it supports the sentence.

The Two-Role Protocol

Verification must be adversarial and it must be **separate**. A researcher who proposes a source and then checks their own source will confirm it, because they already believe it. Run two roles:

**Proposer.** Given one claim, find sources that support it. Return, per source: the identifier, the metadata, and **the sentence from the source that carries the claim, quoted verbatim**.

**Skeptic.** Given the claim and the proposed source, **try to refute the pairing**. Run this in a fresh context that has never seen the proposer's reasoning.

Three rules make the skeptic work:

1. **The default verdict is reject.** Uncertainty resolves against the source, not in its favour. A skeptic that needs to be convinced to reject will confirm almost everything. 2. **No verbatim supporting sentence, no confirmation.** If the skeptic cannot quote the line that carries the claim, the pairing is rejected regardless of how plausible it looks. This single rule produced most of the 29 `does_not_support_claim` rejections. 3. **Spend more on verification than on research.** The skeptic reads the actual source; the proposer often reads an abstract. Give the verify stage the higher reasoning budget, not the lower one.

If the claim can fail in more than one way, give the skeptics different lenses rather than running three identical ones: does the source say this, is the population right, is this the primary record. Diversity catches failure modes that redundancy cannot.

`references/adversarial-protocol.md` carries the prompt shapes and the structured output schema.

Four Outcomes, Never Two

Confirmed-or-rejected is not enough, because it forces two different situations into one bucket and loses the distinction that matters most.

• **CONFIRMED.** The source supports the claim, the metadata is right, and the

supporting sentence has been quoted. The claim can go into the manuscript with this citation.

• **PARTIAL.** The finding holds, but at least one proposed source has to be

replaced first. **Do not apply. Do not remove the marker.** A PARTIAL that gets written up as done is how an unsupported sentence enters a draft while the audit trail says it was verified.

• **REJECTED.** The claim is not supported by anything proposed. Either it is

narrowed to what the surviving sources do support, or the sentence goes.

• **BLOCKED.** The answer exists but is behind a paywall, in proprietary vendor

data, or requires a decision only the authors can make. This is a real state and it needs to be reported as one, not left to look like unfinished work.

In the run above, 44 markers were researched. Eight were discharged outright, two were narrowed to what the surviving sources support, and **25 came back PARTIAL**. That was the largest bucket by a wide margin. Plan for it: most of the work after a verification pass is replacing individual sources, not rewriting claims.

Named Traps

Each of these is a real rejection from the run described above. They repeat across projects, so check for them by name.

**Derived value laundered into a measurement.** A proposal reported an equivalent input noise of 31 dB(A) as a datasheet value. The datasheet gives only a