Verify citations resolve and support document claims
Verifies citations in a document actually resolve and support their claims, recomputing figures and flagging unsupported claims via the Stipple API.
Why it matters
Ensure every citation in a research report, whitepaper, or LLM-generated document actually resolves to a real source and supports the claim it's attached to, so you can publish with confidence or catch fabricated references before they cause reputational damage.
Outcomes
What it gets done
Check that each citation URL or reference resolves to an accessible source
Flag citations where the source content doesn't match the claim made
Recompute arithmetic and financial figures to catch decimal shifts or calculation errors
Identify unsupported claims that have no citation backing them at all
Install
Add it to your toolbox
Run in your project directory:
curl -fsSL https://spark.entire.vc/get/ag-verify-citations | bash Overview
Citation Verification
This skill verifies that citations in a document resolve and actually support their attached claims, using the Stipple API to recompute figures and flag unsupported claims. Use it before publishing a research report, tender response, or whitepaper, or when reviewing an LLM-generated document whose citations may be plausibly formatted but wrong.
What it does
This skill verifies that citations in a document actually resolve and support the claims they're attached to, using the Stipple API for citation resolution, arithmetic recomputation, and unsupported-claim detection.
When to use - and when NOT to
Use it before submitting or publishing a research report, tender response, or whitepaper; when reviewing an LLM-generated document, since LLM citations are plausibly formatted and frequently wrong; for due diligence on third-party reports; or for academic reference checking. Report results as verification coverage, not a truth verdict - an unresolved or unsupported claim is not the same as a false one, and a human reviewer stays responsible for the publication or academic decision.
Inputs and outputs
Input is a document by URL, local file (PDF, DOCX, Markdown), or pasted text.
curl -X POST https://www.stipple.sh/v1/verify-references \
-F "file=@report.pdf" \
-H "Authorization: Bearer $STIPPLE_API_KEY"
Deep mode, ?deep=true, costs more credits but cross-checks citations against live web sources. Output is a verification_coverage percentage such as 78%, a per-citation status of resolved-and-matching, resolved-but-mismatched, or unresolvable with the issue explained, an arithmetic list recomputing stated figures against the source (catching things like a revenue figure stated at $4.2m against an actual $3.7m), and an unsupported_claims list for claims carrying no citation at all, such as an "industry-leading accuracy" line with no source in the document. A resolved-but-mismatched citation might look like a claimed 340% adoption growth where the actual source states 34%, a decimal-shift error the recomputation step is specifically built to catch, alongside cases where a citation simply matches its source cleanly, such as an unemployment figure that lines up with the statistics agency it cites.
Integrations
The Stipple API's anonymous free tier works without an API key for standard verification, drawing on a shared weekly allowance; a free key raises the metering ceiling for dedicated use. Deep verification resolves citations against live web sources at a higher credit cost, trading speed for a more thorough resolution pass. It pairs well with the verify-document skill, run first, to check whether the source document itself is authentic before trusting what it cites.
Who it's for
Authors and reviewers of research reports, tender responses, whitepapers, or LLM-generated documents who need to know which specific claims are backed by a real, matching citation before publishing or submitting, and which ones still need a source, a corrected figure, or removal before the document goes out.
FAQ
Common questions
Discussion
Questions & comments · 0
Sign In Sign in to leave a comment.