---
name: statistical-claim-checker
description: Validate every quantitative claim in an article against the source data it cites, flagging numbers that are unsupported, outdated, or selectively quoted.
title: Statistical Claim Checker
category: research-analysis
difficulty: advanced
author: admin
icon: 🧮
input: mixed
output: markdown
phase: post
domain: research
tags: fact-checking,statistical-validation,claim-verification,source-attribution,quantitative-analysis,document-comparison,data-provenance,freshness-detection,cherry-picking-detection,markdown-annotation,pdf-csv-extraction,evidence-retrieval
best_for:
  - Editorial fact-checking before publication
  - Academic paper validation against cited sources
  - Regulatory compliance document review
  - Audit trails for quantitative claims in reports
  - Journalism/news article verification
---

## Description

Reads a draft article and a folder of source PDFs or CSVs. Extracts every numeric claim (percentages, dollar amounts, counts, rates), links each to its cited source, and returns a per-claim verdict: supported / unsupported / outdated / cherry-picked. The output is a line-by-line annotation you can diff against the draft.

## Why it works

Most factual errors in published writing are not invented numbers — they are stale numbers, context-stripped numbers, or numbers that cite a source that technically says the opposite. A general proofread misses these because the claim *sounds* plausible. A targeted extract-then-verify loop catches them because each number is checked against the exact sentence it came from.

## How it works

1. Claim extraction: walk the draft sentence-by-sentence and emit (sentence, number, unit, cited_source) tuples. 2. Source chunking: index the cited sources as retrievable passages. 3. For each claim, retrieve the 3 most similar source passages and ask the LLM to judge support with required verbatim evidence quote. 4. Freshness check: compare the source's publication date to the claim's tense — 'as of 2024' against a 2019 source fails. 5. Cherry-pick check: for supported claims, ask whether adjacent passages contradict or qualify the number. 6. Emit markdown annotations keyed to draft line numbers.