Preview
What it is.
Ours, and we sell the service it runs inside. Three phases run with a different model in each: one extracts every checkable claim from the draft as a list, one searches the live web for each claim, and one flags what is false and rewrites it with correct figures and citations.
You get a claim list a writer can review line by line, which an inline pass never gives you. The reason for three models is that live search and editorial judgment on conflicting sources are different strengths, and one model doing both answers confidently and wrongly.
What you get.
- Every checkable claim in a draft, extracted as a list
- A live web search per claim, so the check is against current sources
- False claims flagged and rewritten with correct figures
- Citations attached to the corrections
- A pass a writer can review instead of a verdict they have to trust
How to set it up.
Split the job into three phases before you build. Extraction, search and synthesis have different failure modes and different best models.
Extract claims first as a list. A claim list is something a writer can review line by line, which an inline pass never is.
Use a search-capable model for the lookup phase, since this is where a general model invents a source.
Give the synthesis phase the editorial judgment work: which of two conflicting sources to trust, and how to rewrite the sentence.
Require a citation on every correction, so a writer can check the checker.
Run it on finished drafts and not on outlines. A claim that is still a placeholder wastes a search.
Pricing Plans
Not sold as a skill.
It runs inside client engagements. The published account dates from April 2026.
Use cases
Pre-client stat audit
A report full of figures is due to a client tomorrow. Hand it the finished draft and every checkable claim comes back as a list, each one searched against live sources. The false ones arrive rewritten with citations attached.
Cleaning an AI-researched draft
A writer built the piece from model output and the figures arrived confident and unsourced. The extraction phase lists them all and the search phase looks each one up, so the corrections come back with sources a reviewer can follow.
Regulated-category legal review
A fintech draft needs a source behind every claim before legal will clear it. Feed it the draft and the output is a line-by-line claim list with a citation on each correction, so the reviewer can check the checker.
Best for
Anything with numbers going to a client
A wrong statistic in a published piece is the expensive kind of error, and a claim list plus citations is the cheapest insurance.
Content written from AI drafts
Confident unsourced figures are the standard failure mode, and extraction catches them as a list.
Regulated categories
Fintech and healthcare drafts need a source per claim, and the citation requirement makes that reviewable.