SIA
FactcheckIQ
How it works · Verification pipeline
Join the EMOS Platform
Inside the EMOS Platform
FACTCHECKIQ

How a draft becomes a graded, sourced report in a single pass

INTERACTIVE
Sample scenario · Fairground (illustrative)
10 STEPS · 2 MODES · 6 VERDICTS · 5 PRINCIPLES · 11 METHODS
Raw draft in

Paste, Markdown, PDF, or DOCX. Every checkable claim is pulled out, word for word.

Fairground's draft blog post + 2 files
nielsen-trust.pdf fairground-earned-media-post.docx

Nielsen surveyed more than 28,000 people across 56 countries, finding that 92 percent of consumers trust earned media, meaning press and media coverage, above every other form of advertising. The same study found online reviews ranked as the least trusted source of brand information. Content marketing costs 62 percent less than traditional marketing while generating three times the leads, and companies that blog get 67 percent more leads, which proves that blogging drives growth. In fact, a 2023 Harvard Business School study found that PR-driven leads convert five times better than paid ones.

The ten step pipelinethe box below shows the step it is parked on. drive it with the player, or click any number
SPEED
MODE
STEP 01 / 10 · PLAYING

Sequential mode: one claim verified at a time, in the order shown below. Default for shorter drafts.

STEP 01

Intake & scope

Read the whole piece, then state plainly what is being checked and how many claims were found.

PASTEMARKDOWNPDFDOCXCLAIM COUNT
Graded report outone verdict per claim, each with a real source, never a placeholder
fact-check-report.mdFULL AUDIT · 2026 · 6 CLAIMS
1 Verified1 Partly accurate1 Misleading1 Unverifiable1 Inaccurate1 Fabricated
Not ready as written. Fix the Fabricated, Inaccurate, and Misleading claims first.
#
Claim as written
Verdict
What verification found
Source
1
Nielsen surveyed more than 28,000 people across 56 countries
Verified
Nielsen's 2012 Global Trust in Advertising surveyed 28,000+ internet respondents in 56 countries.
2
92% trust earned media (meaning press coverage) above all advertising
Partly accurate
The 92% figure is real, but Nielsen defines it as recommendations from friends and family (word of mouth), not press coverage, and it is a 2012 number.
3
Online reviews were the least trusted source of brand information
Inaccurate
The same Nielsen study found online consumer reviews were the second most trusted source, at 70%.
4
Content marketing costs 62% less and generates 3x the leads
Unverifiable
Traces only to a DemandMetric infographic with no published sample, method, or year. Widely repeated, never sourced to primary data.
5
Companies that blog get 67% more leads, which proves blogging drives growth
Misleading
The 67% benchmark is correlational and variously attributed (HubSpot, DemandMetric, InsideView). “Proves blogging drives growth” is causation claimed from correlation.
6
A 2023 Harvard Business School study found PR-driven leads convert 5x better than paid
Fabricated
No such study exists. A search of HBS faculty research and the web returns nothing matching this claim.
none found
The verdict scaleCONFIRMED INVENTED
Verified

A primary source, ideally two, confirms it as written.

Partly accurate

The core is right, but a detail is off, dated, or loosely rounded.

Misleading

Traceable, but stripped context or framing claims more than the source supports.

?Unverifiable

We searched and found no reliable source either way. It may be true, but it cannot be confirmed as written. Action: add a citation or cut it.

Inaccurate

A credible source directly contradicts it.

Fabricated

We checked, and the cited study, source, or quote does not exist, or the real source says something different. Action: remove it.

The difference: Unverifiable means we could not find evidence either way (absence of evidence). Fabricated means we found evidence it is invented or wrong (evidence of absence). Never treat one as the other.

Verification frameworks5 cross-cutting principles · the heart of the skill

The ten steps are the sequence claims move through. These five principles are the standards applied at every step, plus six more woven directly into steps 04, 05, 06, 08 and 09 above, and one (parallel verification) behind the 2 MODES stat up top. Eleven methods, one skill.

1Lateral reading / SIFT

Judge a claim by opening other sources, not by reading further down the same page.

Every load-bearing claim is checked away from the document, following SIFT's four moves: Stop, Investigate the source, Find better coverage, Trace to the original.

In plain terms: Don't trust a page to vouch for itself. Open up what independent sources say, then decide.

APPLIES AT STEPS 3, 4, 5
2IFCN / Poynter standards

The professional code real fact-checking organisations operate under.

Non-partisanship, transparency about sources and method, and visible corrections, borrowed from the International Fact-Checking Network.

In plain terms: The rulebook serious fact-checkers sign onto. The skill works to that standard.

APPLIES AT STEPS 1 TO 10
3Primary-source tracing

Follow every claim back to where it actually originated.

Statistics and quotes are traced past any secondary report to the original dataset, paper, or transcript.

In plain terms: A primary source is the original, not someone repeating it.

APPLIES AT STEPS 4, 5, 6
4Source-reliability tiers

Rank every source before trusting it.

T1Primary — dataset, paper, transcript
T2Reputable secondary — major news, institutions
T3Weak — blogs, marketing, Wikipedia
T4Unreliable — farms, AI pages, spam

In plain terms: A claim is only “Verified” on T1 or T2. The skill won't bless anything lower.

APPLIES AT STEPS 4, 5, 9
5Prompt-injection hygiene

A safety guard for a tool that reads the open web.

Any instruction found inside a fetched page is treated as suspicious data to report, never as a command to obey.

In plain terms: Bad actors hide commands inside web pages to trick AI. The skill reads those as content to flag, not orders to follow.

APPLIES AT STEPS 4, 5, 7
6Parallel verification · the 2nd mode

Rigour that is too slow to use does not get used.

Many claims can be verified concurrently with helper agents, then aggregated, applying the same rules to each. This is “Parallel” mode, the second of the 2 MODES above, alongside default Sequential mode.

In plain terms: For a long piece, the skill checks many claims at once instead of one by one. Same standards, less waiting.

THE 2ND OF THE 2 MODES

Rigour that is too slow to use does not get used.

The design principle behind all 11 methods
11 ×
What this is

A safety net, not a ghostwriter

Spell-check caught typos. Grammar-check caught grammar. FactcheckIQ catches the false claim and the invented source, before your reader does. It works the same whether a person or an AI wrote the draft: it verifies, and it never writes for you. Like the typewriter, spell-check, and Grammarly before it, this is a tool that makes writers more trustworthy, not obsolete.

For human writers and AI drafts alike
FACTCHECKIQCatch the made up statistic before your reader does.