What it is.
Stage one extracts the checkable claims from a draft. Stage two verifies each one with a web search model. Stage three writes a synthesis report marking what verified, what failed, and what could not be sourced at all.
It runs 22 nodes across four AI models, and a 2,000-word article completes in under 3 minutes. The same article checked properly by hand is documented at 45 to 60 minutes.
TripleDart's own guide, and it names our own platform. The three stages work in any tool.
What you get.
- A claim extraction stage that lists every statistic and attributed statement in the draft.
- A verification pass per claim, with a source attached or an unverifiable verdict.
- A synthesis report written so every correction can be made in one pass.
- Source priority lists by vertical, so fintech and cybersecurity claims get checked against different authorities.
- A three-call starting version, for building it yourself before committing to 22 nodes.
How to set it up.
Start with the three-call version and check that claim extraction finds everything before you build further.
Write your source priority list per vertical, since what counts as authoritative differs by field.
Run stage one alone on a finished draft and read the extracted claims, because a missed claim never gets checked.
Point verification at a web search model, so it reads live pages instead of recalling them.
Treat an unverifiable verdict as a decision point: rewrite the sentence or cut the number.
Put the pipeline before publishing, since a correction after the fact costs more.
Use cases
List every claim in a draft
Run the extraction stage so each statistic and quote is on the table.
Verify one claim at a time
Let the verification pass check each one against its own source.
Check a long article quickly
Run a full draft through the stages before the editor reads it.
Best for
Content with statistics in it
Every borrowed number is a small liability, and nobody remembers where half of them came from.
AI-assisted drafts
A model that writes a plausible statistic is the exact reason this stage has to exist.
Auditing an archive
Old posts carry old numbers, and a batch run tells you which ones now read as wrong.