Your AI has never once said I don't know

The invented figure arrives wearing the same confidence as every true thing the model told you that morning, which is exactly why a read-through never catches it.

So this stops trusting the read-through. One prompt pulls every checkable claim out of the draft and ranks it by what being wrong would cost you, and one script goes and tests the links and quotes itself.

• The claims worth checking, ranked by what being wrong would cost you
• Why a real source attached to an invented number is the hardest failure to spot
• The check that proves a quote exists rather than that it sounds plausible
• How to tell a dead link from one that now points somewhere else
• The category of claim that is fabricated most often and read past most easily
• What to do with a number you cannot source but do not want to cut

I've spent eight years publishing claims that had to survive a client's legal review, including conversion copy for NordVPN, PandaDoc and OpusClip.