Why every draft is checked by a separate step

After the writer drafts a message, a second step reads it with one question: can every claim in it be traced to a fact that was actually found? It sees the draft and the facts, and it was never asked to make the message good. If a claim is not in the facts, the line does not ship.

Why the writer cannot do it

A model that has just written a sentence believes it. Asked to review its own draft, it tends to explain why the sentence is fine. Given only the facts and the finished text, and told that its whole job is to find what is not supported, a separate step reads the same sentence very differently.

How it decides

It lists every claim the draft makes about the company and marks each one supported, unsupported or contradicted by the facts. When it is unsure, the claim counts as unsupported. Whether the draft passes is not the model's call: one unsupported or contradicted claim fails it, and that rule is plain code, so no wording in a draft can talk its way past it.

What it is there to catch

• A number that is close to the real one but is not it.
• A claim that sounds like research and is really what companies like this one usually do.
• Praise that implies something was read when it was not.
• The sentence that slipped in from general knowledge because it made the message flow.

What it does not do

It does not make a message better. Tone, length and the shape of the opening belong to the writer and the rules it writes under, the same rules behind the guide on messages that read as AI. The check exists for one failure only, because a check that tries to improve a message starts rewriting it, and a rewrite is a new place for an invented fact to appear.

What it means for you

You still read every draft before it goes, and you should. But the draft that reaches you has already been read once by something whose only job was to find the sentence that is not true.