Pre-publish guardrail validator

Most AI publishing failures are not judgement calls. A draft went out with no citation because nobody counted the citations; a guarantee survived because nobody was responsible for removing it; a piece was 10 percent edited and described as reviewed. Each of those is a rule that could have been checked in ten seconds and was not, because the check lived in a style guide rather than in the workflow.

The validator makes the rules explicit and mechanical. Human edit share must reach a floor, at least one citation has to be supplied, prohibited claims must be at zero, a named person has to sign, and the attestation has to be given deliberately. When a rule fails the form refuses to record, which is the entire point: a guardrail that can be talked past is a suggestion.

The numbers on the form describe the draft, so the tool measures what it can. Press measure and it counts the citations it can see — numbered references and named links — and scans for the phrases that usually get a piece pulled: guarantees, absolutes, clinically proven, regulator approval, cures. What it finds fills the declared fields, and you can correct them when the tool's count and your judgement disagree.

That disagreement is worth looking at. If the scanner finds three citations and you believe there are six, the other three are probably in a format a reader will not recognise either.

It cannot tell whether a citation supports the sentence it is attached to, whether the human edit improved the draft, or whether a claim is true. It enforces the presence of the conditions under which review can be trusted — a person, sources, an edit, no forbidden language — and records that someone asserted them. The judgement stays with the reviewer, which is why the certificate names one.