Guardrails: the rules a draft must pass before it publishes

Most 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 described as reviewed had eleven percent human edits. None of those are cases where somebody weighed the options and chose badly; they are cases where nobody checked, because the check lived in a style guide rather than in the workflow.

That distinction matters because it tells you what kind of control to build. Judgement calls need experienced people. Rules need a mechanism, and a mechanism is cheaper, faster and more reliable than a person remembering.

The validator is the mechanism: a short list of conditions that must be true before a piece is allowed out, each one reported pass or fail as you fill the form. It does not make the piece good. It makes it ineligible to publish while obviously unfinished, which is a different and very valuable guarantee.

The list is deliberately short. Six rules that are checked every time beat forty that are checked never, and a rule nobody enforces is worse than no rule because it implies a control that does not exist.

The six rules and why each one is there

A minimum human edit share, because the difference between generated and written is the human pass, and a percentage is the only cheap proxy for it. A floor of twenty-five percent is a judgement about what constitutes a real edit rather than a restyling.

At least one citation, because an uncited factual statement is a debt and a document with no citations is a document that has not engaged with any evidence. The rule is crude on purpose: counting citations is easy, judging their quality is not, and the easy version catches the common case.

Zero prohibited claims, because guarantees, absolutes and regulator approvals are the sentences that turn into complaints. The scanner looks for the vocabulary and the form asks you to confirm the count.

A named reviewer, because an unattributed review is nobody's review. The name is the accountability, and it is also what makes the certificate meaningful to a third party.

A deliberate attestation, because a form filled in by reflex is not a decision. One checkbox that says a person read this is worth more than four fields that auto-fill.

And a recorded validation, because an audit that exists only in memory cannot be shown to anyone.

• Human edit share at or above the floor.
• At least one citation supplied by the draft.
• No guarantees, absolutes or regulator approvals.
• A named reviewer with the right discipline.
• An explicit attestation, given rather than assumed.
• A recorded result with a date and a draft identifier.

Measure first, declare second

The numbers describe the draft, so the tool measures what it can. Press measure and it counts citations it can see — numbered references, named links — and scans for the phrases that get a piece pulled. What it finds fills the declared fields.

You can correct the figures, and you should when the measurement and your judgement disagree. But the disagreement is itself information. 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, and that is worth fixing in the text rather than in the form.

The same applies to the prohibited-claim count. A scanner that finds nothing when you know the draft says guaranteed means the phrase is spelled or spaced in a way the rule does not match, which means it is also in a form a reader might miss. Fix the text, then measure again.

Measuring before declaring also stops the small dishonesty of filling in the numbers you want. The fields are pre-filled from the draft, so the work is correcting a measurement rather than producing an opinion.

What a passing validation is worth

It is worth exactly what it says: at a point in time, this draft had a certain edit share, at least one citation, no prohibited claims, and a named person who attested to reading it. That is a process record, and process records are what procurement and legal teams ask for.

It is not a quality judgement. A draft can pass every rule and be dull, badly structured or wrong in a way no rule detects. Rules are a floor, and a floor is not a ceiling.

It is not a fact-check. Citations are counted, not read. A draft can carry six citations that do not support the sentences they are attached to, and the validator will pass it.

Which is why the certificate is a chain rather than a single document. The guardrails prove the conditions were met; the claim matrix proves the evidence was recorded; the certificate binds them together under a name and a timestamp. Any one of them alone is thinner than it looks.

The prohibited list for a regulated client

In health, finance and legal work the prohibited words are not a style preference, they are the boundary between marketing and a regulated claim. Cures, guarantees, risk-free, approved by, clinically proven: each has a definition and each has been litigated.

Build the list with the client's compliance contact rather than from a template, and record the date it was agreed. A list built from a generic source will miss the phrasing their regulator specifically watches, and a list with no provenance is impossible to defend when somebody asks who decided.

Treat the scanner as a first pass, not the control. It matches vocabulary; a compliance review reads for the claim behind the words. A sentence can avoid every prohibited term and still promise an outcome, which is the failure mode that a word list cannot see and an experienced reviewer can.

Keep the findings rather than only the pass. A record of seven prohibited phrases removed before publication is evidence of a control working, and that record is worth more in an audit than a green tick with nothing behind it.

When a rule fails late

Late failures are the interesting ones, because they tell you which rule was doing real work. A draft that fails the citation rule the day before publication is a draft somebody believed was finished, and the conversation about it is more valuable than the fix.

Distinguish between a rule that caught a real problem and a rule that is miscalibrated. A prohibited phrase found in a quotation from a regulator is a false positive that argues for a narrower pattern; a prohibited phrase in the client's own claim is the rule doing exactly its job.

When a rule is waived, record the waiver with a name and a reason. A silent exception teaches everyone that the rules are advisory, and rules that are advisory stop being checked within a month.

Then look for the process that let it get this far. If the citation rule only fails at the end, the draft was never measured at the start, and moving the measurement earlier is a smaller change than adding a reviewer to the final day.

Where guardrails sit in a CMS workflow

A validator that lives in a separate tab is a validator that gets used on Fridays. The rules have to sit at the last gate before publication, which usually means a checklist on a review screen, a ticket template or a staging step.

The practical pattern is a link and a required field. The ticket links to the validator; the ticket requires the recorded result before the publish step is allowed. Nothing about that is technical, and it is the difference between a rule and an intention.

For teams with a content calendar, run the measurement early instead of only at the end. A draft measured on Monday can be fixed on Tuesday; the same measurement on Thursday produces an argument about the deadline.

The rule that catches the most

If a team can enforce one thing, it should be the citation rule rather than the edit share. Edit share is a proxy and a soft one; a citation is binary, visible in the document and directly tied to whether a claim can be defended.

The second most valuable is the prohibited-claim scan, because the vocabulary involved is short, stable and expensive to get wrong. Both are cheap to check and both fail in ways that are visible to a reader.

Setting the floor for your own team

Publish the rules somewhere everyone can read them, and make the numbers visible before work starts. A writer who knows the edit-share floor on Monday writes differently from one who discovers it on Friday.

Raise the floor rather than adding rules. Teams that add a seventh and eighth condition quickly produce work that satisfies none of them, because each one costs attention and attention is the constraint.

Review the rules once a quarter against the things that actually went wrong. A rule that has never failed in a year is either unnecessary or not being measured, and both are worth knowing.