When you ask whether something is required

<!-- Reviewed 2026-09-10 (draft first — owner decision; the Change Management Procedure is the pilot): the covered system-prompt.ts gained a second build-mode prompt variant and flow.ts picks it only when the document being built starts from a complete draft. The obligation rule this page describes — every "is it required" answer names its clause, and the guardian sends a sourceless one back — is untouched: both build variants retire the same bracket-interview words and keep the same rulebook around them. Stands as written. -->

Half the questions people ask while tailoring a document are the same question wearing different clothes: do we actually have to?

Do we have to keep this register. Do we have to name an owner. Is it a problem that we do not have one. It is the question that decides how much work the next month holds, so it is the one an answer has to be right about — and "right" here does not mean confident. It means traceable.

Every answer names where it came from

When the assistant tells you something is required, mandatory, optional, or fine to leave out, it names the clause that says so. Not as a flourish at the end — as the source the sentence rests on.

That is a rule the product enforces, not a habit it has. If a reply reaches that point without a clause behind it, it does not go out: the assistant reads the clause first and answers again from the text. You will occasionally notice the difference as a slightly longer pause on that kind of question. That pause is the clause being read.

Where there is nothing to read — the question is about a framework that is not in your catalogue, or the text does not settle it — you get told exactly that, and no verdict. An honest "I cannot tell you from what I hold" is the answer the rule asks for. A confident guess is the thing it exists to prevent.

Required to do it, required to write it down

These are different, and the difference is where most of the confusion lives.

A clause can require an outcome — that you have worked something out, that you have decided it, that you know who is responsible — without requiring any particular document to show it. The outcome is not negotiable. The register, the log, the spreadsheet is the usual way people demonstrate the outcome, and usually not the mandated one.

So the answer to "is this register mandatory?" is often two answers:

• the thing the clause actually requires, which you do have to do;
• the artefact you were asking about, which is expected rather than demanded, and which you may replace with something else that shows the same thing.

The assistant is expected to give you both halves. Collapsing them into one — "yes, it is mandatory" or "no, it is not" — is how a true sentence turns into a misleading one.

Three levels, and one that means "we have not checked"

Where a decision on a template has been written up against a framework you have added, the answer arrives with the question rather than after it. You will see it on the question card and in the answer, in one of three words:

A decision with no level shown is not a fourth level. It means nobody has written one up for that decision against that framework yet — so the assistant reads the clause and answers from the text, exactly as it would have. It never reads a missing level as "not required". These write-ups are authored and reviewed like any other content in your catalogue, framework by framework, so a framework can be fully covered or partly covered, and the label tells you which.

When you push back

Tell the assistant it is wrong and it will not simply agree with you.

It re-reads the clause and answers from the text again. It may agree with you, disagree, or tell you that you are each right about a different half — which is the most common outcome, because the two halves above are usually what the disagreement is about. What it will not do is reverse its position because you sounded certain. A changed answer comes with a clause you can check; it does not come with an apology.

This is deliberate, and it is worth knowing about, because the alternative is worse than it looks. An assistant that folds under pressure gives you the answer you already believed — which means it has told you nothing, at exactly the moment you were checking.

If you think it has the clause wrong, quote the text at it. That is a new source, and a new source is precisely the thing that is allowed to change the answer.

What this does not do

• It does not predict what an auditor will decide. It tells you what the framework requires and where you stand against it. Two auditors can read the same evidence differently and neither of them works here.
• It is not legal advice. Where a question turns on law rather than on a framework, you get the compliance view and a line saying so.
• It cannot tell you about a framework you have not added. The catalogue is what it reads; nothing else is in scope, and it will say so rather than improvising.

<!-- Reviewed 2026-09-10 (the rulebook diet, workstream S7): the covered services/alchex-ai/src/agents/chat/system-prompt.ts lost three passages that repeat instructions the assistant already receives elsewhere in the same turn. The pushback paragraph this page describes — re-read rather than agree, never reverse without a clause — was deliberately NOT one of them: the check that enforces it names a measurement nobody has taken yet as the condition for dropping the words beside it, and an untaken measurement is not a reason. Everything this page promises holds as written. -->

See also

• AI review and assistance — how the assistant works generally, and what every AI change leaves behind.
• Templates — where these decisions come from, and what the label on a question card means while you are answering it.
• Standards knowledge — the requirements, guidance and regulations it reads, and how it keeps the three apart.