Module 01 Activity

Scenario

Your team has been asked to put a question-answering assistant in front of the company's support policies. Before anyone builds anything, you have to say what it answers and what it must never attempt.

What you produce

A boundary document for a corpus you actually have access to - your team's runbooks, a public policy set, or your own notes.

Steps

1. Collect fifteen real questions people ask about that material. Support tickets, chat history, or a colleague's list - not questions you invent.
2. Sort each into one of the four kinds from Unit 03: documented fact, aggregate over records, judgement call, or never recorded. Count them.
3. For every question that is not a documented fact, write one sentence on what the assistant should do instead of answering.
4. Write the boundary document: what it answers from, what it never answers, whether every claim must be cited, the exact wording when nothing matches, and when it escalates.
5. Pick two entries from the never-answers list and describe the case you would use to check each one is honoured.

Evidence to hand in

• The fifteen questions with their kind, as a table.
• The completed boundary document.
• Two check cases drawn from the never-answers list.

Review checklist

• Every question came from a real source, not from imagination.
• The kind counts are reported, including how many are not retrieval-shaped.
• The no-match wording is fixed and exact, so a check could match it.
• The escalation list is non-empty and each entry could be recognised without judgement.
• The document says what proportion of real questions this assistant would decline.

Mark complete