Build a One-Document FAQ That Shows Its Evidence

Can a small team make a useful FAQ without blending pages or hiding what the source never says? Yes: use one approved public document and an answer-level evidence ledger.

This tutorial proposes, not reports, a completed FAQ. The document owner must select and approve reuse; that approval is hypothetical here. Use DOC-1 as a neutral label.

Current guidance treats questions as design, not headings. GOV.UK guidance, updated June 24, 2026, says to ask understandable questions and allow “I’m not sure” or “I do not know.” Designing good questions NIST says high-integrity information distinguishes fact from inference, acknowledges uncertainty, links original sources, and states expiry. NIST AI 600-1

Materials and roles

Prepare one owner-approved public policy, help page, or guidance PDF; exact URL; permitted text copy or snapshot; five to eight questions; an evidence sheet; timer; and reviewer who can hold publication. The source owner reviews the handoff.

Record title, URL, visible publication or update date, access date, and DOC-1. A snapshot can reopen the input. This is a one-time record, not repeated fetching or version comparison.

00:00–00:05 — Lock the source boundary

Write “Answers may use DOC-1 only” at the top. Note headings, pages, tables, footnotes, exceptions, and explicit omissions. Do not add a related article. If the document is inaccessible, unapproved, or too long for the time box, stop and choose a narrower accessible official source with owner approval.

Write one user need beside each question. Split “Who is eligible, what does it cost, and can I appeal?” into separate questions. Do not smuggle in an assumption such as “Why does the policy guarantee ...?” when the document may state no guarantee.

00:05–00:12 — Make the question set answerable

List five to eight plain-language questions, one fact or action each, including one boundary question that may expose an omission. Examples: “What does the document say about access?” “Which exception is named?” “Does it state a response time?”

Mark each answer expected, partial possible, or unknown acceptable before asking AI. If the owner cannot explain why a question matters, move it to the future-question queue.

00:12–00:24 — Draft from DOC-1, with evidence fields required

Give the model only the approved document and the question list. Use an instruction like this:

Use DOC-1 and no outside knowledge. For each question, return question_id, status (answered, partial, or unknown), a concise answer, one locator (heading plus page or paragraph), a short exact excerpt for answered or partial rows, a caveat, and a follow-up owner. For an unknown row, leave the excerpt empty and return a coverage note listing the sections checked. If DOC-1 is silent, say so and do not infer a policy, deadline, exception, or guarantee. Preserve “may,” “usually,” “subject to,” and “not specified.”

Do not ask for a final page yet. Required evidence fields expose unsupported confidence. A structured response can enforce fields and statuses, but still inspect the evidence: structure is a filing aid, not proof.

00:24–00:31 — Review every answer against the page

The reviewer opens DOC-1 and checks each row. The locator must reach the relevant heading, page, paragraph, table cell, or footnote; the excerpt must be exact and short; scope and qualifiers must survive. A generic locator fails even if page one appears to answer.

Keep negative evidence honest. “The source does not state a retention period” is an unknown supported by a coverage note such as “headings 1–6 and table A checked,” not a made-up quotation. An unknown row should have no excerpt. If a sentence is close but narrower, use partial with an exact excerpt and explain the gap.

Edge case: suppose a synthetic training document says “requests are usually handled within five business days,” while the question asks, “Is five days guaranteed?” Retain “usually” and say DOC-1 does not state a guarantee. This wording is illustrative, not a real policy.

00:31–00:36 — Package the handoff

Deliver four artifacts together: readable FAQ, evidence sheet, source record, and unknowns queue. Keep the source record beside the FAQ. NIST lists sources, dates, modifications, and history as provenance; here record a locator and source check for every answer.

Use this result sheet:

The handoff is complete when every answer has a status, each answered or partial row has a locator and excerpt, each unknown has a coverage note and reason, and a reviewer has named follow-up ownership. Do not turn the unknowns queue into a second source packet.

00:36–00:40 — Apply the failure rule and choose the next decision

Fail answered or partial rows lacking a locator, with inexact excerpts, outside claims, or lost qualifiers. Fail unknown rows lacking coverage notes, containing guesses, or lacking reproducible search boundaries. Also fail questions asking for decisions DOC-1 does not own or unapproved sources. Mark hold for source owner.

If every row passes, decide on editorial review of wording, accessibility, and destination. If the source changes, reopen its record and decide whether to refresh; no freshness claim extends beyond the checked date.

The contribution is a four-part handoff: one source boundary, answer-level evidence, an unknowns queue, and a named reviewer decision. It makes a compact FAQ useful without treating silence as fact. No model, approval, publication, or live result is claimed.

Sources and limitations

All sources below were checked September 2, 2026.

1. GOV.UK Service Manual, “Designing good questions” — supports understandable one-purpose questions, valid “I’m not sure” or “I do not know” responses, and the page’s June 24, 2026 update. Limitation: this is government service-question guidance, not a universal FAQ template or evidence that a particular public document is approved.
2. National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile” — supports distinguishing fact, opinion, and inference; acknowledging uncertainty; linking original sources; and recording provenance such as dates, modifications, and sources. Limitation: the profile is voluntary risk-management guidance, published July 2024, not a required FAQ production standard.