Practical guide

AI policy brief verification checklist

A policy brief is ready only when a reviewer can trace its decision, evidence, options, recommendation, and residual uncertainty without reconstructing the author’s work. Use the TRACE checklist below before approval. It is an editorial control, not a substitute for policy expertise, legal review, consultation, or an organisation’s own clearance process. A failed item should produce a named.

By Two Prune

A policy brief is ready only when a reviewer can trace its decision, evidence, options, recommendation, and residual uncertainty without reconstructing the author’s work. Use the TRACE checklist below before approval. It is an editorial control, not a substitute for policy expertise, legal review, consultation, or an organisation’s own clearance process. A failed item should produce a named repair, owner, and recheck date rather than a vague request to “strengthen the brief.”

T — Target the decision and reader

Begin with the decision the brief is meant to support. Record the named decision-maker or decision forum, the decision date, the jurisdiction or operating context, and what is outside scope. The title and opening should let a busy reader distinguish the problem from a general topic.

The reviewer should be able to answer: “What could the reader do differently after reading this?” If the answer is merely “learn about the topic,” the brief still needs a sharper decision target.

  • State one decision as a verb: approve, choose, sequence, fund, pause, or investigate.
  • Name the affected population and material constraints.
  • Define the time horizon and the point at which evidence becomes stale.
  • Confirm that the opening explains why action or deliberate inaction is timely.
  • Remove background that does not change the decision.

R — Register every material source

Create a compact evidence register rather than relying on endnotes alone. For each material claim, capture the source owner, publication date, retrieval date, population, method, geography, and limitation. Separate direct evidence from author inference and stakeholder judgement.

The WHO evidence-use checklist is designed to support routine evidence use in policy-making. TRACE adapts that discipline to the pre-approval review of one brief; it does not claim that a checked brief is correct.

  • Trace every number, threshold, and causal statement to a source.
  • Recalculate derived figures from the cited inputs and preserve the working.
  • Mark evidence that is superseded, indirect, non-comparable, or incomplete.
  • Record meaningful contradictory evidence rather than averaging it away.
  • Check whether confidential, personal, or restricted material is handled under the appropriate policy.

Sources: [1]

A — Appraise options on the same basis

A useful brief makes alternatives comparable. Include the status quo or “do nothing now” position. Apply the same criteria to every option: expected benefit, affected groups, implementation burden, cost range, reversibility, evidence strength, risks, and dependencies.

If the evidence supports problem recognition but not option selection, the brief should recommend a bounded investigation or pilot rather than presenting false precision.

  • Verify that options are genuinely distinct rather than renamed versions of one proposal.
  • Test whether the preferred option receives more generous assumptions than the alternatives.
  • Identify the conditions under which a non-preferred option would become better.
  • Distinguish evidence about the problem from evidence that a particular response will work.
  • Flag distributional effects that an overall average would conceal.

C — Calibrate claims and recommendation

Read each consequential sentence as a claim. Label it descriptive, comparative, causal, forecast, or normative. Then ask whether the cited evidence supports that claim type and strength. Words such as “will,” “prevents,” and “causes” require stronger support than “is associated with” or “may help.”

The WHO policy brief template calls for a concise standalone brief with a clear problem, options, evidence, and actionable recommendations. TRACE adds a verification pass around those elements.

  • Match recommendation strength to evidence strength and decision reversibility.
  • Show material assumptions next to the recommendation, not only in an appendix.
  • Express ranges when point estimates imply unsupported precision.
  • Separate stakeholder preferences from empirical findings.
  • Include stop, review, or escalation conditions for uncertain recommendations.

Sources: [2]

E — Execute clearance and preserve the audit trail

Assign reviewers by risk, not prestige. A policy owner checks decision fit; a subject specialist checks domain meaning; an analyst checks evidence and calculations; legal, privacy, finance, accessibility, or communications reviewers join when the content triggers their remit.

A green checklist records what was checked and by whom. It is not a promise that every future consequence has been predicted.

  • Resolve comments with a disposition: accepted, partially accepted, or not accepted with reason.
  • Confirm links, attachments, tables, and labels survive the final export.
  • Record the approved version, approver, approval time, and next review date.
  • Preserve the evidence register and calculation file with the released brief.
  • Reopen review when a source, assumption, option, or implementation condition changes materially.

Worked example — adding a mobile-library route

A county team drafts a brief proposing a weekly mobile-library route to four rural settlements. TRACE exposes four problems. The decision target omits the annual operating ceiling and the date of the vehicle-replacement decision. The evidence register contains resident survey responses but no current travel-time or borrowing data. The options table compares the four-stop route only with no service, ignoring a two-stop pilot and scheduled collection lockers. The recommendation says the route will reduce digital exclusion although the local evidence establishes access barriers, not the effect of this particular service design.

The repair changes the decision to whether the county should run a twelve-week, two-stop pilot alongside a locker comparison. The team adds journey-time estimates, historic request data, vehicle and staffing availability, an accessibility review, and explicit evidence gaps for residents who did not answer the survey. The recommendation becomes conditional: proceed only if the vehicle, safeguarding plan, and accessible-stop design pass review; monitor visits, loans, unmet requests, service incidents, and cost per operating hour; review at weeks four and twelve; stop if vehicle or safeguarding controls fail. The approval record names library operations, finance, accessibility, and data-protection reviewers. TRACE does not decide the policy. It turns an overconfident brief into a bounded, reviewable decision artifact.

Copyable TRACE review record

For each item, record pass, repair, not applicable, or escalate. Add an owner and evidence link for every repair. The final record should contain the decision statement, evidence-register location, options table, claim-calibration notes, unresolved uncertainties, reviewer dispositions, approved version, review trigger, and next review date.

Two Prune assesses how a participant frames a task, navigates evidence, uses AI support, verifies outputs, exercises judgement, and communicates a deliverable in a realistic simulation. It does not author policy for the customer, certify a brief, or make employment decisions. Customers remain responsible for policy expertise, approvals, and real-world action.

Sources

  1. 1.WHO evidence-use checklist · www.who.int
  2. 2.WHO policy brief template · iris.who.int