Practical guide

AI knowledge base article verification checklist

Use this checklist after an AI-assisted knowledge base article has been drafted and before publication. It tests whether a defined reader can complete one bounded task from current, approved evidence. It is not a writing prompt, a style score, or a substitute for product, policy, security, legal, accessibility, or support review. A named human owns every acceptance decision. The checklist is.

By Two Prune

Use this checklist after an AI-assisted knowledge base article has been drafted and before publication. It tests whether a defined reader can complete one bounded task from current, approved evidence. It is not a writing prompt, a style score, or a substitute for product, policy, security, legal, accessibility, or support review. A named human owns every acceptance decision. The checklist is deliberately evidence-based. A polished sentence does not pass because it sounds plausible. The reviewer must be able to show the source, reproduce the procedure, observe the expected result, and route a failure safely.

Set up the verification packet

Freeze the article revision and record its title, URL, audience, task, owner, review date, and source versions. Attach the current product or policy sources, permission matrix, interface references, known exceptions, and escalation route. Mark unresolved contradictions before testing. Do not ask an AI system to resolve authority by choosing the most confident wording.

Digital.gov recommends audience-focused plain language, structured and findable content, regular accuracy review, clear authority, accessible formats, and human validation of machine translation. This is U.S. federal digital-content guidance rather than a universal standard, but it provides a useful primary-source basis for checking whether task content is usable and maintained. Digital.gov: An introduction to content

Select at least two representative readers who did not author the article. Use a safe environment that matches the documented permissions and interface closely enough to test the procedure. Seed one expected failure, such as a missing control or validation error, so recovery instructions are tested rather than merely read.

Sources: [1]

The TRACE checklist

TRACE separates five verification decisions. Record pass, fail, not applicable, or blocked for every item, with an evidence link or observation. “Looks good” is not an acceptable result.

### T: Task and audience fit

Evidence to retain: reader task statement, audience definition, permission source, and a short findability test. Hold the article if the task actually contains several unrelated outcomes or if readers cannot tell which path applies.

### R: References and claim lineage

Evidence to retain: a claim-to-source table and the disposition of every contradiction. A missing source is not permission to generalize from memory.

### A: Actions, results, and recovery

Evidence to retain: test recording or notes, screenshots where permitted, observed results, defects, and repaired revision. If the documented result cannot be reproduced, hold publication until the source or procedure is corrected.

### C: Clarity and inclusive use

Digital.gov’s plain-language design guidance recommends putting key information first, using consistent heading levels, descriptive headings, lists for scanning, and tables for genuinely complex relationships. These checks support understanding but do not prove that every reader or accessibility need has been covered. Digital.gov: Design for understanding

Evidence to retain: accessibility review outcome, reader comprehension notes, and unresolved findings with owners. Do not convert an accommodation or assistive-technology need into a reader failure.

### E: Evidence, ownership, and expiry

Evidence to retain: test roster by role, environment, results, defect log, acceptance decision, owner, review date, and triggers. A page without a maintenance owner and expiry rule is a snapshot, not a controlled support artifact.

  • [ ] The title names one recognizable reader outcome rather than a broad topic.
  • [ ] The opening states who the article is for, the starting condition, prerequisites, and the intended result.
  • [ ] Exclusions and stop conditions appear before a reader can take an unsafe or irrelevant action.
  • [ ] Required permission is named using the current product or policy label.
  • [ ] The procedure does not assume that every person with a role label has the same authority.
  • [ ] Search terms and headings match language the intended reader is likely to use.
  • [ ] A reader can decide within the opening section whether the article applies.
  • [ ] Every consequential claim has a current approved source.
  • [ ] Product behavior, policy, permissions, deadlines, prices, data handling, and irreversible actions are traced separately.
  • [ ] Each source has an owner, version or verification date, and authority status.
  • [ ] Conflicting sources are escalated to the accountable owner rather than blended.
  • [ ] Examples are labeled as examples and do not silently become rules.
  • [ ] AI-generated additions are checked against the frozen packet; unsupported details are removed.
  • [ ] External links are clearly labeled and still resolve to the intended authoritative material.
  • [ ] Steps begin from the documented starting state and follow the reader’s actual sequence.
  • [ ] Each step contains one action, its expected visible result, and the next decision.
  • [ ] Interface labels match the tested version.
  • [ ] Warnings appear before the action they qualify.
  • [ ] Alternate paths exist only where evidence changes the action.
  • [ ] Likely failures include a safe check, a stop condition, and an escalation route.
  • [ ] The escalation packet requests only necessary information and never asks for secrets.
  • [ ] A reader can recover from the seeded failure without guessing.
  • [ ] Headings form a logical hierarchy and describe the decision below them.
  • [ ] Instructions use active voice and name the responsible actor.
  • [ ] Sentences and sections are short enough to scan without losing necessary context.
  • [ ] Uncommon terms and abbreviations are defined at first use.
  • [ ] Tables are used only when relationships are clearer than prose.
  • [ ] Links make sense out of context and do not rely on “click here.”
  • [ ] Meaning is not carried by color, position, or an image alone.
  • [ ] The article works with keyboard navigation and the organization’s required assistive-technology checks.
  • [ ] Any machine-translated version remains held until a qualified human review is complete.
  • [ ] At least two representative readers can find the article and determine whether it applies.
  • [ ] Readers complete the bounded task in the safe test environment.
  • [ ] The test includes the most consequential normal path and one recovery path.
  • [ ] Defects are repaired and retested against the same frozen acceptance conditions.
  • [ ] A publication owner accepts the residual risks.
  • [ ] Source owners, article owner, review date, and change triggers are recorded.
  • [ ] Triggers include material product, policy, permission, interface, escalation, or accessibility changes.
  • [ ] The evidence packet is stored with the published revision.

Sources: [2]

Copyable verification record

| Check | Result | Evidence | Defect or uncertainty | Owner | Retest date | | --- | --- | --- | --- | --- | --- | | Task applies to intended reader | pass / fail / blocked / N/A | task statement and reader test | note | name | date | | Consequential claims traced | pass / fail / blocked / N/A | claim-to-source map | note | name | date | | Normal path reproduced | pass / fail / blocked / N/A | observation | note | name | date | | Recovery path reproduced | pass / fail / blocked / N/A | seeded-failure result | note | name | date | | Accessibility checks completed | pass / fail / blocked / N/A | review record | note | name | date | | Ownership and expiry recorded | pass / fail / blocked / N/A | evidence card | note | name | date |

Do not collapse these rows into one percentage. A critical unsupported claim or unsafe recovery path cannot be averaged away by strong formatting.

Worked test: domain-change access article

An AI-assisted article tells workspace administrators how to restore access after an approved domain change. The reviewer freezes revision 7 and tests it with two administrators and one support agent.

Task: all three readers identify that the page applies only after the domain change has been approved. One ordinary member is used for a boundary test and correctly routes to an administrator rather than seeing privileged steps.

References: the claim map exposes an unsupported “allow 24 hours” statement and an old button label. The timing sentence is removed. The product owner confirms the current label and files a correction for the stale runbook. The article stays blocked until that authority decision is recorded.

Actions: the first administrator completes the normal path. The second receives the seeded validation error, performs the documented safe retry once, then escalates with the error text and timestamp. The draft initially requests an authentication code; the reviewer removes that request because support does not need it.

Clarity: the support agent cannot choose between two alternate paths. The author replaces the prose with a two-row decision table based on permission state. Keyboard review also finds that an image-only callout lacks equivalent text, which is repaired before retest.

Evidence: the retest passes the bounded normal and recovery paths. The product owner accepts publication, assigns a 60-day review, and adds triggers for permission, domain-flow, and interface-label changes. The decision is “accept revision 8 for this task,” not “all access articles are verified.”

Publication and hold rules

Publish only when critical claims trace to approved sources, representative readers can reproduce the task, recovery is safe, accessibility findings are resolved or explicitly dispositioned, and a named owner accepts maintenance. Hold when authority is disputed, permission is unknown, the test environment cannot represent the procedure, a serious failure has no safe route, unnecessary sensitive data is requested, or a translation lacks human review.

Two Prune can use this checklist as an artifact in a realistic assessment and describe how a participant frames a reader task, navigates evidence, uses AI, verifies claims, exercises judgment, and communicates instructions. It does not certify support quality, promise resolution, validate the checklist across every domain, or make employment decisions.

Sources

Digital.gov: An introduction to content

  • Digital.gov: Design for understanding

Sources: [1] [2]

Sources

  1. 1.Digital.gov: An introduction to content · U.S. General Services Administration
  2. 2.Digital.gov: Design for understanding · U.S. General Services Administration