Practical guide
AI Capability Programme Verification Checklist
Use this checklist to decide whether an AI capability programme is ready to launch, continue, or revise. It verifies the programme specification and its evidence trail; it is not a substitute for running the programme or observing later workplace performance. Complete it against one named programme version and attach the cited artifacts.
Use this checklist to decide whether an AI capability programme is ready to launch, continue, or revise. It verifies the programme specification and its evidence trail; it is not a substitute for running the programme or observing later workplace performance. Complete it against one named programme version and attach the cited artifacts.
Set the verification record
Record the programme name, version, owner, reviewer, review date, covered workflows, participant group, AI systems, and next review trigger. Link the approved programme specification rather than a presentation that merely describes it. List exclusions explicitly, including workflows, locations, or systems outside scope.
NIST's AI RMF Core calls for documented roles, responsibilities, communication lines, duty-enabling training, and periodic review. Those outcomes make useful verification headings, but they do not turn a checklist into evidence that a programme works. NIST AI RMF Core
Mark each item met, partly met, not met, or not applicable. Every “met” answer needs an artifact reference, version, and owner. A statement such as “covered in training” is not evidence unless the reviewer can locate the relevant material and connect it to the workflow.
Sources: [1]
Verify purpose, work, and ownership
The U.S. Office of Personnel Management advises using needs assessment to determine what must be done and why it is not being done, noting that training is not always the best or only intervention. Use that principle to reject programmes that treat every performance gap as a course request. OPM planning and evaluating guidance
- [ ] The programme names the business tasks and decisions it supports.
- [ ] Each priority workflow has an observable capability contract: inputs, allowed AI assistance, human decisions, expected artifact, checks, principal failure, and escalation route.
- [ ] A needs assessment distinguishes skill gaps from access, policy, process, workload, or management problems.
- [ ] Scope includes the participant group, systems, permitted information, source requirements, and material error consequences.
- [ ] One accountable owner is named for programme scope, source material, practice cases, assessment, operational support, and revisions.
- [ ] Reviewers and escalation contacts have accepted their responsibilities.
Sources: [2]
Verify learning, practice, and support
For every checked item, capture the artifact ID and a short verification note. For example: SIM-03 v2.1 — conflicting policy sources present; expected escalation recorded in scoring guide. Do not accept a blank template as completed evidence.
- [ ] Orientation covers permitted use, data handling, model limitations, accountability, and escalation.
- [ ] Demonstrations show a complete task, including a rejected or corrected AI output.
- [ ] Practice varies ambiguity, evidence quality, and failure conditions rather than repeating one ideal case.
- [ ] At least one realistic simulation requires a work artifact under stated constraints.
- [ ] Participants can access current sources, systems, checklists, and review routes.
- [ ] Facilitator and participant instructions use the same programme version.
- [ ] Accessibility and reasonable adjustment needs have an owner and route.
Verify evidence and claims
OPM's Training Evaluation Field Guide recommends aligning evaluation to organizational purpose, critical behaviours, and workplace support, and planning evaluation before delivery. Use it to test the evaluation design, not to certify a particular programme. OPM Training Evaluation Field Guide
- [ ] Evaluation questions were written before delivery and map to programme decisions.
- [ ] Attendance, knowledge checks, task observations, telemetry, workplace review, and operating outcomes remain separate evidence layers.
- [ ] The assessment method defines observable evidence, scoring rules, reviewer guidance, and missing-data treatment.
- [ ] Calculations can be reproduced from retained inputs.
- [ ] Claims state their population, task, conditions, time window, and limitations.
- [ ] Contradictory or incomplete evidence is retained and explained.
- [ ] The programme does not infer capability from attendance alone.
- [ ] The programme does not present a bounded assessment as proof of workplace transfer, control operation, or business impact.
Sources: [3]
Verify change, approval, and residual uncertainty
[ ] Source documents and programme artifacts have version control.
- [ ] Material changes identify an initiator, impact review, approver, effective date, and affected artifacts.
- [ ] Release criteria and stop conditions are written before launch.
- [ ] Exceptions have an owner, expiry date, and review route.
- [ ] The next review is triggered by relevant system, policy, task, evidence, or failure-pattern changes.
- [ ] Residual uncertainties are recorded with the decision they constrain.
- [ ] Final approval names who accepted scope and unresolved risks.
Worked verification: customer-guidance programme
A reviewer applies the checklist to programme version 2.1 for 24 service advisers. The programme has three workflow contracts, a policy-source exercise, two practice cases, and a synthetic assessment. Ownership and learning-design items pass. The reviewer finds that the scoring guide refers to policy library version 4 while the participant pack links version 5, and that two missing event streams would be scored as poor behaviour rather than unavailable evidence.
The reviewer marks source consistency and missing-data treatment “not met,” records the conflicting artifact IDs, and blocks launch. The programme owner aligns the policy versions, changes the scoring rule, and supplies a two-case test showing that a complete stream produces a workflow observation while a missing stream produces a confidence limitation. The second review passes those items. The record still notes that the simulation does not establish later workplace transfer.
Make the disposition reproducible
Summarize totals by status, but base the decision on materiality rather than a percentage. One missing approval or invalid assessment rule may matter more than several minor documentation gaps. Choose ready, ready with time-bound actions, revise before launch, or stop and redesign. Name each blocking item, owner, due date, evidence needed to close it, and the version that must be rechecked.
Two Prune can provide a controlled work simulation and structured evidence about observed performance in that task. It does not operate the full capability programme, certify AI literacy or legal compliance, prove transfer or operating outcomes, or make employment decisions.
Sources
- 1.NIST AI RMF Core · National Institute of Standards and Technology
- 2.OPM planning and evaluating guidance · U.S. Office of Personnel Management
- 3.OPM Training Evaluation Field Guide · U.S. Office of Personnel Management