Practical guide
AI compliance matrix verification checklist
Verify a compliance matrix as evidence: test its scope, sources, applicability, status, proof, owners, exceptions, claims, and release controls before operational use in any decision.
Verify a compliance matrix as evidence: test its scope, sources, applicability, status, proof, owners, exceptions, claims, and release controls before operational use in any decision.
Start with the matrix's declared job
Record the decision the matrix is meant to support, the system or workflow in scope, jurisdictions, business units, lifecycle stage, source editions, assessment date, and named reviewer. A matrix built for pilot readiness is not automatically suitable for a legal opinion, certification, audit conclusion, or production approval.
The source set must match that declared job. NIST describes the AI RMF Playbook as voluntary, adaptable guidance rather than a universal checklist or ordered implementation sequence. If a matrix turns every Playbook suggestion into a mandatory requirement, it has already lost source fidelity. NIST AI RMF Playbook FAQ
Sources: [1]
Run the TRACE-PROOF checklist
Apply each check to the whole matrix and to a sample of individual rows.
### T — Target and boundaries
Fail this check when readers could reuse the matrix for a broader decision without noticing that its evidence was collected for a narrower one.
### R — Reference fidelity
Record the sample and result. “Links reviewed” is not reproducible evidence.
### A — Applicability reasoning
AI can extract candidate text. It cannot establish jurisdiction, contractual coverage, or legal applicability merely from wording similarity.
### C — Control and action mapping
NIST's AI RMF Core calls for documentation that supports relevant actors' decisions, documented application scope, human oversight processes, and identified internal controls. Those outcomes support explicit mappings; they do not prove that a mapped control operates. NIST AI RMF Core
### E — Evidence sufficiency
The U.S. Government Accountability Office's AI Accountability Framework links governance, data, performance, and monitoring principles to questions and audit procedures. Use that structure to challenge self-attestation and request reviewable evidence, without presenting the matrix as a GAO audit. GAO AI Accountability Framework
### P — Population and sampling
A sample drawn only from rows already labelled complete cannot reveal omitted requirements or silent bypasses.
### R — Review and contradiction handling
### O — Ownership, approval, and expiry
### O — Output and claims
### F — Final release package
- Does the header name the matrix purpose and decision?
- Are the system, process, owner, geography, population, and dates explicit?
- Are exclusions and unresolved scope questions visible?
- Does each source record its edition, effective date, retrieval date, and refresh trigger?
- Does every row resolve to a stable source and section?
- Does the quoted text or bounded summary preserve conditions, exceptions, and cross-references?
- Are legal requirements, contractual terms, voluntary guidance, internal rules, and reviewer interpretations labelled separately?
- Has a reviewer checked samples in both directions: source to matrix for omissions, and matrix to source for unsupported rows?
- Is every row marked in scope, out of scope, conditional, or unresolved?
- Does the row state why, using the facts of the declared context?
- Is the applicability reviewer qualified for that source and decision?
- Are unresolved interpretations routed to a named owner rather than silently marked not applicable?
- Does the row translate the source statement into a specific process, decision, control, or evidence request?
- Is the accountable owner different from a person who merely contributes?
- Is the implementation status separate from the quality of proof?
- Are dependencies, compensating measures, and exceptions visible?
- Does each “implemented” or “effective” statement link to an actual controlled artifact?
- Is the evidence contemporaneous with the assessed period?
- Does it show operation, not only design or intention?
- Can a reviewer repeat the row's evidence test and reach the recorded result?
- Are contrary results, missing records, and residual uncertainty retained?
- Is the source population defined independently of the matrix statuses?
- Does the reviewer reconcile that population to matrix rows?
- Is the sampling method, sample size, period, and selection rationale recorded?
- Were both apparently compliant and non-compliant cases tested?
- Has a second reviewer challenged material rows?
- Are duplicate or conflicting sources resolved without deleting the disagreement trail?
- Are calculation formulas, scoring rules, and status roll-ups tested?
- Does a single failed material row remain visible rather than disappearing inside an overall percentage?
- Does every open action have one accountable owner and due date?
- Does the matrix show author, reviewer, approver, version, and approval state?
- Are time-limited exceptions given expiry and revocation conditions?
- Is the next review triggered by both a date and material source or system change?
- Does the summary distinguish design, implementation, operation, and tested effectiveness?
- Are percentages accompanied by denominators, exclusions, and weighting rules?
- Are claims confined to the tested scope and period?
- Does the output avoid “certified,” “compliant,” or “effective” unless the required qualified review and operating evidence actually support that wording?
- Is the reviewed matrix locked or versioned?
- Are sources, evidence links, test notes, exceptions, and approvals retained together?
- Can a new reviewer reproduce material conclusions without relying on oral explanation?
- Does the release note list unresolved items and prohibited downstream uses?
Sources: [2] [3]
Worked artifact: defect log for a synthetic matrix
A programme team reviews a 42-row matrix for an internal AI support assistant. It selects all four rows marked “complete,” five conditional rows, and five rows sampled directly from the source index.
| Check | Observed defect | Disposition | Release effect | | --- | --- | --- | --- | | Reference fidelity | One row cites a framework category but omits the limiting subcategory text | Rewrite citation and interpretation; repeat reverse trace | Material hold | | Applicability | A privacy row is marked out of scope with no jurisdiction or reviewer | Route to privacy owner; status becomes unresolved | Material hold | | Evidence sufficiency | “Human review implemented” links only to a draft procedure | Change implementation to planned; request three operating records | Status correction | | Population | Matrix lists 11 approved uses; the system inventory contains 13 | Reconcile two unmatched uses and test their controls | Material hold | | Expiry | Temporary vendor exception has no expiry or revocation trigger | Add expiry, owner, monitoring, and closure evidence | Conditional hold | | Output claim | Summary reports “93% compliant” using rows with unequal materiality | Remove percentage; report tested statuses and gaps | Claim correction |
The reviewer does not “fix” these defects by changing cells to green. The artifact is released only after material holds are resolved or explicitly accepted by authorized owners, with unresolved legal or regulatory questions routed to qualified review. The example is synthetic and does not demonstrate any customer's compliance or control effectiveness.
Record the verification result
Save the completed checklist with matrix version, reviewer, date, sample, defects, dispositions, unresolved issues, and approval state. A downloadable checklist is useful only if its execution leaves evidence.
AI may help extract source text, compare versions, find missing fields, and format the defect log. People remain responsible for source selection, applicability, evidence sufficiency, acceptance of residual risk, and release claims. Two Prune may assess how a participant performs this evidence work in a realistic simulation; it does not provide legal advice, certify compliance, operate customer controls, or make employment decisions.
Sources
NIST: AI RMF Playbook
- NIST: AI RMF Playbook FAQ
- NIST: AI RMF Core
- U.S. GAO: Artificial Intelligence Accountability Framework
Sources: [4] [1] [2] [3]
Sources
- 1.NIST AI RMF Playbook FAQ · National Institute of Standards and Technology
- 2.NIST AI RMF Core · National Institute of Standards and Technology
- 3.GAO AI Accountability Framework · U.S. Government Accountability Office
- 4.NIST: AI RMF Playbook · National Institute of Standards and Technology