Practical guide

What evidence shows high-impact use restrictions are operating?

Show restrictions operate through evidence: current rules, a use inventory, configured guards, decisions, blocked events, exceptions, monitoring, bypass tests, closure records, and change reviews together.

By Two Prune

Show restrictions operate through evidence: current rules, a use inventory, configured guards, decisions, blocked events, exceptions, monitoring, bypass tests, closure records, and change reviews together.

Define the control before testing it

State the restriction objective in observable terms. A useful form is: “Prevent or route defined AI uses with potentially material effects on health, safety, rights, access, or consequential decisions unless the approved authority, safeguards, and evidence conditions are present.”

Then define:

Do not copy a label such as “high impact” from one framework and treat it as universal. Legal classifications depend on jurisdiction, actors, intended purpose, and facts. The EU AI Act, for example, separately defines prohibited practices and classification rules for high-risk systems, with conditions and exceptions that must be read in their actual scope. EU AI Act, Regulation (EU) 2024/1689

  • the systems, models, workflows, users, geographies, and lifecycle stages covered;
  • the categories that are prohibited, restricted, approval-gated, or unresolved;
  • the authority that can classify a use, approve it, suspend it, or revoke an exception;
  • the inventory and event populations used to test operation;
  • the assessment period and review date.

Sources: [1]

Test a RESTRICT evidence chain

### R — Rules are current and attributable

Retain the approved restriction policy, source register, definitions, decision rights, version, effective date, and change history. Link each category to its source or internal risk decision. Record whether it is a legal requirement, contractual constraint, voluntary framework outcome, or organization-specific rule.

### E — Entire use population is reconciled

Build the source population independently from the restriction register. Combine model gateway records, procurement and vendor inventory, application registry, data-flow records, workflow catalogues, access groups, and change tickets. Reconcile every discovered use to a classification record.

The test fails if the policy register contains 18 uses but operational inventories reveal 21 and the three unmatched uses are left outside the sample.

### S — Scope decisions are reproducible

Each classification record should contain intended purpose, users, affected people, decisions influenced, data categories, deployment context, geography, system dependencies, reviewer, source version, outcome, rationale, and next review trigger.

NIST's AI RMF Core calls for targeted application scope, context, risks, human oversight, and internal controls to be defined and documented. Use those outcomes to challenge vague classifications, while recognizing that the voluntary framework does not determine legal status. NIST AI RMF Core

### T — Technical and workflow restrictions match the rule

Trace the rule into actual implementation:

A policy statement is design evidence. It is not proof that the restriction executes.

### R — Records show attempted uses and decisions

Sample both approved and denied activity. Retain request, classification, authority, evidence reviewed, decision, conditions, implementation, timestamp, and requester notification. Include incidents where a blocked use was attempted and cases where the restriction should have triggered but did not.

### I — Independent tests challenge bypasses

Construct tests from the underlying workflow rather than asking owners to nominate good examples. Attempt representative routes through direct interfaces, integrations, bulk imports, service accounts, vendor tools, copied data, and changed configurations. Verify the expected block, approval request, or escalation and retain the event record.

### C — Conditions and exceptions remain controlled

For every exception, require scope, accountable owner, rationale, compensating measures, monitoring, expiry, revocation trigger, and closure evidence. Reconcile active exceptions to technical settings and observed events. An expired exception with continuing access is a control failure even if no harmful event is known.

### T — Trends, changes, and closure are reviewed

Review unmatched uses, attempted bypasses, classification changes, false blocks, incidents, expiring exceptions, control changes, and overdue remediation. Reperform affected tests after material policy, model, system, vendor, or legal changes.

The NIST AI RMF Playbook offers adaptable actions across Govern, Map, Measure, and Manage and explicitly is not a one-size-fits-all checklist. Select relevant practices for the context, document the choice, and test the resulting control rather than claiming framework alignment from a completed worksheet. NIST AI RMF Playbook

  • access groups and role permissions;
  • model or tool allowlists and denylists;
  • prohibited data filters;
  • approval gates and release checks;
  • environment separation;
  • user notices and escalation routes;
  • logging and alert configuration;
  • vendor and procurement controls.

Sources: [2] [3]

Worked evidence artifact: restricted-use review

A synthetic organization prohibits workplace emotion inference, requires enhanced review for AI that materially influences access to essential services, and requires ordinary approval for internal drafting tools. The review period is one quarter.

| Evidence layer | Population or artifact | Test | Result | | --- | --- | --- | --- | | Rule | Policy v4 and source register | Confirm category, scope, authority, effective date, and change approval | Pass | | Inventory | 27 discovered AI-enabled uses | Reconcile gateway, vendor, application, and procurement records to classification register | Two unmatched uses | | Classification | 25 existing records plus two new reviews | Reperform five high-impact and five ordinary-use classifications | One rationale lacks affected-person analysis | | Configuration | Access groups, release gates, model filters | Trace all restricted categories to settings and approvals | One legacy service account bypass | | Events | 14 blocked attempts, 9 approvals, 3 exceptions | Match events to decisions and requester notification | One blocked event absent from case register | | Exceptions | Three active records | Check scope, monitoring, expiry, revocation, and technical alignment | One exception expired while access remained active | | Independent test | Eight planned bypass routes | Exercise interfaces, API integration, vendor upload, and service account | Seven behave as expected; legacy account fails |

The reviewer records four linked findings: two inventory gaps, one incomplete classification, one unregistered blocked event, and one expired exception with live access. The legacy account is disabled immediately, the affected event population is expanded, and the organization repeats the bypass and reconciliation tests after remediation.

The conclusion is not “restrictions effective.” It is: “For the quarter and tested scope, seven of eight independent routes produced the expected block or gate; material exceptions remain open for inventory reconciliation, classification rationale, event registration, and expired access.” That wording preserves the observed evidence and its limits.

Reject weak evidence

Do not accept these as proof of operation on their own:

AI can help compare inventories, detect missing fields, draft test cases, and summarize event records. Accountable people must set the rules, decide applicability, authorize restrictions and exceptions, evaluate evidence, and close findings. Two Prune may assess a participant's governance reasoning in a synthetic work simulation; it does not provide legal advice, operate customer controls, certify compliance, or make employment decisions.

  • an approved policy without a reconciled use population;
  • training completion without observed restriction events;
  • a clean exception register built only from self-reported exceptions;
  • screenshots without controlled configuration exports or change history;
  • zero incidents without evidence that detection and logging work;
  • a single successful block without alternate-route testing;
  • a vendor assurance that does not cover the organization's configured use;
  • a legal label assigned without qualified, current jurisdiction-specific review.

Sources

European Union: Regulation (EU) 2024/1689 (AI Act)

  • NIST: AI RMF Core
  • NIST: AI RMF Playbook
  • NIST: AI RMF Playbook FAQ

Sources: [1] [2] [3] [4]

Sources

  1. 1.EU AI Act, Regulation (EU) 2024/1689 · eur-lex.europa.eu
  2. 2.NIST AI RMF Core · National Institute of Standards and Technology
  3. 3.NIST AI RMF Playbook · National Institute of Standards and Technology
  4. 4.NIST: AI RMF Playbook FAQ · National Institute of Standards and Technology