Practical guide

What evidence shows AI quality assurance is operating?

An AI quality-assurance policy is not operating evidence. Evidence must connect a versioned requirement to an executed check, a reproducible result, an exception and its disposition, and ongoing monitoring in the real workflow. A collection of screenshots, test counts, or approval signatures without that chain cannot show what was tested or what decision followed.

By Two Prune

An AI quality-assurance policy is not operating evidence. Evidence must connect a versioned requirement to an executed check, a reproducible result, an exception and its disposition, and ongoing monitoring in the real workflow. A collection of screenshots, test counts, or approval signatures without that chain cannot show what was tested or what decision followed.

Define the control objective and scope

Write a bounded objective such as: “Released AI-assisted customer responses meet the approved accuracy, source, privacy, and escalation criteria for the support workflow.” Name the system or workflow version, users, deployment setting, data classes, excluded uses, accountable owner, independent reviewer, and review date.

Quality criteria must come from the intended use and consequences. A grammar score cannot substitute for factual accuracy; a model benchmark cannot substitute for a workflow acceptance test; and a sampled pass cannot establish performance outside the sample or version.

NIST’s AI Risk Management Framework Core says AI systems should be tested before deployment and regularly during operation. Its Measure function calls for documented test sets, metrics, tools, deployment-relevant conditions, independent review, uncertainty, production monitoring, and evaluation of the measurement process itself. These outcomes support an evidence chain, but adoption of the framework alone does not establish that a control operates. NIST AI RMF Core

Sources: [1]

Build the RATED evidence chain

RATED links the assurance requirement to the release or operating decision.

### R: Requirement and risk

Record each acceptance criterion, why it matters, the failure consequence, and its source. Distinguish product requirements, workflow policy, legal or contractual obligations, and local risk decisions. Assign severity and define any zero-tolerance condition before testing.

### A: Assessment design

Document the method, test population, source data, expected result, comparison, reviewer, sampling rule, and uncertainty. Include ordinary, boundary, adversarial, and recovery cases where relevant. Preserve version identifiers for the model, prompts, retrieval sources, tools, policies, and test assets.

### T: Test execution

Capture who or what ran the check, when, in which environment, with which configuration, and whether the procedure deviated. Retain enough input, output, calculation, and trace evidence to reproduce a material finding without unnecessarily retaining sensitive content.

### E: Exceptions and evaluation

Record every material failure, missing result, reviewer disagreement, and out-of-scope case. Link it to severity, affected versions, root-cause status, compensating measure, owner, due date, and retest. Missing evidence is not a pass. A changed criterion requires a visible rationale and fresh approval rather than silent threshold movement.

### D: Disposition and durability

Connect results to release, repair, hold, stop, rollback, or restricted-use decisions. Name the authority and residual uncertainty. Add production indicators, complaint or incident routes, drift or change triggers, sample cadence, and the date the assurance conclusion expires.

Minimum operating evidence packet

| Artifact | What it must show | Failure signal | | --- | --- | --- | | requirement register | criterion, source, severity, owner, version | vague or retrofitted acceptance rule | | test plan | representative and challenge cases, method, threshold | convenience sample or missing consequence | | execution record | environment, configuration, input set, result, trace | irreproducible screenshot or aggregate only | | exception register | defects, missingness, disagreement, remediation | deleted failures or no owner | | decision record | disposition, authority, rationale, residual risk | approval detached from results | | monitoring record | live indicators, incidents, changes, review date | release treated as permanent assurance |

The U.S. Government Accountability Office’s AI Accountability Framework organizes practices across governance, data, performance, and monitoring. Its questions for managers and assessors reinforce that performance evidence depends on documented objectives, relevant data, evaluation, human oversight, and continued monitoring. Use it as an accountability reference, not as a certificate for a particular system. GAO AI Accountability Framework

Sources: [2]

Worked example: AI-assisted support responses

This synthetic workflow drafts customer-support responses from an approved knowledge base. The quality objective requires correct source use, no invented account facts, appropriate privacy handling, and escalation when the knowledge base does not resolve the issue. The release owner is the support operations lead; a quality reviewer independent of the prompt author samples results; the security owner reviews data handling.

Requirement: the team defines four acceptance criteria and treats disclosure of another customer’s information as a stop condition. Unsupported policy statements are high severity. Style preferences are lower severity and cannot offset a factual failure.

Assessment design: the frozen test set contains routine questions, outdated articles, conflicting articles, an account-specific request, an unanswerable question, and a prompt-injection attempt embedded in retrieved text. Each case has an expected evidence path and escalation rule. The record identifies the model, retrieval index, prompt, tool permissions, and knowledge-base snapshot.

Test execution: two reviewers independently inspect forty cases. Thirty-six meet all criteria. One response uses an outdated refund rule, one invents an account status, one fails to escalate an unanswerable request, and one is correctly blocked after the injected instruction requests an external action. The aggregate “90% pass” is recorded, but the three material defects remain visible and control the decision.

Exceptions: the invented account status traces to a prompt that invites completion when evidence is missing. The refund error traces to stale retrieval metadata. Reviewer disagreement on the escalation case is adjudicated against the written criterion. Owners, repairs, affected versions, and retests are recorded.

Disposition: the owner chooses hold for autonomous sending but permits a restricted drafting mode with mandatory human review. Release requires passing the repaired challenge cases, a fresh representative sample, and confirmation that stale articles are excluded. Monitoring then tracks overrides, complaints, source failures, and unexpected external actions. The conclusion expires on any material model, prompt, retrieval, tool, or policy change.

Test the evidence itself

An independent reviewer should select a critical requirement and reconstruct the path from source to test, execution, exception, decision, and monitoring. They should reproduce key calculations, verify the tested version, inspect excluded cases, and confirm that failures remain in the denominator. Expand the review if any result lacks provenance or an approval predates the evidence.

Useful evidence supports three different statements:

Do not collapse these into a claim that the workflow is universally effective. Real operating evidence accumulates over time and can also show that a control failed.

  • the assurance procedure is designed;
  • the procedure was executed for the recorded scope and version;
  • results support the stated disposition under tested conditions.

Hold conditions and claims boundary

Hold release or the assurance conclusion when requirements are undefined, the tested version is unclear, data is unrepresentative, critical cases are absent, execution cannot be reproduced, failures disappear from reporting, the reviewer lacks independence for the stated risk, exceptions have no owners, or production monitoring is missing.

Two Prune can include quality-assurance work in a realistic AI-enabled assessment and evaluate how a participant defines criteria, tests outputs, handles exceptions, verifies evidence, and communicates a disposition. It does not operate or certify an organization’s quality controls and does not make employment decisions.

Sources

National Institute of Standards and Technology: AI Risk Management Framework Core

  • U.S. Government Accountability Office: Artificial Intelligence Accountability Framework

Sources: [1] [2]

Sources

  1. 1.NIST AI RMF Core · National Institute of Standards and Technology
  2. 2.GAO AI Accountability Framework · U.S. Government Accountability Office