How to develop and assess iterative refinement in AI-enabled work

Set a stopping standard before revising

Define the audience, decision, required evidence, constraints, output form, and unacceptable failures. Then translate those requirements into a small acceptance table. Without a stopping standard, a person can keep prompting until the prose sounds different while the underlying error remains.

The NIST TEVV-Athlon Framework page describes an August 2026 initial public draft intended to distinguish and organize testing, evaluation, verification, and validation across varied AI applications. Its draft status matters. It supports selecting a check that fits the question, not treating repeated generation as evaluation.

The UK government's AI Playbook advises teams to understand system limitations, test outputs, retain meaningful human control, and use evidence throughout the lifecycle. The playbook is guidance for UK government organizations. It does not establish a universal measure of individual skill.

Sources: [1] [2]

Run the diagnose-change-check loop

Use four fields for every material revision:

Preserve a version only when it improves the deliverable. Record regressions as well as gains. A revision that fixes tone while deleting a key limitation is not an improvement.

• Diagnose: identify the exact defect and the requirement it violates.
• Change: choose the smallest intervention capable of resolving it, such as adding a missing source, changing a calculation, narrowing a claim, restructuring an option, or asking for a counterexample.
• Check: test the affected item against the original evidence and acceptance standard.
• Dispose: accept the revision, revise again, restore an earlier version, hold the item, or escalate it.

Worked artifact: a synthetic budget recommendation

A fictional programme analyst asks an AI assistant to recommend which of three internal pilots should receive next quarter's budget. The first draft recommends Pilot B because it has the highest satisfaction score and the lowest reported cost.

The analyst's acceptance table requires comparable cost periods, observed operational evidence, explicit uncertainty, and an option that fits the approved budget ceiling. The refinement ledger records:

| Version | Diagnosis | Targeted change | Check and disposition | | --- | --- | --- | --- | | V1 | Pilot B cost covers two months; A and C cover six | Normalize cost period from finance records | B is no longer lowest cost; revise | | V2 | Satisfaction is treated as evidence of work quality | Relabel the survey and add observed error-review data | Confidence claim removed; accept evidence section | | V3 | Recommendation exceeds the ceiling after support cost | Ask for a constrained option using verified totals | Produces phased B option; recalculate | | V4 | Phased option omits vendor exit dependency | Add contract exit condition and downside case | Dependency visible; escalate contract point |

The final brief does not claim that Pilot B is proven superior. It recommends a bounded phase with a cost ceiling, names the unresolved contract question, and identifies what would trigger a stop. The example is synthetic and does not establish that this revision sequence fits a real investment decision.

Check for false progress

Compare each version with the acceptance table, not only with the preceding version. Look for four common forms of false progress:

Use an independent calculation, source trace, counterexample, or alternative method when the decision warrants it. Ask what evidence would reverse the recommendation. Stop when the acceptance criteria are met, remaining uncertainty is visible, and further revision would not materially improve the decision. Escalate when the required evidence, authority, or specialist judgment is unavailable.

• surface change: words or structure change while the unsupported conclusion stays;
• evidence drift: a new claim appears without a source or a source is stretched beyond its scope;
• constraint loss: a revision improves one criterion by dropping a limit, audience need, or required caveat;
• regression: a resolved error returns after a later rewrite.

Develop and assess the behavior

Practice should include an initially useful output with a small number of material defects. Ask the learner to submit the ledger, intermediate versions, and final artifact. Feedback should address whether the diagnosis was specific, the intervention matched the failure, the check could detect the problem, and the stopping decision was proportionate.

A bounded assessment can observe version control, evidence use, regression detection, constraint preservation, and escalation. One successful revision does not prove a general capability. Report the task, opportunity, observed evidence, and limitations.

Two Prune can create a controlled simulation in which refinement behavior is visible for human interpretation. It does not validate an AI model, promise error-free work, establish mature workforce benchmarks, or make employment decisions.