Revision quality in AI-enabled work: improve the artifact without losing its evidence

Diagnose the revision trigger

Start by classifying why the artifact needs revision. Common triggers include a factual error, missing evidence, unclear explanation, changed requirement, contradictory source, unusable structure, reader misunderstanding, or new risk. Convert each trigger into a testable issue statement.

For example, “make the recommendation stronger” is ambiguous. The problem could be weak evidence, an unstated decision criterion, buried uncertainty, or merely timid wording. Ask what the reader could not decide or do, which passage caused the problem, and what evidence would show it is resolved.

Digital.gov's guidance on testing for understanding recommends testing early, correcting content from feedback, and testing again. It describes paraphrase testing for shorter content and usability testing when finding information matters. These are useful models for workplace artifacts: test the reader's understanding or action rather than relying only on the author's judgment that a revision “reads better.”

Sources: [1]

Use the TRACE revision ledger

Record six fields for every material revision:

The ledger prevents two common failure modes. The first is mechanical compliance, where the writer accepts wording without checking whether it is accurate. The second is invisible drift, where repeated AI rewrites alter scope, numbers, qualifications, or ownership while making the text sound more confident.

The UK Cabinet Office's periodic-review checklist for functional standards uses a master review sheet, explicit dispositions such as accept, modify, reject, note, or defer, clear versioning, and recorded rationale. That process is designed for government standards, not every workplace draft. Its transferable lesson is that consequential feedback should receive a visible disposition and rationale before text is changed.

• Trigger: What feedback, evidence, test result, or requirement initiated the change?
• Requested outcome: What should the reader understand, decide, verify, or do after revision?
• Analysis: What is the underlying problem, and what valid content must be preserved?
• Change: Which passage, calculation, structure, source, or artifact changed?
• Evidence: What source or test supports the new version?
• Effect check: How will the team confirm that the revision solved the issue without creating another one?

Sources: [2]

Worked artifact: revising a synthetic programme brief

A fictional learning team prepares a brief about a small AI-workshop pilot. The first draft says, “The workshop improved AI capability by 28% and should be expanded company-wide.” The source file shows a 28% increase in self-reported confidence among 18 attendees, no comparison group, and no post-work task.

The reviewer flags three issues: the construct is mislabeled, the conclusion exceeds the evidence, and the expansion decision lacks operational criteria. The author records three TRACE entries:

| Trigger | Analysis | Change | Evidence | Effect check | | --- | --- | --- | --- | --- | | Construct mismatch | Confidence is not demonstrated capability | Replace the capability claim with the measured survey result | Survey instrument and response table | A reviewer can identify exactly what was measured | | Unsupported causal language | No comparison or baseline supports attribution | State that attendees reported higher confidence after the workshop; do not claim causation | Study design notes | No sentence implies the workshop caused task performance | | Premature scaling recommendation | Cost, transfer, and manager follow-through are unknown | Propose a bounded follow-up cohort with a controlled work task and decision criteria | Pilot budget and draft assessment plan | Sponsor can see the evidence gate before expansion |

The revised brief keeps the 28% figure only with its population, measure, and limitation. It adds the response count, removes company-wide impact language, and asks for approval of a follow-up test rather than full expansion. The AI assistant helps compare versions and flag changed qualifiers, but the author checks the survey, decides the disposition, and owns the final statement.

This example is synthetic. It does not establish a training effect, validated measure, benchmark, or customer outcome.

Compare meaning, evidence, and usability

Review the old and new versions at three levels. First, compare meaning: numbers, scope, recommendation, confidence, obligations, and ownership. Second, compare evidence: source links, calculations, quotations, dates, and limitations. Third, compare usability: can the intended reader find the decision, understand the rationale, inspect uncertainty, and take the next action?

Use a fresh reviewer for the effect check when stakes justify it. Ask the reviewer to paraphrase the conclusion and identify its supporting evidence and limits. If the revision fixes clarity but changes the claim, or fixes a claim while breaking the handoff, it is not complete. Record rejected and deferred feedback so it does not return later as an unexplained request.

Assess revision-quality behavior

A controlled assessment can provide an initial artifact, source pack, conflicting feedback, and a later evidence update. Observable evidence includes whether the participant diagnoses the real issue, checks the source, gives feedback a reasoned disposition, protects valid content, updates dependent sections, compares versions, and retests reader understanding.

Do not reward accepting every comment or producing the largest diff. Strong judgment may reject a confident but unsupported suggestion, defer a change until evidence arrives, or make a small correction with large decision value. Report only the observed revision process and artifact within the scenario. Two Prune provides task-bounded assessment evidence for human interpretation and does not make employment decisions.