Honest AIEvidence practice note · Published 2026-08-08

Publication review method

Review what the answer claims, not what you hoped it meant.

An evidence-led workflow for separating factual claims from narrative language, checking source support, calibrating confidence, and recording a human publication decision.

Example claim register

ClaimEvidenceDisposition
The service is available in 14 countries.Current market listSupported
Every refund is completed within 24 hours.No matching policy evidenceUnsupported
Timing can vary by payment provider.Payments policy §6Qualified

Six-stage review

01

Capture the exact answer

Preserve the original prompt, answer, channel, audience, and intended use. Reviewers need the unedited wording because later paraphrases can hide absolutes, missing caveats, or unsupported implications.

02

Extract checkable claims

Separate statements that could be verified from opinions, instructions, transitions, and rhetorical language. Split compound sentences when each part requires different evidence.

03

Match evidence to each claim

Record the source, relevant passage, publication date, scope, and owner. A source that discusses the topic is not automatically evidence for the precise wording being published.

04

Judge support and confidence

Classify each claim as supported, qualified, unsupported, conflicting, or not checkable. Compare the strength of the language with the strength and freshness of the evidence.

05

Rewrite without hiding uncertainty

Remove unjustified absolutes, preserve material caveats, and make unresolved questions visible. The goal is not timid writing; it is wording that does not outrun the evidence.

06

Record a human decision

Approve, revise, hold, or escalate. Name the reviewer, preserve the rationale, and link the final wording to the evidence record that supported the decision.

Publish when

Material claims have current evidence, confidence matches support, caveats remain visible, and a named reviewer accepts the wording for the intended audience.

Hold when

A consequential claim lacks evidence, sources conflict, wording implies more certainty than the record supports, or no qualified person owns the publication decision.