Fact-check and revise in loops

AI can help find claims, but it cannot reliably verify its own output by rereading it. Use two distinct steps.

First, extract a claim ledger:

List every factual claim, number, date, name, quotation, attribution,
causal statement, comparison, and externally testable recommendation.
Copy the exact wording and note its location. Do not judge accuracy.

Second, verify each entry against an authoritative source you actually inspect. A model may miss claims or classify them incorrectly, so add anything it overlooked.

A practical ledger records:

• claim identifier and exact wording;
• claim type;
• source required;
• source location or link;
• support status: supported, contradicted, partial, or unresolved;
• scope and date limits;
• revision needed;
• reviewer and review date.

“Supported” means the source justifies the draft’s wording—not merely that both discuss the same topic.

Suppose a small internal pilot found fewer support escalations after a checklist was introduced. The evidence may support “Escalations decreased during this pilot.” It may not support “The checklist permanently reduces escalations for every team.” The second statement adds causality, permanence, and generality.

Check these dimensions:

• Population: who or what was actually studied?
• Time: when was the information true?
• Magnitude: does the number match the source and denominator?
• Causality: is the relationship causal or only associated?
• Certainty: does the source say may, likely, or will?
• Scope: does a local example support a broad conclusion?
• Conditions: what exceptions or prerequisites apply?

Preserve uncertainty. Qualifiers are not clutter when they define the truth.

Prefer original, current, authoritative sources appropriate to the claim: official documentation for product behavior, primary research for study findings, statutes or qualified counsel for legal interpretation, and direct records for organizational facts. A search snippet, generated citation, or unsourced summary is not enough.

Open the source. Confirm the title, author, date, relevant passage, and context. Check that a quotation is exact and that omitted text does not reverse its meaning. Verify that a link resolves to the intended material and that readers have appropriate access.

AI systems can invent plausible titles, authors, URLs, statistics, and quotations. Never “complete” incomplete citation details by generation. Use [SOURCE NEEDED] until a real source is found. If a central claim cannot be supported, narrow it, label it as opinion or hypothesis, or remove it.

For changing information, record the verification date. Product features, prices, officeholders, policies, and scientific guidance can become stale.

Revision improves decisions, not just sentences. Use a loop with explicit passes:

1. Brief fit: Does the document still serve purpose and audience?
2. Argument: Do conclusions follow from evidence?
3. Structure: Is information ordered for the reader?
4. Claims: Is each statement supported and proportionate?
5. Clarity: Can the reader follow actors, terms, and conditions?
6. Voice: Does the expression fit the author and context?
7. Proof: Are grammar, links, formatting, and references correct?

Running separate passes makes changes inspectable. A single “improve everything” prompt can repair grammar while introducing factual drift.

After each significant revision, rerun checks affected by the change. Moving a sentence can detach it from its citation. Shortening can remove a condition. Adding a new example creates new claims. Verification is iterative.

For consequential documents, record:

• requested change and reason;
• section affected;
• source or reviewer behind the request;
• claims added, removed, or modified;
• verification repeated;
• unresolved decision;
• person who accepted the change.

This need not be bureaucratic. A simple checklist or version note can prevent reviewers from debating old text and help the author explain why a claim changed.

Ask the model to propose changes, not silently apply them:

Against the brief and claim ledger, identify the three highest-impact
revision needs. For each, quote a short excerpt, explain the reader problem,
and propose an edit. Do not add evidence or mark claims verified.

Prioritize substantive defects over cosmetic preferences. An unsupported recommendation matters more than a repeated adjective.

Invite a skeptical pass:

• What would a knowledgeable reader challenge?
• Which claim relies on ambiguous wording?
• What counterexample weakens the conclusion?
• Which stakeholder or consequence is omitted?
• Where does the draft confuse observation with inference?
• What action could a reader take incorrectly?

Red-teaming should strengthen accuracy, not manufacture false balance. Not every established fact needs an invented opposing view. Evaluate objections using evidence and relevance.

High-stakes writing—medical, legal, financial, safety, employment, public policy, or security—requires qualified human review. AI can organize a checklist, but it cannot assume professional accountability or know every applicable rule.

Revision can continue forever. Define completion criteria from the brief: all required claims supported, unresolved items disclosed, structure tested, voice reviewed, links checked, required approvers signed off, and no placeholders remaining.

Then freeze a candidate and conduct a final proof against the actual rendered document. Check headings, captions, footnotes, links, tables, names, numbers, and version dates. Do not use the model’s statement “everything is accurate” as a completion criterion.