Four common wrong-answer patterns
The riskiest pages are not always obviously broken. They can be well-written, recently viewed, and individually plausible while disagreeing with another source.
- Contradictory rules: two pages give different approval limits, escalation steps, or eligibility criteria.
- Duplicated guidance: copied runbooks drift apart after only one version is updated.
- Stale authority: an old page keeps its links and visibility after a replacement is published.
- Prompt-like content: page text attempts to redirect an assistant or override the task context.
Evidence a reviewer needs
A useful finding should name the affected page, category, severity, confidence, short evidence, rationale, and suggested action. For contradictions, both claims need to be visible so a reviewer can choose the source of truth.
Do not automatically rewrite the page from an AI suggestion. The reviewer may need business context, policy authority, or legal approval that the scanning system does not have.
- Page title, space, URL, owner context, and relevant evidence location.
- The competing claim or duplicated source when the risk spans pages.
- A concrete action: update, merge, deprecate, restrict, or accept with explanation.
- A Jira link or audit event showing who acted and when.
How DocsTrust limits the blast radius
DocsTrust starts with selected spaces, uses bounded AI budgets, validates AI output, and does not allow the model to execute cleanup actions. Full page bodies are not persisted in the V1 audit result.
These controls reduce product and data risk, but no audit can guarantee that every possible wrong answer is found. High-consequence AI workflows still need user guidance, monitoring, fallback paths, and human accountability.
- Selected-space scope instead of an uncontrolled tenant-wide rollout.
- Deterministic checks plus bounded semantic analysis.
- Human review before Jira cleanup tickets.
- Re-scan after remediation and monitor recurrence.