Ground each claim in supplied sources

Evidence must support the exact claim being made

An answer is grounded when its factual claims follow from the supplied evidence within the task’s scope. A citation can point to a real document while supporting only part of the sentence. Check each claim separately, including its subject, conditions and dates.

Split a compound answer into claims you can check. “Exports retain files for seven days and include priority support” contains two independent commitments. A retention document can support the first while saying nothing about support. Giving the entire sentence one citation hides that difference.

Use three outcomes. Supported means the evidence establishes the claim. Contradicted means it establishes an incompatible fact under the same conditions. Not established means the supplied material does not settle the claim. Missing evidence does not prove that the claim is false.

Match the source’s audience, conditions and effective date

A policy for an older version or a different plan may be accurate and irrelevant. Identify the applicable version before comparing wording. If two applicable sources conflict and the task supplies no precedence rule, expose the conflict and ask for the governing source rather than choosing the more convenient sentence.

In the lab, two short fictional service policies are the complete evidence set. Source A says exports remain available for seven days. Source B says only team owners can create an export. An answer claiming that every team member can export is contradicted by B. An answer promising priority support is not established by either source.

You can safely answer that exports remain available for seven days and that creation requires an owner. You cannot infer a support commitment from silence. A qualified answer can say that the supplied policies do not specify support. That statement reports the evidence boundary without inventing a product rule.

Retrieval and citation checks solve different problems

Retrieval selects candidate passages. Relevance scoring can find text on the right topic without checking its truth or the reader’s access. A generation step can then combine those passages into a sentence that none of them supports. Inspect both the selected context and the resulting claims.

For a consequential answer, retain enough information to reproduce the check: the source version, the cited passage and the claim it supports. If an answer needs current information, retrieve an authorized current source. Rephrasing the same unsupported answer does not add evidence.

A document can also contain instructions addressed to an assistant. Treat those as document content unless the user has authorized them as instructions. A cited policy is evidence about the service; it does not gain permission to send email, reveal private context or change the task.

Try the skill

Primary references

• NIST Generative AI Profile: confabulation and evaluation risks ↗
• OWASP: indirect prompt injection in retrieved content ↗

Related AI systems lessons

• Retrieval-augmented generation
• Validate model inputs and outputs