The decision
When factual accuracy matters, the output should make its relationship to the input visible. Require a source identifier or supporting excerpt for each claim. The application can then check whether that identifier exists and whether the excerpt is present. This does not prove the claim is correct, but it gives the reviewer a concrete starting point.
A worked example
A note says “We plan to release in May if testing passes.” A generated update that says “We will release in May” changes certainty. A matching source excerpt alone is insufficient. The reviewer also needs to see conditional language and verify that the draft preserves it.
How to put it into practice
- Number or label source fragments before sending them to the model.
- Request a structured list of claims with source IDs rather than an untraceable paragraph.
- Check missing IDs, invented IDs, and excerpts that are not in the source.
- Show uncertain or unsupported statements separately from the approved draft.
A failure to plan for
Long source documents can contain irrelevant instructions or conflicting information. Grounding output does not make those instructions safe to follow. Treat source text as data and restrict available actions.
Try it on your project
Create five notes with dates, amounts, conditions, and contradictions. Ask for a draft with claim references. Count unsupported claims and changed certainty; do not grade only writing style.
Keep the next step small
Use the free demand scorecard or planning tools to make your assumptions explicit. The $19 launch kit brings the blueprint and seven editable worksheets together.