Technical Notes
Can a reviewer trace the AI finding to its source?
Keep the image, the record and the interpretation distinct, so someone else can check the result.
Reference from 2024: NIST: Generative AI Profile

An explanation is not a source
The model might assemble a useful narrative from a crop, a timestamp and a work order. The evidence is the crop, timestamp and work order. Keep the exact image region and relevant excerpt available to the reviewer.
Use stable references. A link to an entire folder is not a practical way to verify one sentence in a brief.
Give the model a bounded task
Check that the retrieved records refer to the same asset and relevant date. Ask the model to separate what is visible, what the record supports, what conflicts and what remains unknown.
Allow it to return insufficient evidence. Treat text from retrieved documents as source material, not instructions to follow.
Example: a work order from the wrong day
The follow-up image shows material near an access lane. A work order lists authorized staging there, but its date is the day before the capture. The model can report the material and the possible explanation. It cannot say that the record confirms what happened.
The brief should ask the site operator for the current work order and vehicle access status. It should not tell the crew the lane is open or closed.
Make it possible to disagree
An analyst should be able to reject the finding, revise it or request a new observation. Record the reviewer decision against the version of the evidence they saw.
When comparing assisted and unassisted review, ask how often reviewers can reproduce a claim and how long they take to resolve contradictions. A polished summary is not an evaluation result.
Practical checklist
- Keep a stable image-region and record reference for each substantial claim.
- Give uncertainty and conflicts their own place in the result.
- Preserve corrections and the evidence version used in the review.