A claim needs a reference.
A neatly presented inventory may list an item that is not there. A description may omit a visible defect. Reading the text can reveal some inconsistencies; examining the property or object establishes what is actually present.
Illustrative example: an AI output lists three items of equipment, while the expert finds two on the inspected object. The original output is preserved, the missing item is identified and the discrepancy is linked to dated evidence. This is not a client audit result.
What the report needs to distinguish.
An established finding, an unresolved point and an out-of-scope issue are different statuses. Missing evidence does not demonstrate an error; it also does not justify treating the output as accurate.
The Hu-score summarises the checks possible on the initial output, identified object and date. If a determinant fact cannot be established, no overall grade is assigned. The human correction remains separate from the AI-generated output.
Source and scope
NIST AI 600-1 describes confabulation risk: generative output can contain false assertions. This reference explains the need for checking; it does not measure Huxpert’s effectiveness. NIST AI 600-1.
Related reading
Have an AI-generated result checked
Identify the output, actual object and checks required before requesting a quote.