Catch the bad ones early.

"Acceptable" is far easier to demonstrate than to define. Show examples of good and bad, and the model picks up the distinction — including the parts nobody wrote into the spec.

Is this item acceptable, and if not, what's wrong with it?

Images, text, or structured records — the same mechanic applies to each.

What makes this hard by hand.

The spec is incomplete
Every quality standard has tacit knowledge in it that never made the document. Examples carry that knowledge; prose doesn't.
Inspecting everything is the goal
Sampling catches problems late. A check cheap enough to run on every item catches them at the source.
Borderline needs a human
Confidence scores let you pass the clear cases and escalate the genuinely marginal ones rather than guessing.

Three steps.

1

Collect good and bad examples

Photos of acceptable and unacceptable items, or records with and without the defect.

2

Decide what happens at each confidence level

Reject confidently bad, pass confidently good, queue the rest.

3

Add detection if you need location

Classification tells you whether there's a defect. Detection tells you where it is and how many.