AI Systems

Building Vision Audits That Can Explain Their Work

A useful field audit must do more than identify an object. It must preserve where the evidence came from, explain uncertainty, and make omissions visible.

Yakir Yaacov · August 2026 · 5 min read

The first version of many vision systems is deceptively simple: attach photographs, ask a model what it sees, and save the answer. That can make a good demo. It is not enough for an operational audit.

Attribution comes before detection

If a photograph belongs to the wrong room, even a correct detection creates an incorrect report. I treat room matching as its own pipeline: normalize external labels, preserve source identifiers, record unmatched rooms, and refuse to silently convert missing attribution into missing equipment.

“Not detected” and “not evaluated” are different states. A trustworthy report never merges them.

References need structure

A reference library becomes useful when each example has an equipment identity, a clear visual description, and known edge cases. Every evaluated room receives the relevant references, while the output retains the claim photograph that supports each finding. This turns a model response into traceable evidence.

Human review is part of the design

Blurred frames, obstructed equipment, and similar-looking components produce unavoidable uncertainty. The system should expose confidence and route borderline images to review. Labels must remain editable because approved knowledge improves later evaluations.

The practical test

If those answers are yes, the system is no longer just recognizing images. It is producing an auditable operational decision.