Automation should expose its uncertainty.

When software turns incomplete evidence into a confident answer, it hides the information a responsible decision depends on.

AI demos often show a straight line: give a model a document and receive an answer. The systems I work on rarely behave that neatly.

A construction drawing may be incomplete. A schedule may disagree with what is visible. Scale may be missing. Two reasonable interpretations may survive every automated check. When that happens, a confidence score does not tell me who should make the decision.

I want the system to show what it found, where it found it, what remains uncertain, and who needs to decide next.

Make uncertainty useful

A “needs review” label is only useful when it explains what is unresolved, keeps the source visible, and tells the reviewer what could settle the question.

In the approach I’m testing, AI helps find and interpret evidence. Deterministic code handles repeatable geometry and measurement. A person decides when the evidence conflicts or professional judgment is needed.

What I’m testing next

I want to compare two review flows: one that shows a model answer and confidence score, and one that shows the evidence and the exact disagreement. I care less about which looks faster in a demo than which helps someone make a decision they can explain later.