AI Data Guard

The product runs in the window beside this column, and every control in it works.

Step 4 of 10

The error the queue is willing to show you

120 of the window’s detections adjudicated by two officers, sixty each, giving 93.3% precision and a 6.7% false-positive rate counted from those verdicts and from nothing else — with the next 16 drawn by the same rule, waiting in the section above this card and carrying no verdict at all. The card in front of you is one of the eight errors: “Project Sycamore” is an M&A codename at this bank and a tree everywhere else, and it fired on a landscaping vendor’s planting schedule.

A detection product that shows no error is a product measuring nothing. This one sits in the 6.7% rather than being excused out of it. The page also says how the sample was drawn, that reported errors are in it by construction, and that recall is not measured at all — because the two questions an examiner asks about a precision figure are how it was sampled and what it is missing. Mark a row yourself and the two figures at the top do not move: they are the record, counted from what two named officers wrote. Your marks are laid over that record on a line of their own, labelled as your session’s and persisted nowhere, because this demo has no database and a figure a visitor can change is not the institution’s measurement.

Look atThe detector behind it: the institution’s own dictionary, firing exactly where the bank told it to look, and still wrong.