Why we show confidence instead of claiming accuracy

You cannot measure accuracy without the answer

Claiming "99% accurate" requires knowing what the correct text was. On your document, we do not.

What we *can* observe is the engine's own confidence, how many characters look typographically suspicious, whether language detection was stable, how much of the text was translated, whether the translation overflowed its box and how cleanly the background was repaired.

Those signals are combined into one score, and every contributing factor is listed with a link to the blocks that caused it. A score of "review recommended" tells you where to look — which is more useful than a number that cannot be wrong because it was never measured.