Per-field accuracy on this page is real — it comes from Aito's _evaluate with select: ["cases"]. We hold out each row, predict the field from supplier + description + amount, compare to ground truth, and bucket by confidence band. Predictions ≥ 0.85 are the auto-approve zone; lower bands flag review work.

Unlike traditional ML, aito.. needs no feature engineering, no model selection, no deployment — predictions come directly from the database, and so does this evaluation.