Your AI scorer no longer agrees with your trained raters, and nobody can say where it drifts. I build a 60-response probe set from your own task types, score it blind against your band descriptors, compare it with your model output, and show you exactly where the disagreement sits: by task type, by band, by language feature. You get the agreement analysis, an error taxonomy, a prioritised fix list and a 30-minute readout with your team.