Your judge holds 78% macro-F1 on the in-distribution gold set, but drops to 61% on a new document domain, though accuracy is 89%. The PM wants to ship anyway. Adjudicate.

The high accuracy figure is deceptive because it likely stems from the model defaulting to the majority label in a domain where it lacks sufficient exemplars. The significant drop in macro-F1 indicates a genuine failure in out-of-distribution performance. Shipping should be blocked until the exemplar set is expanded to cover the new domain and the OOD macro-F1 score is improved.