Gold-standard annotation is slow and expensive, and large language models (LLMs) as judges have been proposed to expedite it. Existing methods raise accuracy by sampling, asking the same question repeatedly at one evaluation pass per answer, and by deferring their least certain instances to a human. An annotation method must therefore be judged on three quantities, the accuracy attained, the evaluation passes expended, and the share of the corpus deferred to a person. This study asks whether all three can be improved by structuring the labels rather than re-examining them. The proposed method decomposes labels into a hierarchy of granularities, declares their relations as logical rules, evaluates each granularity in a mutually blind pass, and applies the rules deterministically at no model cost. It was validated on 700 building-regulation provisions for automated design compliance checking and 300 out-of-domain GoEmotions comments, against four state-of-the-art LLM-assisted annotation methods and three human experts annotating independently. On accuracy, for the regulation provisions, even a single structured pass of the proposed method reached 94.0% accuracy, above the best baseline at 89.7% and the majority verdict of three human annotators at 81.9%. At three passes, the proposed method reached its highest accuracy, 97.2%, while deferring only 9.0% of the corpus, far below the 20.0% deferral ratio of the two routing baselines. The out-of-domain corpus reproduced the pass and deferral results, with the proposed method ahead of all four baselines in paired comparisons by at least 1.7 %p at two passes, significantly so against two of them. A gold standard can therefore be built at a review load a small team can absorb.