When general-aviation aircraft leave a training airspace and several return to the circuit together, the tower controller must assign one of six allocation schemes to each aircraft within seconds. Historical records, however, resolve identical situations in different ways, so a purely supervised classifier trained on them saturates: on the SR-20 sample used here (= 20,000, six schemes) it reaches only 87.69 % accuracy and 86.85 % macro-F1. This paper moves the supervision target from the records to an explicit, auditable matching-rule model that encodes the operating regulations as a ternary structure with three-source evidence and a trend-strength gate; only its continuous coefficients are estimated by a conditional-logit fit. The rule posterior is distilled into a LightGBM student as a soft label. Against the deterministic rule standard of the held-out test split the student reaches 95.48 % accuracy, 94.99 % macro-precision and 95.04 % macro-F1, i.e., meets deployment requirements. The single-modality controls fail - a BiLSTM’s macro-F1 on the flight frames reaches 81.94 % and a Chinese-RoBERTa text classifier’s macro-F1 reaches 50.48 % - so the two modalities carry complementary information; a random forest’s macro-F1 reaches 90.89 % but is less auditable. Because every training label is generated by a named rule, the deployed matcher stays traceable, and retraining costs under one minute on a desktop CPU, which is the property a safety-critical tower function requires.