<p>Food additive compliance assessment is an important procedure for formulation review, enterprise quality control, and regulatory auditing. Under GB 2760-2024, this task requires grounding the product to the correct food category, checking additive permissions along the lawful category path, and treating missing authorization conservatively under whitelist regulation. This study proposes FoodAudit-AG, an auditable procedure for food additive compliance-state assessment in formulation review. The procedure combines a regulatory knowledge graph, verified category anchoring, hierarchical applicability checking, whitelist-aware decision control, and traceable audit output. Each additive-level judgment is linked to explicit evidence, category-path support, applicability conditions, and quantity-related rationale before a recipe-level conclusion is produced. Experiments on 450 manually verified recipes show that FoodAudit-AG achieves a recipe-level all-correct accuracy of 0.9533 (95% CI: [0.9200, 0.9800]); for regulatory risk-state recognition, it obtains Precision, Recall, and F1 scores of 0.9910, 0.9483, and 0.9692, respectively. These results indicate that formulation review benefits from combining product-to-category mapping with lawful applicability checks and conservative whitelist control in a reviewable quality-control workflow.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

FoodAudit-AG: An auditable procedure for food additive compliance-state assessment under GB 2760-2024

  • Zhihao Huang,
  • Zhiran Liang,
  • Huadi Huang,
  • Xueru Zhang,
  • Song Shen,
  • Haohan Ding

摘要

Food additive compliance assessment is an important procedure for formulation review, enterprise quality control, and regulatory auditing. Under GB 2760-2024, this task requires grounding the product to the correct food category, checking additive permissions along the lawful category path, and treating missing authorization conservatively under whitelist regulation. This study proposes FoodAudit-AG, an auditable procedure for food additive compliance-state assessment in formulation review. The procedure combines a regulatory knowledge graph, verified category anchoring, hierarchical applicability checking, whitelist-aware decision control, and traceable audit output. Each additive-level judgment is linked to explicit evidence, category-path support, applicability conditions, and quantity-related rationale before a recipe-level conclusion is produced. Experiments on 450 manually verified recipes show that FoodAudit-AG achieves a recipe-level all-correct accuracy of 0.9533 (95% CI: [0.9200, 0.9800]); for regulatory risk-state recognition, it obtains Precision, Recall, and F1 scores of 0.9910, 0.9483, and 0.9692, respectively. These results indicate that formulation review benefits from combining product-to-category mapping with lawful applicability checks and conservative whitelist control in a reviewable quality-control workflow.