A DIKWP-3No-4S Semantic White-Box Evaluation Framework for Chinese Dialect Recognition by Large Language Models

Chinese dialect recognition by large language models (LLMs) cannot be evaluated adequately by terminal-label accuracy alone because correct labels may still rely on weak evidence, faulty knowledge mapping, or poorly calibrated uncertainty. We present DIKWP-3No-4S as a semantic white-box evaluation framework in which externally coded semantic relations and scoring rules are inspectable, while the evaluated LLM remains a black box. The original controlled experiment used six foundational cases across Clean and three 3-No conditions, generating 270 repeated observations. To strengthen prospective assessment of the framework, we added 18 new independent cases with reference standards frozen before response collection. Three LLMs were tested under Clean and one preassigned 3-No perturbation, yielding 108 responses independently blind-rated by two experts. Case-level permutation tests, case-clustered bootstrap intervals, and Holm correction showed stable terminal-label accuracy but significant changes in judgment-boundary and condition-adaptation scores. Inter-rater analysis showed strong consistency for judgment-boundary and condition-adaptation scores, whereas several fine-grained semantic measures, including evidence-explanation quality, D → D activation, path quality, and SemA, showed low or near-zero inter-rater reliability and are therefore treated as exploratory rather than validated measures. These results support DIKWP-3No-4S for external semantic auditing and fault localization, but do not establish an independent terminal-performance advantage from DIKWP structuring.