Beyond Task-driven Evaluation: An Intrinsic Framework for Unified Abstract Syntax Trees in Cross-language Code Representation

Cross-language code representations are commonly evaluated through downstream tasks, making it difficult to separate representation quality from model architecture and training effects.This study proposes a diagnostic intrinsic framework for Unified Abstract Syntax Trees (UASTs) based on Programming Element Completeness (PEC), Cross-Language Structural Alignment (CLSA), and Representational Element Simplification Ratio (RESR), with Computational Overhead (CO) as a feasibility measure.The framework is validated using controlled degradation and lightweight retrieval, clone-detection, and non-neural summarization proxies on CodeXGLUE, Code2Code Java-C#, and HumanEval-X.UAST maintains high programming-element preservation, with dataset-level PEC-all and PECcore reaching 99.48% on Code2Code Java-C# and 98.92% on HumanEval-X, while CLSA improves for all 15 HumanEval-X language pairs, increasing on average from 0.50 to 0.59, and from 0.68 to 0.72 on Java-C#.RESR values of 0.5903 and 0.4550 indicate substantial simplification.Under deterministic corruption, PEC and CLSA decline consistently (Spearman ρ=-0.7413 to -0.7684), while CLSA shows a moderate positive association with toprank retrieval.UAST also remains competitive in clone detection and non-neural summarization.These findings show that UAST provides a transparent function-level structural abstraction whose preservation, alignment, simplification, and practical utility can be diagnosed before task-specific modeling; however, category-level OOP preservation should not be interpreted as complete preservation of object-oriented semantics.