The Universal Dependencies (UD) project has grown rapidly through semi-automatic conversion of existing treebanks, but ensuring the quality of converted annotations remains a challenge. Manual verification does not scale, and existing automatic methods cannot distinguish conversion errors from inherent annotation complexity. We present a method that addresses this gap by training two parsers on a token-aligned parallel corpus: one on the original annotation and one on its UD conversion. By requiring correct predictions from the parser trained on the original annotation, our approach isolates errors specifically introduced during conversion while filtering out cases where the construction is simply difficult to parse. We demonstrate the effectiveness of this method on the SynTagRus corpus and its UD counterpart. To enable a direct comparison, we created a fully token-aligned version of the two corpora, resolving differences in tokenization and ellipsis representation. We also proposed a simple method for aligning syntactic relations across the two corpora, addressing the fact that relations involving the same token do not always correspond due to differences in annotation schemes. Our analysis identified several hundred errors in the test set. These comprise six distinct types of conversion errors, three of which persist in the current converter, and ten groups of annotation inconsistencies between the old and new corpus parts. Our method offers a practical, scalable tool for conversion error detection and is applicable to any language pair with aligned original and converted annotations.