Zheng & Xu's arXiv audit finds component collapse in 2 of 6 corpora

Jiuyi Zheng and Gan Xu have released a new arXiv paper (2608.08892) that systematically audits "component collapse" in hierarchical procedural corpora. The paper builds on earlier work by Guvenilir and Dogan that showed merging relation types can create a giant component that obstructs data splitting. Rather than claiming a new phenomenon, the authors constructed a fixed panel of six hierarchical procedural datasets and ran a symmetric audit across three layers: a content-near-duplicate layer, a common-container layer, and their union.

Two of the six corpora, MyFixit and Doc2Dial, exhibited the same failure pattern under the same operational criteria: both layers passed individually, but the union failed. That 2-of-6 fraction is a deliberate panel metric, not a prevalence estimate. The authors also tested a prespecified bridge-specific predictor, but it was not statistically distinguished from a registered union-density control, so the panel could not identify a bridge-specific mechanism. Secondary diagnostics bound threshold sensitivity, annotation coverage, and lexical cues without overgeneralizing. The paper contributes a bounded measurement and audit methodology for keeping relation families visible and reporting negative cases, but proposes no new splitting algorithm or broad causal claim.