The semantic ceiling of OpenUSD for engineering CAD ingestion: a specification analysis with an open measurement framework on NIST test models

Industrial digital-twin platforms increasingly ingest engineering CAD data into OpenUSD. That format was never designed to carry boundary representation (B-rep) geometry, product and manufacturing information (PMI), or assembly constraints. Practitioners report semantic loss at this step, yet no peer-reviewed assessment exists. The available evidence is vendor self-reporting, which we show to be structurally silent about semantic loss. This paper makes three contributions. First, we derive the semantic ceiling of OpenUSD from its specification and the standardization state of the Alliance for OpenUSD (AOUSD). The result is a two-level attribution taxonomy separating degradation imposed by the specification from degradation attributable to converter implementations. Second, we present an open, kernel-neutral measurement framework built on public-domain NIST MBE PMI test models. It uses analytic B-rep surface sampling as ground truth and a registration protocol that reports coordinate-convention changes instead of absorbing them. It spans five degradation dimensions and adds a downstream measurement task as an external validity anchor. Third, we apply the framework to a fully open-source reference pipeline and to the NVIDIA Omniverse CAD converter (published with NVIDIA’s written permission), and report where each degradation class originates. All 838 top-level semantic PMI annotations across 16 test models arrived in USD with no machine-readable carrier, in every pipeline. Nearly every converted model collapsed to a single mesh. The open-source chain lost its unit declaration in 16 of 16 models. The commercial pipeline preserved declarations faithfully yet produced unit-heterogeneous assets. The commercial pipeline silently discarded all supplemental (non-solid) surface geometry, shrinking one part’s extent by 34 mm, yet reported success with zero warnings. The loss hid behind per-model mean deviations below 0.01 mm until spatially localized and identified at face level. Hole-diameter extraction, a routine metrology task, succeeded on about 5% of the 1,866 cylindrical features in the set for both pipeline families. Finally, we show the ceiling is thin. A minimal carrier profile using only standard USD mechanisms, implemented in a prototype writer, moves semantic PMI and face identity from lose-all to keep-all under the same audit. We offer it as input to the ongoing AOUSD standardization. All scripts, protocols, and test data are public and re-runnable.