Every object in an IFC building model carries a 22-character GlobalId. The standard says no two may share one (IfcRoot, uniqueness rule UR1), and everything downstream of a model check relies on that. A BCF issue points at an element by its GlobalId. A model-comparison tool matches this week's export to last week's by GlobalId. IDS Model Check's own GitLab report keys each failing element on it. If two elements share one, the BCF issue selects both and the diff matches the wrong pair.
Until today the checker read GlobalIds and never said when two elements shared one. Now it does, in every output: a block on the page, a warning in the text report, CSV rows, a BCF topic per GlobalId, and in CI an annotation on the element's line, a job-summary entry and a GitLab issue (--fail-on-duplicate-globalids makes it fail the build). It does not change the IDS verdict, because the IDS never asked about it.
How common is it?
We ran it on 163 public files: the HHS Office and Schultz Residence exports the earlier tests used, all 35 IFC files in buildingSMART's sample repository, and all 121 in xBIM's test folders. Together they hold 179,721 rooted objects.
- 3 files have a duplicate. In HHS Office
construction2.ifca column has the GlobalId of a floor slab. In an xBIM alignment sample two spaces share one. An xBIM parser test repeats one GlobalId six times on purpose. - The six original HHS and Schultz exports have none. An earlier write-up said they "already had 1 to 3". A text search over the file does find a repeated 22-character string, but it is the property name
TreadLengthAtInnerSide, not a GlobalId. That note is corrected. - On the largest file (HHS MEP, 665,224 entities) the check adds 130 to 155 ms to a 1.5 to 1.9 s parse.
To make sure we weren't grading our own work, we counted the same thing with IfcOpenShell. On the 133 files IfcOpenShell opens, the two tools give identical groups, GlobalId by GlobalId and element by element. IfcOpenShell does not open the other 30: most are schema versions it does not ship (IFC4X1, IFC4X3 drafts), and a few are xBIM's deliberately broken test files.
The bug the cross-check found
The first comparison was not 133 of 133. On xBIM's parser test IfcOpenShell found six group objects sharing the GlobalId and we found five. The sixth ends its name with \S\'. In the STEP format, \S\ takes the next character and adds 128 to its code, so \S\' is "§" and the quote after it closes the string. Our reader saw two apostrophes in a row, read them as an escaped quote, and ran the string on into the next line. That lost the group and the project object after it. Both now read correctly. Only 2 of the 163 files contain \S\', so this was a narrow bug. Without a second tool to compare against it would have gone unnoticed.
Measurements, commands and the cross-check script: eval/duplicate-globalids-2026-11-17.md. The checker is free, MIT licensed and runs in your browser: nothing is uploaded. It is a check against what the IDS and the IFC schema say, not a design review or professional advice.