IDS Model Check checks a building model against a buildingSMART Information Delivery Specification in a browser tab, with nothing uploaded. Until today it only read STEP .ifc files. IFC comes in two other forms that people actually get sent: ifcXML, the same model written as XML, and ifcZIP, an .ifc or ifcXML file zipped. Both now work, on the page and on the command line (node cli.mjs model.ifczip spec.ids).
Two layouts, one model
ifcXML has no attribute order, and STEP is all order, so the reader goes through the IFC schema to put each value in the slot a STEP file would use. After that the checker runs on it unchanged. There are two layouts in real files:
- IFC2X3 ifcXML writes each attribute as a child element (
<Name>Basic Wall</Name>), every entity at the top level, references as<IfcOwnerHistory ref="i2" xsi:nil="true"/>. - IFC4 ifcXML takes shortcuts. Simple attributes become XML attributes (
PredefinedType="solidwall"), and entities sit inside whatever uses them. An attribute element can be the value itself (<RelatingObject xsi:type="IfcProject" ref="p1"/>, or with no type at all, an instance of the attribute's own class). Relationships can sit under an inverse attribute (<IsDefinedBy>inside a wall) without their own pointer back to that wall.
That last shortcut needed something the checker's schema tables didn't have: for each inverse attribute, the relationship class and the attribute it points back through ("IsDefinedBy: IfcRelDefinesByProperties for RelatedObjects"). The schema generator now pulls that out of the EXPRESS files. All 153 IFC4 inverses resolve, and the existing tables come out byte for byte as before.
How it was checked
The test that counts is the same model in both forms. xBIM's test files include a Revit 2014 export, 4walls1floorSite, as a STEP file and as xBIM's ifcXML export of it. Both read to 579 entities. After mapping references, 2,205 of 2,205 attribute values are identical, and the starter IDS gives the same verdicts on both.
No public IFC4 ifcXML with a STEP twin turned up, so I wrote one: the checker's sample house as IFC4 ifcXML, by hand, using every shortcut above. It matches the STEP sample on 319 of 319 values, and both starter IDS files give the same verdicts.
Across all 16 xBIM ifcXML and ifcZIP test files (9 distinct ifcXML documents from about seven exporters, plus 7 zips) the reader reported one problem. The file really has it: a ref to an id it never defines. Every zip entry was byte-identical to Info-ZIP's unzip -p. One zip, written the way .NET writes them, stores its sizes as 0xFFFFFFFF with the real values in a zip64 extra field. The first version of the zip reader refused it, and the fix went in before this post. buildingSMART's 334 IDS test pairs still score 334 of 334. In Chromium the page unzips with the browser's own DecompressionStream.
What is not shown
No real IFC4X3 ifcXML has been through it. The reader recognises the namespace, but I didn't find a file to test on. And nine documents isn't every exporter: if one of yours reads wrong, the page's feedback link is the place to say so.
Source, tests and the measurement: ventures/ids-model-check (ifcxml.mjs, zip.mjs, eval/ifcxml-xbim-2026-11-16.md). MIT licensed.