Earlier today IDS Model Check started comparing GlobalIds across the models of a federated set. If the architect's file and the structural file hold the same GlobalId, a viewer, a clash test or a BCF topic treats two objects as one, and a quantity takeoff counts the element twice. On buildingSMART's ten-file Simple-Scene (IFC4) it reported 35 such GlobalIds.
The obvious next step was an allow list, because some sharing is deliberate: Simple-Scene copies a "geo-reference" proxy into all ten discipline files and an "origin" marker into the four building files. Before writing it, I counted what the 35 were. Only 8 are elements. Allowing the origin marker by name would still leave three warnings about it: its type, the relationship that assigns the type, and its material link.
Followers
So the check now looks inside shared relationships. If a shared IfcRelDefinesByType or IfcRelAssociatesMaterial names, in every file, only objects that are shared too, it follows those objects. It is listed under the element ("with it: IFCROOFTYPE "house - roof", IFCRELDEFINESBYTYPE ...") rather than reported as a separate finding. The same applies to a type that such a relationship names next to a shared element. Simple-Scene's 8 shared elements carry 14 followers this way.
One GlobalId, different relationships
Twelve shared relationships did not follow anything, and they turned out to be a different and worse fault. Take 0bat_vVbL5e9xkiecCrRuV, a material link that appears in three files:
- in
Infra-Bridge.ifcit links four "rail bridge - filler" proxies; - in
Building-Landscaping.ifc, the house's underground site; - in
Infra-Landscaping.ifc, four park, river and site elements.
It looks as if the exporter derives one GlobalId per material. To a tool that merges models by GlobalId these are one relationship, so it keeps one set of links and drops the others. The check now names that case ("one GlobalId for relationships that link different objects in each model") instead of calling it "the same object in two files".
The allow list
--shared-globalids-allow FILE (Action input shared-globalids-allow) takes one rule per line:
* "geo-reference" # one georeferencing proxy in all ten files
IFCBUILDINGELEMENTPROXY "origin" # the building's origin marker
IFCGRID # no grid is shared here: reported as unused
3XrBtx9eX7mQE6EqWHPk1R # or a single GlobalId
A shared element is allowed when every file's copy matches a rule, and its followers go with it. A GlobalId's first character is always 0 to 3, which keeps a 22-letter class name like IFCBUILDINGELEMENTPART from being read as an id. A rule that matched nothing is listed in the job summary. Allow lists outlive the models they were written for, and a stale rule can hide the next real mistake.
On Simple-Scene IFC4 the result goes from 35 findings to 23 with no list, and to 21 with the two marker rules, which clear 8 GlobalIds. The roof modelled in both architecture and structure is still reported, as are the two manholes in plumbing and electrical. HHS Office's three discipline exports are unchanged (one wall in two files), and one class-and-name rule clears it.
The numbers, the rule format and how to rerun them are in eval/federation-allow-2026-11-17.md. Everything runs on the build machine, and no model leaves it.