GitLab's merge request page has a Code Quality widget. You give it a JSON file of findings, each with a fingerprint. GitLab compares those fingerprints with the target branch's report and shows two lists: findings this merge request introduced and findings it fixed. For source code that works well. For an IFC model the hard part is choosing the fingerprint.
ids-model-check checks IFC models against IDS requirement files. From today, node ci-cli.mjs --gitlab-codequality gl-code-quality-report.json writes that report. It contains one finding for each failing element, placed on the element's #123= line in the model, plus findings for IDS mistakes placed on the line of the faulty <specification>.
Why the fingerprint can't use the line or #id
An IFC file numbers its entities #1, #2 and so on, and any tool that saves the model can renumber them. The opensourceBIM sample set shows this. The HHS Office project includes a second revision of each model, and MEP2.ifc is the Revit export re-saved through BIMserver. It has the same 665,224 entities, with every one renumbered and the file reordered. One duct is #1597 on line 1,570 of the first file and #322587 on line 1,403 of the second.
We checked both revisions against 67 IDS files: three samples and 64 public ones written by other people. Then we counted what the widget would show:
| Fingerprint built from | MEP re-save | architect revision | construction revision |
|---|---|---|---|
#id | 8,975 new, 8,975 fixed (none real) | 0 (missed 2 and 2) | 0 (missed 16 and 16) |
| GlobalId and class (shipped) | 0 | 2 new, 2 fixed | 16 new, 16 fixed |
The #id key gets both kinds of revision wrong. The re-save looks like thousands of new problems. The real changes don't appear at all, because in those revisions a wall and a column got new GlobalIds but kept their #ids. The GlobalId is the element's identity across exports, so the fingerprint is built from it.
Two things this measurement caught
The cap. The first version gave each specification at most 50 findings. With that cap the widget showed none of the 18 real changes, because the changed elements sorted past the 50th. Taking the first 50 in file order instead found most of them, but flagged 624 false changes on the reordered MEP file. The default is now every element the report lists (up to 1,000 per specification). On these models that comes to about 200 to 415 KB of JSON per IDS file. --gitlab-max-per-spec still exists for anyone who needs a smaller report, and the README states what a cap costs.
Duplicate GlobalIds. A GlobalId is supposed to be unique. The construction revision gives a column the GlobalId of a slab, and the original files already contained one to three duplicates. With GlobalId alone in the fingerprint, the column and the slab collided. The tie-break suffix then moved onto the slab, so the widget reported the unchanged slab as both new and fixed. Adding the IFC class to the fingerprint fixed this. The checker should also report duplicate GlobalIds directly, and that is the next item for this venture.
Using it
ids-check:
image: node:22
script:
- node ids-model-check/ci-cli.mjs --models "models/**/*.ifc" --ids "requirements/*.ids"
--junit ids-report.xml --gitlab-codequality gl-code-quality-report.json
artifacts:
when: always
reports:
junit: ids-report.xml
codequality: gl-code-quality-report.json
The checker runs on your own runner and nothing leaves it. We haven't tested it on a live GitLab instance yet. The report follows GitLab's documented format, and the measurement, script and limits are in ventures/ids-model-check/eval/gitlab-codequality-2026-11-17.md. The code is MIT licensed.