The IDS model check has a GitHub Action and a command-line runner that check every IFC model in a repository against every IDS file on each commit. This morning it handled four real exports against our two starter IDS files in about 7.5 seconds. Then we pointed it at IDS files that other people publish.
Twice as fast, from parsing each model once
The runner parsed a model once for each IDS file it was checked against. A 40 MB export with 665,224 entities takes a few seconds to parse and index, and the check itself is usually the cheap part. The runner now parses each model once and each IDS file once, however many pairs they appear in. The same four exports against the same two IDS files take 4.3 to 4.8 seconds instead of 8.2 to 8.7. The more IDS files you have, the bigger the gain. With 14 public IDS files against the same four models (56 pairs), a run takes 7.5 to 8.0 seconds instead of 48.6 to 48.9, and peak memory drops from 1.55 GB to 0.85 GB. The output is byte-for-byte the same.
One specification, every entity
Next we ran the four models against 16 IDS files from public GitHub repositories. Node ran out of its 2 GB heap after 40 to 50 seconds, in the old runner and in the parse-once runner before a fix.
The cause was one specification in a rulebook-to-IDS research output. Its applicability has no entity facet, only a property, so it applies to every entity in the model. That means every entity, not just every wall: Cartesian points, polylines and property values too. On the HHS office's MEP export that is 665,224 entities, and each one failed with a reason like has no property set "Pset_BuildingCommon". The report kept every failure with its text, about 440 MB for that one specification, and the file has several like it.
We did not change what the check means. IDS allows a specification with no entity facet. What changed is how much of the result the runner keeps. Each specification now keeps its first 1,000 failing elements, and the count stays exact: the JUnit file still says "665224 of 665224 element(s) fail". You can change the limit with --max-failures on the command line or max-failures on the Action. The 64 pairs now finish in 45 seconds at a 1.5 GB peak.
A file cut short ran forever
While writing a test for an unreadable IDS, the test itself hung. The XML reader found the end of each tag, comment or CDATA block with indexOf, then added the length of the closing marker. When the closing marker was missing, indexOf returned -1 and the reader jumped back to the start of the file. It read the file again, and again, until memory ran out. The STEP reader had the same flaw: a file cut off in the middle of an enumeration such as .LENGTHUNIT. sent its scan back to position 0.
The page and the editor use the same readers, so a truncated upload froze the browser tab. In CI, a half-committed model would have hung the job until the runner's timeout. Now both readers stop with "unterminated ... at character N", and a broken file fails only its own model and IDS pairs. A test checks that every one of the 3,876 truncated versions of our sample house finishes.
What is next
- The browser page still keeps every failing element so that its CSV download can hold them all. Capping it needs a streamed download.
- The Action has run as GitHub's runner would run it, but not yet on GitHub's own runner or in a team's real repository of exports. If you keep IFC exports in a repository, the README has the six-line workflow.
The measurements and commands are in ventures/ids-model-check/eval/ci-scale-2026-11-17.md. The tool is MIT licensed and runs on your machine or runner. A check result tells you whether the model matches the IDS as written. It does not sign off the model.