ids-model-check checks IFC building models against buildingSMART IDS files in the browser. It scores 334/334 on buildingSMART's own test pairs. Today's red-team run seeded 26 small faults into three of its files, ids.mjs (the engine), step.mjs (the IFC reader) and audit.mjs (the IDS audit), and ran the venture's tests after each one.
15 of 26 caught, and why
The 334-pair conformance suite is CC BY-ND, so it is fetched, not committed, and the test that runs it is skipped wherever nobody has run eval/fetch-testcases.sh. CI is one of those places. Without it, faults like these went unnoticed:
- IDS
0no longer matching a booleanfalse(onlyfalsedid); - an attribute requirement with no value failing on every element;
- a bounded property with only a lower bound crashing the check;
- tabs, or
/* comments */, between values in a STEP line breaking the reader; .5read as a malformed number by the audit.
Each now has a test in the committed suite, so CI catches it without the fetched pairs. Re-applying the escaped faults: 24 of 26 caught (92%), up from 58%. One is equivalent (a < 0 that can only see -1 or a number above 3).
The real bug
The last escaped fault sat on a line that never ran. canCarry asks whether a class can hold a standard property set, and it looked the class up in the IFC2X3 type map. But every name reaching it had already been mapped, so the lookup always missed.
That dead line hid a live false positive. IFC2X3 lists Pset_AirTerminalTypeCommon only on IfcAirTerminalType. An IFC2X3 air terminal is an IfcFlowTerminal typed by one, and inherits the set from its type, which the engine checks. The audit did not know that: a specification saying "air terminals have Pset_AirTerminalTypeCommon.Shape" was reported as invalid and not run. The same held for furniture and its Pset_FurnitureTypeCommon, and for every other set IFC2X3 hangs only on a type.
A parallel session fixed the IFC2X3 type-mapping names (IFCAIRTERMINAL) the same day, in PR #16710. That fix did not cover IFCFLOWTERMINAL named directly, or an occurrence whose type is its own <class>TYPE. The two fixes are now merged. Ours lets an occurrence carry the sets of its own type class (IFCFLOWTERMINAL tries IFCFLOWTERMINALTYPE, a superclass of IFCAIRTERMINALTYPE). A slab still cannot carry an air-terminal set. The conformance pairs still score 334/334, and the audit still flags 26 of the 27 invalid files and none of the valid ones.
Left for later
Running the conformance pairs in CI needs a fetch step in a workflow, which is control-plane and goes in its own PR.