The IDS model checker tests a building model (IFC) against an IDS file, a list of rules like "every wall has a name" or "every element's description is one of these values". It is free, runs in the browser or in CI, and is MIT licensed.
We check it against 68 IDS files, 3 of ours and 65 published on GitHub. One of them, a Dutch BIM fixture, writes its attribute names in a way that matches every attribute of every element. Its 12 rules apply to all 665,224 entities of a 40 MB office MEP export, most of them geometry. It also asks about the same attributes four times: once for "is it set" and once each for three different values.
Most of the answer doesn't depend on the value
Before an attribute rule compares a value, it asks some questions that have nothing to do with the value. Does this class have the attribute? Is it set? Is it empty? Is it a list or a reference, which IDS cannot compare by value?
On this model the last question settles nearly everything. A point's coordinates are a list. A face's bounds are references. Every rooted object has an OwnerHistory reference. Only 190,924 tests needed the value at all.
The checker already kept one byte per element per rule. That cache is keyed on the whole rule, value included, so it could not tell that four rules share a name. Over the 68 files it ran the attribute test 2,851,787 times.
Spare bits
A verdict needs two bits: passes, absent or fails. The byte has eight. So an attribute rule's byte now also holds that value-free reading. A rule with the same attribute name reads it from any other kept rule's byte before looking at the element itself. That is one reading per element (665,224 of them), and no new memory.
We tried a separate array per name first. It was faster still, but it put our 219 MB stress model over a 1 GB heap, where the old version fits. We dropped it.
The leak the memory check found
The version that shares bits also ran out of memory at 1 GB. Its live heap was no higher than before, yet a heap snapshot showed 644,045 extra strings kept alive.
Our STEP reader decoded every text value one character at a time. In V8 a string built that way stays a chain of one-character pieces, about 32 bytes a character, until something flattens it. Running a regular expression over it is one thing that does. The old value tests did that by accident: every rule ran its pattern over every name and description. Once those tests stopped running, the chains stayed, 20 MB of them on the MEP model.
The decoder now returns text that has no escapes unchanged and copies plain runs in one piece. We fuzzed it against the old decoder on 300,000 random strings made of every STEP escape, and the outputs matched every time.
Results
Every report hashes the same as before on all five models.
| model | before | after |
|---|---|---|
| 40 MB MEP export | 4.1-4.5 s | 2.9-3.1 s |
| 22.5 MB residence | 2.1-2.3 s | 1.6-1.7 s |
| 219 MB federated model | 35.7-36.7 s | 28.4-29.6 s |
| same, 1 GB heap | 49.3-51.5 s | 35.4-36.6 s |
The method, the 12 planted faults the new test catches and the one it doesn't (it changes no verdict) are in eval/attribute-shape-2026-11-18.md.
The lesson we're keeping: a cache keyed on a whole rule misses what rules share, and memory you can't explain is worth a heap snapshot. The leak had been there all along. A faster path took away the accident that was hiding it.