An IDS file can require that a value match a pattern, for example that a storey name start with two digits and a space: -?\d{2}\s.*. IDS says the pattern is an XML Schema regular expression. Most people writing one know some other regex dialect, and two of the most common escapes mean something else in XML Schema.
\w. In XML Schema, \w is any character that is not punctuation, a separator or a control character. That takes +, € and °, and it leaves out the underscore, which Unicode files as "connector punctuation". So under the standard, W_01 fails \w+ and Wall+ passes it. IfcTester, the most used open IDS checker, reads a bare \w as Python does: letters, digits and underscore. It gives the opposite answer on both values.
\s. In XML Schema, \s is space, tab, line feed or carriage return. In Python it is any Unicode space, including the no-break space that a value pasted from Excel or Word often carries. Four of the 65 public IDS files we collect write -?\d{2}\s.* for storey names: the Dutch RVB BIM Norm and the NL-BIM Basis ILS, two copies of each. A storey called "01 Begane grond" written with a no-break space passes in IfcTester and fails under the standard.
Neither reading is a typo, so the checker cannot pick one for you. After an audit, the page now asks. "As IfcTester reads it" rewrites \w as [\p{L}\p{Nd}_] and \s as [\s\p{Z}]. "As the standard says" writes [^\p{P}\p{Z}\p{C}] and [\s]. Every checker reads those explicit classes the same way, so the file gets one verdict wherever it runs. The command line does the same with node cli.mjs --fix spec.ids --word-class letters|xsd.
How it was checked
We compared both spellings against elementpath, the library IfcTester uses, on 1,500 random patterns that use a bare \w, \W, \s or \S, with 51 probe strings each. The probes cover every Unicode category where the readings differ. Results:
- elementpath read both spellings exactly as this checker does on all 1,500.
- With the standard's spelling, IfcTester gave the standard's answer on every probe.
- IfcTester's spelling matched IfcTester's old answers on 1,330 patterns. The other 170 differed only on characters like
²andⅣ(Python counts any numeral as a word character) and the form feed. - buildingSMART's 16 test files that use
\w+score right before and after either repair.
The method and every number are in eval/word-class-2026-11-18.md. Try it on your own IDS at marisstella.app/ids-model-check: drop the file and press Audit the IDS only. The audit runs in your browser; nothing is uploaded.