An IDS file is a list of rules a building model (IFC) has to meet. People usually agree those rules in a spreadsheet first. This morning our free IDS Model Check learned to turn such a spreadsheet into IDS 1.0 and back, one row per rule. We tested it on workbooks we wrote ourselves and on workbooks LibreOffice saved, with every cell stored as text.
A BIM manager's workbook usually comes from Microsoft Excel, and Excel doesn't store what you type as text. Type 2024-01-01 and it saves the number 45292 with a date format. Type true and it saves a boolean, which our reader gave back as TRUE. In IDS a boolean is written true, and TRUE matches no IfcBoolean in any model, so a person who typed the right value got a rule that failed every wall. Excel also saves some fractions to 17 digits (0.93827160493827155) where it shows 15.
What we measured
Apache POI, the Java library for Office files, keeps 353 .xlsx files from its bug reports in its test data. 289 of them were saved by Excel (desktop, Mac or Online). We read each file's first worksheet with our reader and with openpyxl, cell by cell, as the text Excel shows: dates by the cell's number format as ISO 8601, numbers to 15 digits, booleans as true/false.
| before | after | |
|---|---|---|
| cells read as Excel shows them | 62,182 of 65,548 | 65,537 of 65,548 |
| date and time cells | 0 of 1,149 | 1,139 of 1,149 |
| boolean cells | 0 of 262 | 262 of 262 |
| workbooks with every cell right | 255 of 321 | 316 of 321 |
Most of the 11 cells left are ones where no date exists. A day count of 0 under a date format is Excel's 1900-01-00, so we keep the number. Six are DD.MM.YYYY dates, which openpyxl reads as a time because it only checks lower-case format codes; Excel doesn't care about case.
On the way, we found three more problems. One workbook names its parts with backslashes (xl\workbook.xml); Excel opens it, and openpyxl refuses it. Excel stores a carriage return inside a string as _x000D_. And an .xls or a password-protected workbook is an older container format that our page used to read as CSV, giving nonsense rows. The reader now handles the first two and names the third, with how to save a plain .xlsx.
Typing is the other half
Reading a typed cell correctly can't undo the typing. A number cell holds 42, whether you typed 42 or 42.0. That matters to an IDS: buildingSMART's test suite has a case where 42.0 must be refused as an integer, and once a spreadsheet has typed it, it passes.
We measured that two ways in LibreOffice Calc, since there's no Excel on our build machine. First, we opened each IDS's CSV with type detection on, the way a double-clicked CSV opens. With the new reader, 47 of 56 public IDS files came back with the same content (24 before), and 302 of 334 buildingSMART test files (288 before). What is left is mostly numbers whose spelling changed, which no reader can recover.
Second, we took the workbook our page downloads and retyped every cell a spreadsheet could turn into a number, date or boolean, through Calc's own input line, as a person editing it would. With the cells formatted as General, as they were, 24 of 56 public files came back the same. So the download now formats every cell, and the empty cells of each column, as Text:
| retyped download | before | after |
|---|---|---|
| public IDS files, same content | 24 of 56 | 56 of 56 |
| buildingSMART test files, same content | 288 of 334 | 334 of 334 |
| buildingSMART verdicts unchanged | 325 of 334 | 334 of 334 |
What you type into the downloaded sheet now stays the characters you typed.
The details and scripts are in eval/excel-workbooks-2026-11-19.md. We had no copy of Excel itself to type into. Excel documents the same rule for Text cells, but we haven't measured it. If you have a real requirements table, we'd like to run it.