Excel in most of Europe saves a CSV with semicolons between fields, because the comma is already taken: it is the decimal point. A price of nineteen euros ninety-nine is 19,99, and one thousand two hundred and forty-nine is 1.249,00.
Several of our tools were built to accept that format. Each one checks whether the header has more semicolons than commas and splits on whichever wins. Then the number cleaner, written for American exports, stripped every comma as a thousands separator. 19,99 became 1999.
Nothing failed and no warning appeared. The numbers were just wrong:
- wasserstein-distance compared two groups whose means were 2 and 4 and reported 20 and 40, with every distance ten times too large.
- clearance-markdown-planner fitted price elasticity to prices 100 times too high.
- liquidation-pallet-bid-calculator read a 19,99 retail as 1,999 and a 1.249,00 retail as 1.249, so a pallet's retail total and suggested bid were off in both directions.
- gaussian-process-regression, hdbscan-clustering, isolation-forest and ad-incrementality-test-analyzer had the same bug in their x/y values, features or store series.
How it was found
This was a scheduled red-team run. It seeds small faults into a file and counts how many the tests catch: 15 of 17 non-equivalent faults this time, and 17 of 17 after two new tests. While reading wasserstein-distance's parser as an attacker would, the run typed a four-row semicolon file and got means of 20 and 40. The same numText line turned up in three sibling engines.
A search for the pattern (a parser that sniffs ;, then a currency strip whose character class holds a comma) found three more. Two other parsers matched a looser search but handle the decimal mark on purpose (bank-csv-to-qbo-converter and eu-vat-oss-calculator), so the final pattern skips them.
The fix
The repository had already met this bug. On 2026-11-14 a red-team run found 29 report analyzers that took the decimal mark from the delimiter alone, which misreads a tab-delimited German export and a semicolon file with dot decimals. They now share a detectDecimalComma(rows, delimiter) helper that votes from the values, and two checks test every copy. These seven tools never had any decimal-comma handling, so neither check matched them.
They now use the same helper. When the values vote for a decimal comma, a cell like 1,5 or 1.234,5 (optionally with a currency sign, a percent sign or accounting brackets) is read as 1.5 or 1234.5. Otherwise commas still group thousands, so a quoted "1,234" in a comma file is still one thousand two hundred and thirty-four. A tab-delimited European export now works too. Each venture got a test that fails without its fix.
A first version decided from the delimiter alone. The existing check_decimal_format.py caught that before merge, which is the point of having it.
One case stays ambiguous: a semicolon file with no header row. 1,5;0,25 has more commas than semicolons, so the delimiter sniff picks comma. That ends in a visible error, not a wrong number, so it was left alone.
The guard
platform/ci/check_decimal_comma.py flags any venture file that offers ; as a delimiter and strips commas with a currency-style character class but never handles the decimal comma. That is the gap the older checks left: they test the detector, and cannot see a parser that has none. It needs no workflow change, because unwired_checks.py runs every unwired check_*.py on every venture pull request. On today's tree it passes. Before the fixes it named all seven files.