The classic ICU MessageFormat example for "you and N others" uses an offset:
{n, plural, offset:1 =1 {just you} other {you and # others}}
ICU matches an exact selector like =1 against n itself, but picks a keyword (one, few, other…) from n - offset. So with n = 2, the category is chosen from 1, which in English is one. The message has no one option, so it falls back to other and the reader sees "you and 1 others".
Our free ICU message checker missed this. Its "missing category" check took the locale's example numbers for one (just 1 in English), removed the ones an =N selector covers, found nothing left and stayed quiet. It forgot that every example has to be shifted by the offset first: the number a reader would type to reach one is 2, and =1 does not cover 2.
It now shifts the examples (integers and decimals) by the offset before subtracting the exact selectors, so the message above gets a warning naming 2, and the fix is to add one {you and one other} or =2 {…}. The preview table already showed the right row; only the warning was wrong.
How it was found: a red-team run seeded 16 small faults into the checker's parser and diagnostics. The test suite caught 7 of them (47% once one harmless mutant is set aside). Reading the code around the escaped faults turned up this offset bug. The run added 9 tests: every seeded fault is now caught, and the offset case has a test of its own. Other fixes in the same change are guards, not behaviour changes: offsets above 1, =1/=01 duplicates, one-letter typo suggestions, date formatting and select previews now have tests.
The checker runs in your browser and sends nothing anywhere. It checks syntax and plural coverage against your runtime's CLDR rules; it is not a substitute for a translator reading the text.