Nothing in your build fails when your README starts lying. Mine lied for three weeks, and I was the one quoting it. Documentation is code that nobody runs.
Sentence stood in my README, on the store page and in the release notes:
"A company that ships one model and coasts is insolvent inside three years, and an automated test fails the build if that stops being true."
Three numbers are involved there, and I published the only one nothing had measured. Measured one is five.
A company running the probe's coasting style, with four staff, goes under in its fifth year. That is what harness what produces when you run it.
The other two are boundaries, not results. The build fails if that company is still trading after eight years, and it fails if it dies before its second. One catches an economy gone soft, the other catches a collapse so fast the test would pass for the wrong reason.
Three years was none of the three. I wrote it once, it sounded right, and I quoted myself for three weeks across three websites.
That is a worse feeling than finding a bug, because a bug is something you got wrong and this was something I was telling people.
Six more numbers
I found it while correcting six other numbers that had drifted. The badge at the top of the README, the first thing anybody sees, said 793 tests. There are 888. It said 77 files. There are 88. Commits, lines, save version and the count of unreachable mechanisms were all stale, and the last one was stale by four.
Same shape every time. A number written once keeps living, and nothing ever runs it.
The fix is not proofreading. Every figure in that README now comes from a script that reads the repository, and the sentence about coasting names the fifth year, with the staff count that produced it.
I am 22, in Radom, driving a taxi, looking for my first job in software. Post above is the work: I go hunting for these in my own output, and I would rather publish them than wait for somebody else to find them.
What is the oldest number in your documentation that nothing verifies?