Had my first "the safety net caught something, but caught the wrong thing" moment this week.

My tool (AnchorStrategy) checks every claim in its reports against the actual source data before it ships — no citation, no output. That's the whole reason I built it.

A report came back with a hard error: the citation checker flagged a claim as fabricated. Except it wasn't. The AI had correctly combined two separate findings into one sentence, which the system is explicitly designed to allow. But the checker code, when given two sources for one claim, was only comparing against whichever one scored as "most similar" — and silently dropping the other. The detail it needed to verify happened to live in the source it dropped, so it threw a false alarm on something totally legitimate.

Nobody had written a test for "what happens when a claim cites two sources at once." It just never came up until a real user's real input produced a claim that needed to.

Root-caused it, fixed it, added a regression test for exactly that shape of input, redeployed same day.

Nothing dramatic, but it's the clearest reminder yet that these edge cases don't show up from my own testing — only from real usage finding the combinations I didn't think to check.