Verifying AI Output Before You Trust It

This is the largest cluster in the whole Build Log, and it took writing 170 uncovered posts to notice it: a huge amount of the actual work of running an AI-assisted shop is not building anything, it's refusing to believe a claim of done. A saved record, a green deploy, a passing dry run, a plausible-looking mockup, a status field that says LIVE, an AI result that looks finished. Every one of those can be wrong in the exact same way: something reported success without anyone checking that the real, observable state matches the report.

The posts below are grouped by what kind of claim got checked. Some are about AI output specifically: a mockup that missed a design token, a checkout button that grew four pixels, a plausible answer that dropped a whole class of information. Some are about backlog and status hygiene: an epic marked Not Started with six tickets already shipped under it, a priority label that quietly became a history marker instead of a fact. Some are about the review itself, a second model, a second person, a recorded dissent, because the fastest way to catch a wrong answer is to make a different mind check it before it ships.

None of this is abstract caution for its own sake. Every post here traces back to a real moment where trusting the report instead of the result cost something: a live mailer overwritten, a customer-facing bug hidden behind a stale label, a queue worker that was never actually built while the button said it was working. The fix is nearly always the same one line, reworded: go look at the actual thing, not the thing that claims to describe it.