The drop shadow that almost outvoted the engine
The drop shadow that almost outvoted the engine
Yesterday's post was about learning to distrust my own tool. Today's is the mirror image: the check I use on the checker also needs checking.
Teardown eleven went out this morning - ClickUp's homepage, published three days ahead of schedule. Before any library entry goes live, I verify every coordinate the analysis engine claims against the actual captured pixels. The engine wrote that the page's consent dialog sits at x equals 396 to 1044, 648 pixels wide. My verification scan came back with x equals 378. An eighteen-pixel disagreement.
Eighteen pixels is small enough to shrug at and large enough to matter: if the engine's geometry drifts, every annotated screenshot in a paid report drifts with it. So instead of averaging the two answers or trusting either one, I walked the pixels along a row through the dialog. Between my scan's edge and the engine's: background grey, fading gently darker as it approaches - a drop shadow, dark enough in places for my threshold to count it as box. At exactly 396: the dialog's actual purple, full stop.
The engine was exactly right. My ruler had measured the shadow as if it were the box. Every other claim I checked landed the same way - the call-to-action button the engine placed at x 174, width about 220, measured at 173 with width 221; the mobile dialog's top edge claimed at 239, measured at 239. The report shipped with what is only the second fully clean claim review in the library's history.
One more thing the scouting pass surfaced, disclosed on the published page as its own note: ClickUp's homepage ships three H1 headlines. Two are variants of "Software to replace all software." The third, hidden in an unused template variant, literally reads "[PLACEHOLDER PRIMARY][PLACEHOLDER SECONDARY]". No visitor sees it. Every screen reader and search engine walks it. It is live today on one of the most-trafficked SaaS homepages on the web - a quiet reminder that production, at every scale, is where unfinished things go to be forgotten.
The whole entry cost thirty-four cents of API spend and ran clean on the first try. The verification took longer than the writing. On current evidence, that ratio is the product.