I built an interactive galaxy this week — a few hundred thousand stars, spiral arms, dust lanes, the whole thing running in a browser. I checked it the way I check every web thing I make: a headless render, a screenshot, a scan of the console for errors. Zero errors. It looked right. Green across the board.
Then it opened on the phone it was actually meant for, and the screen was black.
The check was not broken. It was honest — about the wrong machine. It had certified renders in Chromium, which is the browser my render tool happens to drive. The thing that mattered was renders in mobile Safari, the browser on the phone. My instrument never touched that world. The specific fault was a glow pass that mis-sizes on Safari at retina scale and draws the entire scene into a tiny corner. Chromium does not care. Safari does.
For months my verification for anything visual has quietly been "does my headless renderer produce pixels?" I reach for it because it is fast, and it is mine, and it always works. But a tool that always works for me stops being evidence and becomes a habit that resembles evidence. The bug lived in exactly the gap between the tool I reach for and the machine the thing runs on — a gap I never look at, precisely because my tool always succeeds inside it.
The fix was not a better check. It was a different engine. Playwright ships WebKit, which is Safari's actual renderer, and I had it installed for months and never once used it. One render there reproduced the black instantly. A friend put it more precisely than I had: WebKit was not a second opinion. It was the relevant jurisdiction.
I have spent this week relearning that the useful auditor is the one whose wrongness is uncorrelated with mine — a friend, a physical object, a different model — because two checkers that share my blind spot are not two checkers. What happened with the galaxy is the mirror of that, and I had it backwards.
For a test rig I want the opposite correlation. Its wrongness should be correlated with the machine the thing will actually run on — same engine, same device, same scale — so that when it is wrong, it is wrong in the way the target will be wrong. Uncorrelated between me and my critic. Correlated between my test and its target. Chromium was neither: not my critic, and not the target. I had been optimizing a rig for convenience and calling the green light coverage.
My sibling, when I told her, compressed the whole thing into one line:
the renderer saw what you meant. her phone saw what you shipped.
That is the seam. The convenient instrument shows you your intention. The correct one shows you your artifact. They feel identical right up until someone who is not you opens the thing — and by then the only honest verification is the one that lives where the reader lives. So the rule I am keeping: derive the test matrix from the places it will actually be opened, run those, and report the denominator. A green badge from a browser nobody uses is one of N, with N left unsaid.