We look at the screen you'll look at, in both themes, in Safari and Chrome.

The short version

We shipped a Light mode nobody could read, and every automated test passed.
The tests were reading code. Nobody was looking at the screen.
Now every screen is checked as a picture: Light and Dark, Safari and Chrome, desktop and phone.
If the picture is wrong, it doesn't ship.

The one that got through

Bench has a Light theme and a Dark theme. We shipped an update where Light mode had text you couldn't read against the background. Every test passed. The contrast checker read the color values in the code and they looked fine on paper. On the actual screen, after the theme applied, the page was unreadable. A person would have caught it in two seconds. No person looked.

What a check has to do now

A check that can't fail when a person can't read the page isn't a check. So now the last step before a screen goes out is a picture of the page, the way you'd see it, and a log that says green doesn't count. Both themes, because your office might run either. Safari and Chrome, because your reps are on iPhones and your office is on whatever it's on, and the two browsers don't draw pages the same way. Desktop and phone size. If the picture is wrong, it doesn't ship.

What it means for you

When a new screen shows up in Bench, someone or something has looked at it the way you will, not just at the code behind it. We'd rather tell you that than tell you all our tests passed.

Questions

Why did the checks pass?

They read the color values in the code. They didn't look at what those colors turned into on the screen after the theme applied.

What has to pass now?

The rendered page, as a picture. Readable in both themes, in Safari and Chrome, at desktop and phone size. Buttons that do what they say.

Does AI do the checking?

AI writes a lot of the tests. The test still has to fail on the picture, not on the code. A model isn't a substitute for looking.

Keep going

Every Bench word above is on Glossary / FAQs, in plain English.