The green light the AI handed me

I use an AI agent every day in design work. It counts, converts, propagates and checks. The first thing it taught me was not productivity, it was distrust: when the same agent writes the measurement and the thing being measured, green does not say the product works. It says the two halves agree.

The cheapest case to tell is the most expensive one I lived. I had a measurement that checked a component's color by reading Flutter's component tree. It passed. The screen came out wrong. The measurement read the declared color, and the question was the pixel: blue at 18% opacity has a different final value than full blue, and the tree cannot know that because compositing happens after it. I had asked for that measurement. The agent wrote exactly what I asked for.

Weeks later came the worse version. Six green measurements on a sign-up flow, and the submit button was dead. Not one ruler looked at the tap. Two taps with a finger on the screen found in fifteen seconds what the program had not found in any run.

That is where the rule I use today came from. Every new measurement is born with two assertions: one that proves it flags a defect built on purpose, and one that proves it does not flag the healthy case. Without both, what exists is green decoration. And the agent does not close a task on green: it hands me the green, I open the screen.

The denominator is part of the measurement. The first number I published about finished components was false, because the program doing the counting silently skipped the cases it could not classify. The total looked better than it was. Today the first thing I ask of a new repository is not the answer, it is the denominator, and I read the denominator before I read the answer.

The system I work on runs 2,062 measurements on every version and 24 screenshots that have to come out identical. That is worth something for one reason only: every one of those measurements has already failed on purpose once, in front of me.