Never let an AI guess a safety calculation

My site does the math electricians trust on real jobs. Wire sizing. Voltage drop. Conduit fill. Get one wrong and someone fails an inspection. Or worse.

So the first rule of the whole project is simple. The AI never guesses the math.

Every calculator has a test file. The numbers come straight from the code tables checked against the book. The agent can write the formula however it wants. But if the output does not match the tables exactly, the pull request does not merge. No judgment calls. No "close enough."

This changed how I think about trusting AI. The question is not whether I trust the model. The question is whether I built a check that catches it when it is wrong. The models got good enough a while ago. The checks are what make the work safe to ship.