36 out of 36 is not 100%
Why the textbook confidence interval gives impossible answers at small sample sizes, with real coverage as low as 3.9% at 16 trials, and the one-line fix.
What breaks when software ships faster than people can check it, how to find it yourself, and what we have measured about our own tool.
Why the textbook confidence interval gives impossible answers at small sample sizes, with real coverage as low as 3.9% at 16 trials, and the one-line fix.
An automated checker ships around a hundred rules. Counted against the spec they land on 29 success criteria, and 29 of those rules cover one criterion between them.
In a research dataset of ten thousand simulated users, 27% write code, 12% are digital natives and one in five arrives trusting your software.
Seven signs of the house style many AI-built sites now share, in the order a visitor meets them, each one visible without developer tools.