Field notes from active builds.

On an AI build, "works correctly" is not a contract term. What replaces it: a test set the client wrote, a threshold with a number in it, and a named judge.

A build/no-build memo is four documents and one uncomfortable sentence. What goes in each, and why the no-build verdict is the one worth paying for.

Working demo by end of week two, or we refund the sprint fee. What that promise forces, and the failures it catches while they are still cheap.

Nobody in our peer set puts a number on the homepage. We do. The reasons are mostly about whose time we refuse to waste.

A field triage from scoping calls: four features that reach production, six that die in the demo phase.

Demo-day accuracy numbers die on contact with real users. Evals that keep meaning something after launch.

Twelve-week engagements pretend to be twice as much scope as six. In practice they rarely are.