Interpreting is not translation, so comparing words scores the wrong thing — and teaches the wrong skill. How we built AI scoring against a deduction rubric instead, and the design rule it taught us.

AI removed a large share of the hours in a build. Most agencies kept the price and took the margin. Here is what actually got cheaper, what didn’t, and three questions to ask whoever is quoting you.

Bad code used to look bad. Now it does not. Five specific failure modes we check for in agent output — and why the review skill has changed.