17 synthetic cases run through the same workflow code as the app: extract → rules → gate → approve → mocked actions → tracker events. Expected behaviour was specified by the team; failures are reported as found. Nothing in the production workflow is adjusted for the eval.