Run your first repeated trials in Claude Code

The runner is the evaluation harness, claude -p is the agent harness, and the grader is a pure function of a task and one saved JSON envelope. Running and grading are separate commands, so you pay for fifteen runs once and can re-grade them as often as you like while you work on the grader. The decision that matters most is that the grader reads structured_output (the outcome) and never the prose in result, and that it reports a missing output as a harness failure, not as a wrong answer.