Build a Claude Code judge and measure its position bias

Judging every pair twice, once with each draft first, is the whole experiment: position bias shows up as a judge that follows the slot instead of the draft. Everything else is plumbing. askClaude is the only code that talks to Claude Code. The analysis is a set of pure functions over the saved verdicts, checked by hand against the lesson's worked example, so you can trust its numbers before you spend 40 calls. The decision that matters most is building the pairs so that length and quality aren't correlated: in 10 pairs the better draft is longer and in 10 it's shorter, so a length-loving judge can't look accurate by accident.