Test-time search for terminal agents assumes you can score a half-finished run. A quarter of the way in, a critic reading the transcript sorted eventual winners from losers 49% of the time — a coin flip — which is why wrapping the agent in tree search bought 0.6 points for 3.8× the money. Wait until the agent calls itself done, hand the critic a throwaway fork of that state and of the state it started from, and let it run commands against them: misjudged endings fell from 36–39% to 22%, and the pass rate went from 63.4% to 72.4% at a little over half the cost of the search.