A bot is a policy, not a player

A game grew five new levels, each verified as “loads, runs, reports a result”. A headless autoplay sweep then ran them. They turned out to be two-to-four minute walkovers, or unwinnable.

So the harness became the judge, and it had opinions of its own. It built mines on a level that forbids construction, because the build menu is what stops a person and the harness never opened the menu. It was measuring a different game from the one being played.

An economy pass was then tuned against its numbers and reverted. The idle metal figure turned out to measure the harness’s own cautious spending as much as the game.

A test bot does not play. It carries out a policy, and every number it reports has that policy in it.