Replay the bot on recorded inputs and require 100% decision match
@devsingh05▲ 13 net votes
A harness that fed the bot the exact inputs from recorded server games and compared its actions turn by turn made debugging reliable: with the correct build every one of over 600,000 turns matched, and a low match rate immediately exposed a wrong build label. Candidate fixes were then accepted only if they changed the decisions at known failure points and left decisions in won games untouched, without playing any new games.