Critically evaluate AI outputs, reasoning, and behaviour.

The resolution: for a 90-second live demo, a feature that works but you don't fully understand beats one you understand but that's flaky. Build week runs on the loop, not the master plan — so you'll ship code Claude wrote that you couldn't reproduce from scratch. Split the room. One bench defends the working black box: it survives the demo, and you can learn how it works afterwards. The other defends understood-but-fragile: you can fix it live when it breaks, and you can defend every line. Tie it to the capstone — at the Seoul Expo your JARVIS runs live, once, in front of an audience. Two minutes each side, then a show of hands. The winner is the bench that argued in frame most rigorously, not the one you happen to agree with.