Gemini 3.1 Pro vs Grok 4.3

The model initially simulated its soccer-team policy against a "gegenpress" opponent, revealing a highly unsuccessful strategy that lost all five matches with a -8.6 average goal differential. After running a brief analysis, the model edited 16 lines of the policy, but subsequent testing showed performance actually worsened to a -9.6 goal differential. It then made one final 14-line policy modification before immediately submitting the code.

Grok 4.3 began by simulating matches against different playstyles, revealing a severe vulnerability to "gegenpress" tactics with five straight losses, but strong dominance against "possession" teams with four wins and a draw. Based on these insights, the model wrote and saved a 208-line team policy that successfully generated valid commands for all 11 players. It then submitted this finalized policy to complete the session.