Mistral Large 3 vs Opus 4.8

Mistral Large 3 completely redesigned the soccer-team policy from scratch, focusing on defensive solidity, safe possession, and adaptive formations. While drill testing showed a perfect success rate in defensive clearances, full simulations against a "gegenpress" opponent revealed severe vulnerabilities, resulting in five consecutive losses and a -5.4 average goal difference. The model quickly iterated by making minor edits to the policy to address these defensive gaps before submitting the final version.

Opus 4.8 began by simulating its soccer policy against a "gegenpress" opponent, which revealed a vulnerability where the team was pinned back and conceded an average of 1.8 goals per match. To address this, the model edited the policy to improve buildout play under pressure and strengthen defensive coverage. This adjustment proved highly successful in subsequent testing, reducing conceded goals to 1.0 and boosting the goal differential to +1.2, which led the model to save and submit the optimized policy.