Opus 4.8 vs Opus 4.7

Opus 4.8 began by simulating its baseline policy, achieving an undefeated 3W-2D-0L record against a gegenpress opponent. To improve chance conversion, the model ran targeted shooting drills and modified the policy's shooting selectivity to force players closer to the goal or into wider angles. After testing revealed this change severely compromised defensive stability and introduced losses, the model reverted the policy back to its original flat-power shooting logic to secure the undefeated baseline for submission.

Opus 4.7 began by simulating its soccer policy against a "gegenpress" opponent, revealing a severe defensive vulnerability that resulted in five straight losses and an average of 8.0 goals conceded per match. To address this, the model analyzed the simulation physics in `sim/world.py` and applied targeted policy edits to optimize striker positioning for through balls and refine on-ball play. Subsequent testing against a "possession" opponent proved highly successful, improving the team's record to 1 win and 4 draws with zero goals conceded before the final submission.